Multi-variety variable batch size co-linear manufacturing variable working condition process compensation real-time parameter adjusting method

Abstract

This invention provides a real-time parameter tuning method for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions, relating to the fields of intelligent manufacturing and online optimization of process parameters. This invention solves the technical challenge of data-driven deep learning quality prediction models being difficult to adapt to dynamic production conditions, effectively avoiding the problems of condition perception deviations and process decision failures caused by fluctuations in the production environment and material batch changes in multi-variety, variable-batch manufacturing scenarios. This invention constructs a complete closed loop of "perception preprocessing – self-supervised decoupling – continuous updating – rolling compensation," achieving a leap from passive monitoring to proactive intervention, ensuring continuous and stable production yield under varying operating conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and online optimization of process parameters, specifically to a real-time parameter adjustment method for process compensation under varying operating conditions in multi-variety, variable-batch, co-line manufacturing. Background Technology

[0002] With the continued growth in consumer demand for personalized products and the profound adjustments in the global supply chain, the traditional manufacturing model centered on large-scale production of the same model is rapidly evolving into a multi-variety, variable-batch co-line manufacturing model. In the mobile phone precision assembly industry, the same production line needs to frequently switch between multiple models, such as standard, Pro, and customized versions. Each production change involves not only changing fixtures and resetting process formulas, but also dealing with fluctuations in physical properties caused by batch differences in assembled components. In the new energy vehicle assembly industry, key processes such as battery pack assembly and electric drive system integration also face the challenge of multi-model mixed-line production. These scenarios collectively exhibit the typical characteristics of variable operating conditions, high-frequency production changes, and sparse data, posing unprecedented challenges to precise quality control in the manufacturing process.

[0003] In the complex multi-variety, variable-batch manufacturing scenarios described above, production conditions exhibit dynamic changes, primarily driven by two types of physical mechanisms. First, frequent production line changes and material batch shifts trigger discrete, abrupt changes in operating conditions, visually manifested as step jumps in sensor signal distribution. Second, microscopic wear of assembly tools, equipment aging, and fluctuations in workshop temperature and humidity induce continuous, gradual changes in operating conditions, characterized by slow, time-varying drift of the feature baseline. To address the quality control needs of multi-variety, variable-batch co-line manufacturing, existing research mainly focuses on data-driven deep learning quality prediction models. These models are typically trained once on offline static datasets and then deployed for production line quality prediction and parameter optimization.

[0004] However, when data-driven deep learning quality prediction models are deployed to actual workshops, their prediction performance drops drastically when faced with shifting operating conditions, making them unable to adapt to dynamically changing production conditions. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a real-time parameter adjustment method for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions, which solves the technical problem that data-driven deep learning quality prediction models cannot adapt to dynamically changing production conditions.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for real-time parameter adjustment of process compensation for multi-variety, variable-batch, co-line manufacturing under varying operating conditions, comprising: Acquire and standardize multi-source sensor data to obtain a high-dimensional sensing feature matrix; Perform different types of enhancement operations on the high-dimensional perceptual feature matrix; define the views generated by different enhancements of the same original segment as positive sample pairs, and define the views from different processing beats or feature space distances exceeding a threshold as negative samples, thus constructing a sample space; A dual-branch encoder model, including an online encoder and a momentum encoder, is trained based on the sample space. After the model is trained, the network parameters are fixed, and the process decoupling of the sample space features is performed through the online encoder to separate the gradual degradation features and abrupt operating condition features. Drift detection is performed by measuring the statistical divergence between the current time window's gradual degradation and abrupt operating condition feature distributions and the historical reference distributions. When significant operating condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to the shift of important weights. At the same time, the core set samples selected by the maximum coverage distance criterion are used for scenario memory playback. The online encoder is jointly optimized to update the gradual degradation features and abrupt operating condition features to adapt to the new operating conditions. The updated gradual degradation features and abrupt change features are concatenated as auxiliary inputs to the model, constructing a lightweight proxy mapping model between process parameters and quality features. Within the constrained neighborhood of the current baseline process parameters, the optimization objective is to minimize the predicted value of quality deviation. The finite-time domain constrained optimization problem is solved in a rolling manner to generate a small optimal process compensation amount, which is then sent to the underlying programmable logic controller of the production line in real time to perform online closed-loop adjustment of key process parameters.

[0007] Preferably, the step of acquiring and standardizing multi-source sensor data to obtain a high-dimensional sensing feature matrix includes: Acquire multi-source sensor data, wherein the multi-source sensor data refers to the continuous acquisition of multi-dimensional time-series signals of the assembly process through a multi-source sensor network consisting of D sensors deployed at key assembly stations in a high-frequency sampling manner. Denoising filtering and local smoothing are performed sequentially on the signals of each sensing channel to eliminate periodic interference and random impulse noise, thereby obtaining smooth timing signals for each channel. To address the issue of inconsistent sampling frequencies among different sensors, a preset common sampling frequency is used as a benchmark to resample and align the smoothed time-series signals of each channel, ensuring that each channel has the same number of sampling points T on a unified time axis. The aligned signals of the D sensing channels are then cascaded along the channel dimensions to construct a high-dimensional sensing feature matrix X with dimensions D×T.

