A multi-stage manufacturing process quality prediction method based on cross-modal phase perception process memory

Abstract

The application discloses a kind of multi-stage manufacturing process quality prediction methods based on cross-modal phase perception process memory.The method utilizes segmented historical production cycle data to train quality prediction network, so that only input preset process parameters can obtain quality prediction value in inference process.Process memory library explicitly saves phase quality formation evidence in historical sensor track;Dynamic programming based on channel phase breakpoint consistency suppresses error breakpoint caused by single noise channel;Process parameter guided cross-modal attention filters important sensor response according to different parameter conditions;Auxiliary quality supervision is applied to cycle-level process memory vector, so that process memory is more oriented to final quality.The method can predict quality before production, without current cycle sensor data, and without real trial for each set of set process parameters, thereby reducing production consumables and test time.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing quality prediction technology, specifically relating to a multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory. This method is applicable to injection molding, die casting, and other multi-stage batch manufacturing processes. It can predict product quality based solely on preset process parameters before the start of a new production cycle and assist in optimizing those parameters. Background Technology

[0002] In intelligent manufacturing, product quality prediction is a crucial foundation for stable production and process optimization. Taking multi-stage batch manufacturing processes such as injection molding and die casting as examples, product quality is jointly affected by stages such as injection, holding pressure, cooling, and mold opening. The process signals such as pressure, speed, position, temperature, and force in each stage exhibit significant differences. Traditional parameter optimization typically relies on manual experience and multiple rounds of mold trials. After each parameter adjustment, actual production must be completed and the weight, dimensions, transmittance, or appearance quality must be tested. This results in long feedback cycles and high material consumption, making it difficult to meet the demands of rapid prototyping and stable mass production.

[0003] Existing quality prediction methods largely rely on sensor trajectories collected in the current cycle. These methods can utilize real process information from the current cycle, but they can only operate after production begins, making it impossible to screen out high-risk parameters before production. Other methods simply use preset process parameters to establish a direct mapping between parameters and quality. While this allows for pre-prediction, the mapping process does not utilize sensor trajectories recorded in historical cycles, easily leading to the loss of crucial evidence regarding the quality formation process.

[0004] Furthermore, historical sensor trajectories from multi-stage manufacturing processes should not be simply compressed into a global sequence. Different stages are governed by different physical mechanisms, and without proper segmentation, stage differences can easily become blurred; relying solely on fixed rules or single-channel segmentation may lead to erroneous breakpoints due to stage time drift or sensor noise. Therefore, a method is needed that can extract staged, quality-sensitive process memories from historical cycles and can be retrieved by new process parameters, enabling the model to perform pre-production quality predictions even without current cycle sensor data. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory. This method divides the multi-channel sensor trajectories in the historical production cycle into physically meaningful stages, uses a cross-modal attention mechanism guided by preset process parameters to filter relevant sensor responses within each stage, and constructs a searchable process memory library through time-adaptive attention and quality supervision. During inference in a new cycle, the model only needs to receive the current preset process parameters to query the historical process memory and output the predicted product quality value.

[0006] A multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory includes the following steps:

[0007] (1) Collect multiple production cycle data of multi-stage manufacturing. Each cycle data includes preset process parameters, multi-channel process sensor trajectories and measured product quality.

[0008] (2) Take any untraversed production cycle as the current cycle, and smooth and standardize each sensor trajectory within it to obtain the corresponding standardized time series;

[0009] (3) Perform single-channel dynamic programming on the standardized time series corresponding to each sensor channel to obtain the corresponding candidate stage breakpoint set;

[0010] (4) Determine the consistency between different channel breakpoints based on the candidate stage breakpoint set, and select a reliable sensor channel set;

[0011] (5) All sensor channels in the reliable sensor channel set are concatenated into a multivariable sequence by column and dynamic programming is performed to obtain the stage boundary of the current cycle; then, according to the stage boundary, the multi-channel sensor trajectory of the current cycle is synchronously extracted into multiple stages and formed into multiple stage data blocks arranged in the process order.

[0012] (6) Repeat steps (2) to (5) until all production cycles have been traversed, and obtain the stage data blocks corresponding to all production cycles arranged in the process sequence;

[0013] (7) The quality prediction network is trained and optimized using the collected preset process parameters, measured quality and stage data blocks obtained in step (6);

[0014] (8) Using the preset process parameters of the cycle to be predicted as input, the corresponding quality prediction value is output by the optimized quality prediction network.

