Workshop production quality online detection system based on machine vision and AI

Abstract

The invention discloses a workshop production quality online detection system based on machine vision and AI, and particularly relates to the field of machine vision detection.The workshop production quality online detection system comprises a multi-modal data acquisition and preprocessing module, product physical signals, process parameters and actual quality values are synchronously acquired and subjected to data cleaning, semantic alignment and standardization processing, and then the product physical signals, the process parameters and the actual quality values are obtained; generating a structured data block associated with the product information; a feature extraction and mixed vector generation module extracts physical information features and multi-modal data driving features and fuses the physical information features and the multi-modal data driving features into a mixed vector, and a network is trained to output a comprehensive quality index and a corresponding time sequence; and finally, calculating a quality toughness coefficient and a process fluctuation conduction coefficient through a quality field domain energy coupling decision module, defining compliance, consistency and risk energy fields, coupling to obtain a comprehensive decision value, and outputting four-level quality decisions of high quality, qualification, to-be-checked and rejection according to the value and field domain characteristics to guide production circulation so as to achieve the purpose of improving production efficiency. And the accuracy and efficiency of quality detection are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine vision detection, more specifically, the present application relates to a workshop production quality online detection system based on machine vision and AI. BACKGROUND

[0002] With the transformation of manufacturing industry to intelligentization and large-scale, the production rhythm of workshop is significantly accelerated, the complexity of product structure and the diversity of process are continuously improved, and the real-time, comprehensiveness and accuracy of quality detection are increasingly demanding; current workshop production quality detection mainly relies on two technical paths: one is traditional manual sampling inspection, relying on the experience of detection personnel to judge the basic indicators such as product appearance and size, although it can cover part of the intuitive defects, but limited by sampling rate, it is difficult to cover the occasional defects in large-scale continuous production, and the detection efficiency decreases significantly with the increase of working hours, and subjective judgment difference also easily leads to non-uniform standards; the second is single equipment detection, which mainly uses 2D vision camera, single-point temperature sensor or pressure sensor, etc., which can only independently collect surface texture, local temperature or single-point process parameters, and cannot synchronously obtain 3D geometric shape, material composition, global process parameter sequence and other multi-dimensional information. Some enterprises try to introduce AI algorithm to assist detection, but most of them focus on single data dimension, and do not realize the deep correlation between multi-source detection data and production process, so the overall detection system is difficult to adapt to the core needs of modern production for quality whole process and multi-dimensional control.

[0003] However, it still has some disadvantages in actual use, such as: 1. Single data acquisition dimension and poor synchronization: the existing technology mainly collects 2D images or single physical signals, lacks the integration of 3D geometric, spectral material, temperature field and other multi-modal data, and the time stamps of each device are inconsistent, the data correlation is low, and the product quality state cannot be fully reflected, which may miss some key defect information; 2. One-sided feature extraction and insufficient fusion: image texture or simple physical features are extracted separately, without combining physical mechanism features such as product heat conduction and structural stress with multi-modal data driven features, the feature dimension is single, it is difficult to accurately describe the nature of complex defects, which may lead to misjudgment and omission in subsequent quality judgment; 3. Subjective quality decision logic and lack of quantization: relying on preset threshold or simple classification model decision, without considering the dynamic influence of process fluctuation on quality, without building a multi-dimensional energy field coupling quantization decision system, the decision result is greatly affected by experience, and the stability and reliability are insufficient under different working conditions; 4. Poor scene adaptability and weak expansibility: for different scenes such as hot processing and assembly, the hardware and algorithm parameters need to be adjusted significantly, and it is difficult to adapt to multi-scene detection needs through a unified framework; when new defect types are added, the model iteration cost is high, and it is difficult to quickly respond to changes in production process. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a workshop production quality online detection system based on machine vision and AI, which solves the problems raised in the above background art through the following scheme.