[0008] Preferably, the enhancement operations performed on the high-dimensional perceptual feature matrix of different types include: Additive white Gaussian noise, whose intensity is constrained by the physical feasible fluctuation range of the assembly process, is superimposed onto the high-dimensional sensing feature matrix. The noise intensity parameter is randomly sampled from an interval with the sensor calibration accuracy as the lower bound and a certain proportion of the physical feasible fluctuation range of the process as the upper bound, to simulate the random perturbation of the workshop environment. The overall amplitude of the high-dimensional perceptual feature matrix is ​​scaled by a random ratio determined by the micro-tolerance range of the assembly materials to simulate the characteristic differences between batches of processed materials. In the high-dimensional sensing feature matrix, a continuous interval with a length not exceeding a preset proportion of the total window length is randomly selected, and all sensing channels within the interval are set to zero to simulate instantaneous packet loss or communication anomalies of the sensor.

[0009] Preferably, a dual-branch encoder model, including an online encoder and a momentum encoder, is trained based on the sample space, comprising: The expression for the loss function of the two-branch encoder model is: in, Total loss; This indicates that the two-branch encoder model exhibits gradual degradation features. InfoNCE contrast loss; This indicates the characteristics of the dual-branch encoder model under abrupt operating conditions. InfoNCE contrast loss; These are the orthogonal constraint weight hyperparameters; Represents the orthogonality constraint term of the Frobenius norm; Let N represent the Frobenius norm of the matrix, where N is the number of samples in the batch. For feature dimensions.

[0010] Preferably, the step of evaluating the importance of historical parameters by calculating the Fisher information matrix includes: in, It is a set of historical samples sampled from the core feature memory; Total loss; Indicates the current network parameters; Indicates the first i historical parameters The importance of; This represents historical samples sampled from the core feature memory.

[0011] Preferably, elastic weight consolidation regularization is introduced to impose a secondary penalty on significant weight shifts, including: in, The total loss on the new operating condition data, This is the hyperparameter for regularization strength; This indicates a secondary penalty loss.

[0012] Preferably, the lightweight proxy mapping model is a lightweight multilayer perceptron proxy mapping. This is used to represent the process parameter vector u and the quality feature prediction output. Nonlinear mapping relationship between them: in, Indicates gradual degradation characteristics; This indicates the characteristics of sudden change in operating conditions.

[0013] Secondly, the present invention provides a real-time parameter adjustment system for multi-variety, variable-batch, co-line manufacturing with variable operating conditions, comprising: The data acquisition module is used to acquire and standardize multi-source sensor data to obtain a high-dimensional sensing feature matrix; The sample space construction module is used to perform different types of enhancement operations on the high-dimensional perceptual feature matrix; it defines the views generated by different enhancements of the same original fragment as positive sample pairs, and defines the views from different processing beats or feature space distances exceeding a threshold as negative samples, thus constructing the sample space; The decoupling module is used to train a dual-branch encoder model, including an online encoder and a momentum encoder, based on the sample space. After the model training is completed, the network parameters are fixed, and the process decoupling of the sample space features is performed through the online encoder to separate the gradual degradation features and abrupt operating condition features. The continuous update module is used to perform drift detection by measuring the statistical divergence between the current time window's gradual degradation and abrupt operating condition feature distributions and the historical reference distribution. When significant operating condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to the shift of important weights. At the same time, the core set samples selected by the maximum coverage distance criterion are used for scenario memory playback. The online encoder is jointly optimized to update the gradual degradation features and abrupt operating condition features to adapt to the new operating conditions. The online parameter compensation and closed-loop control module is used to concatenate the updated gradual degradation features and abrupt operating condition features as auxiliary inputs to the model, constructing a lightweight proxy mapping model between process parameters and quality features. Within the constrained neighborhood of the current baseline process parameters, with the optimization objective of minimizing the predicted value of quality deviation, it continuously solves the finite-time domain constrained optimization problem, generates a small-scale optimal process compensation amount, and sends it down to the underlying programmable logic controller of the production line in real time to execute online closed-loop adjustments to key process parameters.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for real-time parameter adjustment of process compensation for multi-variety variable batch co-line manufacturing under variable operating conditions, wherein the computer program causes a computer to execute the real-time parameter adjustment method for process compensation for multi-variety variable batch co-line manufacturing under variable operating conditions as described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing real-time parameter tuning for process compensation based on multi-variety, variable-batch, collinear manufacturing under variable operating conditions as described above.