[0015] In step (1) above:

[0016] Preferably, the multi-stage manufacturing process is one of injection molding, die casting, or hot pressing.

[0017] Preferably, the preset process parameters include one or more of the following: mold temperature, injection speed, injection pressure, temperature of each zone of the barrel, and hot runner temperature.

[0018] Preferably, each sensor channel corresponds to a process variable, and the collected data is a sensor trajectory; the process variables include one or more of the following: clamping force, injection pressure, injection speed, screw position, screw speed, and nozzle temperature.

[0019] As a preferred option, the measured quality includes one or more of the following: product weight, critical dimensions and dimensional deviations, shrinkage rate, warpage, surface appearance defect level, mechanical properties, transmittance, haze, and color difference.

[0020] In step (2) above:

[0021] Preferably, a rolling median is used to smooth each sensor trajectory, resulting in a smooth sequence for each trajectory consistent with the original sampling length, thereby suppressing transient spikes and isolated outliers (impulse noise). Let the window length be... Where h is a non-negative integer; for the original value at sampling time t A rolling median smoothing was performed using a window centered at time t containing w sampling points, resulting in the smoothed outcome:

[0022]

[0023] in, This indicates a window centered at time t with a length of . The results of local median smoothing; This indicates taking the median.

[0024] As a preferred option, for residual outliers (remaining outlier sampling points) that still exceed the preset statistical threshold after smoothing, linear interpolation of adjacent normal sampling points or replacement with the local median can be used to process the outliers.

[0025] Preferably, each smoothed sequence is independently Z-score standardized to eliminate the influence of different units and amplitude ranges, resulting in a standardized time series. Z-score standardization is a conventional scale normalization process, and its formula is as follows:

[0026]

[0027] in, Represents a standardized time series; and Let represent the mean and standard deviation of the smoothed sequence (the smoothed result at all times), respectively. To prevent extremely small positive numbers with a denominator of zero, it is preferable to select... .

[0028] In step (3) above:

[0029] If a production cycle is divided into K process stages, then each sensor channel is obtained through single-channel dynamic programming. The candidate stage breakpoints are arranged chronologically and formed into a set; the set of candidate stage breakpoints for sensor channel i is denoted as . The elements in the set are all the sampling times or physical times corresponding to the candidate stage breakpoints.

[0030] In step (4) above:

[0031] As a preferred option, the specific steps for obtaining a reliable set of sensor channels are as follows:

[0032] (4.1) For any two sensor channels, calculate the difference between each pair of candidate breakpoints with the same index in the candidate stage breakpoint set for each of them to obtain multiple corresponding time differences, and take the maximum value as the maximum time difference;

[0033] (4.2) Determine the magnitude of the maximum time difference relative to the set threshold:

[0034] If the maximum time difference is less than or equal to the set threshold, the two sensor channels are considered to have the same stage breakpoint and are consistent.

[0035] Otherwise, the breakpoints of the two stages are inconsistent;

[0036] (4.3) Repeat steps (4.1) to (4.2) to find all sensor channel pairs with consistent breakpoints in the current cycle and form a temporary set;

[0037] (4.4) Using each sensor channel in the temporary set as a node and the consistency relationship of stage breakpoints as an edge, an undirected graph is constructed. The subset of channels with the largest number of sensor channels in the graph and whose stage breakpoints are consistent with each other is selected as the reliable sensor channel set.

[0038] As a further preferred option, in step (4.1), the maximum time difference is calculated using the following formula:

[0039]

[0040] in, This represents the maximum time difference between the i-th and j-th sensor channels; and These represent the k-th breakpoint times of sensor channels i and j, respectively; Indicates in Take within range The largest value; This indicates the number of elements in the candidate stage breakpoint set.

[0041] Reflects the worst alignment error of the two sensor channels at all transition moments throughout the entire process; when Not exceeding (less than or equal to) the consistency threshold (set threshold) At that time, it was assumed that the stage breakpoints of the two sensor channels were consistent.

[0042] As a further preferred option, in the reliable sensor channel set obtained in step (4.4), any two sensor channels have consistent stage breakpoints; the selection formula is:

[0043] ,satisfy

[0044] in, This represents the set of reliable sensor channels that meets the pairwise breakpoint consistency constraint and has the largest number of sensor channels. This indicates that the set with the largest number of sensor channels is selected from the candidate channel subset; Represents any subset of candidate channels The number of sensor channels in the process. In step (4.4), by constructing an undirected graph, multiple candidate channel subsets can be obtained, and any two sensor channels in each candidate channel subset can meet the stage breakpoint consistency requirement.

