Method for quality prediction of discrete manufacturing flow line based on mechanism and data joint driving

Abstract

A discrete manufacturing pipeline quality prediction method based on mechanism and data joint driving belongs to the field of intelligent manufacturing and industrial artificial intelligence, and the method comprises the following steps: first, data acquisition and preprocessing under the industrial scene of the discrete manufacturing pipeline; second, state space deduction and single variable nonlinear analysis of the sparse time sequence characteristics of the discrete manufacturing pipeline; third, mechanism joint training and real-time closed-loop intervention of the discrete manufacturing pipeline quality prediction model. The present application significantly improves the physical consistency of the discrete manufacturing pipeline quality prediction, realizes continuous inference of the hidden state under irregular sampling gaps, and enhances the safety decision-making ability and engineering application value of the automatic control of the discrete manufacturing pipeline.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and industrial artificial intelligence, specifically a discrete manufacturing production line quality prediction method driven by both mechanism and data. Background Technology

[0002] Intelligent manufacturing workshops play a crucial role in improving production efficiency and ensuring product quality. In the field of discrete manufacturing, the production quality of various workpieces on the assembly line is the core of industrial manufacturing, and its pass rate directly affects the core competitiveness of enterprises. Establishing a high-precision quality prediction model is fundamental to achieving high-quality production in the collaborative process of multiple discrete processes. However, complex industrial production processes are accompanied by strong dynamic time-varying characteristics, which can easily cause model mismatch when applied across batches. Currently, traditional sampling inspection mechanisms suffer from time lag and are unable to cope with hidden defects caused by condition drift in the early stages of processing. The multi-variety, small-batch discrete manufacturing model requires frequent switching between different specifications and materials on the assembly line, resulting in nonlinear fluctuations in the system's operating state, and discrepancies often exist between historical data distribution and the current production state. The lack of a forward-looking prediction mechanism means that process control systems rely solely on post-processing feedback for compensation, making it difficult to effectively reduce scrap rates. Edge control nodes also cannot implement effective early intervention. Therefore, constructing a closed-loop mechanism for quality prediction and safety intervention with strong generalization and real-time continuous inference capabilities is a key technical challenge that needs to be overcome in the current industrial manufacturing field.

[0003] In the actual production process of multi-equipment collaborative operation on discrete manufacturing production lines, physical variables such as material properties, equipment thermal equilibrium state, and stamping pressure exhibit complex spatiotemporal coupling. As equipment ages, sensor data is prone to long-term non-stationary drift, leading to mismatch in conventional prediction models. Furthermore, existing mainstream time-series prediction models, represented by Long Short-Term Memory (LSTM) networks and standard Transformers, mostly rely on continuous, equidistant, high-frequency sampling data. However, limited by quality inspection costs and production line cycle times, real-world data is mostly sparse, single-piece sampling with irregular time gaps between samples. Existing models, by defaulting to equidistant sampling, often forcibly treat discrete points with extremely large time spans as adjacent events, resulting in distorted time-series characteristics that violate physical laws. This makes it difficult for existing models to accurately capture low-frequency decay trends such as mold micro-wear and machine heat accumulation across batches, and their feature mapping process lacks hard constraints from prior mechanistic knowledge, making it difficult to achieve process traceability for manufacturing deviations and lacking industrial interpretability.

[0004] To achieve quality prediction, existing publicly available technologies have been extensively explored. For extracting spatial mapping features from industrial process data, Yuan et al. provided a quality assessment scheme from a pure data perspective by constructing convolutional kernels of different sizes in parallel. To address the memory decay problem of long sequence data, Yang et al. introduced a self-attention mechanism to capture sequence dependencies across time steps, improving temporal modeling capabilities. Regarding handling data incompleteness under complex operating conditions, Meng et al. constructed a data reconstruction mechanism and a multilayer perceptual fusion network, outputting a single deterministic quality assessment index for feature loss at unsampled time points.

[0005] However, the aforementioned publicly available technologies still have significant limitations when facing real-world industrial scenarios in discrete manufacturing. First, some data-driven methods detach from underlying industrial mechanisms, relying heavily on statistical fitting of historical characteristics. In discrete manufacturing production lines, there are objective energy conversion and matter conservation mechanisms. When faced with complex operational changes or sensor data drift, pure data models, lacking the hard constraints of prior knowledge of these mechanisms, are prone to outputting results that violate basic industrial principles, leading to severe model mismatch in cross-batch applications. Second, deep learning models typically require continuous and equidistant high-frequency sampling data, making them difficult to adapt to the sparse and irregularly spaced single-piece sampling scenarios on real production lines. This non-uniform time interval easily causes temporal discrepancies in the model, making it unable to accurately extract the long-term degradation trend of equipment aging. Furthermore, the feature mapping process of conventional pure data-driven model networks lacks necessary industrial interpretability. Finally, at the closed-loop control level, most existing prediction schemes are limited to a single point prediction mode, which can only output deterministic scalar values. Due to the lack of a quantified risk confidence lower bound, the edge control system cannot assess the reliability of the prediction results and it is difficult to directly issue intervention commands with safety bottom lines to the underlying PLC, resulting in a disconnect between the prediction algorithm and the field equipment linkage control. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes a discrete manufacturing production line quality prediction method driven by both mechanism and data. First, multi-source, heterogeneous, non-uniform, sparse historical sampling data and synchronously mapped energy consumption and output records are obtained from the discrete manufacturing production line. The historical sampling data is then preprocessed using a standard deviation scaling algorithm based on a dynamic sliding time window. Next, a joint inference network based on a state-space model and univariate nonlinear mapping is constructed. This network takes the normalized preprocessed temporal features as input and, while addressing the temporal irregularities caused by non-uniform sampling, continuously infers the evolution trajectory of the hidden states of the production line during processing. This allows for the nonlinear analysis and aggregation of each hidden state variable in an industrially interpretable manner. Finally, when training the network model, a mechanism-constrained loss function bound to the real industrial capacity equation is introduced, along with a learning rate decay and cross-domain dual-validation early stopping mechanism. A conformal prediction algorithm is also used to generate the yield prediction result.