[0016] (III) Beneficial Effects This invention provides a real-time parameter adjustment method for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions. Compared with existing technologies, it has the following advantages: This invention solves the technical challenge of adapting data-driven deep learning quality prediction models to dynamic production conditions, effectively avoiding the problems of condition perception deviations and process decision failures caused by fluctuations in the production environment and material batch changes in multi-variety, variable-batch manufacturing scenarios. This invention constructs a complete closed loop of "perception preprocessing - self-supervised decoupling - continuous updating - rolling compensation," achieving a leap from passive monitoring to proactive intervention, ensuring continuous and stable production yield under changing conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a block diagram of a real-time parameter adjustment method for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides a real-time parameter adjustment method for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions. This method solves the technical problem that data-driven deep learning quality prediction models cannot adapt to dynamically changing production conditions, and achieves continuous and accurate perception of manufacturing quality characteristics and real-time closed-loop compensation of process parameters under varying operating conditions.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: In multi-variety, variable-batch manufacturing environments, production line streaming sensor data faces significant distribution drift issues. The alternating day and night temperatures and humidity within the production workshop, inconsistent upstream and downstream material batches, and microscopic wear and tear on precision assembly tools all contribute to continuous non-stationary drift in time-series signals collected from programmable logic controllers (PLCs) and industrial IoT nodes. Traditional quality control methods based on statistical process control rely on the strong assumption of a steady-state process, making it difficult to achieve forward-looking quality prediction and parameter optimization in the face of distributional shifts caused by frequent production changes. Furthermore, most existing data-driven deep learning quality prediction models are trained once on offline static datasets; once deployed to the physical workshop and encountering operational condition drift, their predictive performance often experiences a precipitous decline.

[0022] Meanwhile, in the acquisition and processing of streaming sensor data, high-frequency sensor signals in assembly workshops are often simultaneously contaminated by power frequency electromagnetic coupling interference and random impulse noise. Traditional single filtering methods are difficult to simultaneously suppress both types of noise. In addition, multi-source sensors usually have different sampling frequencies, requiring efficient and reliable resampling alignment strategies to form a unified multidimensional time series matrix that can be used for subsequent modeling.

[0023] To address the aforementioned issues, this invention provides a real-time parameter tuning method for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions. Without requiring manual labeling of each sample during the feature representation learning stage, this method achieves continuous and accurate perception of manufacturing quality characteristics and real-time closed-loop compensation of process parameters under varying operating conditions through multi-source sensor data acquisition and filtering preprocessing, physical-driven data enhancement and sample construction, self-supervised decoupling of dual-channel process features, elastic weight consolidation and core set memory playback for continuous learning, and online process parameter compensation based on a surrogate model.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] This invention provides a method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions, such as... Figure 1 As shown, it includes: S1. Acquire and standardize multi-source sensor data to obtain a high-dimensional sensing feature matrix; S2. Perform different types of enhancement operations on the high-dimensional perceptual feature matrix; define the views generated by different enhancements of the same original segment as positive sample pairs, and define the views from different processing beats or feature space distances exceeding the threshold as negative samples, and construct the sample space; S3. A dual-branch encoder model, including an online encoder and a momentum encoder, is trained based on the sample space. After the model is trained, the network parameters are fixed, and the process features of the sample space are decoupled through the online encoder to separate the gradual degradation features and abrupt operating condition features. ; S4. Drift detection is performed by measuring the statistical divergence between the current time window's gradual degradation and abrupt operating condition feature distributions and the historical reference distribution. When significant operating condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to the shift of important weights. At the same time, the core set samples selected by the maximum coverage distance criterion are used for scenario memory playback. The online encoder is jointly optimized to update the gradual degradation features and abrupt operating condition features to adapt to the new operating conditions. S5. The updated gradual degradation features and abrupt change conditions features are concatenated as auxiliary inputs to the model to construct a lightweight proxy mapping model between process parameters and quality features. Within the constrained neighborhood of the current baseline process parameters, the optimization objective is to minimize the predicted value of quality deviation. The finite-time domain constrained optimization problem is solved in a rolling manner to generate a small optimal process compensation amount, which is then sent to the programmable logic controller at the bottom of the production line in real time to perform online closed-loop adjustment of key process parameters.

[0026] The embodiments of the present invention achieve continuous and accurate perception of manufacturing quality characteristics under varying operating conditions and closed-loop parameter compensation, effectively suppressing yield degradation caused by operating condition drift.

[0027] In one embodiment, S1, multi-source sensor data is acquired and standardized to obtain a high-dimensional sensing feature matrix. The specific implementation process is as follows: In typical discrete manufacturing scenarios such as precision assembly of mobile phones or assembly of new energy vehicles, a high-precision multi-source sensor network deployed at key assembly stations continuously acquires streaming time-series signals of the production line operation through high-frequency sampling. The collected sensing dimensions include, but are not limited to: the torque and angle of each joint of a multi-axis industrial robot, the displacement and contact force of the end effector, the three-phase current and speed of the servo motor used for screw fastening or dispensing, and the temperature and humidity of the workshop environment. These heterogeneous multi-source signals are aggregated to the edge computing node via industrial Ethernet, where they undergo the following three stages of preprocessing: (a) Adaptive Kalman Filtering Suppresses Structured Measurement Noise and Random Measurement Disturbances. To address the structured periodic disturbances and measurement noise present in the sensor signals from the assembly workshop, a discrete-time state-space model is established and adaptive Kalman filtering is applied to obtain a smooth estimate of the actual process state. The state prediction equation is: in, To make a priori prediction of time k based on the state at time k-1, Here is the state transition matrix. To predict covariance a priori, Let be the process noise covariance matrix. The measurement update equation is: in, Here is the Kalman gain matrix. The actual sensor measurement value at time k. For the observation matrix, To measure the noise covariance matrix, Let I be the identity matrix, representing the posterior estimate of the covariance. To achieve adaptive response to noise level variations under different operating conditions, An online estimation strategy is employed, which involves statistically analyzing the innovation sequence within the current sliding window. Sample covariance Dynamic updates are performed to automatically adjust the filtering strength of the Kalman filter according to the operating conditions.