[0045] As a further optimization, if no sensor channel pairs meet the breakpoint consistency requirement in the current cycle, then for each sensor channel, the corresponding channel quality score is calculated based on the stage fitting residual of the single-channel dynamic programming and the candidate breakpoint interval, and the sensor channel with the highest channel quality score is taken as the reliable sensor channel and proceeds to the next step.

[0046] Among them, the current cycle number Channel quality score for each sensor channel Calculated using the following formula:

[0047]

[0048] In the formula, This indicates the total number of sampling points within the current period; Indicates the number of stages in the current cycle; Indicates the first The sensor channel in the first The sum of squared residuals within each stage, that is, the sum of squared deviations of each sampled value from the mean of that stage; Indicates the first Standardized time series of each sensor channel; Represents the population variance of the standardized time series; The time interval between adjacent stage boundaries is calculated using the following formula:

[0049]

[0050] in, Indicates the first The first sensor channel The boundaries of each stage; and let, , ;

[0051] Calculated using the following formula:

[0052]

[0053] In the formula, The value of the i-th sensor channel at time t; This represents the overall mean of the i-th sensor channel.

[0054] The first term in the above channel quality score calculation formula measures the fitting quality after dynamic programming segmentation; the second term measures the proportion of the shortest stage to the length of the entire production cycle. A higher channel quality score indicates a smaller dynamic programming fitting error and clearer breakpoint intervals for the corresponding sensor channel.

[0055] In step (5) above:

[0056] As a preferred approach, the specific steps for performing dynamic programming on multivariable sequences are as follows:

[0057] For a reliable set of sensor channels, the values ​​of all sensor channels corresponding to each sampling moment constitute the sampling vector for that sampling moment. The number of stages in the current cycle is determined based on the number of candidate stage breakpoints. For each stage, the sum of the squared deviations between the sampling vector at each sampling moment and the mean vector of all sampling moments in that production stage is calculated. With the goal of minimizing the sum of the squared deviations of all production stages, the globally optimal combination of breakpoints is searched under the constraint of maintaining the time order of breakpoints, thereby obtaining the stage boundary of the current cycle.

[0058] Specifically, the sum of squared deviations (sum of squared deviations) for the k-th stage can be expressed as:

[0059]

[0060] in, Let represent the sum of squared deviations in the k-th stage, where k∈[1,K] and K is the stage number of the current cycle; Let i represent the sampling vector at the i-th sampling time in the k-th stage, where i∈[1,N] and N is the number of sampling points in the k-th stage; The mean vector for the k-th stage is represented by the following formula:

[0061]

[0062] Each stage data block contains the time series of all sensor channels corresponding to that stage within the current cycle. For the nth stage of the mth cycle, its sensor trajectory (time series) is denoted as... Dimension is Where S is the number of sensor channels, This represents the number of sampling points in the nth stage.

[0063] In step (7) above:

[0064] Preferably, the quality prediction network includes a historical cycle encoder, a process memory building module, a parameter query decoder, a quality prediction head, and an auxiliary quality prediction head used only during the training phase; the specific process for training and optimizing it is as follows:

[0065] (7.1) For each production cycle, the corresponding process memory vector is output through the historical cycle encoder, with its preset process parameters and corresponding stage data blocks as input.

[0066] (7.2) Take any untraversed production cycle as the current cycle, and construct the current process memory library using the process memory library construction module based on the process memory vectors corresponding to the other production cycles besides the current cycle.

[0067] (7.3) Using the preset process parameters of the current cycle and the current process memory as input, the context vector is obtained by using the parameter query decoder;

[0068] (7.4) Using the context vector as input, obtain the main quality prediction vector for the current cycle through the quality prediction head, and calculate the main loss based on the measured quality of the current cycle; at the same time, input the process memory vector of each production cycle that builds the current process memory into the auxiliary quality prediction head to obtain the corresponding auxiliary quality prediction vector, and calculate the auxiliary loss based on the measured quality of each production cycle that builds the current process memory; and then calculate the total loss based on the main loss and the auxiliary loss.

[0069] (7.5) Perform backpropagation with the goal of minimizing total loss, and update the network parameters of the historical cycle encoder, process memory library construction module, parameter query decoder, quality prediction head and auxiliary quality prediction head;

[0070] (7.6) Take each production cycle sequentially or in batches as the current cycle, and repeat steps (7.1) to (7.5) until the preset training termination condition is met, and obtain the optimized network parameters;

[0071] (7.7) All production cycles are re-encoded using the optimized historical cycle encoder to obtain the corresponding process memory vector, and the final process memory library is constructed using the optimized process memory library construction module.

