Automobile foot mat surface defect defect tracing method

Abstract

The invention provides an automobile foot mat surface defect defect traceability method, which comprises the following steps of: synchronously acquiring and performing time-space alignment on multi-source process parameter time sequence data and a surface image for marking a defect type, constructing a structured and standardized training data set, extracting a defect morphological feature sequence, and extracting a defect morphological feature sequence; establishing a standardized process track library by adopting methods such as dynamic time warping and clustering; in combination with a variational auto-encoder-long and short-term memory network, a process track hidden space model is established, and correlation modeling of defect features and process parameter tracks is realized; a reverse generation type reasoning module is designed, defect features are input into a hidden space and optimized to generate candidate root cause trajectories, and a root cause hypothesis chain and a traceability report with engineering significance are output through physical, time sequence and cooperative constraint verification of a process knowledge base. The process diagnosis accuracy and interpretability are improved, and a scientific basis is provided for process parameter optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of quality management and defect traceability in the automobile manufacturing process, and in particular to a method for tracing surface defects of automobile floor mats. BACKGROUND

[0002] Currently, quality control and defect traceability technology in the automobile manufacturing process has been widely applied in mass production, especially in the surface quality monitoring of interior parts such as automobile floor mats, and certain technical accumulation has been formed. The existing mainstream solutions mainly rely on static process parameter modeling and surface visual detection methods. By collecting key process parameters (such as pressing pressure, temperature, mold closing speed) and product surface images, using discriminant models, correlation analysis or manual rules, the static relationship between abnormal factors and defect types is identified. These methods are usually based on single-point upper and lower limit setting, abnormal statistics or pre-set process formula of process parameters, combined with offline quality traceability and process control, to achieve a certain degree of defect source investigation and quality improvement. The industry development trend gradually moves towards multi-source data fusion and intelligent causal analysis, exploring new quality management models that combine data-driven deep learning models with process knowledge.

[0003] In the prior art, static process parameter modeling is the most common way, that is, it is assumed that product defects are mainly caused by some static parameter values exceeding the standard range, so parameter threshold, abnormal distribution or correlation matrix are established to assist in traceability. However, the real-time dynamic changes of process parameters in the production process are ignored. Product surface defects such as edge burrs, texture breakage, color difference, etc. are often caused by parameter fluctuations and disturbances in the process, rather than a single value exceeding the limit. For example, the abnormal increase of pressing pressure in stages, the instantaneous jitter of temperature curve, the too fast or delayed closing speed of the mold, etc. may all lead to the generation of surface defects of the floor mat, but these dynamic behaviors are difficult to accurately track and infer under traditional static analysis.

[0004] Currently, the research on dynamic process parameter modeling in the industry is still in the initial exploration stage, and most defect traceability methods have not effectively introduced time series modeling or dynamic trajectory analysis. Some technologies attempt to use causal graph, multi-modal tensor, attention mechanism, etc. to realize complex causal relationship reasoning, but often have problems such as complex modeling, high computational cost, and high real-time requirements in the field. In addition, most of these solutions rely on explicit causal structure definition or multi-modal data strong fusion, lack of universality and flexibility, and are difficult to adapt to specific scenarios such as periodic equipment maintenance, process variation between batches, and equipment response lag in actual production environment. SUMMARY

[0005] The present application provides a method for tracing surface defects of automobile floor mats to solve the above technical problems.

[0006] The technical solution of this invention is implemented as follows: A method for tracing the source of defects on the surface of car floor mats, comprising:

[0007] S1: Synchronously collect time-series data of multi-source process parameters and corresponding timestamp surface image data during the production of car floor mats. The process parameters include pressing pressure, temperature curve, and mold closing speed. The image data is labeled with defect type labels to establish a spatiotemporally aligned training dataset.

[0008] S2: Perform morphological feature extraction on the surface image data to generate a defect feature sequence. The features include edge irregularity, texture fracture density and color difference distribution gradient, forming a feature vector sequence aligned with the process parameter time series data.

[0009] S3: Based on multi-source process parameter time series data, a process trajectory alignment model is constructed using the Dynamic Time Warping (DTW) algorithm. By calculating the optimal alignment path of process parameter sequences of different batches, a standardized process action segment library is generated.

[0010] S4: Construct a latent space model of the process trajectory based on a variational autoencoder-long short-term memory network (VAE-LSTM) architecture. The encoder generates latent dynamic features based on the aligned process parameter sequence, and the decoder reconstructs the process trajectory. During training, the correlation mapping relationship between the defect feature sequence and the latent features is optimized simultaneously.

[0011] S5: Design a reverse generative inference module, input specific defect feature sequences into the latent space of the trained VAE-LSTM model, generate dynamic paths of process parameters in reverse through gradient optimization search strategy, and output candidate root cause trajectories containing combinations of key parameter deviations and their time fluctuation patterns.

[0012] S6: Perform feasibility verification on candidate root cause trajectories based on process knowledge constraints, including physical realizability boundary constraints, equipment response delay characteristic matching, and process parameter interaction rule verification, to generate a root cause hypothesis chain with engineering significance;

[0013] S7: Generate a defect tracing report based on the verified root cause hypothesis chain, which includes the dynamic fluctuation range and time axis location information of the key process parameters that caused the defect, and is used to guide parameter adjustment and quality improvement in the production process.

[0014] The present invention provides a method for tracing the source of defects on the surface of car floor mats, which has the following beneficial effects:

[0015] (1) This invention constructs a spatiotemporally aligned multi-source process dataset and introduces dynamic time warping technology to achieve standardized representation of cross-batch process processes, effectively overcoming the asynchronous problem of time series caused by differences in production rhythm; combined with the VAE-LSTM architecture to establish a "process trajectory latent space", it can deeply mine the potential dynamic patterns of key parameters such as pressing pressure and temperature curves in the time dimension, significantly improving the modeling ability of nonlinear and non-stationary process evolution processes, thus providing a high-fidelity feature expression basis for subsequent root cause tracing;

[0016] (2) The present invention innovatively designs a reverse generative reasoning module, which uses the actual observed defect morphological feature sequence as the guiding input. By optimizing the latent vector search strategy, it reverse-engineers the dynamic path of the abnormal parameters most likely to induce the defect in the process latent space. It not only identifies the key parameter combination that deviates from the standard, but also accurately restores its fluctuation sequence on the time axis (such as abnormal increase in heating rate, premature end of the pressure holding stage, etc.), realizing the reconstruction of the causal chain from "result to process". This mechanism breaks through the limitation of traditional diagnostic methods that can only locate isolated parameter over-limit points, and enhances the fine-grained analysis capability of gradual and complex process disturbances.

[0017] (3) The present invention integrates process knowledge constraints (including physical realizability boundaries, equipment response delay characteristics, etc.) to perform feasibility verification and semantic enhancement on the generated abnormal paths, filters out infeasible or unreasonable parameter combinations, and outputs root cause hypothesis chains that conform to the actual situation on site, thus avoiding engineering paradoxes that may occur in pure data-driven models. The entire method framework does not require the construction of explicit causal graphs, attention weight allocation or multimodal tensor fusion structures, nor does it require the deployment of complex real-time closed-loop control systems. It can run efficiently in offline environments, taking into account both computational efficiency and analytical depth, and is particularly suitable for quality traceability tasks in small-batch, multi-variety production scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for tracing the source of surface defects in car floor mats according to the present invention;

[0019] Figure 2 This is a sub-flowchart of a method for tracing the source of surface defects in car floor mats according to the present invention;

[0020] Figure 3 This is another sub-flowchart of a method for tracing the source of surface defects in car floor mats according to the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1 As shown, this invention provides a method for tracing the source of defects on the surface of car floor mats, specifically including:

[0024] S1: Synchronously collect time-series data of multi-source process parameters and corresponding timestamp surface image data during the production of car floor mats. The process parameters include pressing pressure, temperature curve, and mold closing speed. The image data is labeled with defect type labels to establish a spatiotemporally aligned training dataset.

[0025] S2: Perform morphological feature extraction on the surface image data to generate a defect feature sequence. The features include edge irregularity, texture fracture density and color difference distribution gradient, forming a feature vector sequence aligned with the process parameter time series data.

[0026] S3: Based on multi-source process parameter time series data, a process trajectory alignment model is constructed using the Dynamic Time Warping (DTW) algorithm. By calculating the optimal alignment path of process parameter sequences of different batches, a standardized process action segment library is generated.

[0027] S4: Construct a latent space model of the process trajectory based on a variational autoencoder-long short-term memory network (VAE-LSTM) architecture. The encoder generates latent dynamic features based on the aligned process parameter sequence, and the decoder reconstructs the process trajectory. During training, the correlation mapping relationship between the defect feature sequence and the latent features is optimized simultaneously.

[0028] S5: Design a reverse generative inference module, input specific defect feature sequences into the latent space of the trained VAE-LSTM model, generate dynamic paths of process parameters in reverse through gradient optimization search strategy, and output candidate root cause trajectories containing combinations of key parameter deviations and their time fluctuation patterns.