[0028] (b) Moving average filtering to eliminate impulse white noise. Based on the Kalman filter output, a moving average filter with a fixed window length W is further used to locally smooth the state estimation sequence: in, W is the moving average output at time k. W is set according to the typical duration of short-term random fluctuations, usually W=3 to 7, in order to suppress high-frequency disturbances while preserving dynamic characteristics related to the assembly process status as much as possible.

[0029] (c) Interpolation, resampling, alignment, and channel cascading. To address the issues of different sensors having different sampling frequencies, missing samples, and inconsistent sequence lengths, each sensor channel is mapped to a unified time grid. In a preferred embodiment, using a system-preset common sampling frequency as a reference, low-frequency sensor channels are resampled using cubic spline interpolation, while high-frequency channels are processed using window aggregation or decimation, ensuring that each channel has the same number of sampling points T on a unified time axis. All D sensor channels, after the above filtering, smoothing, and resampling processes, are cascaded along the channel dimension to form a high-dimensional sensing feature matrix X of dimension D×T, which serves as the standardized input for physical-driven data augmentation and positive / negative sample construction in step S2.

[0030] In one embodiment, S2, different types of enhancement operations are performed on the high-dimensional perceptual feature matrix; views generated from the same original segment through different enhancements are defined as positive sample pairs, and views from different processing beats or feature space distances exceeding a threshold are defined as negative samples, thus constructing a sample space. The specific implementation process is as follows: Based on the physical feasible domain of the assembly process, several different types of physical-driven enhancement strategies are employed to generate positive and negative samples required for contrastive learning for the high-dimensional perceptual feature matrix. In this embodiment of the invention, random noise injection, random amplitude scaling, and temporal random masking are used as examples. It is understood that in actual operation, other physical-driven enhancement strategies, such as random frequency band masking and local waveform perturbation, can also be employed. These will not be elaborated upon here.

[0031] Additive random noise injection: Additive white Gaussian noise with an intensity constrained by the physically feasible fluctuation range of the assembly process is superimposed onto the high-dimensional perceptual feature matrix. Its noise intensity parameters Random sampling is performed within an interval with the sensor calibration accuracy as the lower bound and a certain proportion of the physical feasibility fluctuation range of the process as the upper bound to simulate random perturbations in the workshop environment.

[0032] Random scaling of amplitude: The overall amplitude of the feature matrix is ​​scaled according to a random ratio determined by the micro-tolerance range of the assembled materials to simulate the characteristic differences between batches of processed materials.

[0033] Temporal random mask: Randomly select a continuous interval [ts, te] in the feature matrix with a length not exceeding a preset proportion of the total window length (length not exceeding 20% ​​of the total window length), and set all sensing channels in the interval to zero to simulate instantaneous packet loss or communication abnormality of the sensor.

[0034] Two views generated after different enhancement operations on the same original working condition segment are defined as positive sample pairs; enhanced views from different working condition segments or those judged by feature space distance measurement to have significant process differences (cosine distance exceeds the working condition difference threshold determined based on historical working condition statistics) are defined as negative samples, thereby constructing the structured sample space required for comparative learning.

[0035] In one embodiment, S3, a dual-branch encoder model including an online encoder and a momentum encoder is trained based on the sample space; after the model training is completed, the network parameters are fixed, and the process features of the sample space are decoupled through the online encoder to separate the gradual degradation features and abrupt operating condition features. The specific implementation process is as follows: It should be noted that in the specific implementation process, the dual-branch encoder model needs to be pre-trained using the historical sample space. The basic pre-training process is an existing technology and will not be described in detail here.

[0036] The dual-branch encoder model includes an inline encoder and a momentum encoder. The inline encoder... Using a temporal convolutional network as the backbone, long-range dependencies of temporal signals are captured through layer-by-layer expanding causal convolutions. The backbone output, after batch normalization and nonlinear activation, is connected to two multilayer perceptron projection heads to compress high-dimensional features into a contrastive feature space. Momentum encoder Structure and Exactly the same, the parameters are adjusted using an exponential moving average. Smooth following: The momentum coefficient m∈(0,1) ensures that the key features provided by the momentum encoder remain highly consistent, preventing representation collapse in contrastive learning.

[0037] Both inline encoders and momentum encoders need to branch into two independent linear mapping subheads after the projection head output: the degradation feature subhead outputs slowly degrading features. Characterizes the gradual quality degradation trend caused by equipment wear and material aging; severe operating condition characteristics of the sub-head output. The two types of features are characterized by sudden changes in operating conditions caused by production changes or batch switching; the comparative loss is calculated separately after L2 normalization.