[0072] (7.8) The final process memory is used together with the optimized parameter query decoder and quality prediction head for quality prediction of the period to be predicted.

[0073] As a further optimization, the specific operation for obtaining a process memory vector for a production cycle in step (7.1) is as follows:

[0074] Each stage data block within the production cycle is subjected to one-dimensional convolution within the stage to extract the stage sensor feature matrix; preset process parameters are projected through a fully connected layer to obtain a query vector; using the query vector as the query and the stage sensor feature matrix as the key and value, cross-modal attention is used to calculate the stage fusion features; then time-adaptive attention is used to calculate the attention weights of each stage, and the fusion features of each stage are weighted and summed to obtain the process memory vector of the production cycle.

[0075] As a further preferred option, the specific process for extracting the stage sensor feature matrix of any stage data block is as follows:

[0076] Using a stage-level multimodal encoder, a stage data block is first input into a two-layer one-dimensional convolutional network. Each layer includes one-dimensional convolution, batch normalization, and ReLU activation to extract local temporal features of each sensor channel. Then, global average pooling and global max pooling are performed on the local temporal features of each sensor channel respectively. Finally, the two pooling results are concatenated and projected onto D dimensions to obtain the stage sensor feature matrix.

[0077] As a further preferred option, the formula for calculating the stage fusion features using cross-modal attention is as follows:

[0078]

[0079] in, This represents the stage fusion feature of the m-th cycle and the n-th stage; The query vector serves as the query Q; Let K be the stage sensor feature matrix for the m-th period and n-th stage, and V be the key and value, respectively; MHA represents multi-head cross-attention; Dropout represents random deactivation; LayerNorm represents layer normalization; residual term Used for stable training;

[0080] The formula for calculating the process memory vector is:

[0081]

[0082]

[0083]

[0084] in, The process memory vector represents the m-th production cycle; n∈[1,N], j∈[1,N], where N represents the total number of stages in the m-th production cycle; This represents the normalized attention weight for the nth stage; This represents the stage fusion feature of the nth stage; and These are the trainable weight matrix and bias vector for time-adaptive attention, respectively. For time-adaptive attention, is a trainable context vector; exp represents the exponential function. This represents the hyperbolic tangent activation function.

[0085] As a further preferred option, in step (7.2), the formula for constructing the process memory is:

[0086]

[0087] in, Represents the procedural memory; This represents the memory vector of M processes. Stacked by row; The weight matrix is ​​trainable. is a trainable bias vector, broadcast over M rows; ReLU represents a non-linear activation function.

[0088] As a further preferred option, in step (7.3), the formula for obtaining the context vector using the parameter query decoder is:

[0089]

[0090] in, This represents the obtained context vector; The query vector is obtained by mapping the current cycle's preset process parameter vector (composed of preset process parameters arranged in a fixed order) through a fully connected layer. The input process memory; MHA represents multi-head cross-attention; Dropout represents random deactivation; LayerNorm represents layer normalization; residual terms Used for stable training.

[0091] As a further optimization, in step (7.4), the formula for calculating the main loss is:

[0092]

[0093] Where L represents the number of quality indicators, ; Indicates the first [number]th ... The measured quality of each quality indicator; Represents the current period principal mass prediction vector The One portion, ;

[0094] Principal mass prediction vector The following formula is used to calculate:

[0095]

[0096] In the formula, and These represent the trainable weight matrices of the quality prediction head; and represents the trainable bias vector of the quality prediction head, respectively; ReLU represents the nonlinear activation function; Dropout represents random deactivation.

[0097] As a further preferred option, in step (7.4), the formula for calculating the auxiliary loss is:

[0098]

[0099] Where L represents the number of quality indicators, ; This indicates the number of production cycles in the current process memory. ; This indicates the m-th production cycle. The measured quality of each quality indicator; The m-th auxiliary quality prediction vector represents the first... One portion, ;

[0100] Auxiliary quality prediction vector The following formula is used to calculate:

[0101]

[0102] In the formula, and These represent the trainable weight matrices of the auxiliary quality prediction head; and These represent the trainable bias vectors of the auxiliary quality prediction head, respectively. Represents the process memory vector for the m-th production cycle; ReLU represents the nonlinear activation function.