[0029] S6: Perform feasibility verification on candidate root cause trajectories based on process knowledge constraints, including physical realizability boundary constraints, equipment response delay characteristic matching, and process parameter interaction rule verification, to generate a root cause hypothesis chain with engineering significance;

[0030] S7: Generate a defect tracing report based on the verified root cause hypothesis chain, which includes the dynamic fluctuation range and time axis location information of the key process parameters that caused the defect, and is used to guide parameter adjustment and quality improvement in the production process.

[0031] Step S1: Synchronously collect time-series data of multi-source process parameters and corresponding timestamp surface image data during the production of automotive floor mats. The process parameters include pressing pressure, temperature profiles, and mold closing speed. The image data is labeled with defect type tags to establish a spatiotemporally aligned training dataset. Specifically, this includes:

[0032] S1.1: Based on the multi-source sensors deployed on the automotive floor mat manufacturing production line, raw time-series data of pressing pressure, temperature curves, and mold closing speed process parameters are collected to obtain key dynamic control variables in the manufacturing process;

[0033] Based on the multi-source sensor array already deployed on the automotive floor mat manufacturing production line, raw time-series data of pressing pressure, temperature curve, and mold closing speed process parameters are collected as input conditions.

[0034] A high-precision strain gauge pressure sensor (range: 0~5MPa, resolution: 0.01MPa) is used to collect the instantaneous pressure value during the pressing process and continuously record it at a sampling frequency of 500Hz to form a high time resolution signal sequence of pressing pressure.

[0035] Furthermore, by using a platinum resistance temperature sensor (Pt100, accuracy: ±0.1℃) in conjunction with a multi-channel temperature acquisition module, temperature curves at different locations in the mold cavity are obtained, with the sampling frequency set to 100Hz, in order to capture subtle fluctuations during the heating, holding, and cooling stages.

[0036] Furthermore, the instantaneous speed of the mold closing mechanism is measured using a grating encoder with a resolution of 0.001 mm / pulse. The closing speed curve is recorded in conjunction with the timestamp of the real-time control system, and the sampling frequency is set to 200 Hz to reflect the dynamic mechanical motion characteristics of the molding process.

[0037] A multi-channel data acquisition card (supporting synchronous trigger mode) is used to perform hardware-level time synchronization triggering on the above pressure, temperature and closing speed signals to ensure strict consistency of the output signals of different sensors in physical time;

[0038] Furthermore, through sensor calibration algorithms (parameters: zero-point offset correction coefficient, proportional coefficient), the amplitude of the original signal is corrected, systematic errors caused by ambient temperature drift or sensor aging are eliminated, and the original data stream of process parameters after physical quantity measurement is obtained.

[0039] The raw data stream is filtered by a bandpass filter (cutoff frequency: 0.1Hz~50Hz) to remove high-frequency noise and low-frequency drift components, thereby obtaining a more stable dynamic control variable record and achieving high-fidelity acquisition of the process trajectory.

[0040] Through the above acquisition and processing chain, the raw signals of various process parameters are transformed into high-precision, multi-source time series data that can be directly entered into the subsequent spatiotemporal alignment steps, so as to realize the complete acquisition of dynamic control variables.

[0041] For example, on a car floor mat pressing production line, the pressure sensor is configured with a range of 0~5MPa and a sampling frequency of 500Hz, recording a peak pressure of 3.25MPa during the pressing stage; the temperature sensor has a sampling frequency of 100Hz, and during the heating stage, the temperature curve rises from 25℃ to 145℃ within 15 seconds, while during the heat preservation stage, it fluctuates between 143~145℃; the closing speed curve of the grating encoder shows a speed of 0.12m / s at the beginning of the pressing stage, decelerating to 0.02m / s at the end of the closing stage to avoid impact. The collected data is synchronously triggered to ensure the consistency of the timestamps of the three types of signals. After zero-point drift correction, the average offset of the pressure signal is reduced from 0.08MPa to 0.005MPa, and after bandpass filtering, the noise amplitude is reduced to 1 / 5 of the original. The final output process parameter raw time series data significantly improves the capture of dynamic fluctuations in subsequent timestamp alignment and feature sequence modeling, supporting the root cause reasoning module to accurately infer the cause of defects.

[0042] S1.2: Use an industrial camera to collect surface image data at a designated inspection station after the foot pad is formed, and label the defect type in the image to form an image sample set with semantic information;

[0043] The input image acquisition object is the output part of the inspection station after the car floor mat is formed. The physical location of the inspection station is precisely configured through the production line layout and process parameter acquisition synchronization signal, so that the image acquisition and process parameter recording are correlated in time.

[0044] A fixed industrial camera acquisition method (parameters: resolution ≥ 5MP, frame rate ≥ 15fps, lens focal length matched with workstation distance) is used to achieve high-precision, low-distortion full-frame imaging of the foot pad surface;

[0045] Furthermore, by using a region gating algorithm (parameter: the range of ROI coordinates of the detection area is obtained by calibration of the workstation reference point), effective pixel acquisition is achieved for the entire surface area of ​​the foot pad, and background pixels of the workstation environment are removed to obtain position-standardized foot pad surface image data;

[0046] Furthermore, a color calibration algorithm (parameters: based on the CIE Lab color space, using a standard color calibration palette to establish a color mapping matrix) is used to correct the color consistency of the acquired images and obtain a corrected image that meets the color accuracy requirements of subsequent defect analysis.

[0047] Furthermore, through a defect labeling algorithm (parameters: combining an automatic detection model based on a convolutional neural network with manual verification; the labeling system includes categories such as scratches, wrinkles, bubbles, and color differences), the contours of defect areas on the collected floor mat surface images are located and their categories are assigned, and an array of defect type labels is generated.

[0048] By using the above-mentioned industrial camera acquisition and defect labeling processing method, the visual observation results of the previous step are transformed into an image sample set with timestamps and semantic information, so as to realize the basic data input effect of image data and process parameter data in the subsequent S1.3 timestamp alignment.

[0049] For example, in a certain production line configuration, a fixed industrial camera is installed at the inspection station. The camera has a resolution of 2448×2048 pixels, a frame rate of 20fps, a lens focal length of 12mm, and is 0.8m away from the foot pad surface. After calibration of the station's reference point, the ROI coordinate range is (200, 150) - (2200, 1900) pixels. Using the CIE Lab color space, a color mapping matrix is ​​established between the measured values ​​of the standard color chart and the camera's original values. The mapping formula is as follows:

[0050]

[0051] in, For the calibrated luminance components, The original luminance component, and The images were obtained by fitting a color calibration chart. The automatic defect detection model is based on the ResNet50 architecture with a confidence threshold of 0.85, and the labels were manually verified. In this example, 500 images were collected from a single batch of production lines, and the defect categories were distributed as follows: scratches (120 images), wrinkles (90 images), bubbles (60 images), and color differences (230 images). All image data were uniformly associated with the corresponding pressing pressure, temperature curves, and mold closing speed parameters timestamps to form a high-quality, spatiotemporally aligned image sample set.

[0052] S1.3: Based on the time synchronization mechanism, the time-series data of the acquired process parameters and the surface image data are timestamped to construct a multimodal data pair that is strictly synchronized in the time dimension;

[0053] S1.4: Perform data cleaning operations on the timestamp-aligned multimodal data, including removing outliers, filling in missing data, and removing duplicate samples, to obtain a well-structured and reliable original training dataset.

[0054] S1.5: Based on the data standardization method, the time series data of the process parameters after cleaning are normalized to eliminate the differences in the dimensions of different sensors and provide a unified data scale basis for subsequent dynamic trajectory modeling;

[0055] S1.6: Perform image enhancement and preprocessing operations on the labeled image data, including grayscale conversion, filtering and noise reduction, contrast enhancement, etc., to improve image quality and reduce the impact of environmental noise on feature extraction;

[0056] S1.7: The normalized process parameter time series data and the preprocessed image data are stored and organized according to timestamps to construct a spatiotemporally aligned and structured training dataset, providing a data input basis for subsequent process trajectory alignment and defect feature sequence modeling.

[0057] Step S2: Perform morphological feature extraction on the surface image data to generate a defect feature sequence. The features include edge irregularity, texture breakage density, and color difference distribution gradient, forming a feature vector sequence aligned with the time-series process parameter data. Specifically, this includes:

[0058] S2.1: Acquire surface image data, which are image samples collected and labeled with defect type tags during the production process of car floor mats, as the input data source for subsequent image feature extraction;

[0059] S2.2: Perform image preprocessing operations on the surface image data, including grayscale conversion, Gaussian filtering noise reduction, and contrast enhancement, to improve the boundary clarity and texture recognizability of defect areas in the image.