[0007] This invention provides the following technical solution: A method for predicting the quality of discrete manufacturing production lines based on a combination of mechanism and data-driven approaches, the method comprising the following steps: The first step is data acquisition and preprocessing in discrete manufacturing production line industrial scenarios. The second step, state-space derivation and univariate nonlinear analysis of the sparse temporal characteristics of the discrete manufacturing pipeline, is as follows: Step (2.1) involves sparse temporal state deduction based on continuous-time discretization, and standardizing the feature tensor containing the underlying equipment operating state and material properties. Input into the deduction module; Step (2.2) analyzes the nonlinear contribution of each hidden state variable to the pass rate in order to achieve process traceability of finished product manufacturing deviations; The third step involves the joint training and real-time closed-loop intervention of the discrete manufacturing production line quality prediction model, as follows: Step (3.1) Calculation of mean square error; Step (3.2) Construction of mechanism error constraints; Step (3.3) Combine total error calculation and model optimization; Step (3.4) Parameter optimization and double check for early stopping; Step (3.5) Quantitative risk confidence interval generation.

[0008] Furthermore, the process of the first step is as follows: Step (1.1) Multidimensional Industrial Time-Series Feature Acquisition: To acquire multi-source heterogeneous manufacturing features, a sensor network deployed on discrete manufacturing production lines is used to capture the underlying operating signals of key processing equipment. Simultaneously, the Manufacturing Execution System (MES) and offline quality inspection terminals are used to capture material properties and form and position tolerance sampling records. The multi-source data is then aggregated to the PLC and industrial control computer. Within the production line control cabinet, Modbus protocol and RS485 communication standard are used for underlying data interaction. Finally, the TCP / IP protocol is used to remotely transmit the data in real time to the control server database for time-series storage. The extracted original input feature tensors are labeled as follows: ,in This indicates the total number of finished product batches collected. This indicates the number of historical consecutive sampling samples included within the sliding time window. This represents the total number of dimensions of the input features. Finally, the actual output vector corresponding to each sample batch is collected synchronously. Actual energy consumption vector and the global true finished product pass rate vector ; Step (1.2) Data normalization preprocessing: The original input feature tensor is normalized using a standard deviation scaling algorithm based on a dynamic sliding time window; Step (1.3) Dataset partitioning: Generate a global sample partitioning index according to the time series order, and apply the preprocessed standardized feature tensor... and its corresponding global true finished product pass rate vector Perform non-overlapping domain partitioning, resulting in a batch size of... The global data is sequentially divided into training, validation, calibration, and test sets. Based on the global sample partitioning index, the actual output vector is simultaneously processed. With actual energy consumption vector Perform mapping and segmentation to generate feature tensors and vector slices of corresponding dimensions for each set, which are used for subsequent network inference, physical mechanism error calculation and confidence boundary quantization.

[0009] The main beneficial effects of this invention are as follows: By constructing an independent mechanism-constrained loss function, the actual energy consumption, output, and finished product qualification rate of discrete manufacturing production lines are coupled together. The energy consumption of a single good product is derived as a boundary condition, and the energy consumption deviation is integrated into the network gradient backpropagation as a hard constraint penalty term. This solves the problem that the output of a pure data-driven model violates physical laws when encountering complex cross-operating conditions or data distribution drift. It makes the evolution of network weights strictly obey the industrial energy conservation principle, and significantly improves the physical consistency of discrete manufacturing production line quality prediction.

[0010] By constructing a dynamic sliding time window at the input to eliminate baseline drift caused by equipment degradation, and introducing a continuous-time discretization mechanism to overcome temporal feature distortion caused by non-uniform sparse sampling, continuous inference of hidden states under irregular sampling intervals is achieved. Furthermore, a univariate mapping network based on B-spline curves is used to analyze the nonlinear relationships in the complex hidden state space. By quantifying the contribution of each hidden state variable to the final pass rate, the model is given industrial interpretability, providing a process traceability basis for manufacturing deviations in finished products.

[0011] A conformal prediction algorithm is introduced to generate a risk quantification confidence interval with statistical probability guarantee. Based on the lower limit of this interval, the speed reduction and load reduction of the production line equipment and the safety warning are automatically triggered. This solves the technical problem that traditional prediction models lack risk quantification indicators as a reliable decision basis, making it difficult to perform protective interventions before continuous scrap is generated. It effectively reduces material loss, mold depreciation and ineffective energy consumption, and further enhances the safety decision-making capability and engineering application value of discrete manufacturing production line automation control. Attached Figure Description

[0012] Figure 1 This is a flowchart of a quality prediction method driven by both mechanism and data. Figure 2 It is a core feature extraction structure diagram based on composite state space and univariate nonlinear mapping; Figure 3 This is a topology diagram of the hardware and software system for edge intervention in the elevator hall door assembly line. Detailed Implementation

[0013] The present invention will now be further described with reference to the accompanying drawings.