[0038] InfoNCE contrast loss. (Regarding...) and InfoNCE contrast loss was calculated independently for each case. For example, for the first batch within a batch Query vectors for each sample Its positive sample key vector is denoted as The remaining 2N-2 vectors within the batch are all considered negative samples. The InfoNCE loss is defined as follows: in, Let τ be the cosine similarity, and τ be the temperature hyperparameter. [k≠ [This is to exclude its own indicator function.]

[0039] Orthogonal decoupling constraint. To ensure that the degenerate feature subspace and the operating condition feature subspace are independent of each other and to prevent mutual leakage of the two types of feature information, a Frobenius norm orthogonal constraint term is introduced: in, Let N represent the Frobenius norm of the matrix, where N is the number of samples in the batch. is the feature dimension, and the denominator is used for normalization. When... and When the column vectors are orthogonal to each other Approaching zero; when linear correlation exists Increase and generate gradients to drive the network to further separate the two types of feature subspaces, achieving deep decoupling between degradation features and operating condition features.

[0040] The total loss of the bi-branch encoder model is the sum of three terms: in, These are the orthogonal constraint weight hyperparameters. Only applicable to online encoders. Gradient backpropagation update, momentum encoder Synchronous exponential moving average update. After training, the momentum encoder is discarded, and the sample space features are decoupled from the process using an online encoder to separate two types of features: one type is the gradual degradation features characterizing equipment wear and material aging. Another category is the abrupt changeover conditions that characterize production line changes and batch switching. After the two are combined, the input is used for continuous updates and proxy modeling in steps S4 and S5.

[0041] In one embodiment, S4, drift detection is performed by measuring the statistical divergence between the current time window's gradual degradation and abrupt change condition feature distributions and the historical reference distribution. When significant condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to significant weight shifts. Simultaneously, core set samples selected using the maximum coverage distance criterion are used for scenario memory playback to jointly optimize the online encoder and update the gradual degradation and abrupt change condition features to adapt to the new conditions. The specific implementation process is as follows: During continuous operation, production lines may experience gradual or abrupt shifts in operating conditions due to changes in environmental temperature and humidity, batch changes in materials, or microscopic degradation of equipment. To enable online encoders to adapt to new operating conditions without forgetting their ability to sense historical operating conditions, the following feature-based incremental evolution mechanism is established.

[0042] (a) Operating condition drift detection and update triggering. A feature distribution difference quantification and evaluation module is established to measure the statistical divergence between the statistical distribution of the gradual degradation features and abrupt operating condition feature vectors within the current time window and the historical steady-state reference distribution in real time. Let the mean vector and covariance matrix of the feature distribution within the current time window be denoted as follows: , The mean vector and covariance matrix of the historical steady-state reference distribution are respectively , The drift statistic is defined as in, For covariance difference weights, The number of consecutive threshold control cycles required to trigger an incremental update. When detected... continuous When the divergence threshold is exceeded for each control cycle, the system triggers the feature network incremental update process. The threshold is adaptively set based on the high quantile statistical results of the historical steady-state distribution to suppress false alarms caused by instantaneous random disturbances.

[0043] (b) Calculation of the empirical Fisher information matrix. Before updating the online encoder parameters using new operating condition data, the current network parameters are used as the basis. Using historical samples in the core feature memory, the diagonal elements of the Fisher information matrix are calculated to assess the importance of each parameter to the historical operating condition perception capability: in, The set of historical samples sampled from the core feature memory M. The total contrastive learning loss defined in step S3, The larger the value, the higher the number of... Parameters The more critical the ability to perceive characteristics that maintain historical operating conditions, the stronger the constraints should be placed on updates. It is the symbol for partial derivatives.

[0044] (c) Elastic Weight Consolidation Regularization. When updating the network using new operating condition data, an EWC regularization term is added to the loss function: in, Compare the total learning loss for step S3 on the new operating condition data. , This is the hyperparameter for regularization strength. These are the historically optimal parameters before the update. The EWC regularization term imposes stronger constraints on parameters with high Fisher values ​​and grants greater update freedom to parameters with low Fisher values, thus protecting the coding weights of historical operating conditions while learning new operating condition features.

[0045] (d) Dynamic maintenance of the core set memory. A core feature memory M with a maximum capacity of C is maintained. Unlike sample redundancy caused by random sampling, a greedy maximum coverage distance criterion is used to adaptively select core samples in the latent space of gradually degrading features and abrupt change features: for candidate samples... Calculate its eigenvector Minimum Euclidean distance to the existing sample set in memory M Select The largest candidate sample is added to the database until the total number reaches the upper limit C. This strategy maximizes the coverage of historical operating condition feature distribution with the fewest stored samples, while taking into account the storage resource constraints of edge computing nodes and the diversity requirements of the memory database.