[0060] S2.3: Based on the Canny edge detection algorithm, perform edge extraction on the preprocessed image to obtain the edge contour image of the defects on the surface of the foot pad, which can be used for subsequent quantitative calculation of edge irregularity features;

[0061] S2.4: Perform morphological skeleton extraction and segmentation on the extracted edge contour image, use Freeman chain code to represent the edge direction, and calculate the edge irregularity index, which is the statistical mean of the chain code direction change frequency to quantify the irregularity of the defect edge;

[0062] S2.5: Based on the texture analysis method, Gabor filter bank response calculation is performed on the preprocessed image to extract local texture features of the defect area and to count the texture break density index. The index is the density of pixels with texture response energy lower than a set threshold, which characterizes the degree of damage to texture continuity.

[0063] S2.6: Perform color difference analysis on the RGB color space of the image, convert the image to the Lab color space based on the CIE Lab color space conversion formula, and calculate the ΔE color difference distribution gradient between the defect area and the normal area to quantify the degree of color deviation in the defect area.

[0064] S2.7: Normalize the three feature indicators of edge irregularity, texture break density, and color difference distribution gradient to form a multi-dimensional feature vector, and align it spatiotemporally with the image acquisition timestamp and process parameter time series data to generate a defect feature sequence for input to the subsequent process trajectory modeling and root cause reasoning module.

[0065] The edge irregularity, texture break density, and color difference distribution gradient indices calculated in steps S2.4 to S2.6 are used as the input feature set. The Z-score normalization method (parameters: mean μ, standard deviation σ) is used to achieve scale uniform processing of feature indices with different dimensions and numerical ranges.

[0066] Furthermore, the normalized eigenvalues ​​are calculated using the following formula. :

[0067]

[0068] in, These are the original eigenvalues. This is the mean of the feature in the sample set. The standard deviation of this feature in the sample set;

[0069] Furthermore, minimum-maximum scaling (parameters: minimum value min, maximum value max) is performed on the three normalized features respectively, mapping the feature values ​​to the [0,1] interval using the following formula:

[0070]

[0071] in, These are the scaled original feature values. and These are the minimum and maximum values ​​of the feature in the sample set, respectively.

[0072] Furthermore, a feature vector concatenation method is adopted (parameters: concatenation order is edge irregularity, texture break density, color difference distribution gradient) to merge the three types of feature data in the same feature space and generate a multi-dimensional feature vector of length n×3, where n is the time series length;

[0073] Furthermore, by using a timestamp index matching algorithm (parameter: time tolerance Δt), the multidimensional feature vector is spatiotemporally aligned with the process parameter time series data constructed in step S1.7, thereby achieving registration processing of different data sources on a unified time axis;

[0074] Through the above normalization, scaling, splicing and spatiotemporal alignment processing methods, the feature extraction results of the previous step are transformed into a defect feature sequence that meets the input requirements of subsequent process trajectory modeling, so as to achieve consistency of feature dimensions and accuracy of time positioning.

[0075] For example, the original mean of edge irregularity for a batch of car floor mat samples was 0.45, with a standard deviation of 0.05; the original mean of texture breakage density was 18.0, with a standard deviation of 2.0; and the original mean of color difference distribution gradient was 5.5, with a standard deviation of 0.5. After Z-score normalization, the edge irregularity feature values... Calculated = Texture fracture density eigenvalue Calculated = ; Color difference distribution gradient eigenvalues Calculated = Then, minimum-maximum scaling is performed. Assuming edge irregularity min = 0.40 and max = 0.60, the normalized value is... = Texture breakage density min=15.0, max=22.0, scaling value is = The color difference distribution gradient is set to min=5.0 and max=7.0, with a scaling value of [value missing]. = Finally, the three scaling values ​​are concatenated in a preset order to form a feature vector of [0.5, 0.714, 0.5], and aligned with the corresponding process parameter data using index matching at a timestamp Δt=0.01s. During the verification phase, this defect feature sequence and the dynamic trajectory of the process parameters achieve strict registration in the time dimension, significantly improving the consistency of feature scales and meeting the input requirements for subsequent VAE-LSTM latent space modeling.

[0076] like Figure 2As shown, step S3 involves constructing a process trajectory alignment model based on multi-source process parameter time-series data using the Dynamic Time Warping (DTW) algorithm. This model generates a standardized process action fragment library by calculating the optimal alignment path for different batches of process parameter sequences. Specifically, this includes:

[0077] S3.1: Normalize variables such as pressing pressure, temperature curve, and mold closing speed in the multi-source process parameter time series data to eliminate the influence of the difference in the dimensions of different sensors on the dynamic time warping (DTW) calculation and obtain a standardized process parameter sequence.

[0078] S3.2: Based on the standardized process parameter sequence, a multidimensional DTW cost function is constructed, in which Euclidean distance is introduced as a point-to-point difference measure, and a time elastic constraint window is combined to control the nonlinear alignment range, and the optimal alignment path of the process parameter sequence between different production batches is calculated.

[0079] Based on the multi-source process parameter sequence data after S3.1 normalization, the multidimensional dynamic time warping (DTW) method (parameter dimensions include pressing pressure, temperature curve, and mold closing speed) is used to achieve nonlinear alignment calculation of cross-batch time sequence data.

[0080] Furthermore, a multidimensional DTW cost function is constructed, where the point-to-point difference measure at each time step is calculated using the Euclidean distance formula:

[0081]

[0082] in For time step and The Euclidean distance between them For the normalized process parameter values, Indicates the feature dimension index. and These are the current aligned time steps for the two time series, respectively.

[0083] Furthermore, a time-elastic constraint window (such as the Sakoe–Chiba bandwidth constraint) is introduced to limit the offset of nonlinear alignment, using window parameters. By controlling the allowable time deviation range, local time step matching constraints of the comparison sequences can be achieved;

[0084] Furthermore, a dynamic programming algorithm is used to search for the globally optimal alignment path in the two-dimensional cost matrix, which is achieved by accumulating the minimum cost function, thereby obtaining the optimal time step alignment index sequence across batches;

[0085] Through the above steps, the non-equal length and asynchronous change process of multi-source process parameter sequences in multi-dimensional space is transformed into a time step comparable alignment path, achieving a standardized matching effect of cross-batch time patterns.

[0086] For example, in two batches of samples from automotive floor mat production, batch A has a pressing pressure sequence of 120 time steps, while batch B has a sequence of 95 time steps. After S3.1 normalization, the pressing pressure unit range is scaled to 0~1, the temperature curve unit range is scaled to 0~1, and the mold closing speed unit range is scaled to 0~1. In the DTW calculation, the dimension k=3 in the Euclidean distance difference metric formula, and the window parameter w=5 to limit the time deviation. The parameter sequences of batch A and batch B are input into the multidimensional DTW cost function to construct a 120×95 cost matrix. During the dynamic programming search process, the DTW cost accumulation formula outputs the total minimum accumulated difference metric value under path optimization. The generated optimal alignment path length is 105, and the mapping rate of the aligned time step index pairs reaches a high level. This path is then converted into a trajectory sequence of a uniform scale through subsequent S3.3 time axis mapping, and the consistency of cross-batch process motion curves is significantly improved in the time domain, providing an accurate alignment benchmark for the generation of a standardized process motion fragment library;

[0087] S3.3: Utilize the optimal alignment path to perform time axis alignment mapping on multiple batches of process parameter sequences, transforming the asynchronous and non-uniform length original process trajectories into aligned trajectory sequences under a unified time scale, so as to obtain a time-aligned multi-source process trajectory dataset;

[0088] S3.4: Based on the time-aligned process trajectory dataset, cluster analysis is used to divide the process motion segments, identify process stages with similar dynamic characteristics, and generate process motion segment prototypes as the basic units of the standardized process motion segment library.

[0089] S3.5: Perform dynamic feature extraction on the prototype of the process action segment, and generate dynamic feature vectors of the process action segment based on time-series statistics (such as mean, variance, and slope) and fluctuation energy density, forming a standardized process action segment library and its feature index.

[0090] Based on the prototype data of the time-aligned process action segments, a time-series statistics calculation method (parameters: segment length L, sampling interval Δt) is used to extract the mean features of a single process variable in order to characterize the steady-state level of the variable within the segment.

[0091] Furthermore, by using variance calculation methods (parameters: unbiased estimation, sample size n), the degree of fluctuation of process variables within a segment is quantified, and data results reflecting stability differences are obtained. The variance calculation formula is as follows:

[0092]

[0093] in, For sample values, The sample mean. The number of samples;

[0094] Furthermore, a linear regression slope calculation method was adopted (parameters: least squares fitting, time series). Value sequence This allows for the extraction of the direction and speed of trend changes in process variables, and the generation of trend slope indicators.

[0095] Furthermore, through the wave energy density calculation method (parameter: frequency component) Fourier transform amplitude This enables the quantification of the fluctuation intensity of process variables within a segment in the frequency domain, where the fluctuation energy density is calculated using the following formula:

[0096]

[0097] in, For the frequency component amplitude, The length of the segment;

[0098] By combining four features—mean, variance, slope, and fluctuation energy density—in sequence, a standardized process action segment feature vector is formed, thereby constructing a segment standardization index.