[0005] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a workshop production quality online detection system based on machine vision and AI, comprising: A multi-modal data acquisition and preprocessing module: synchronously acquires physical signals, process parameters and actual quality values of target products, and ensures consistency of data timestamps through clock synchronization and trigger control; after preprocessing of the acquired raw data, standardized data blocks are generated; A feature extraction and mixed vector generation module: based on the standardized data blocks, physical information features and data-driven features are extracted, partitioned after attention fusion, and then the physical information features and the data-driven features are spliced to generate mixed vectors; based on the mixed vectors, a network model is trained to output a comprehensive quality index CQI and a CQI time sequence; A quality field energy coupling decision module: based on the output of the previous modules, quality resilience coefficients and process fluctuation conduction coefficients are calculated, and compliance energy fields, consistency energy fields and risk energy fields are defined; based on historical data, an energy transfer coefficient matrix and a field phase difference are determined, the total energy is coupled and calculated, and a standardized comprehensive decision value QCI is generated; according to the QCI and the field characteristics, a quality decision is output.

[0006] Technical effects and advantages of the present application: 1. Multi-modal data is synchronously acquired and the dimensions are comprehensive: 2D images, 3D geometry, spectral material, temperature field and process parameter acquisition are integrated, PTP clock synchronization and FPGA trigger control are used to ensure consistency of timestamps, and multi-dimensional quality information of products is comprehensively covered, laying a complete data foundation for accurate detection; 2. Deep fusion of features and accurate characterization: thermal physical and geometric physical features are extracted based on physical mechanisms, multi-modal data-driven features are fused, high-dimensional mixed vectors are formed after attention fusion, the essence of defect attributes is accurately captured, the feature representation ability is effectively improved, and the misjudgment and omission rates are reduced; 3. Quantitative decision system and dynamic adaptation: three types of energy fields, compliance, consistency and risk, are constructed through quality resilience and process fluctuation conduction coefficients, and a comprehensive decision value is coupled and calculated combined with historical data, the influence of process fluctuation on quality is quantified, the decision logic is objective, and it can dynamically respond to changes in production conditions; 4. Multi-scene adaptation and strong expansibility: modular design is adopted, the temperature field acquisition unit is enabled in hot working scenes, and physical feature extraction is adapted through zero padding in other scenes; when new defect types are added, only training samples need to be supplemented, without the need to significantly adjust the framework, and the model iteration cost is low. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 The figure is a schematic diagram of the overall structure of the present application.

[0008] Figure 2 The figure is a schematic diagram of the structure of the multi-modal data acquisition and preprocessing module of the present application.

[0009] Figure 3 The figure is a schematic diagram of the structure of the feature extraction and mixed vector generation module of the present application.

[0010] Figure 4 The figure is a schematic diagram of the structure of the quality field energy coupling decision module of the present application. DETAILED DESCRIPTION

[0011] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0012] Reference Figures 1-4 The machine vision and AI-based workshop production quality online detection system shown includes: The multi-modal data acquisition and preprocessing module synchronously acquires physical signals, process parameters and actual quality values of the target product, and ensures consistent data timestamps through clock synchronization and trigger control. After preprocessing of the acquired raw data, standardized data blocks are generated. It should be further explained that the multi-modal data acquisition and preprocessing module includes a data acquisition sub-module and a preprocessing sub-module. It should be specifically explained in this embodiment that the data acquisition sub-module includes a physical signal acquisition unit, a process parameter acquisition unit and an actual quality value acquisition unit, and the specific acquisition process is as follows: Through the deployment of multiple types of sensors and interfaces at the detection station, product physical signals, process parameters and actual quality values are synchronously acquired to generate a raw data set.

[0013] Physical signal acquisition unit: 2D Surface Image Acquisition: A 5-megapixel, 2592×1944 resolution, 60fps area array camera is deployed 500-800mm directly above the inspection station, equipped with an 8mm fixed-focus lens and a ring LED light source with switchable brightness. When the product enters the station through the photoelectric sensor, the camera simultaneously captures three images with brightness levels of 3000 lux, 6000 lux, and 10000 lux, respectively. The Laplacian variance algorithm is used to select one image with no reflection and complete details in BMP format. This image contains information on the shape and location of the product's surface texture and planar defects. The product texture includes roughness and gloss, and the planar defects include scratches, stains, and missing characters.