[0014] Reference Figures 1-3 A method for predicting the quality of discrete manufacturing production lines based on a combination of mechanism and data-driven approaches, the method comprising the following steps: The first step, data acquisition and preprocessing in a discrete manufacturing production line industrial scenario, is as follows: Step (1.1) Multidimensional Industrial Time-Series Feature Acquisition: To obtain multi-source heterogeneous manufacturing features, a sensor network deployed on the discrete manufacturing production line is used to capture the underlying operating signals of key processing equipment. Simultaneously, the Industrial Manufacturing Execution System (MES) and offline quality inspection terminals are used to capture material properties and form and position tolerance sampling records. The multi-source data is then aggregated to the PLC and industrial control computer. Within the production line control cabinet, the Modbus protocol and RS485 communication standard are used for underlying data interaction. Finally, the TCP / IP protocol is used to remotely transmit the data in real time to the control server database for time-series storage. The extracted original input feature tensors are labeled as follows: ,in This indicates the total number of finished product batches collected. This indicates the number of historical consecutive sampling samples included within the sliding time window. This represents the total number of dimensions of the input features. Finally, the actual output vector corresponding to each sample batch is collected synchronously. Actual energy consumption vector and the global true finished product pass rate vector .

[0015] Step (1.2) Data normalization preprocessing: The original input feature tensor is normalized using a standard deviation scaling algorithm based on a dynamic sliding time window. The process is as follows: Step (1.2.1) Setting the dynamic sliding time window and truncation rules: The Middle The first batch of finished products, the first The first time step The elements of each feature dimension are denoted as Batch index Starting from 1. The preset maximum sample length is... Dynamic sliding time window Batch dimension Perform a sliding operation. Define the actual effective backtracking time window length for this batch as... In the current Strictly less than hour, The value is In the present Greater than or equal to hour, Cut off and fix as The formula is shown below: ; Step (1.2.2) Calculation of dynamic statistics: Calculate the dynamic arithmetic mean covering all historical batches within the current sliding time window. With dynamic standard deviation This will rewind all data within the current sliding time window to the past. The first consecutive batch, in the... The time step and the first The values ​​of each feature dimension are summed and then divided by . ,get ; Calculate the feature values ​​of each historical batch at the same time step and dimension. The difference, squared and summed, then divided by In addition to the preset minimal positive numbers And extract the arithmetic square root to obtain The formula is shown below: ; ; in, Indicates in In the middle, with the current number The finished product batch serves as the anchor point for tracing back through history along the batch dimension. In the first batch, the first The first time step The original feature elements of each feature dimension; Step (1.2.3) Temporal feature normalization mapping: ... Each of them Subtract the corresponding Then divide by the corresponding The normalized feature elements after mapping are obtained. The formula is shown below: ; The final generated Together they constitute the standardized feature tensor .

[0016] Step (1.3) Dataset partitioning: Generate a global sample partitioning index according to the time series order, and apply the preprocessed standardized feature tensor... and its corresponding global true finished product pass rate vector Perform non-overlapping domain partitioning, resulting in a batch size of... The global data is sequentially divided into training, validation, calibration, and test sets. Based on the global sample partitioning index, the actual output vector is simultaneously processed. With actual energy consumption vector Perform mapping and segmentation to generate feature tensors and vector slices of corresponding dimensions for each set, which are used for subsequent network inference, physical mechanism error calculation and confidence boundary quantization.

[0017] The second step involves state-space derivation and univariate nonlinear analysis of the sparse temporal characteristics of a discrete manufacturing production line, such as... Figure 2 The diagram shows the core feature extraction structure based on a composite state space and a univariate nonlinear mapping. From left to right, this network structure comprises three core components: a discretized state space derivation module, a temporal feature concatenation module, and a univariate nonlinear mapping module. Input features... After entering the discretized state-space derivation module, the data is transmitted in two paths. The first path enters the dynamic step-size extraction network to obtain... This is then mapped to generate a discretized state transition matrix. With discretized input matrix Subsequently Compared to the previous hidden state The input is multiplied by the first multiplication node, and at the same time... Input features directly introduced along the second path The input is multiplied by the second multiplication node, and the outputs of these two multiplication nodes are combined and summed at the addition node to update and generate the current hidden state. In the temporal feature concatenation module, the hidden state sequences at each time step are combined into a single-batch temporal feature tensor. And extract its final time step slice Finally, we proceed to the univariate nonlinear mapping module, which will... Decomposed into multiple independent channels, each channel is passed through a corresponding learnable B-spline basis function. Nonlinear feature extraction is performed, and the extraction results of all channels are aggregated by the terminal summation node to output the final single-batch prediction result. .