[0099] For example, in a pressing process segment of a car floor mat production line, the sampling length L is 120 seconds, the sampling interval Δt is 1 second, the mean value is calculated to give a steady-state pressing pressure of 200 kPa, and the variance is calculated to give a fluctuation level of... (Unit: kPa^2), the pressure is calculated from the trend slope. The rate of decrease is kPa / s, and the frequency domain energy level calculated from the wave energy density is: (Unit: kPa^2 / s). In another temperature curve segment, the mean is 180℃ and the variance is... (°C^2), slope is ℃ / s, wave energy density is (℃^2 / s). After combining the above four feature values ​​by dimension and performing Z-score normalization, a feature vector of length 4 is obtained. This vector is indexed and stored in the normalized process motion fragment library. The verification results show that this feature index significantly improves the similarity search accuracy and matching speed when matching process trajectories across batches.

[0100] like Figure 3As shown, step S4 involves constructing a latent space model of the process trajectory using a variational autoencoder-long short-term memory (VAE-LSTM) network architecture. The encoder generates latent dynamic features based on the aligned process parameter sequence, and the decoder reconstructs the process trajectory. During training, the correlation mapping between the defect feature sequence and the latent features is optimized simultaneously. Specifically, this includes:

[0101] S4.1: Based on the aligned multi-source process parameter time series data, a VAE-LSTM network architecture is constructed, in which the encoder is composed of LSTM to extract the potential dynamic features of the process trajectory, and the decoder is also composed of LSTM to reconstruct the original process parameter sequence in order to establish the latent space representation of the process trajectory.

[0102] The time-aligned multi-source process parameter time series data obtained through step S3 are used as input for constructing the latent space model of the process trajectory.

[0103] The variational autoencoder-long short-term memory (VAE-LSTM) architecture is adopted (parameter: number of encoder units). =128, number of decoder units =128, potential spatial dimensions =32, the time step length T is equal to the aligned trajectory length), to realize the nonlinear feature abstraction and multi-step dependency capture of process trajectory time series data;

[0104] Furthermore, by configuring a multi-layer LSTM network in the encoder section (parameter: number of layers) =2, activation function is tanh, and recurrent weights are initialized using Xavier), to extract the hidden state of the process trajectory time series and encode the dynamic evolution mode of variables such as pressing pressure, temperature curve, and mold closing speed in the hidden state;

[0105] Furthermore, a multi-layer LSTM network is configured through the structured decoder section (parameter: number of layers). =2, the activation function is tanh, and the output layer adopts a linear mapping to the original variable dimension), to realize the time-step reconstruction of the latent space representation to the original process parameter time series;

[0106] By adopting a network weight sharing strategy, the temporal correlation parameters of the encoder and decoder are optimized in tandem to achieve a synergistic improvement in latent space representation and reconstruction capabilities;

[0107] By using a structured fully connected layer at the end of the encoder, the hidden state sequence is mapped to the mean of the latent space. With variance The parameter vector is used to support subsequent sampling and latent space distribution modeling;

[0108] By using the VAE-LSTM architecture, the time-aligned trajectory of the previous step is transformed into a latent space representation containing potential dynamic features, thereby achieving temporal feature compression and semantic decoupling of the process trajectory.

[0109] For example, the time-series data of a batch of car floor mat process parameters, after DTW time alignment, has a length of T=200, including three-dimensional variables: pressing pressure (MPa), mold closing speed (mm / s), and temperature profile (°C). The encoder is configured with a two-layer LSTM, with 128 cells per layer, an input dimension of 3, a time step of 200, and an output hidden state dimension of 128. The fully connected layer at the end of the encoder is mapped to a latent space dimension of 32, generating a mean vector. With variance vector ,in and The calculation formula is:

[0110]

[0111]

[0112] in This is the hidden state at the end of the encoder's time step. , This is the weight matrix. , This is the bias vector. The sampling vector is generated using reparameterization techniques. ,in It follows a standard normal distribution. The decoder is configured with two LSTM layers, each with 128 units, and the input sampling vector is processed step-by-step. The model outputs a 3D reconstructed sequence, and the mean squared error is compared with the original trajectory. The error of the reconstructed process trajectory is lower than the set threshold, verifying the model's feature compression and temporal reconstruction capabilities. The application results show that the latent space representation can significantly improve the accuracy of inverse inference from defect features to process trajectories, while maintaining the model's interpretability and stability.

[0113] S4.2: Perform LSTM encoding on the sequence of process parameters input by the encoder to obtain the hidden state and cell state of the time series. Then, map them to the mean vector and variance vector of the latent space through a fully connected layer to generate a latent variable representation that conforms to a Gaussian distribution, so as to capture the dynamic evolution characteristics of the process trajectory.

[0114] The aligned multi-source process parameter time series data is input to the input layer of the VAE-LSTM encoder, and the Long Short Time Memory (LSTM) network encoding method is used (parameters: 128 hidden layer units, time step equal to the alignment trajectory length) to realize the dynamic feature capture of the process trajectory in the time series;

[0115] Furthermore, the hidden state vector at each time step is obtained through the state transfer mechanism of the LSTM unit. With the unit state vector By combining the weight parameters of the input gate, forget gate, and output gate, the state update at the current time step is calculated, forming a complete time series hidden state matrix;

[0116] Furthermore, a fully connected layer mapping method is employed (parameter: the dimension of the weight matrix is ​​the hidden state dimension × the latent space dimension) to compress the hidden state matrix into the mean vector of the latent space. With variance vector The mean vector obtains the intuitive center position of the process trajectory through linear mapping, and the variance vector ensures that its value is positive through exponential mapping.

[0117] Furthermore, through the Gaussian distribution parameterization formula Generate latent variable representations, where It follows a standard normal distribution and is used to introduce sampling randomness to capture the nondeterministic dynamic evolution characteristics of the process trajectory;

[0118] By using LSTM encoding and fully connected layer mapping, the aligned process trajectory from the previous step is transformed into a latent space representation that follows a Gaussian distribution, thereby achieving compressed encoding and feature abstraction of the temporal pattern of dynamic process trajectories.

[0119] For example, in the automotive floor mat pressing process, the aligned process parameter sequence has a length of 200 time steps. Each time step includes three variables: pressing pressure (MPa), mold closing speed (mm / s), and temperature (°C). The normalized input is passed through an LSTM encoder with 128 hidden layer units, and the output hidden state matrix has a shape of 200×128. This matrix is ​​transformed into two 64-dimensional vectors by a mean-mapped weight matrix (128×64) and a variance-mapped weight matrix (128×64), which are used as... and During the latent variable sampling process, the standard normal noise vector Generate a latent representation with 64 dimensions. This means that in the subsequent decoding stage, dynamic change patterns of excessively rapid temperature rise or low compression pressure can be reconstructed, significantly improving the ability of the defect root cause model to capture dynamic fluctuations.

[0120] S4.3: Based on the latent variable representation, a sampling vector is generated using the reparameterization technique as the initial input of the decoder. The sampling vector is then decoded step by step by the LSTM decoder to output the reconstructed process parameter sequence, thereby minimizing the mean square error between the original process trajectory and the reconstructed trajectory and realizing unsupervised learning of the latent space of the process trajectory.

[0121] S4.4: The defect feature sequence is used as an auxiliary supervision signal and mapped to the expected mean vector of the latent space through a multilayer perceptron. The distribution of latent variables is constrained by the contrastive loss function to enhance the semantic correlation between defect features and the latent representation of process trajectory, thereby optimizing the feature mapping capability of the VAE-LSTM model.

[0122] S4.5: Introduce the KL divergence loss term during model training to constrain latent variables to follow a standard normal distribution, prevent overfitting of the latent space, and combine reconstruction error and contrast loss to construct a joint optimization objective function to improve the generalization and interpretability of the latent space of the process trajectory.

[0123] The training process of the VAE-LSTM model introduces a KL divergence loss term (parameters: latent variable distribution, standard normal distribution) to achieve regularization constraints on the latent variable distribution to approximate the standard normal distribution;

[0124] Furthermore, by jointly optimizing the reconstruction error term (parameters: original process trajectory sequence, reconstructed process trajectory sequence) and the comparison loss term (parameters: defect feature latent vector, process trajectory latent vector), we can simultaneously improve the trajectory reconstruction accuracy and the correlation between defect features and trajectory latent space, and obtain comprehensive performance evaluation data that includes KL divergence regularization, reconstruction accuracy and feature correlation.

[0125] Furthermore, the gradient descent method (parameters: learning rate α, batch size B, number of iterations E) is used to optimize the joint loss function, thereby achieving global updates of parameters in the latent space and during the decoding mapping process, and generating convergence performance metrics.

[0126] Furthermore, batch normalization (parameters: batch mean μB, batch variance σB²) is used to scale the distribution of latent variables within each training batch, reducing the drift effect of the latent space and generating stability evaluation indicators.