[0014] 3D geometric shape acquisition: A line laser scanner with a point cloud density of 100 points / mm, a scanning speed of 500 lines / second, and a laser wavelength of 660nm is deployed at a 30° and -45° angle to the side of the camera, and is triggered synchronously with the 2D camera. The scanner performs line scanning on the product surface, generating single-frame point cloud data containing three-dimensional coordinates x, y, z in mm, with an accuracy of ±0.01mm, and can extract spatial features of surface height differences and three-dimensional defects.

[0015] Hyperspectral material acquisition: A hyperspectral imager is deployed on the other side of the camera. The parameters of the hyperspectral imager are: wavelength 400-1000nm, spectral resolution 5nm, 120 bands, spatial resolution 640×512, and the detection angle is perpendicular to the product surface. The hyperspectral cube of the product surface is acquired simultaneously in ENVI format, which contains the original reflectance values ​​at different wavelengths, and can reflect the spectral characteristics of material composition differences and hidden sensory defects.

[0016] Temperature field acquisition: An infrared thermal imager is deployed 1-2m to the side of the welding or heat treatment station. Its parameters are: 640×512 resolution, temperature measurement range of 20-500℃, 30fps frame rate, and error of ±2℃. Ten frames of temperature field matrix are continuously acquired. The peak temperature moment of the molten pool or heat treatment area is located by the peak detection algorithm. The selected frame of data includes the temperature characteristics of heat input uniformity and thermal defects.

[0017] Process parameter acquisition unit: An OPC UA protocol interface module is deployed in the industrial control box next to the testing station. It is connected to the equipment PLC via hardwire to collect process parameters that are highly relevant to quality. The sampling frequency is 30Hz, and time series data is generated with a total of 100 sampling points, including timestamps, accurate to milliseconds. The data is stored according to product ID, parameter name, timestamp, and numerical format to reflect the dynamic fluctuations of the process.

[0018] Actual quality value acquisition unit: The off-line station is provided with an on-line detection device: a three-coordinate measuring instrument is used to detect the dimensional error of the machined part; a tensile testing machine is used to detect the tensile strength of the welded part; an insulation resistance meter is used to detect the insulation performance of the electronic component. The detection personnel record the product ID, detection item, measured value and time as the reference true value for quality judgment.

[0019] It should be further explained that the synchronous data acquisition process is as follows: A PTP precision clock server is deployed beside the station, with an accuracy of ±100 ns, which binds the clock sources of all sensors, PLCs and acquisition equipment, to ensure that the timestamp error is ≤1 ms; a 20 ms period trigger signal is generated by the FPGA to control the start and stop synchronization of the sensor and parameter acquisition.

[0020] It should be further explained that the specific process of the preprocessing submodule is as follows: Data cleaning: 2D image cleaning: Zhang's calibration method (10x10 chessboard, grid distance 20mm) is used to correct distortion (pixel error ≤0.5); Otsu algorithm is used to segment the foreground (product area extraction accuracy ≥99%); CLAHE algorithm is used to enhance the contrast of low light area.

[0021] 3D point cloud cleaning: statistical filtering (neighborhood point = 50, standard deviation = 3) is used to remove outliers (integrity ≥98%); voxel filtering (0.05mm³) is used to downsample to 500,000 points; based on the reference hole coordinates, it is aligned to the design coordinate system (error ≤0.02mm).

[0022] Hyperspectral data cleaning: whiteboard with reflectivity of 99% and dark current are used to correct reflectivity; wavebands with SNR≥30 are selected, and 80-100 are retained; Gaussian filtering (σ=1.0) is used to smooth the spectral curve and remove high-frequency noise.

[0023] Temperature field cleaning: subtract the ambient temperature field to eliminate interference; threshold segmentation to retain the effective area.

[0024] Semantic alignment: Defect area positioning: manually annotate the bounding box (x1, y1, x2, y2) and type label (10 typical defects) of the defect in the 2D image; train the RPN network (IOU≥0.85) to automatically generate defect region coordinates, and guide the 3D point cloud, hyperspectral and temperature field data to the same physical area (spatial deviation ≤1mm).