[0018] The second step of this embodiment is as follows: Step (2.1) is based on the sparse temporal state deduction of continuous-time discretization. In order to restore the continuous processing state of the discrete manufacturing line during the non-uniform sampling gap of this batch, the standardized feature tensor containing the underlying equipment operating state and material properties is used. The input is fed into the deduction module, and the process is as follows: Step (2.1.1) Construction of continuous-time state space: The Middle The finished product batch in the first The discrete input feature vector extracted at each time step is denoted as . Based on the zero-order hold assumption, it is kept constant over adjacent sampling intervals, thus transforming it into a continuous-time function. To complete the continuous input features between adjacent sampling intervals; The first derivative of the hidden state vector is calculated by multiplying the basic state transition matrix by the hidden state vector in continuous time, and then adding the product of the input projection matrix and the continuous input signal vector. To reflect the rate of change of the hidden state of the discrete manufacturing pipeline during the processing of this batch, the formula is as follows: ; in, The hidden state vector; Represents the basic state transition matrix; Indicates the input projection matrix; The number of channels representing the hidden state features; Step (2.1.2) Dynamic step size operator extraction: For the first step size operator... The first batch of finished products will be the first batch of finished products. Discrete time steps Multiplying the result by the internal weight matrix and adding the bias vector, then inputting it into a smooth nonlinear activation function, the output has... Dynamic step size vector of each channel dimension The actual processing time span corresponding to the current discrete sampling data is dynamically calculated using the following formula: ; in, Represents the weight matrix; Represents the bias vector; For smoothing nonlinear activation functions; Step (2.1.3) Discretization mapping of state parameters: The vector... Transform into a diagonal matrix and merge with After multiplying, the matrix exponent is obtained to generate the discretized state transition matrix. This represents the evolution of the hidden state of the discrete manufacturing pipeline during the current batch processing over the current actual sampling time span; based on the zero-order hold discretization derivation, the matrix exponent calculated above is subtracted from the identity matrix, and then multiplied by... The inverse matrix, then combined with the input projection matrix. Multiplying them together yields the simplified discretized input matrix. To quantify the impact of the current multidimensional physical features on the update of the hidden state, the formula is as follows: ; ; in, Represents the matrix exponential function; This represents the operation of transforming an input vector into a diagonal matrix with diagonal elements; Represents the identity matrix; express The inverse matrix, to ensure that the inverse matrix in the formula The mathematical existence of the fundamental state transition matrix supports the commutative law of multiplication in the subsequent discretization derivation. It is strictly constrained to be a structured diagonal matrix with negative real part diagonal elements during network initialization; Step (2.1.4) Discrete State Update and Feature Concatenation: Initialize the first... The hidden state vector of each finished batch at time step 0 The vector is all zeros, representing the initial hidden state of this batch at the start of processing; the hidden state vector from the previous time step is... and Multiply, add and The product of these two vectors yields the hidden state vector updated at the current time step. This is to reflect the superposition result of the decay of the hidden state at the previous time step and the current physical feature input, as shown in the following formula: ; After extrapolating to the end of the time window, all The latent state vectors generated at each time step are concatenated along the time series dimension to generate a single-batch time series feature matrix. This is used to record the hidden state evolution trajectory of the discrete manufacturing pipeline throughout the entire sampling cycle of that batch. Subsequently, for all The above simulation and splicing operations are performed in parallel in several independent batches, combining each... The global hidden state feature tensor is formed by stacking along the batch dimension. This leads to the construction of a global hidden state feature space that covers multiple batches of production conditions.

[0019] Step (2.2) analyzes the nonlinear contribution of each hidden state variable to the pass rate in order to achieve process traceability of finished product manufacturing deviations. The process is as follows: Step (2.2.1) Basis feature extraction of independent variables: from Extract the last time step Corresponding feature vector To reflect the final hidden state accumulated at the end of the current time window of the batch of finished products, and to serve as the basis of independent variables for subsequent univariate nonlinear mappings, the extracted first... The first batch of finished products The elements of each feature dimension are denoted as As an independent variable input for subsequent quantification of the contribution of a single hidden state variable to the pass rate, it provides a basis for process traceability of finished product manufacturing deviations; Step (2.2.2) Univariate nonlinear mapping: Enter them separately In the B-spline basis function, the corresponding control point weight coefficients are multiplied and summed to calculate the univariate nonlinear mapping output value. To quantify the contribution of this hidden state variable to the final product qualification rate of this batch of finished products, and to give this nonlinear mapping process industrial interpretability, the formula is as follows: ; in, The number of B-spline bases; For the cumulative index of the B-spline basis; Let be the order of the B-spline curve; For the first The hidden state feature channel corresponds to the first... Weighting coefficients of control points for each base; For the first indivual B-order spline basis functions; Step (2.2.3) Aggregate prediction and tensor splicing: Combine all The univariate nonlinear mapping output values ​​corresponding to the hidden state feature channels are accumulated and summed to obtain the th... The predicted pass rate value for each batch of finished products The formula is shown below: ; For all After executing each batch in parallel, all Vectors are concatenated along the batch dimension to finally output the global predicted pass rate vector. .

[0020] The third step involves the joint training and real-time closed-loop intervention of the discrete manufacturing production line quality prediction model, as follows: Step (3.1) Mean Square Error Calculation: Calculate the true finished product pass rate vector from the training set. The first in element With the training set predicted pass rate vector The corresponding number in element Subtract and square the difference, then apply the result to the total number of batches in the training set. After summing all batches, divide by The mean squared error is derived to quantify the global prediction bias of the finished product qualification rate, as shown in the following formula: ; in, This represents the mean square error.

[0021] Step (3.2) Mechanism error constraint construction: construct the actual energy consumption vector of the training set. The first in element Divide by the actual output vector of the training set The first in element and The maximum value between the product of the two factors and the preset minimum positive constant is used to obtain the true energy consumption vector of a single good product. The One element; will Divide by and The maximum value between the product of the two products and the same minimal positive constants is used to obtain the predicted energy consumption vector per unit of good product. The Each element is subtracted from the sum of its elements to obtain the error value. To avoid the error being non-differentiable at zero, a smoothed L1 loss function is used to approximate the absolute value calculation. This approximation is applied to all elements. After summing all batches, divide by The physical mechanism error is calculated to constrain the predicted yield rate to follow the energy consumption and capacity conservation mechanism of discrete manufacturing production lines, as shown in the following formula: ; ; ; in, express The One element; express The One element; This is a function to find the maximum value. It is a very small positive number; Indicates error in physical mechanism; This represents the smoothed L1 loss function.