[0127] Through the above joint optimization method, the learning results of the latent variables in the previous step are transformed into a latent space representation of the process trajectory with high generalization and strong interpretability, so as to realize a robust correlation between defect features and dynamic process trajectories.

[0128] Using the KL divergence formula, the regularization constraint of the latent variable distribution is achieved through the following calculations:

[0129]

[0130] in, For the KL divergence term, Let be the mean vector of the latent variables. Let the standard deviation vector of the latent variables be denoted as .

[0131] The joint optimization objective function is expressed as:

[0132]

[0133] in, For the reconstruction error (MSE) term, To compare the loss items, and These are the KL divergence and the contrastive loss weighting coefficients, respectively.

[0134] For example, in the dynamic trajectory modeling of automotive floor mat manufacturing processes, a latent variable mean vector is set. Given a vector of length 32, with element values ​​between [−0.3, 0.4], and a standard deviation vector... Given a vector of length 32, with elements ranging from [0.8, 1.2]. In the KL divergence calculation, the weighting coefficients... Set it to 0.1 and compare the loss weight coefficient. The target value for reconstruction error is set to 0.05, and the target value is set to be less than 1 / 2000 under the normalized process parameter scale. Gradient descent optimization of the joint loss function was performed using the Adam optimizer with a batch size of B=64, a learning rate of α=0.002, and 5000 iterations E. During training, the batch normalization module stabilized the mean of each latent variable dimension to near zero and the variance to near 1. Upon model convergence, the KL divergence value stabilized at approximately 0.32, the reconstruction error at approximately 0.0008, and the contrastive loss at less than 0.004. The resulting latent space representation of the process trajectory effectively reflects the defect characteristics and corresponding dynamic paths of process parameters in subsequent inverse generation inference, improving the accuracy and interpretability of root cause inference.

[0135] S4.6: Based on the trained VAE-LSTM model, extract the latent variable representation of the encoder output and construct the process trajectory latent space database for the subsequent inverse generative inference module based on defect features for latent path search and root cause trajectory generation.

[0136] Based on the trained VAE-LSTM model structure, the latent variable representations output by the encoder are selected as the basic data objects for the latent space of the process trajectory. A batch extraction method (parameters: batch size 64, sequence length equal to the alignment trajectory duration) is adopted to efficiently obtain the mean vector and variance vector of the latent variables of all training samples.

[0137] Furthermore, the latent space is clustered using a latent variable distribution analysis algorithm (parameter: Gaussian mixture model components are set to 8), and the center vector and covariance matrix of each cluster are obtained for subsequent index structure construction.

[0138] Furthermore, by using the principal component analysis algorithm (parameters: retain the first 10 principal components with a cumulative contribution rate of over 0.9), we can achieve dimensionality reduction of the high-dimensional representation of latent variables and generate a dataset of coordinates of the dimensionality-reduced latent variables, thereby reducing storage overhead and search computational complexity.

[0139] Furthermore, the KD-Tree index construction method (parameter: node splitting threshold is 32) is adopted to realize the spatial organization of the dimensionality-reduced latent variable data and generate a process trajectory latent space index structure that supports nearest neighbor search;

[0140] By indexing, the latent variable results obtained in the previous step are transformed into a searchable latent space database, achieving the expected technical effect of supporting fast search and reverse mapping of latent paths based on defect features.

[0141] For example, in a quality inspection application scenario for automotive floor mat production, the trained VAE-LSTM model encoder is used to process time-series data of process parameters aligned by DTW. Each sample sequence is 120 steps long, the batch size is 64, and the latent space mean vector is extracted. Sum of variance vectors A Gaussian mixture model was used to cluster the latent variable set, with a component count of 8, resulting in the center vectors of 8 clusters, which were recorded in the index metadata. Principal component analysis was used to calculate the covariance matrix, and the first 10 principal components were selected to construct a projection matrix, projecting the original 8-dimensional latent vectors into a 10-dimensional space to improve subsequent search accuracy. An index was built using the KD-Tree method, with a node splitting threshold of 32, forming a latent space database that supports nearest neighbor search based on defect feature vectors within millisecond response times. In validation testing, queries using known defect features quickly located the corresponding potential process trajectory paths, significantly improving the retrieval efficiency and matching accuracy in the reverse inference stage.

[0142] Step S5: Design a reverse generative inference module, input specific defect feature sequences into the latent space of the trained VAE-LSTM model, and generate dynamic paths of process parameters in reverse through a gradient optimization search strategy, outputting candidate root cause trajectories containing combinations of key parameter deviations and their temporal fluctuation patterns. Specifically, this includes:

[0143] S5.1: Based on the trained VAE-LSTM model structure, construct an inverse mapping function from the latent space to the process parameter sequence, wherein the latent space is composed of the hidden state vectors output by the encoder, and the inverse mapping function is implemented through the decoder network to support reverse reasoning from defect features to process trajectory.

[0144] The input conditions are the VAE-LSTM process trajectory latent space model structure trained by step S4 and the set of latent variable representations generated by the encoder.

[0145] A method based on network weight parsing (parameters: encoder output hidden state matrix H, cell state matrix C) is used to construct the initial structure of the inverse mapping function from the latent space hidden state vector to the process parameter sequence.

[0146] Furthermore, by using the decoder network parameter loading method (parameters: trained LSTM unit weight matrix W, bias vector b), the time-step decoding mechanism of the inverse mapping function on the latent space sampling vector is realized, and the feasibility verification data of the original process parameter sequence reconstruction result is obtained.

[0147] Furthermore, a fully connected layer dimension adjustment algorithm is adopted (parameter: latent space dimension). Process sequence dimension This enables the matching of the latent space to the decoder input size and generates the decoder input tensor, ensuring dimensional consistency during reverse inference.

[0148] Furthermore, by optimizing the serialization mapping function (parameters: regularization factor λ, learning rate η), the adaptability of the inverse mapping function to different defect feature embeddings is improved, and the coupling response curve from parameters to defect features is generated.

[0149] By constructing and optimizing the above-mentioned inverse mapping function, a stable inverse inference channel is established between the latent spatial variables and the process parameter sequence, thereby achieving the technical effect of controllable generation of specific defect features into process trajectories.

[0150] For example, targeting the defect feature of "high edge irregularity" on the surface of car floor mats, a trained VAE-LSTM model is selected, with the encoder output latent variable dimension being... The length of the process parameter sequence is Each time step contains three-dimensional parameters: pressing pressure, temperature profile, and mold closing speed. First, the hidden state matrix H of the encoder is read, with the shape as follows: × The shape of the unit state matrix C is × The shape of the transformation matrix M of the fully connected layer is... × This maps latent variables to the decoder input size, ensuring that the input dimension at each time step corresponds to the three-dimensional process parameter structure. The weight matrix W of the decoder LSTM unit is loaded, with a size of... × The size of the bias vector b is × This enables the reverse generation of the process parameter trajectory from the gradually decoded potential space. In the reverse mapping test, the inverse mapping function can generate a trajectory on the temperature curve where the heating rate abnormally increases; the temperature curve of this trajectory at time step... - The slope of the interval increased significantly, while the pressing pressure fluctuated within the same interval, forming a dynamic pattern strongly correlated with the defect characteristics. Feasibility verification showed that this pattern conforms to the constraints of equipment temperature rise limits and pressure regulation response time, can explain the causes of defects, and provides a basis for process adjustments.

[0151] S5.2: Perform normalization processing on specific defect feature sequences, and perform standardization transformation based on the statistical distribution characteristics of feature vectors to match the scale range of the input layer of the VAE-LSTM model, so as to obtain normalized defect feature vectors that are adapted to the model input interface.

[0152] For the received specific defect feature sequence data, the Z-score normalization method (parameters: mean μ, standard deviation σ) is used to realize the zero mean and unit variance transformation of the feature components, so as to eliminate the influence of the original dimensional differences on the accuracy of the potential spatial mapping.

[0153] Furthermore, the min-max normalization method (parameter: minimum value of eigencomponents) is used. maximum value This enables linear scaling of each component of the feature vector within the [0,1] interval, and yields a normalized feature matrix, providing basic data for subsequent scaling consistency with the input scale of the VAE-LSTM model.

[0154] Furthermore, a robust scaling method (parameters: median m of feature components, interquartile range IQR) is adopted to adjust the scale of outlier components and generate an outlier robust feature vector set to reduce the distortion effect of outliers on the activation distribution of neurons in the model input layer.

[0155] Furthermore, a vector reconstruction algorithm is used to achieve the weighted fusion of the feature components processed by multiple methods (weights based on the variance contribution rate of each feature in the training set), and a single normalized defect feature vector is generated to ensure that the results of different preprocessing methods retain the optimal distribution characteristics after fusion.

[0156] By using normalization and weighted fusion processing, the defect feature sequence results from the previous step are transformed into data that is compatible with the input interface of VAE-LSTM latent space, thereby achieving scale matching and distribution consistency between defect features and latent patterns of process trajectories.