[0025] It should be further explained that after preprocessing the original data collected, standardized data blocks are generated, which specifically include: 2D image block: cropped to 256x256x3, pixel value normalized to 0-1; 3D point cloud patches: sampled to 1024 points, coordinates normalized to [-1, 1]; Hyperspectral cube: cropped to 64x64x30, reflectance normalized to 0-1; Temperature field matrix (hot working scenario): cropped to 256x256, temperature values normalized to [-1, 1].

[0026] Additional information: product ID, timestamp, defect label, process parameter time series P(t), actual mass value, packed as an HDF5 file, directly input into the feature extraction and hybrid vector generation module.

[0027] Feature extraction and hybrid vector generation module: based on the standardized data block, physical information features and data-driven features are extracted, and after attention fusion, they are partitioned and then spliced with physical information features and data-driven features to generate a hybrid vector; based on the hybrid vector, a network model is trained, and a comprehensive quality index CQI and a CQI time series are output; It needs to be further explained that the feature extraction and hybrid vector generation module includes a physical information feature extraction unit, a data-driven feature extraction unit, and a hybrid vector generation unit. It needs to be specifically explained in this embodiment that the physical information feature extraction unit includes thermal physical feature extraction and geometric physical feature extraction. It needs to be further explained that the thermal physical feature extraction process is as follows: Input the temperature field matrix of 256x256 after preprocessing by the multi-modal data acquisition and preprocessing module, with temperature values normalized to [-1, 1], for 10 consecutive frames.

[0028] Preprocessing operation: perform two-dimensional Fourier transform on each frame of temperature field matrix to obtain frequency domain features (including spatial frequency components of temperature distribution); calculate the temperature gradient (horizontal direction ∂T / ∂x, vertical direction ∂T / ∂y), peak temperature (the highest temperature value in the region), and temperature uniformity index (standard deviation / mean) of each frame.

[0029] Network training: use a 3D convolutional network, with the input being a 4D tensor (256x256x10x1) stacked by 10 frames of temperature field matrix; Network structure: 4 layers of 3D convolutional layers, with a convolution kernel of 3x3x3, a step size of 1, padding=1, followed by a BN layer and a ReLU activation function after each layer, and finally outputting a 128-dimensional feature vector through global average pooling; Training data: 1000 sets of temperature field sequences of hot working products (including 500 sets of normal and defective samples); Physical constraints: add a heat conduction equation regularization term to the loss function, and its specific mathematical function is as follows: Here, 0.1 is a weighting coefficient used to balance the ratio of physical constraint loss to other model losses, preventing physical constraints from excessively suppressing the learning of data features. Let T be the partial derivative of temperature T with respect to time t, which physically represents the rate of change of temperature over time. The thermal diffusivity of the material, Let T be the Laplace operator for temperature T.

[0030] Output features: 128-dimensional thermophysical feature vectors, including features with clear physical meanings such as heat conduction efficiency (temperature diffusion rate), molten pool stability (peak temperature fluctuation), and heat input uniformity (gradient distribution).

[0031] It needs to be further explained that the specific process of geometric physical feature extraction is as follows: Input: 3D point cloud fragments output by the multimodal data acquisition and preprocessing module; Physical parameter calculations: Extract crack length a, depth c, dimensional deviation δ, and surface roughness, and calculate the stress concentration factor. , The radius of curvature at the tip; Network training: A fully connected network is used, with inputs including crack length, depth, dimensional deviation, surface roughness, stress concentration factor, and point cloud density. The network structure consists of 3 fully connected layers (6, 64, 128, 128), each followed by a BN layer and a ReLU activation function. Training data: 3D point clouds of 1000 products and corresponding online detection physical parameters; Loss function: mean squared error (the deviation between predicted and measured parameters); Training metric: Validation set parameter prediction error ≤ 5%.