[0022] Step (3.3) involves the joint calculation of the total error and model optimization, as follows: Step (3.3.1) Construction of the joint total error: Multiplied by the first fixed weighting coefficient, and The products of these products, multiplied by the second fixed weighting coefficient, are summed to construct the joint total error, thus achieving joint optimization of data prediction accuracy and consistency with physical mechanisms. The formula is shown below: ; in, Indicates the total joint error; Indicates the first fixed weight coefficient; This represents the second fixed weighting coefficient; Step (3.3.2) Network Weight Update: Utilizing the backpropagation algorithm combined with the Adam adaptive moment estimation optimizer, based on... Calculate the gradient matrix and update the network weights to make the update of model parameters conform to the objective physical laws of discrete manufacturing production lines; Step (3.3.3) Root Mean Square Error Extraction: For The root mean square error is calculated by extracting the arithmetic square root. The numerical unit that is restored to be the same as the actual pass rate is used as the triggering criterion for the subsequent cross-domain double-check early stop mechanism, as shown in the following formula: .

[0023] Step (3.4) involves parameter optimization and dual-checked early stopping, executing the learning rate decay mechanism (LRS) and the cross-domain early stopping mechanism (CDSC). The process is as follows: Step (3.4.1) Validation set error extraction and first-order difference calculation: The current training iteration number is denoted as... Using the same computational rules as the training set, the root mean square error on the validation set for the current round is calculated and extracted synchronously. Simultaneously, extract the joint total error on the validation set. The first-order difference is calculated by subtracting the joint total error of the previous iteration from the convergence trend of the model on unseen new batches. The formula is shown below: ; Step (3.4.2) Dynamic Learning Rate Decay Mechanism: LRS is a classic optimization strategy in deep learning used to assist model convergence. It dynamically reduces the learning rate step size to avoid local oscillations near the optimum. LRS applies if and only if the training iterations... If the first-order difference value is greater than zero and persists for three consecutive rounds, it indicates that the optimization process has experienced local oscillations, triggering the LRS mechanism to adjust the network learning rate. Multiply by the attenuation factor to reduce the value; otherwise, keep it unchanged. The formula is as follows: ; in, This indicates the preset attenuation factor; For logic and symbols; It is a universal quantifier symbol; This represents the historical round offset index used to determine consecutive round conditions; Step (3.4.3) Cross-domain Dual Validation Early Stopping Mechanism: CDSC is an overfitting prevention strategy based on the principle of early termination. It triggers forced truncation of network training by extracting and validating the error change trend on the independent validation set. The early stopping flag that triggers the CDSC mechanism is denoted as... When training rounds And the joint total error of the training set in this round If the value is less than the preset tolerance threshold, and the minimum value of the root mean square error on the validation set within the past ten consecutive rounds is greater than or equal to the root mean square error value from ten rounds ago, then the early stopping condition is triggered and... Assigned value Otherwise, assign a value The formula is shown below: ; in, This represents a function that takes the minimum value. This is the tolerance threshold. When for At this time, the network weight update is forcibly terminated and the network weight matrix is ​​fixed to avoid the model from overfitting to a specific batch of finished products and to ensure its generalization ability when applied to actual production lines.

[0024] Step (3.5) generates the confidence interval for risk quantification, and the process is as follows: Step (3.5.1) Inconsistency Score and Quantile Extraction: The conformal prediction algorithm is an uncertainty quantification mechanism independent of the underlying network structure. Based on the empirical error distribution of a finite number of samples, it outputs prediction confidence boundaries with strict statistical probability guarantees. The conformal prediction algorithm is applied to extract the inconsistency score vector on the calibration set. The actual finished product pass rate vector of the calibration set is then used... With the calibration set predicted pass rate vector The first in The non-consistency score of a sample is calculated by subtracting each element from the others and taking the absolute value, as shown in the formula below: ; in, Indicates the first in the calibration set The non-consistency score of each sample; express The Middle One element; express The Middle Each element.

[0025] After calculating the inconsistency scores of all samples in the calibration set, they are combined into a complete inconsistency score vector. ,in This indicates the total number of batches of the calibration set samples. Subsequently, for... The internal elements are arranged in ascending order of numerical value, generating an ordered fraction vector. ,in Indicates the first position after ascending order. A non-consistent fractional scalar, which strictly satisfies the monotonically increasing constraint. ; Based on the finite sample calibration criterion of the conformal prediction algorithm, calculate the discrete quantile index corresponding to the 95% confidence level. The formula is shown below: ; in, This represents the floor function. Ultimately, it will... The first in Each element is used as the error quantile. Extract the output, that is .

[0026] Step (3.5.2) Construction of confidence intervals with forced boundary truncation: Predict the pass rate vector for the test set. Add each of its elements separately and with Take the minimum value to form the upper limit vector of the interval. At the same time, subtract each of its elements separately. and with Take the maximum value to construct the lower bound vector of the interval. To ensure the predicted pass rate is within the range of [0,1], the calculation formula is as follows: ; ; Step (3.5.3) Real-time closed-loop control based on confidence boundaries: Let As a trigger for underlying protective interventions, For the tolerance safety pass rate threshold scalar, when the lower limit of the confidence interval is strictly lower than the preset tolerance safety pass rate threshold, The value is assigned as 1 if the condition is met, and 0 otherwise. The determination formula is as follows: ; in, This indicates that a preset speed reduction and load reduction strategy for the production line equipment has been triggered, along with a safety warning strategy. The edge control terminal takes over the underlying controller, reduces the operating cycle time of the associated equipment to a safe baseline value, and simultaneously sends a high-risk warning signal to the manufacturing execution system, achieving real-time closed-loop control of the discrete manufacturing production line.