[0157] For example, in the scenario of tracing defects in car floor mats, the input feature sequence is an edge irregularity of 0.35, a texture breakage density of 0.12, and a color difference distribution gradient ΔE value of 18.6. The Z-score normalization method is used to calculate the normalized value of each component. The formula is as follows:

[0158]

[0159] in, This is the original value of the i-th defect feature. This is the mean of the feature in the training set. The standard deviation is given. The calculated normalized values ​​are: edge irregularity 0.48, texture breakage density -0.27, and color difference distribution gradient 1.15.

[0160] The min-max normalization formula is as follows:

[0161]

[0162] The minimum and maximum values ​​were obtained from historical process defect sample data. The normalized values ​​were: edge irregularity 0.56, texture breakage density 0.21, and color difference distribution gradient 0.69.

[0163] The robust scaling formula for the gradient feature components of the color difference distribution is as follows:

[0164]

[0165] in, This is the median of the characteristic component. The interquartile range is 0.73, calculated from historical samples. After processing by each method, the three feature components are weighted and fused using variance contribution rates (0.45, 0.35, and 0.20, respectively) to obtain the final normalized defect feature vector [0.54, 0.18, 0.65]. After being projected into the VAE-LSTM latent space, the component distribution of this vector is highly consistent with the latent variable distribution output by the encoder during the training phase, ensuring the stable convergence of the gradient descent of the loss function during the subsequent reverse process trajectory generation.

[0166] S5.3: Embed the normalized defect feature vector into the latent space of the VAE-LSTM model, use the hidden state vector in the latent space as the starting point for optimization, and initialize the latent representation of the candidate process parameter sequence based on the coupling mapping relationship between defect features and process trajectory.

[0167] Using the normalized defect feature vector as input data, the latent space embedding method (parameter: the mapping matrix W between the defect feature vector dimension and the latent space dimension) is adopted to project the defect features onto the latent space coordinate system of the trained VAE-LSTM model.

[0168] Furthermore, by embedding a multilayer perceptron into a network (parameters: number of hidden layer nodes, activation function type), a nonlinear mapping from the defect feature vector to the initial hidden state vector in the latent space is achieved, and the hidden state is obtained. and unit state As a starting point for optimization;

[0169] Furthermore, a feature-trajectory coupling mapping matrix (parameters: matrix coefficients are derived from the joint training in step S4) is employed to encode the conditional constraints of defect features on process parameter trajectory patterns in the latent space, and to generate initial latent representations of candidate process parameter sequences. ;

[0170] Furthermore, the latent space normalization transformation method is used (formula:

[0171] ), to achieve normalization adjustment of the initial latent representation on the standard normal distribution, where The mean of the potential space, For potential spatial standard deviation;

[0172] By embedding and normalization, the normalized defect feature vector from the previous step is transformed into an initial path representation that adapts to the latent space search, thereby providing a stable and optimizable starting point for dynamic trajectory search during the reverse generation process.

[0173] For example, in a case involving a batch of car floor mats with excessively high pressing pressure and significantly abnormal texture fracture density, the normalized defect feature vector has a dimension of 3. This vector is mapped to a 16-dimensional latent space coordinate system via an embedding matrix W (3×16). The number of hidden layer nodes in the embedding network is set to 64, and the activation function is ReLU. The output hidden state... With unit state These are vectors of length 16. The initial latent representation is calculated using the defect-trajectory coupling mapping matrix (derived from the S4 network weights). After normalization, the values ​​of μ=0.05 and σ=0.92 were substituted into the formula to obtain the result. The initial potential representation is uniformly distributed within the range of [-1.1, 1.3]. When this initial potential representation is input into the gradient optimization search module of S5.4, it quickly converges in the reverse-generated dynamic path of process parameters to the candidate root cause trajectory containing the pressing pressure that is consistently higher than the standard value and the temperature curve that cools down prematurely, effectively improving the back-reasoning accuracy and enhancing the interpretability of defect causes;

[0174] S5.4: Based on the initialized latent representation, a gradient optimization search strategy is used to iteratively adjust the paths in the latent space. By minimizing the reconstruction error and feature consistency loss function, the dynamic path of process parameters that can generate target defect features is gradually approximated.

[0175] S5.5: Perform decoding operation on the optimized potential path, and convert the hidden state sequence in the potential space into a time sequence of process parameters based on the decoder network structure to obtain candidate root cause trajectories containing dynamic change patterns of pressing pressure, temperature curve and mold closing speed.

[0176] S5.6: Post-process the candidate root cause trajectories, extract the key parameter deviation combinations and their time fluctuation patterns, and form an interpretable set of candidate root cause hypotheses to provide input for subsequent feasibility verification based on process knowledge constraints.

[0177] Step S6: Perform feasibility verification on candidate root cause trajectories based on process knowledge constraints, including physical realizability boundary constraints, equipment response delay characteristic matching, and verification of process parameter interaction rules, to generate a root cause hypothesis chain with engineering significance. Specifically, this includes:

[0178] S6.1: Based on the combination of key process parameters in the candidate root cause trajectory, perform physical realizability boundary constraint modeling, and use the preset upper and lower limits of parameters in the process knowledge base to eliminate parameter combinations that exceed the equipment's capability range in order to obtain a set of feasible parameters that conform to physical laws.

[0179] Based on the combination of key process parameters in the candidate root cause trajectory, a physical realizability boundary constraint modeling method is adopted (parameter source: preset upper and lower limits of the process knowledge base) to achieve boundary condition matching and preliminary screening of the parameter set;

[0180] Furthermore, through the interval constraint checking algorithm (input: dynamic sequence of pressing pressure, temperature curve, mold closing speed, etc.), the value of each moment in the parameter sequence is compared with the corresponding physical boundary value of the process point by point, and the violation point index dataset is obtained.

[0181] Furthermore, the comprehensive boundary violation judgment formula is used to identify combinations of parameters exceeding the limit:

[0182]

[0183] in, As an indicator of the violation rate, Let be the function for determining the limit, when Returns 1 if the value exceeds the upper or lower limit range, otherwise returns 0. The total number of tests for the parameter;

[0184] Furthermore, a multi-dimensional parameter combination filtering algorithm is adopted (constraint condition: all parameters satisfy their respective physical feasible intervals) to eliminate illegal combinations and generate an index table of feasible parameter sets that conform to physical laws;

[0185] By using the boundary constraint algorithm, the candidate root cause trajectory of the previous step is transformed into a set of parameters that have been screened for physical feasibility, so as to achieve the expected technical effect of process condition compliance and subsequent dynamic response verification.

[0186] For example, in the automotive floor mat pressing manufacturing scenario, the candidate root cause trajectory includes a pressing pressure sequence [4.5MPa, 5.8MPa, 7.2MPa], temperature curve peak values ​​[155℃, 170℃, 185℃], and mold closing speeds [0.48m / s, 0.52m / s, 0.65m / s]. The process knowledge base presets allowable pressing pressure ranges as [4.0MPa, 6.0MPa], temperature curve peak value ranges as [150℃, 180℃], and closing speed ranges as [0.45m / s, 0.60m / s]. Comparing the parameter values ​​at each moment using an interval constraint check algorithm, it was found that the pressure of 7.2MPa, temperature of 185℃, and closing speed of 0.65m / s in the third set of data all exceeded the corresponding ranges, resulting in a violation judgment function. Return 1 for each.

[0187] The violation rate is calculated as follows:

[0188]

[0189] This refers to the proportion of the number of times the three parameters exceeded the limit out of the total number of tests (9). This filters out non-compliant combinations, retaining only the first two sets of parameters as the feasible parameter set for input into the subsequent device response delay characteristic matching verification in S6.2, significantly improving the authenticity and physical feasibility of the tracing reasoning.

[0190] S6.2: Based on the equipment response delay characteristic model, perform dynamic response matching processing on the time fluctuation patterns in the feasible parameter set. Utilize the time delay relationship between parameter changes and equipment response in historical process data to verify whether parameter changes in candidate trajectories conform to the actual response behavior of the equipment, so as to select candidate trajectories with time consistency.

[0191] S6.3: Based on the rule base of process parameter interaction, the parameter coupling relationship is verified on the candidate trajectory of time consistency. The parameter coordination change rules defined in the expert knowledge are used to identify and eliminate parameter combinations that violate the process logic in order to obtain a set of process trajectories that meet the multi-parameter coordination constraints.

[0192] Based on the input of the time-consistent candidate trajectory set, the process parameter interaction rule base matching method is adopted (parameter source: predefined parameter coordination change rule set in expert knowledge base) to realize the one-to-one mapping retrieval between candidate trajectories and rule templates;

[0193] Furthermore, by constructing a parameter coupling matrix (parameters: time series of pressing pressure, temperature curve, and mold closing speed in the candidate trajectory), the synergistic change coefficients between each pair of process parameters are calculated, and a coupling coefficient matrix reflecting the dynamic correlation strength of multiple parameters is obtained.