[0032] Output features: 128-dimensional geometric and physical feature vectors, including fatigue risk factor R and structural integrity index (based on the ratio of dimensional deviation to design tolerance, with a value range of 0-1). Physical feature fusion operation: Concatenate the 128-dimensional thermophysical feature vector with the 128-dimensional geometric physical feature vector to form a 256-dimensional physical information feature vector. ; Special handling: In non-thermal processing scenarios, only geometric and physical features are used, and the 256 dimensions are supplemented by zero padding; Failure acceleration feature extraction: from Extract failure acceleration features from the last 64 dimensions It reflects the crack propagation rate and structural failure trend, with a value range of 0-1 (learned from historical failure data through an SVM model). It should be further explained that the fatigue risk factor is calculated as follows: Calculation of actual working stress amplitude based on process parameters acquired by the multimodal data acquisition and preprocessing module Further adjust the actual stress amplitude: Based on the material SN curve formula: Calculate the theoretical fatigue life under the corresponding stress amplitude: Then calculate the life safety factor. Finally, the fatigue risk factor was obtained: , where m and C are the fatigue strength index and material fatigue constant, respectively, both of which are inherent properties of the material and are measured experimentally, and N is the number of stress cycles.

[0033] This embodiment requires a detailed explanation of the data-driven feature extraction unit process as follows: Dataset construction: 10,000 samples were selected from the output of the multimodal data acquisition and preprocessing module, including 10 types of defects: scratches, cracks, dents, bumps, stains, missing characters, uneven plating, assembly misalignment, dimensional deviation, and material impurities. 1,000 samples were selected for each type and divided into training set and validation set in an 8:2 ratio. Data augmentation: Randomly rotate (±15°), randomly scale (0.8-1.2 times), and add Gaussian noise (σ=0.01) to the training set images to expand the sample size to 20,000.

[0034] Multi-branch network training: 2D Feature Branch (ResNet18): Input a 2D image block (256×256×3) into the multimodal data acquisition and preprocessing module; Network structure: Layer 1: A 7×7 convolutional layer with 64 channels, stride 2, and padding 3; followed by a BN layer and ReLU activation function with 3×3 max pooling and stride 2. Layers 2-5: 4 residual blocks, each containing 2 layers of 3×3 convolutions, with the number of channels being 64, 128, 256, and 512 respectively; Layer 6: Global average pooling and 256-channel 1×1 convolution; Output: 256-dimensional texture feature vector, including edge gradient direction, color distribution entropy, local binary mode (LBP) statistics, etc. Training objective: Defect classification using cross-entropy loss, validation set accuracy ≥ 98%.

[0035] 3D feature branches: Input: 3D point cloud fragments preprocessed by the multimodal data acquisition and preprocessing module (1024 points × 3 coordinates, normalized to [-1, 1]); Network structure: Layer 1: T-Net, 3x3 transformation matrix, align point cloud pose; Layer 2-4: MLP, 64, 128, 256 channels respectively, ReLU activation; Layer 5: Global max pooling with MLP (512, 256 channels, ReLU activation); Output: 256-dimensional geometric feature vector, including point cloud distribution entropy, normal vector direction, curvature change rate, etc. Training target: defect classification cross-entropy loss, verification set accuracy ≥97%.

[0036] Hyperspectral feature branch (3D convolutional network): Input hyperspectral cube of multi-modal data acquisition and preprocessing module, specification: 64x64x30, reflectivity normalized to 0-1; Network structure: Layers 1-3: 3D convolution (3x3x3 kernel, channel number 32, 64, 128, step 1, padding=1) connected with BN layer and ReLU activation function; Layer 4: 3x3x30 convolution, global average pooling, 256 channel 1x1 convolution; Output: 256-dimensional spectral feature vector, including feature band reflectivity, spectral angle mapper (SAM), and band correlation; The training target is the same as the 2D branch.

[0037] The embodiment needs to be specifically described as follows: Feature fusion and partitioning: Train attention weight network, input is 2D, 3D, and hyperspectral branch feature vectors, each 256-dimensional; Network structure: 3 layers of full connection (768, 256, 3), Softmax activation, output 3 weights w1, w2, w3, sum to 1; Training data: 1000 groups of sample branch features and corresponding defect labels; Training target: classification cross-entropy loss of fused features, make weight adaptive to defect type.