[0103] As a further optimization, in step (7.4), the total loss is obtained by the following formula:

[0104]

[0105] in, Indicates the total loss; λ represents the auxiliary supervision weight, used to balance the main loss and auxiliary loss, and is preferably determined based on the validation set; specifically, λ is set as a hyperparameter before model training, and is then determined by training the model on the candidate value set and comparing the performance of the validation set.

[0106] In step (7.6), each production cycle is taken sequentially or in batches as the current cycle, and steps (7.1) to (7.5) are repeated. The network parameters are updated according to the total loss until the preset training termination condition is met. The network parameters are updated once after each training iteration, and the next training is carried out based on the network parameters updated in the previous iteration.

[0107] For the training target sample (current cycle), only its preset process parameters are used as queries for the parameter query decoder; the preset process parameters, sensor trajectories and measured quality of the other training samples are used to form the current process memory through the historical cycle encoder and process memory library construction module. The leave-one-out method is used to avoid the target sample from retrieving itself, and the network parameters are jointly updated according to the total loss.

[0108] Preferably, the trained and optimized quality prediction network can be embedded as a surrogate model into a multi-objective optimization algorithm, with the optimization objective being to minimize the deviation between the predicted values ​​of each quality indicator and the target values, thereby obtaining recommended process parameters. In addition to pre-production quality prediction, the method of this invention can also be used for pre-production risk identification and process parameter recommendation.

[0109] Among them, the NSGA-II algorithm can be selected for multi-objective optimization.

[0110] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0111] This invention presents a multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory. By segmenting and processing historical production cycle data across multiple stages, and using the processed data to train a quality prediction network, the predicted quality can be obtained by inputting only preset process parameters during subsequent quality prediction. That is, the inference stage does not require current cycle sensor data, allowing quality prediction to be performed before production without the need for actual trial runs for each set process parameter, effectively reducing production material consumption and testing time. The process memory library explicitly stores staged quality evidence in historical sensor trajectories, which is superior to simply learning the direct mapping from parameters to quality. Dynamic programming based on channel stage breakpoint consistency can suppress erroneous breakpoints caused by single noisy channels. Process parameter-guided cross-modal attention can filter important sensor responses based on different parameter conditions. Auxiliary quality supervision is applied to the cycle-level process memory vector, making the process memory more oriented towards final quality, which is more conducive to improving prediction accuracy and process optimization reliability. This invention achieves parameter-only pre-production quality prediction for multi-stage manufacturing processes and can be used for pre-production risk identification, process parameter recommendation, and intelligent manufacturing quality decision-making. Attached Figure Description

[0112] Figure 1 This is a schematic diagram of the overall structure of a multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory;

[0113] Figure 2 The flowchart shows the dynamic programming stage segmentation process based on the consistency of channel stage breakpoints. (a) represents the standardized time series corresponding to different sensor channels (clamping force CF, injection pressure IP, screw position SP, etc.); (b) is a diagram showing whether two sensor channels have stage breakpoint consistency based on the maximum time difference between them. This diagram is a matrix diagram. In the matrix, CF (clamping force), IP (injection pressure), IV (injection speed), SP (screw position), SV (screw speed), and NT (nozzle temperature) in the rows and columns represent different sensor channels. The elements (values) in the matrix represent the maximum time difference between the two sensor channels in the corresponding row and column. TOL is the consistency threshold; (c) is a diagram showing how the multi-channel sensor trajectory of the current cycle is synchronously extracted into multiple stages according to the stage boundary and formed into multiple stage data blocks arranged in the process sequence. The vertical dashed lines in the diagram represent the stage boundaries. Detailed Implementation

[0114] The present invention will be further described below with reference to the accompanying drawings and injection molding process, but the present invention is not limited to this embodiment. For manufacturing scenarios with staged process characteristics such as die casting, hot pressing, and composite material molding, as long as the preset process parameters of historical production cycles, multi-channel process sensor trajectories, and corresponding product quality inspection values ​​(measured quality) can be obtained, the method of the present invention can be used.

[0115] This method employs an encoder-decoder workflow. During the training phase, the historical cycle encoder generates a process memory vector using preset process parameters from historical production cycles and multi-channel sensor trajectories, and implements auxiliary supervision using corresponding measured quality data. The process memory library construction module, parameter query decoder, and quality prediction head are jointly trained with the historical cycle encoder. During the inference phase, the parameter query decoder only receives preset process parameters for the current cycle and outputs the quality prediction value for the current cycle by querying a fixed historical process memory library, without requiring sensor data from the current cycle.