[0194] Furthermore, a rule consistency judgment function (parameters: coupling coefficient matrix, expected collaborative mode in rule base) is adopted to evaluate the conformity between the candidate trajectory parameter change mode and the predefined process collaborative mode, and generate a conformity score vector.

[0195] Furthermore, by using a threshold filtering algorithm (parameter: the minimum compliance threshold corresponding to the rule base), candidate trajectories below the threshold are eliminated to obtain a set of process trajectories that satisfy the collaborative constraints;

[0196] Furthermore, a conflict detection algorithm (parameters: a set of process trajectories that satisfy collaborative constraints, and process logic conflict rules) is adopted to identify cases that violate process logic in parameter combinations and remove conflicting trajectories from the set.

[0197] By verifying the parameter coupling relationship, the time consistency candidate trajectory of the previous step is transformed into a set of process trajectories that meet the multi-parameter collaborative constraints, thereby enhancing the adaptability of the candidate root cause trajectory in the process logic dimension.

[0198] For example, in the process monitoring system of a car floor mat production line, the time-consistent candidate trajectory includes the pressing pressure sequence P(t), the temperature curve T(t), and the mold closing speed V(t). The cooperative change rule defined by the expert knowledge base is: when the heating rate of T(t) is greater than 100°C during the heating stage... At ℃ / s, P(t) should be in to The value increases slowly within the kPa range, and the rate of decrease of V(t) does not exceed [a certain value]. mm / s. Perform parameter coupling matrix calculation on the candidate trajectory, such as the coordination coefficient between T and P. The synergy coefficient between P and V is All values ​​are higher than the 0.75 threshold set by the rule base, meeting the requirements for the intensity of collaborative change. A compliance determination function is used to align the actual coefficient matrix with the expected collaborative pattern to calculate the overall compliance, yielding a comprehensive score. Higher than the minimum compliance The threshold screening process is used to match the process logic conflict rules during the conflict detection stage. This confirms that the trajectory does not contain an illegal combination of rapid temperature rise and rapid pressure drop. Finally, the trajectory is retained and enters the stability evaluation stage of S6.4, which significantly improves the logical rigor and engineering feasibility of the final root cause hypothesis chain.

[0199] S6.4: Based on the obtained set of process trajectories that satisfy multi-parameter collaborative constraints, perform trajectory stability assessment processing, and use the stability index distribution in the process history database to calculate the stability score of candidate trajectories in terms of parameter fluctuation amplitude, duration and change rate, so as to screen out the root cause trajectories with high stability.

[0200] S6.5: Based on the highly stable root cause trajectory, a root cause hypothesis chain with engineering significance is generated. Using the causal relationship description in the process knowledge base, parameter fluctuations and defect characteristics are semantically mapped, and a structured root cause hypothesis chain containing parameter deviation range, time fluctuation pattern and process influence mechanism is output.

[0201] Based on a highly stable root cause trajectory dataset, a causal mapping method (parameters: causal relationship rule table, defect feature encoding table) from the process knowledge base is used to construct the correspondence between process parameter fluctuation patterns and defect features.

[0202] Furthermore, through the parameter semantic parsing algorithm (parameters: parameter deviation amplitude range, time fluctuation curve), the dynamic deviation range of each key process parameter in the trajectory is transformed into structured semantic labels, and the semantic description result of the parameter abnormal range is obtained.

[0203] Furthermore, by using a time pattern matching algorithm (parameters: standard process stage time template, equipment response delay compensation value), the mapping between abnormal fluctuation patterns in the trajectory and production process stages is realized, and a time fluctuation pattern association index is generated.

[0204] Furthermore, a process mechanism reasoning engine (parameters: physical mechanism formula library, equipment characteristic model) is adopted to realize the reasoning from parameter deviation and time pattern to defect cause mechanism, and generate process influence mechanism description vector;

[0205] By using a joint structured coding method, the semantic labels of parameter deviation intervals, the time fluctuation pattern association index, and the description vector of process influence mechanism obtained in the previous step are integrated into a unified root cause hypothesis chain code, thereby realizing the structured output of the root cause hypothesis chain.

[0206] For example, in the automotive floor mat lamination process, root cause trajectories with a stability score greater than 80 are obtained, with the key parameter being the lamination pressure. MPa, the temperature curve appears during the heating stage. Overshoot, mold closing speed deviates from standard value m / s. The above parameter range is input into a semantic parsing algorithm to generate a tag combination of "insufficient compression - excessively rapid heating - excessively rapid closure". Combined with time pattern matching, it is confirmed that temperature overshoot corresponds to the end of the heating stage, while insufficient pressure and abnormal closure speed occur simultaneously in the initial stage of compression. The process mechanism inference engine matches the following formula:

[0207]

[0208] in This represents the overall energy anomaly value resulting from deviations in process parameters. This refers to the actual pressing pressure. Standard pressure value, This represents the actual peak temperature. Standard temperature value, For closing velocity, As a standard value, the calculation results indicate that the defect cause is related to the superposition effect of abnormal energy input. Finally, the root cause hypothesis chain output is: "Insufficient pressing pressure, accompanied by overshooting of temperature at the end of the heating stage and excessively high mold closing speed, causes stress concentration on the material surface, resulting in texture fracture and irregular edge defects." This achieves an engineering interpretable description of the defect cause and provides structured input for the generation of subsequent defect tracing reports.

[0209] Step S7: Generate a defect tracing report based on the verified root cause hypothesis chain, including the dynamic fluctuation range and time axis location information of the key process parameters leading to the defect, to guide parameter adjustments and quality improvements in the production process. Specifically, this includes:

[0210] S7.1: Based on the verified root cause hypothesis chain, extract the dynamic fluctuation range of key process parameters, wherein the fluctuation range includes the amplitude range and duration of the parameter deviation from the standard value, so as to form a structured description of the parameter anomaly pattern.

[0211] S7.2: Perform time axis positioning mapping on the extracted parameter anomaly patterns, align the fluctuation range with the corresponding process stage in the foot mat production process based on timestamp information, and generate an anomaly event sequence in the time dimension;

[0212] S7.3: Construct a process disturbance causal map based on the abnormal event sequence, and generate a visual description of the abnormal fluctuation propagation path by utilizing the interaction rules and physical correlation between process parameters, so as to enhance the interpretability of defect causes;

[0213] S7.4: Prioritize the key disturbance nodes in the cause-effect graph, and output a list of priority adjustment parameters and their influence weights based on the correlation strength between defect features and the dynamic path of process parameters.

[0214] S7.5: Integrate the above structured description, timeline positioning information, cause-effect graph and parameter suggestions to generate a defect tracing report, and output it to the quality management system or human-computer interaction interface in a standardized format for production engineers to make process optimization decisions.

[0215] Based on the structured description, timeline location information, causal graph and parameter suggestion list generated by steps S7.1 to S7.4, a data structure fusion algorithm (parameters: hash index rules, timestamp mapping table) is used to realize the construction of a unified data model for multi-source information.

[0216] Furthermore, through a multi-domain information association and matching method (parameters: node identifier, event sequence order matrix), a bidirectional binding relationship between time axis abnormal events and causal graph nodes is realized, and a verified parameter deviation-defect feature mapping table is obtained.

[0217] Furthermore, through the report generation template engine (parameters: formatting rule set, field mapping table), the automatic conversion of structured data into report text and graphical elements is realized, and chapter-based output content containing process stages, dynamic fluctuation ranges and influencing weights is generated.

[0218] Furthermore, through a visualization layout algorithm (parameters: graph node coordinate matrix, color coding mapping table), the two-dimensional layout of the causal graph in the report is optimized, and a clear and readable graphic file of the abnormal fluctuation propagation path is generated;

[0219] Furthermore, through the export interface module (parameters: JSON, XML, PDF format selection, UTF-8 encoding), the generated report is encapsulated into a standardized data stream, ensuring that it can be seamlessly called by the quality management system and the human-computer interaction interface;

[0220] Through the above multi-stage fusion processing method, the analysis results of the previous step are transformed into readable and usable data reports that can directly support production optimization, thereby enabling rapid transmission of defect causes and efficient support for process parameter adjustment decisions.

[0221] For example, in an embodiment of an automotive floor mat production line, the input structured description includes deviation ranges for three key process parameters, such as the pressing pressure deviating from the standard value. kPa, pressure holding time shortened s. Increased mold closing speed m / s; the time axis positioning information locates the above deviations in the early, middle, and late stages of the molding process, respectively; the cause-effect diagram consists of 12 nodes, including nodes related to raw material status, equipment parameter adjustments, and temperature curve changes; the parameter suggestion list includes adjusting the pressing pressure to... kPa, extend the holding time s. Slow down the mold closing speed m / s. The above data is bound to nodes and time events through a fusion algorithm to generate a mapping table. The report template engine arranges it into three parts: an anomaly event summary, a causal chain diagram, and a list of optimization suggestions. The layout algorithm presents the 12-node graph in the form of a radial hierarchy diagram, with node colors encoded according to their influence weights. Finally, a standardized report file containing text, graphics, and tables is generated through the PDF export interface. After being imported into the quality management system, production engineers can make targeted adjustments to the parameters before the next batch of production, significantly reducing the defect rate and greatly improving process stability.