[0046] (e) Memory playback joint optimization. Optimization using new operating condition data. Simultaneously, historical core samples are sampled from memory M according to operating conditions, and their application in the current network parameters is calculated. Contrast learning loss And jointly optimize the total loss: in, The weights for memory playback loss are used. After the update, the Fisher information matrix is ​​recalculated, and the representative core samples of the current new working condition are updated to the memory bank M according to the maximum coverage distance criterion, so as to realize the self-maintained continuous evolution of the online encoder in multi-working-condition scenarios, and continuously update the gradual degradation features and abrupt working-condition features.

[0047] In one embodiment, S5, the updated gradual degradation features and abrupt change condition features are concatenated as auxiliary inputs to the model, constructing a lightweight proxy mapping model between process parameters and quality features; within the constrained neighborhood of the current baseline process parameters, the finite-time constrained optimization problem is solved in a rolling manner with the optimization objective of minimizing the predicted quality deviation value, generating a small-scale optimal process compensation amount, which is then sent to the underlying programmable logic controller of the production line in real time to perform online closed-loop adjustment of key process parameters. The specific implementation process is as follows: (a) Feature proxy model construction. Based on the gradually degrading features and abrupt change condition features continuously updated in step S4 ( , Using concatenated representations as input, a lightweight multilayer perceptron proxy mapping is constructed. Establish the process parameter vector u and the quality feature prediction output. Nonlinear mapping between: The surrogate model is subjected to supervised calibration based on a small amount of online quality inspection sampling data. This supervised calibration is only used for the surrogate model and does not participate in the sample-by-sample manual annotation training of the online encoder in step S3.

[0048] (b) Constraint optimization and rolling compensation. Under the current baseline process parameters... Within the constrained neighborhood, with the objective of minimizing the predicted quality deviation and control increment, a finite-time constrained optimization problem is constructed and solved in a rolling manner: using the preset quality deviation weight matrix Q and control increment penalty matrix R as the objective weights, and the upper limit under the physical constraints of process parameters [ , and maximum adjustment range per step As a constraint, the dual-channel eigenvectors are approximately assumed to remain piecewise constant within a single prediction time domain, and are re-extracted and updated in each control cycle. A rolling time-domain strategy is adopted, compensating only the first step of the optimal solution sequence with incremental changes. Implemented, resulting in real-time compensation instructions. .

[0049] (c) Effect feedback and closed-loop iteration. Compensation instructions. The system transmits data in real-time via industrial Ethernet to the production line's underlying programmable logic controller (PLC), driving the servo system to perform precise micro-step online adjustments to key process parameters such as dispensing pressure, tightening torque, and assembly feed speed. This effectively offsets quality deviations caused by operational drift. At the start of the next control cycle, the adjusted production line sensor data is collected in real-time. After filtering, smoothing, and resampling preprocessing in step S1, the data is fed into the online encoder in step S3. The latest dual-channel feature vector is updated to the surrogate model input, and the surrogate model is incrementally fine-tuned using online sampling data, forming a complete online closed-loop control system encompassing sensor acquisition, feature extraction, parameter optimization, and command issuance. When significant operational drift is detected again in step S4, the continuous learning mechanism is triggered synchronously, incrementally updating the online encoder. This ensures the entire methodology operates stably over the long term in scenarios involving multiple product types, varying batch sizes, and continuous production changes, achieving continuous and accurate perception of manufacturing quality characteristics and real-time compensation and optimization of process parameters.

[0050] An embodiment of the present invention provides a real-time parameter adjustment system for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions, comprising: The data acquisition module is used to acquire and standardize multi-source sensor data to obtain a high-dimensional sensing feature matrix; The sample space construction module is used to perform different types of enhancement operations on the high-dimensional perceptual feature matrix; it defines the views generated by different enhancements of the same original fragment as positive sample pairs, and defines the views from different processing beats or feature space distances exceeding a threshold as negative samples, thus constructing the sample space; The decoupling module is used to train a dual-branch encoder model, including an online encoder and a momentum encoder, based on the sample space. After the model training is completed, the network parameters are fixed, and the process decoupling of the sample space features is performed through the online encoder to separate the gradual degradation features and abrupt operating condition features. The continuous update module is used to perform drift detection by measuring the statistical divergence between the current time window's gradual degradation and abrupt operating condition feature distributions and the historical reference distribution. When significant operating condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to the shift of important weights. At the same time, the core set samples selected by the maximum coverage distance criterion are used for scenario memory playback. The online encoder is jointly optimized to update the gradual degradation features and abrupt operating condition features to adapt to the new operating conditions. The online parameter compensation and closed-loop control module is used to concatenate the updated gradual degradation features and abrupt operating condition features as auxiliary inputs to the model, constructing a lightweight proxy mapping model between process parameters and quality features. Within the constrained neighborhood of the current baseline process parameters, with the optimization objective of minimizing the predicted value of quality deviation, it continuously solves the finite-time domain constrained optimization problem, generates a small-scale optimal process compensation amount, and sends it down to the underlying programmable logic controller of the production line in real time to execute online closed-loop adjustments to key process parameters.