[0038] Fusion calculation: wherein, is a 256-dimensional texture feature vector, is a 256-dimensional geometric feature vector, is a 256-dimensional spectral feature vector; Feature partitioning: Divided into 3 sub-regions according to function: First 64 dimensions Process and quality correlation features, including features strongly correlated with process parameters of the multi-modal data acquisition and preprocessing module, screened by Pearson correlation coefficient, |r|≥0.7; Middle 128 dimensions Visual discrimination features, including discriminative features related to pure image texture, color, and shape; Last 64 dimensions Spectral material features, including spectral features related to material composition and oxidation degree; Mixed feature splicing: splice 256-dimensional data-driven features with 256-dimensional physical information features to obtain 512-dimensional mixed features ; Through t-SNE dimensionality reduction visualization, it is verified that the mixed features can be obviously clustered according to defect types in low-dimensional space, and the clustering purity of defects of the same type is ≥95%.

[0039] It should be further explained that the network model is trained based on the mixed vector, and the comprehensive quality index CQI is output, and the specific calculation process is as follows: Input data: 512-dimensional mixed features and process parameter time series of the multi-modal data acquisition and preprocessing module The process parameter time series is k-dimensional, and k is the number of parameters; Network training: Network structure: 3-layer fully connected network (512+k, 128, 64, 1), output layer with Sigmoid activation, mapped to 0-100; Training data: 8000 groups of samples of , and actual quality values of the multi-modal data acquisition and preprocessing module; Training target: mean square error (MSE) of CQI prediction value and actual quality value ≤5, Pearson correlation coefficient ≥0.95; Training optimizer: Adam optimizer, learning rate 0.001, decay rate 1e-5, batch size=32, epoch=50.

[0040] Output result: single-value CQI (0-100, the higher the value, the better the quality); CQI time series, aligned with the timestamp of the multi-modal data acquisition and preprocessing module, 100 points long, reflecting the dynamic changes of quality in the processing process.

[0041] Quality Field Energy Coupling Decision Module: Based on the output of the preceding module, it calculates the quality toughness coefficient and the process fluctuation transmission coefficient, and defines the compliance energy field, the consistency energy field, and the risk energy field; based on historical data, it determines the energy transfer coefficient matrix and the field phase difference, couples and calculates the total energy, and generates the standardized comprehensive decision value QCI; based on the QCI and field characteristics, it outputs the quality decision.

[0042] This embodiment requires specific explanation of the process for defining the compliance energy field, consistency energy field, and risk energy field, as follows: Definition of mass-energy field: Dynamic parameter calculation: Mass toughness coefficient Reflecting the quality's ability to withstand process fluctuations: The variance of the process parameter time series is calculated based on the process parameter time series, denoted as... Process and quality correlation features based on process parameter time series variance, actual quality values, and feature extraction and mixed vector generation modules. Calculate the mass toughness parameters: , ;in, for The Pearson correlation coefficient with the actual quality value is taken as the absolute value, and the range is [0, 1]. Process fluctuation transmission coefficient Quantifying the impact of process fluctuations on quality: Standard deviation of process parameter time series data based on multimodal data acquisition and preprocessing module. CQI standard deviation compared to feature extraction and hybrid vector generation modules as well as right The Granger causality coefficient (0-1) is used to calculate the process fluctuation transmission coefficient, and its specific function is: If ,but ,otherwise , .

[0043] Energy field calculation: Compliant Energy Field : Its value range is [0, 2], and it is related to CQI and process stability. The larger the value, the higher the quality compliance. Consistent energy field This reflects the consistency of product quality with the average level within the same batch and between different batches; a higher value indicates better consistency. Its function is: ,in, This represents the deviation of the CQI from the average of the top 50 products in the same batch. The deviation of CQI from the mean across 3 batches; Risk energy field : The value closer to 0 indicates lower risk.

[0044] It needs to be further explained that the energy coupling and QCI generate its specific process as follows: Calculate the energy transfer coefficient matrix (K): based on the 1000 times of abnormal record history quality accident data training of the multi-modal data acquisition and preprocessing module, the matrix element , represents the energy influence strength between fields: When ≤-5, the risk is strongly weakened by compliance energy; When ≤0.5, the consistency difference amplifies the risk energy; When ≥1.5, high compliance enhances the consistency energy; The rest , , are respectively: 0.3, 0.2, 0.1.

[0045] Calculate the field phase difference : input the , , sequence of the last 10 products, calculate the change rate of each field strength ; The calculation function is: , , reflecting the synchronization of field changes, is in-phase enhancement, is anti-phase cancellation.