[0222] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0223] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for tracing the source of surface defects in car floor mats, characterized in that, Includes the following steps: S1: Synchronously collect multi-source process parameter time-series data and corresponding timestamp surface image data during the production of car floor mats, and label the surface image data with defect type labels to establish a spatiotemporally aligned training dataset; S2: Perform morphological feature extraction on the surface image data to generate a defect feature sequence and form a feature vector sequence aligned with the time series data of process parameters; S3: Based on the multi-source process parameter time series data, a process trajectory alignment model is constructed using a dynamic time warping algorithm. By calculating the optimal alignment path of process parameter sequences of different batches, a standardized process action segment library is generated. S4: Construct a latent space model of the process trajectory based on the VAE-LSTM architecture, where the encoder generates latent dynamic features based on the aligned process parameter sequence, the decoder reconstructs the process trajectory, and the association mapping relationship between the defect feature sequence and the latent features is optimized synchronously during the training process. S5: Design a reverse generative inference module to input specific defect feature sequences into the latent space of the trained process trajectory latent space model, and generate dynamic paths of process parameters in reverse through gradient optimization search strategy, and output candidate root cause trajectories containing combinations of key parameter deviations and their time fluctuation patterns. S6: Perform feasibility verification on the candidate root cause trajectories based on process knowledge constraints to generate a root cause hypothesis chain with engineering significance.

2. The method for tracing the source of surface defects in car floor mats according to claim 1, characterized in that, Following step S6, the following is also included: S7: Generate a defect tracing report based on the root cause hypothesis chain, which includes the dynamic fluctuation range of the key process parameters that caused the defect and the time axis location information.

3. The method for tracing the source of surface defects in car floor mats according to claim 1, characterized in that, Step S1 specifically includes: Based on multi-source sensors deployed on the automotive floor mat manufacturing production line, raw process parameter time-series data including pressing pressure, temperature curve, and mold closing speed process parameters are collected to obtain key dynamic control variables in the manufacturing process; Industrial cameras are used to collect surface image data at a designated inspection station after the foot pads are formed, and defect type labels are marked in the images to form an image sample set with semantic information; Based on the time synchronization mechanism, the original process parameter time series data and the surface image data are timestamped and aligned to construct a multimodal data pair that is synchronized in the time dimension; Perform data cleaning operations on the multimodal data pairs to obtain a well-structured and reliable original training dataset; The time series data of process parameters after cleaning are normalized based on the data standardization method to generate normalized time series data of process parameters. Perform image enhancement and preprocessing operations on the labeled image data to generate preprocessed image data; The normalized process parameter time series data and the preprocessed image data are stored and organized according to timestamps to construct a spatiotemporally aligned and structured training dataset.

4. The method for tracing the source of surface defects in car floor mats according to claim 3, characterized in that, The original process parameter timing data acquisition adopts a multi-source sensor array. The pressure sensor has a range of 0~5MPa and a resolution of 0.01MPa. The temperature sensor has an accuracy of ±0.1℃ and a sampling frequency of 100Hz. The closing speed uses a grating encoder with a resolution of 0.001mm / pulse. All acquired signals are processed by bandpass filtering and time synchronization. The acquisition frequencies are set to 500Hz for pressure, 100Hz for temperature, and 200Hz for speed.

5. The method for tracing the source of surface defects in car floor mats according to claim 1, characterized in that, Step S2 specifically includes: Acquire surface image data, which are image samples collected and labeled with defect type tags during the production process of automotive floor mats; The surface image data is subjected to image preprocessing to obtain preprocessed image data; The edge extraction operation is performed on the preprocessed image data based on the Canny edge detection algorithm to obtain the edge contour image of the defects on the surface of the foot pad. Morphological skeleton extraction and segmentation are performed on the edge contour image, Freeman chain code is used to represent the edge direction, and edge irregularity index is calculated. Based on the texture analysis method, Gabor filter bank response calculation is performed on the preprocessed image data to extract local texture features of the defect area and to statistically analyze the texture fracture density index. Color difference analysis is performed on the RGB color space of the preprocessed image data. The image is converted to the Lab color space based on the CIE Lab color space conversion formula, and the color difference distribution gradient between the defect area and the normal area is calculated. The three feature indices—edge irregularity index, texture fracture density index, and color difference distribution gradient—are normalized to form a multidimensional feature vector. This vector is then spatiotemporally aligned with the image acquisition timestamp and the time-series data of the process parameters to generate a defect feature sequence.

6. The method for tracing the source of surface defects in car floor mats according to claim 1, characterized in that, Step S3 specifically includes: Normalization is performed on the variables in the multi-source process parameter time series data to obtain a standardized process parameter sequence; Based on the standardized process parameter sequence, a multidimensional dynamic time warping cost function is constructed to calculate the optimal alignment path of the process parameter sequence between different production batches. The optimal alignment path is used to perform time axis alignment mapping on multiple batches of process parameter sequences, transforming the asynchronous and non-uniform original process trajectories into aligned trajectory sequences under a unified time scale, thus obtaining a time-aligned multi-source process trajectory dataset. Based on the time-aligned process trajectory dataset, cluster analysis is used to divide the process action segments, identify process stages with similar dynamic characteristics, and generate process action segment prototypes. Dynamic feature extraction is performed on the prototype of the process action segment. Based on time-series statistics and fluctuation energy density calculation, dynamic feature vectors of the process action segments are generated, forming a standardized process action segment library and its feature index.

7. The method for tracing the source of surface defects in car floor mats according to claim 6, characterized in that, Step S3 further includes selecting Euclidean distance as a time-series point-to-point difference measure and using a time elastic window to constrain the time offset between batches, finally outputting an aligned process trajectory dataset, and forming a standardized process action segment library with four statistical features: mean, variance, slope, and fluctuation energy density.

8. The method for tracing the source of surface defects in car floor mats according to claim 1, characterized in that, Step S4 specifically includes: Based on the aligned multi-source process parameter time series data, a VAE-LSTM network architecture is constructed, in which the encoder is composed of LSTM to extract the potential dynamic features of the process trajectory, and the decoder is also composed of LSTM to reconstruct the original process parameter sequence. The process parameter sequence input by the encoder is processed by LSTM encoding to obtain the hidden state and cell state of the time series. The hidden state and cell state are then mapped to the mean vector and variance vector of the latent space through a fully connected layer to generate a latent variable representation that conforms to a Gaussian distribution. Based on the latent variable representation, a sampling vector is generated using the reparameterization technique, which serves as the initial input to the decoder. The sampling vector is then decoded step-by-step to output a reconstructed sequence of process parameters. The defect feature sequence is used as an auxiliary supervision signal and mapped to the expected mean vector of the latent space through a multilayer perceptron. The distribution of latent variables is constrained by the contrastive loss function. During model training, a KL divergence loss term is introduced to constrain the latent variables to follow a standard normal distribution. At the same time, a joint optimization objective function is constructed by combining reconstruction error and contrastive loss. Based on the trained VAE-LSTM model, the latent variable representations of the encoder output are extracted to construct a latent space database of the process trajectory.

9. A method for tracing the source of surface defects in car floor mats according to claim 8, characterized in that, In step S4, the VAE-LSTM model encoder and decoder are both configured as at least two-layer LSTM structures, with a latent space dimension of 32. During training, the loss function jointly considers reconstruction error, contrast loss, and KL divergence. The Adam optimizer is used for training, and the latent space represents the feature compression and defect feature association mapping used for process trajectory.

10. A method for tracing the source of surface defects in automotive floor mats according to claim 1, characterized in that, Step S5 specifically includes: Based on the trained VAE-LSTM model structure, an inverse mapping function from the latent space to the process parameter sequence is constructed. Normalization is performed on a specific defect feature sequence. Based on the statistical distribution characteristics of the feature vector, a standardization transformation is performed to match the scale range of the input layer of the VAE-LSTM model, thereby obtaining a normalized defect feature vector. The normalized defect feature vector is embedded into the latent space of the VAE-LSTM model. The hidden state vector in the latent space is used as the starting point for optimization. Based on the coupling mapping relationship between defect features and process trajectory, the latent representation of candidate process parameter sequence is initialized. Based on the initialized latent representation, a gradient optimization search strategy is used to iteratively adjust the paths in the latent space. By minimizing the reconstruction error and feature consistency loss function, the dynamic path of process parameters that can generate target defect features is gradually approximated. Decoding operations are performed on the optimized potential paths. Based on the decoder network structure, the hidden state sequence in the potential space is converted into a time sequence of process parameters to obtain candidate root cause trajectories containing dynamic change patterns of pressing pressure, temperature curves and mold closing speed. The candidate root cause trajectories are post-processed to extract key parameter deviation combinations and their temporal fluctuation patterns, forming a set of candidate root cause hypotheses with interpretability.

Images