[0051] It is understood that the real-time parameter adjustment system for multi-variety variable batch co-line manufacturing with variable operating conditions provided in this embodiment of the invention corresponds to the above-mentioned real-time parameter adjustment method for multi-variety variable batch co-line manufacturing with variable operating conditions. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the real-time parameter adjustment method for multi-variety variable batch co-line manufacturing with variable operating conditions, and will not be repeated here.

[0052] This invention also provides a computer-readable storage medium storing a computer program for real-time parameter adjustment of process compensation for multi-variety, variable-batch, co-line manufacturing under varying operating conditions. The computer program causes a computer to execute the real-time parameter adjustment method for multi-variety, variable-batch, co-line manufacturing under varying operating conditions as described above.

[0053] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs include methods for performing real-time parameter adjustment for multi-variety, variable-batch, co-line manufacturing process compensation under variable operating conditions as described above.

[0054] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention addresses the technical challenge of adapting data-driven deep learning quality prediction models to dynamic production conditions, effectively avoiding the problems of condition perception deviations and process decision failures caused by fluctuations in the production environment and material batch changes in multi-variety, variable-batch manufacturing scenarios. This invention constructs a complete closed loop of "perception preprocessing - self-supervised decoupling - continuous updating - rolling compensation," achieving a leap from passive monitoring to proactive intervention, ensuring continuous and stable production yield under changing conditions.

[0055] 2. The embodiments of the present invention propose a multi-stage cascaded filtering preprocessing scheme, which suppresses structured measurement noise and high-frequency random pulse disturbances in manufacturing flow sensing signals in a layered manner, and maps asynchronous, multi-frequency, multi-source time-series signals to a comparable time grid, providing a high-quality input foundation for subsequent self-supervised feature learning, and solving the limitation of traditional single filtering methods that cannot take into account different types of noise.

[0056] (2) This invention proposes a dual-channel orthogonal decoupling framework for degradation features and operating condition features, overcoming the limitation of general contrastive learning frameworks in manufacturing sensing data that cannot distinguish features of different physical mechanisms. By introducing orthogonal decoupling constraints, the two types of feature subspaces are made independent of each other, so that slow degradation features and severe operating condition features serve different downstream tasks respectively. This provides a high-quality foundation of gradual degradation features and abrupt operating condition features for refined process parameter compensation without the need for manual annotation.

[0057] (3) The embodiments of the present invention integrate elastic weight consolidation regularization and maximum coverage distance core set memory playback to construct an efficient anti-forgetting continuous learning mechanism for edge computing power-limited scenarios. By complementing each other with parameter space constraints and data space playback, the dual-branch encoder model can continuously retain the accurate perception capability of historical multi-conditions while adapting to new working condition features, and realize the self-maintained evolution of the feature base.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time parameter adjustment in multi-variety, variable-batch, co-line manufacturing process compensation under varying operating conditions, characterized in that: include: Acquire and standardize multi-source sensor data to obtain a high-dimensional sensing feature matrix; Perform different types of augmentation operations on the high-dimensional perceptual feature matrix; Views generated from the same original fragment through different enhancements are defined as positive sample pairs, and views from different processing beats or feature space distances exceeding a threshold are defined as negative samples, thus constructing a sample space; A dual-branch encoder model, including an online encoder and a momentum encoder, is trained based on the sample space. After the model training is completed, the network parameters are fixed, and the process decoupling of the sample space features is performed by an online encoder to separate the gradual degradation features and abrupt working condition features. Drift detection is performed by measuring the statistical divergence between the current time window's gradual degradation and abrupt operating condition feature distributions and the historical reference distributions. When significant operating condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to the shift of important weights. At the same time, the core set samples selected by the maximum coverage distance criterion are used for scenario memory playback. The online encoder is jointly optimized to update the gradual degradation features and abrupt operating condition features to adapt to the new operating conditions. The updated gradual degradation features and abrupt change features are concatenated as auxiliary inputs to the model, constructing a lightweight proxy mapping model between process parameters and quality features. Within the constrained neighborhood of the current baseline process parameters, the optimization objective is to minimize the predicted value of quality deviation. The finite-time domain constrained optimization problem is solved in a rolling manner to generate a small optimal process compensation amount, which is then sent to the underlying programmable logic controller of the production line in real time to perform online closed-loop adjustment of key process parameters.

2. The method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions as described in claim 1, is characterized in that... The acquisition and standardization of multi-source sensor data to obtain a high-dimensional sensing feature matrix includes: Acquire multi-source sensor data, wherein the multi-source sensor data refers to the continuous acquisition of multi-dimensional time-series signals of the assembly process through a multi-source sensor network consisting of D sensors deployed at key assembly stations in a high-frequency sampling manner. Denoising filtering and local smoothing are performed sequentially on the signals of each sensing channel to eliminate periodic interference and random impulse noise, thereby obtaining smooth timing signals for each channel. To address the issue of inconsistent sampling frequencies among different sensors, a preset common sampling frequency is used as a benchmark to resample and align the smoothed time-series signals of each channel, ensuring that each channel has the same number of sampling points T on a unified time axis. The aligned signals of the D sensing channels are then cascaded along the channel dimensions to construct a high-dimensional sensing feature matrix X with dimensions D×T.