[0027] To achieve the aforementioned real-time closed-loop control based on confidence boundaries, such as Figure 3 The diagram shows the topology of an edge intervention hardware and software system designed using an elevator hall door assembly line as an example. The system is divided into three layers from bottom to top in terms of physical hierarchy: the physical equipment layer, the edge control layer, and the information and decision-making layer. Key processing equipment located in the physical equipment layer extracts underlying operating signals through a sensor network and uploads them to the PLC and industrial control computer nodes in the edge control layer via the Modbus protocol and RS485 communication standard. Simultaneously, the form and position tolerance sampling records output by the offline quality inspection terminal, as well as the material properties and actual output issued by the industrial manufacturing execution system, are synchronously aggregated to the PLC and industrial control computer. After being aggregated by the industrial control computer, the aforementioned multi-source heterogeneous time-series data is remotely transmitted to the control server database in the information and decision-making layer using the TCP / IP protocol for time-series storage. The control server database provides raw features to the pass rate prediction and risk quantification modules. Mechanism and Tag Vector , , When this module outputs a flag... When a protective intervention command is issued, the command is sent to the edge control terminal. This edge control terminal takes over the lower-level controller and outputs control commands to the key processing equipment to reduce speed, load, and operating cycle time. Simultaneously, it sends high-risk early warning signals to the upper-level industrial manufacturing execution system, thereby constructing a closed-loop mechanism for quality prediction and safety intervention, realizing real-time closed-loop processing of the elevator hall door production line.

[0028] The performance comparison process of different prediction models in this embodiment is as follows: Step 1: Define the baseline contrast model as follows: Multilayer Perceptron (MLP): As a conventional, purely data-driven feedforward neural network, it relies solely on historical feature data for mapping and fitting. Due to the lack of an internal memory control unit, this network cannot effectively capture long-range dependencies in time-series data.

[0029] Long Short-Term Memory (LSTM) network: This is a typical variant of the Recurrent Neural Network (RNN) architecture. This network controls the flow of information by introducing input gates, forget gates, and output gates, and is specifically designed to process and capture global long-term temporal dependencies in sequential data.

[0030] The Transformer model abandons traditional recurrent network entities and builds a sequence modeling framework entirely based on the standard multi-head self-attention mechanism. This architecture can directly compute and obtain global feature weights, and performs excellently when dealing with complex parallel spatial mappings and long-range temporal features.

[0031] The improved self-attention model (Informer) is a network architecture specifically optimized for long-sequence time-series prediction tasks. Its core lies in utilizing a probabilistic sparse self-attention mechanism to significantly reduce the time complexity and memory consumption during long-sequence inference.

[0032] Step two, define the experimental dataset and evaluation metrics, as follows: Taking a real hall door processing production line of a large elevator manufacturing company as an example, we collected the time-series signals of the operating status of key processing equipment and the corresponding offline quality sampling records for nearly four years, and divided them into training, verification, calibration, and test sets in sequence. For the evaluation of prediction accuracy, the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (COP) were used. In terms of reliability evaluation of industrial protective interventions, the model’s ability to detect anomalies under high-risk conditions on the production line is rigorously measured by using the false alarm rate (MAR) and false alarm rate (FAR) in combination with the preset safety warning threshold.

[0033] Step 3: Analyze and compare the results, as follows: To verify the effectiveness of the method of this invention in non-uniform sparse sampling environments, it was compared with four baseline models, and the experimental results are shown in Table 1. The results in the table show that the pure data-driven baseline models, due to the lack of underlying physical mechanism constraints, exhibit temporal characteristic biases when handling complex conditions. Among them, attention-based networks such as Transformer... While offering some improvement over LSTM, the predicted values ​​exhibit local fluctuations due to the lack of physical boundary constraints, resulting in an FAR of 13.4%. In contrast, the method of this invention adapts to the non-uniform sampling time intervals by introducing continuous-time discretized state-space derivation, and simultaneously utilizes physical mechanism error constraints to limit the divergence trend of network weights. Final test results show that the RMSE of the method of this invention is 0.018%. The accuracy is 0.992, and the MAR and FAR are 1.2% and 1.8%, respectively. In summary, the method of this invention exhibits superior performance in both prediction accuracy and early warning reliability.

[0034] Table 1 shows the performance comparison results of different prediction models;

[0035] This embodiment demonstrates an ablation experiment using the method of the present invention, including the following steps: Step 1: Define the ablation experiment variant model as follows: Based on the complete technical solution of this invention, the following three variant models were constructed by stripping specific modules and compared with the complete model of this invention.

[0036] Removing the continuous-time discretization mechanism (w / o CT-SSM): The lack of a sparse temporal state extrapolation process based on continuous-time discretization forces non-uniform sparse sampling data as an equidistant input sequence, and fails to address the temporal feature distortion caused by non-uniform sparse sampling.

[0037] Remove physical mechanism error constraints (w / o Phy-Loss): The construction lacks mechanism error constraints, and only the mean square error is used for backpropagation during the network weight update stage, thus eliminating the constraints that follow the pipeline energy consumption and capacity conservation mechanism.

[0038] Remove cross-domain double-validation early stopping mechanism (w / o CDSC): The cross-domain double-validation early stopping mechanism is lacking. It only uses the root mean square error of a single validation set to truncate the training and does not perform joint determination with the cross-domain trend of the joint total error.

[0039] Step two, define the experimental dataset and evaluation metrics, as follows: The actual dataset and data partitioning benchmark of the elevator manufacturing workshop in Specific Implementation Method 2 are used, and the same evaluation index system is adopted for performance evaluation.