[0046] Total energy and QCI calculation: Total coupling energy: ; QCI standardization: statistics of the minimum value of the total energy of the history and the maximum value of the best quality ; , , the larger the value, the better the comprehensive quality; The embodiment needs to be specifically explained as follows: Based on the QCI and energy field characteristics, the output is a clear quality judgment result, guiding the production circulation.

[0047] Decision rules: High-quality level: QCI≥85: ≥1.8, >=1.5, >=-1.0, output direct release instruction, marked as benchmark product; Qualified level: 70<=QCI<85: >=1.2, >=1.0, >=-3.0, output normal release instruction, generate quality report; To be checked level: 50<=QCI<70: there is , indicating poor consistency amplification risk, output suspension circulation instruction, and directional review; Reject level: QCI<50: , output scrap instruction, and start root cause tracing.

[0048] Secondly: the drawings in the disclosed embodiments of the application only involve structures related to the disclosed embodiments of the application, other structures can refer to general design, and in the case of no conflict, the same embodiment and different embodiments of the application can be combined with each other; Finally: the above only describes preferred embodiments of the application and is not used to limit the application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application should be included in the protection scope of the application.

Claims

1. A machine vision and AI based online detection system for quality of production in a workshop, characterized in that, Comprise: Multimodal data acquisition and preprocessing module: synchronous acquisition of physical signals, process parameters, actual quality values of target products, time stamp consistency is ensured through clock synchronization and trigger control; After preprocessing the collected raw data, standardized data blocks are generated; Feature extraction and mixed vector generation module: based on the standardized data blocks, physical information features and data-driven features are extracted, partitioned after attention fusion, and then the physical information features and data-driven features are spliced to generate mixed vectors; Train the network model based on the mixed vectors, output the comprehensive quality index CQI and the CQI time series; Quality field energy coupling decision module: based on the output of the previous module, calculate the quality resilience coefficient and process fluctuation conduction coefficient, and define the compliance energy field, consistency energy field and risk energy field; Based on historical data, determine the energy transfer coefficient matrix and field phase difference, coupled calculation of total energy and generation of standardized comprehensive decision value QCI; According to QCI and field characteristics, output quality decision.

2. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The multimodal data acquisition comprises: Based on the 2D surface image acquisition unit, the surface texture information and the shape and position information of the planar defects of the target product are collected; based on the 3D geometric morphology acquisition unit, the three-dimensional geometric morphology information and the spatial feature information of the three-dimensional defects of the target product are collected; based on the hyperspectral material quality acquisition unit, the surface spectral reflectivity information and the spectral feature information of the material composition difference and the hidden sensory defects of the target product are collected; based on the temperature field acquisition unit, the temperature distribution information and the thermal input uniformity and thermal defect temperature feature information of the target product hot processing area are collected; based on the process parameter acquisition unit, the process parameter time series information closely related to the quality of the target product is collected; based on the actual quality value acquisition unit, the quality detection value information reflecting the actual performance of the target product is collected. 3.The machine vision and AI based online detection system for production quality in a factory as claimed in claim 1 wherein: The preprocessing comprises: Clean each acquisition unit data respectively: 2D image distortion correction, foreground segmentation and contrast enhancement; 3D data outlier removal, downsampling and coordinate alignment; hyperspectral data reflectance correction, effective band selection and smoothing; temperature data environment temperature elimination and effective region segmentation; then based on artificial annotation training region proposal network to locate defect area, guide multimodal data spatial semantic alignment; finally, the aligned data is cropped and sampled to fixed dimension and normalized to generate standardized data blocks.

4. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The physical information features comprise: Based on the temperature field data in the standardized data blocks, two-dimensional Fourier transform is performed on each frame of temperature field data, and temperature gradient, peak temperature and temperature uniformity index are calculated; 128-dimensional thermal physical features are extracted through a 3D convolution network containing a thermal conduction equation regular term, which includes thermal conduction efficiency, molten pool stability and thermal input uniformity; Based on the 3D point cloud segment in the standardized data blocks, crack length, depth, size deviation, surface roughness and stress concentration coefficient are calculated; 128-dimensional geometric physical features are extracted through a fully connected network, which includes fatigue risk factor and structure integrity index; The thermal physical characteristics and the geometric physical characteristics are spliced to form a 256-dimensional physical information characteristic vector, the non-thermal processing scene only uses the geometric physical characteristics and is supplemented to 256 dimensions by zero padding, and a failure acceleration feature reflecting crack propagation rate and structure failure trend is extracted from the physical information characteristic vector.

5. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The data-driven feature includes: Samples are selected from the standardized data blocks output by the multi-modal data acquisition and preprocessing module, training sets and validation sets are divided, and data augmentation is performed; Texture features are extracted from 2D image blocks, geometric features are extracted from 3D point cloud segments, and spectral features are extracted from hyperspectral cubes through corresponding modal feature extraction networks; An attention weight network is trained to weight and fuse the above three types of features to obtain a data-driven feature vector; the data-driven feature vector is divided into functionally partitioned process and quality associated features that are strongly related to process parameters, visual discriminant features that reflect image discriminability, and spectral material features that reflect material properties, providing data-driven feature inputs for hybrid vector generation.

6. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The comprehensive quality index includes: A 3-layer fully connected network is trained with the hybrid vector generated in the feature extraction stage and the process parameter time series output by the multi-modal data acquisition and preprocessing module as input; the output layer of the network maps the result to the range of 0-100 through Sigmoid activation; The training process aims to make the mean square error between the CQI prediction value and the actual quality detection value less than or equal to 5, and the Pearson correlation coefficient greater than or equal to 0.95; The final output is a single-value CQI and a CQI time series aligned with the process parameter timestamp.

7. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The defined compliance energy field, consistent energy field, and risk energy field include: First, based on the process parameter variance, actual quality value output by the multi-modal data acquisition and preprocessing module, and the process and quality correlation characteristics output by the feature extraction and mixed vector generation module, the quality toughness coefficient is calculated ; Based on the process parameter standard deviation output by the multi-modal data acquisition and preprocessing module, and the CQI standard deviation and the Granger causality coefficient of process parameters on CQI output by the feature extraction and mixed vector generation module, the process fluctuation conduction coefficient is calculated ; Recombine , and CQI, define a compliance energy field reflecting quality compliance; combine , and CQI intra-batch bias Z, inter-batch bias , define a consistency energy field reflecting quality consistency; In combination A risk energy field quantifying the quality risk is defined in combination with the fatigue risk factors and failure acceleration features output by the feature extraction and mixed vector generation module.

8. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The comprehensive decision value includes: Based on historical quality accident data, an energy transfer coefficient matrix is determined to depict the influence strength between each energy field; the change rates of the compliance energy field, the consistent energy field, and the risk energy field of the continuous target product are calculated to obtain the field phase difference reflecting the change synchronization of each field; then the compliance energy field, the consistent energy field, and the risk energy field are coupled with the energy transfer coefficient matrix and the field phase difference to obtain the total energy; finally, the optimal value and the worst value of the historical total energy are calculated, and the current total energy is normalized to generate a comprehensive decision value.

9. The machine vision and AI-based online detection system for workshop production quality according to claim 1, characterized in that: The output quality decision includes: with the integrated decision value QCI, the compliance energy field , the consistent energy field , the risk energy field , the energy transfer coefficient matrix in and the historical total energy extreme value as the basis for judgment, the quality grade is divided and the corresponding decision is output: If QCI ≥ 85 and ≥ 1.8, ≥ 1.5, ≥ -1.0, it is determined as high quality level, output direct release instruction and mark as benchmark product; if 70 ≤ QCI < 85 and ≥ 1.2, ≥ 1.0, ≥ -3.0, it is determined to be a qualified level, a normal release instruction is output, and a quality brief is generated; If 50≤QCI<70 and there exists , it is determined as the level to be investigated, the suspension circulation instruction is outputted and the directional review is triggered, is and the field phase difference; If QCI < 50 and , the decision is to reject level, output scrap instruction and start root cause tracing; is the current total energy, is the historical total energy worst value, is the historical total energy optimal value.

Images