3. The method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions as described in claim 1, characterized in that... The different types of enhancement operations performed on the high-dimensional perceptual feature matrix include: Additive white Gaussian noise, whose intensity is constrained by the physical feasible fluctuation range of the assembly process, is superimposed onto the high-dimensional sensing feature matrix. The noise intensity parameter is randomly sampled from an interval with the sensor calibration accuracy as the lower bound and a certain proportion of the physical feasible fluctuation range of the process as the upper bound, to simulate the random perturbation of the workshop environment. The overall amplitude of the high-dimensional perceptual feature matrix is ​​scaled by a random ratio determined by the micro-tolerance range of the assembly materials to simulate the characteristic differences between batches of processed materials. In the high-dimensional sensing feature matrix, a continuous interval with a length not exceeding a preset proportion of the total window length is randomly selected, and all sensing channels within the interval are set to zero to simulate instantaneous packet loss or communication anomalies of the sensor.

4. The method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions as described in claim 1, characterized in that... A dual-branch encoder model, including an online encoder and a momentum encoder, is trained based on the sample space, including: The expression for the loss function of the two-branch encoder model is: in, Total loss; This indicates that the two-branch encoder model exhibits gradual degradation features. InfoNCE contrast loss; This indicates the characteristics of the dual-branch encoder model under abrupt operating conditions. InfoNCE contrast loss; These are the orthogonal constraint weight hyperparameters; Represents the orthogonality constraint term of the Frobenius norm; Let N represent the Frobenius norm of the matrix, where N is the number of samples in the batch. For feature dimensions.

5. The method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions as described in claim 4, characterized in that... The method of evaluating the importance of historical parameters by calculating the Fisher information matrix includes: in, It is a set of historical samples sampled from the core feature memory; Total loss; Indicates the current network parameters; Indicates the first i historical parameters The importance of; This represents historical samples sampled from the core feature memory.

6. The method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions as described in claim 5, is characterized in that... The introduction of elastic weights reinforces regularization by imposing a secondary penalty on significant weight shifts, including: in, The total loss on the new operating condition data, This is the hyperparameter for regularization strength; This indicates a secondary penalty loss.

7. The method for real-time parameter adjustment and compensation of multi-variety, variable-batch, co-line manufacturing processes under varying operating conditions, as described in any one of claims 1 to 6, is characterized in that... The lightweight proxy mapping model is a lightweight multilayer perceptron proxy mapping. This is used to represent the process parameter vector u and the quality feature prediction output. Nonlinear mapping relationship between them: in, Indicates gradual degradation characteristics; This indicates the characteristics of sudden change in operating conditions.

8. A real-time parameter adjustment system for process compensation in multi-variety, variable-batch, co-line manufacturing under varying operating conditions, characterized in that: include: The data acquisition module is used to acquire and standardize multi-source sensor data to obtain a high-dimensional sensing feature matrix; The sample space construction module is used to perform different types of enhancement operations on the high-dimensional perceptual feature matrix; it defines the views generated by different enhancements of the same original fragment as positive sample pairs, and defines the views from different processing beats or feature space distances exceeding a threshold as negative samples, thus constructing the sample space; A decoupling module is used to train a dual-branch encoder model, including an online encoder and a momentum encoder, based on the sample space. After the model training is completed, the network parameters are fixed, and the process decoupling of the sample space features is performed by an online encoder to separate the gradual degradation features and abrupt working condition features. The continuous update module is used to perform drift detection by measuring the statistical divergence between the current time window's gradual degradation and abrupt operating condition feature distributions and the historical reference distribution. When significant operating condition drift is determined, the importance of historical parameters is evaluated by calculating the Fisher information matrix. Elastic weights are introduced to reinforce regularization and apply a secondary penalty to the shift of important weights. At the same time, the core set samples selected by the maximum coverage distance criterion are used for scenario memory playback. The online encoder is jointly optimized to update the gradual degradation features and abrupt operating condition features to adapt to the new operating conditions. The online parameter compensation and closed-loop control module is used to concatenate the updated gradual degradation features and abrupt operating condition features as auxiliary inputs to the model, constructing a lightweight proxy mapping model between process parameters and quality features. Within the constrained neighborhood of the current baseline process parameters, with the optimization objective of minimizing the predicted value of quality deviation, it continuously solves the finite-time domain constrained optimization problem, generates a small-scale optimal process compensation amount, and sends it down to the underlying programmable logic controller of the production line in real time to execute online closed-loop adjustments to key process parameters.

9. A computer-readable storage medium, characterized in that, It stores a computer program for real-time parameter adjustment of process compensation for multi-variety variable batch co-line manufacturing under variable operating conditions, wherein the computer program causes the computer to execute the real-time parameter adjustment method for process compensation for multi-variety variable batch co-line manufacturing under variable operating conditions as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing real-time parameter tuning for process compensation based on multi-variety variable batch collinear manufacturing under variable operating conditions as described in any one of claims 1 to 7.

Images