[0040] Step 3: Analyze the ablation comparison results, as follows: To verify the effectiveness of the core mechanisms of this invention, ablation experiments of the core modules were designed. The performance metrics of the complete model and three variant models of this invention were compared on the test set. The experimental results are shown in Table 2. The results in the table show that removing the continuous-time discretization mechanism (w / o CT-SSM), by forcing non-uniform sampling to be equidistant, cannot effectively overcome the temporal feature distortion caused by non-uniform sparse sampling, leading to its... The error rate dropped to 0.945, and the baseline drift caused by equipment degradation increased to 6.2%. Removing the physical mechanism error constraint (w / o Phy-Loss) resulted in an abnormal jump in network output due to the lack of supervision from the energy consumption and capacity conservation mechanism, making the network output susceptible to local noise interference. The FAR increased to 12.8%. Removing the cross-domain double-validation early stopping mechanism (w / o CDSC) caused local overfitting due to early stopping on a single validation set, leading to a decrease in generalization ability during long-term testing and an increase in errors for all prediction metrics. Experimental data show that the complete model of this invention, after combining the above core mechanisms, exhibits better performance across all metrics.

[0041] Table 2 shows the performance indicators of the core module ablation experiment.

[0042] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A discrete manufacturing pipeline quality prediction method based on joint mechanism and data-driven approach, characterized in that, The method includes the following steps: Step 1: Data acquisition and preprocessing in discrete manufacturing assembly line industrial scenarios; The second step, state-space derivation and univariate nonlinear analysis of the sparse temporal characteristics of the discrete manufacturing pipeline, is as follows: Step (2.1) involves sparse temporal state deduction based on continuous-time discretization, and standardizing the feature tensor containing the underlying equipment operating state and material properties. Input into the deduction module; Step (2.2) analyzes the nonlinear contribution of each hidden state variable to the pass rate in order to achieve process traceability of finished product manufacturing deviations; The third step involves the joint training and real-time closed-loop intervention of the discrete manufacturing production line quality prediction model, as follows: Step (3.1) Calculation of mean square error; Step (3.2) Construction of mechanism error constraints; Step (3.3) Combine total error calculation and model optimization; Step (3.4) Parameter optimization and double check for early stopping; Step (3.5) Quantitative risk confidence interval generation.

2. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 1, characterized in that, The process of the first step is as follows: Step (1.1) Multi-dimensional industrial time-series feature acquisition: The underlying operating signals of key processing equipment are captured by the sensor network deployed on the discrete manufacturing production line. The material properties and form and position tolerance sampling records are captured simultaneously by combining the industrial manufacturing execution system (MES) and offline quality inspection terminal. The multi-source data is then aggregated to the PLC and industrial control computer. The Modbus protocol and RS485 communication standard are used for low-level data interaction in the production line control cabinet. Finally, the TCP / IP protocol is used to remotely transmit the data to the control server database in real time for time-series storage. The original input feature tensor is extracted, and finally the actual output vector, actual energy consumption vector and global real finished product qualification rate vector corresponding to each sample batch are collected synchronously. Step (1.2) Data normalization preprocessing: The original input feature tensor is normalized using a standard deviation scaling algorithm based on a dynamic sliding time window; Step (1.3) Dataset Partitioning: Generate a global sample partitioning index according to the time series order. Perform non-intersecting domain partitioning on the preprocessed standardized feature tensor and its corresponding global true finished product pass rate vector, dividing the total number of batches into... The global data is divided into training set, validation set, calibration set and test set in sequence. Based on the global sample partitioning index, the actual output vector and the actual energy consumption vector are mapped and segmented simultaneously to generate feature tensors and vector slices of corresponding dimensions for each set, which are used for subsequent network inference, physical mechanism error calculation and confidence boundary quantization.

3. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 2, characterized in that, The process of step (1.2) is as follows: Step (1.2.1) Setting the dynamic sliding time window and truncation rule: Set the original input feature tensor The Middle The first batch of finished products, the first The first time step The elements of each feature dimension are denoted as Batch index Starting from 1, the preset maximum sample length is... Dynamic sliding time window Batch dimension Perform sliding; define the actual effective backtracking time window length for this batch as... In the current Strictly less than hour, The value is In the present Greater than or equal to hour, Cut off and fix as ; Step (1.2.2) Calculation of dynamic statistics: Calculate the dynamic arithmetic mean covering all historical batches within the current sliding time window. With dynamic standard deviation ; Step (1.2.3) Temporal feature normalization mapping: ... Each of them Subtract the corresponding Then divide by the corresponding The normalized feature elements after mapping are obtained. The final generated Together they constitute the standardized feature tensor .

4. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in any one of claims 1 to 3, characterized in that, The process of step (2.1) is as follows: Step (2.1.1) Construction of continuous-time state space: The Middle The finished product batch in the first The discrete input feature vector extracted at each time step is denoted as . Based on the zero-order hold assumption, the function is kept constant within adjacent sampling intervals, thus transforming it into a continuous-time function. The first derivative of the hidden state vector is calculated by multiplying the basic state transition matrix with the hidden state vector in continuous time, and then adding the product of the input projection matrix and the continuous input signal vector. ; Step (2.1.2) Dynamic step size operator extraction: For the first step size operator... The first batch of finished products will be the first batch of finished products. Discrete time steps Multiplying the result by the internal weight matrix and adding the bias vector, then inputting it into a smooth nonlinear activation function, the output has... Dynamic step size vector of each channel dimension ; Step (2.1.3) Discretization mapping of state parameters: The vector... Transform into a diagonal matrix and merge with After multiplying, the matrix exponent is obtained to generate the discretized state transition matrix. ; Based on the discretization derivation of the zero-order hold, the simplified discretized input matrix is ​​calculated. ; Step (2.1.4) Discrete State Update and Feature Concatenation: Initialize the first... The hidden state vector of each finished batch at time step 0 It is an all-zero vector, representing the initial hidden state of this batch at the start of processing; Calculate the hidden state vector updated at the current time step. After extrapolating to the end of the time window, all The hidden state vectors generated at each time step are concatenated along the time series dimension to generate a single-batch time series feature matrix; subsequently, for all The above simulation and splicing operations are performed in parallel in several independent batches, combining each... The batches are stacked along the batch dimension to form a global hidden state feature tensor.

5. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 4, characterized in that, The process of step (2.2) is as follows: Step (2.2.1) Basis feature extraction of independent variables: from Extract the last time step The corresponding feature vector will be the extracted first... The first batch of finished products The elements of each feature dimension are denoted as , as the input of the independent variable for subsequent quantification of the contribution of a single hidden state variable to the pass rate; Step (2.2.2) Univariate nonlinear mapping: Enter them separately In the B-spline basis function, the corresponding control point weight coefficients are multiplied and summed to calculate the univariate nonlinear mapping output value; Step (2.2.3) Aggregate prediction and tensor splicing: Combine all The univariate nonlinear mapping output values ​​corresponding to the hidden state feature channels are accumulated and summed to obtain the th... The predicted pass rate value for each batch of finished products For all After executing each batch in parallel, all Vectors are concatenated along the batch dimension to finally output the global predicted pass rate vector. .

6. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 5, characterized in that, In step (3.1), the training set's actual finished product pass rate vector is... The first in element With the training set predicted pass rate vector The corresponding number in element Subtract and square the difference, then apply the result to the total number of batches in the training set. After summing all batches, divide by The mean square error is derived to quantify the global prediction bias of the finished product qualification rate.

7. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 6, characterized in that, In step (3.2), the mechanism error constraint is constructed by: converting the actual energy consumption vector of the training set... The first in element Divide by the actual output vector of the training set The first in element and The maximum value between the product of the two factors and the preset minimum positive constant is used to obtain the true energy consumption vector of a single good product. The One element; will Divide by and The maximum value between the product of the two products and the same minimal positive constants is used to obtain the predicted energy consumption vector per unit of good product. The One element; Subtracting the two values ​​yields the error value. To avoid the error being non-differentiable at zero, a smoothed L1 loss function is used to approximate the absolute value calculation, and this is applied to all... After summing all batches, divide by The physical mechanism error is calculated to constrain the predicted yield rate to follow the energy consumption and capacity conservation mechanism of discrete manufacturing production lines.

8. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 7, characterized in that, The process of step (3.3) is as follows: Step (3.3.1) Construction of the joint total error: Multiplied by the first fixed weighting coefficient, and The products of the two products, multiplied by the second fixed weighting coefficient, are added together to construct the joint total error, thereby achieving joint optimization of data prediction accuracy and consistency with physical mechanisms. Step (3.3.2) Network Weight Update: Utilizing the backpropagation algorithm combined with the Adam adaptive moment estimation optimizer, based on... Calculate the gradient matrix and update the network weights to make the update of model parameters conform to the objective physical laws of discrete manufacturing production lines; Step (3.3.3) Root Mean Square Error Extraction: For The root mean square error is calculated by extracting the arithmetic square root. The numerical unit that is restored to be the same as the actual pass rate is used as the triggering basis for the early stop mechanism of cross-domain dual verification.

9. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 8, characterized in that, In step (3.4), the learning rate decay mechanism and the cross-domain early stopping mechanism are executed, as follows: Step (3.4.1) Validation set error extraction and first-order difference calculation: The current training iteration number is denoted as... Using the same computational rules as the training set, the root mean square error on the validation set for the current round is calculated and extracted synchronously. Simultaneously, extract the joint total error on the validation set. The first-order difference is calculated by subtracting the joint total error from the previous iteration. ; Step (3.4.2) Learning rate dynamic decay mechanism: if and only if the training iteration rounds If the first-order difference value is greater than zero and persists for three consecutive rounds, it indicates that the optimization process has experienced local oscillations, triggering the LRS mechanism to adjust the network learning rate. Multiply by the attenuation factor and adjust downwards; otherwise, keep it unchanged. Step (3.4.3) Cross-domain Dual Check Early Stop Mechanism: Record the early stop flag that triggers the CDSC mechanism as... When training rounds And the joint total error of the training set in this round If the value is less than the preset tolerance threshold, and the minimum value of the root mean square error on the validation set within the past ten consecutive rounds is greater than or equal to the root mean square error value from ten rounds ago, then the early stopping condition is triggered and... Assign a value of 1 if the value is 1, otherwise assign a value of 0.

10. The discrete manufacturing pipeline quality prediction method based on joint mechanism and data driving as described in claim 7, characterized in that, The process of step (3.5) is as follows: Step (3.5.1) Inconsistency score and quantile extraction: Apply the conformal prediction algorithm to extract the inconsistency score vector on the calibration set, and then extract the true finished product pass rate vector from the calibration set. With the calibration set predicted pass rate vector The first in Subtract each element from the others and extract the absolute value to get the non-consistency score for that sample. After calculating the non-consistency scores of all samples in the calibration set, they are combined into a complete non-consistency score vector. Subsequently, the elements within the non-consistency score vector are sorted in ascending order of numerical value to generate an ordered score vector. Based on the finite sample calibration criterion of the conformal prediction algorithm, the discrete quantile index corresponding to the set percentage confidence level is calculated. ; Step (3.5.2) Construction of confidence intervals with forced boundary truncation: Predict the pass rate vector for the test set. Add each of its elements separately and with Take the minimum value to form the upper limit vector of the interval. At the same time, subtract each of its elements separately. and with Take the maximum value to form the lower limit vector of the interval; Step (3.5.3) Real-time closed-loop control based on confidence boundaries: Let As a trigger for underlying protective interventions, For the tolerance safety pass rate threshold scalar, when the lower limit of the confidence interval is strictly lower than the preset tolerance safety pass rate threshold, Assign a value of 1 if the value is 1, otherwise assign a value of 0.

Images