[0116] Example:

[0117] like Figure 1 As shown, a multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory includes the following steps:

[0118] Step 1: Collect data from multiple historical production cycles across multiple manufacturing stages. Each historical cycle includes preset process parameters, multi-channel process sensor trajectories, and the measured product quality (measured quality) corresponding to that cycle. Each sensor channel corresponds to a process variable, and the data are combined at a unified sampling time to form a complete time series, i.e., the sensor trajectory.

[0119] Multi-stage manufacturing includes injection molding, die casting, or hot pressing.

[0120] Taking injection-molded products as an example, process variables may include clamping force, injection pressure, injection speed, screw position, screw speed, and nozzle temperature. Preset process parameters may include mold temperature, injection speed, injection pressure, barrel temperatures in each zone, and hot runner temperature. Quality indicators for ordinary injection-molded products may include product weight, critical dimensions and dimensional deviations, shrinkage, warpage, surface appearance defect level, and mechanical properties; for transparent or optical products, transmittance, haze, and color difference may also be included.

[0121] For the m-th production cycle, the preset process parameters are arranged in a fixed order to form a preset process parameter vector, denoted as . The multi-channel sensor trajectory is recorded as The measured quality indicators are arranged in a fixed order to form a measured quality vector, denoted as . .in, The dimension is the number of process parameters. It consists of complete time series corresponding to each process variable. The dimension is the number of quality indicators.

[0122] Step 2: Take any untraversed production cycle m as the current cycle, and smooth and standardize the trajectory of each sensor within it to obtain a standardized time series suitable for stage segmentation and network modeling.

[0123] Specifically, for Each sensor trajectory in the process undergoes rolling median smoothing to produce a smooth sequence consistent with the original sampling length, thereby suppressing transient spikes and isolated outliers.

[0124] Subsequently, the smoothed sequences corresponding to each sensor trajectory were independently Z-score standardized to eliminate the influence of different dimensions and amplitude ranges.

[0125] Set window length Where h is a non-negative integer; for the original value of the sensor trajectory at sampling time t By using a window centered at time t and containing w sampling points for rolling median smoothing and standardization, we obtain:

[0126]

[0127] in, This indicates a window centered at t with a length of t. The results of local median smoothing; This indicates taking the median; Represents a standardized time series; and These represent the mean and standard deviation of the smoothed sequence of the sensor trajectory (the smoothed result at all times), respectively. To prevent extremely small positive numbers with a denominator of zero, it is preferable to select... .

[0128] Optionally, for residual outliers that still exceed a preset statistical threshold after smoothing, linear interpolation of adjacent normal sampling points or replacement with the local median can be used.

[0129] In this step, rolling median smoothing is responsible for suppressing impulse noise, outlier replacement is used to handle the remaining outlier sampling points, and Z-score normalization is a conventional scale normalization process.

[0130] Step 3: Standardize the time series corresponding to each sensor channel ( Figure 2 In (a) each channel performs single-channel dynamic programming to obtain the corresponding candidate stage breakpoint set, and calculates the consistency between breakpoints of different sensor channels to select a reliable sensor channel set.

[0131] Suppose a production cycle is divided into K process stages, then each sensor channel is obtained through single-channel dynamic programming. There are candidate stage breakpoints arranged chronologically. The set of candidate stage breakpoints for channel i is denoted as . The elements in the set are the sampling time or physical time corresponding to the breakpoint.

[0132] The specific steps for selecting a reliable set of sensor channels are as follows:

[0133] (3.1) For any two sensor channels i and j, calculate the difference between each pair of candidate breakpoints k with the same index in the candidate stage breakpoint set corresponding to them, and obtain multiple corresponding time differences. Take the maximum value among them as the maximum time difference. The calculation formula is as follows:

[0134]

[0135] in, This represents the maximum time difference between the i-th and j-th sensor channels; and These represent the k-th breakpoint times of sensor channels i and j, respectively; Indicates in Take within range The largest The maximum value is calculated by comparing the time differences of all corresponding breakpoints, rather than repeatedly taking the maximum value for a single fixed difference.

[0136] (3.2) As Figure 2 As shown in (b), the maximum time difference is compared with a set threshold:

[0137] If the maximum time difference is less than or equal to the set threshold, the two sensor channels are considered to have the same stage breakpoint and are consistent.

[0138] Otherwise, the breakpoints of the two stages will be inconsistent.

[0139] Reflects the worst alignment error of the two sensor channels at all transition moments throughout the entire process; when Not exceeding the consistency threshold (set threshold) At that time, it was assumed that the stage breakpoints of the two sensor channels were consistent.

[0140] (3.3) Repeat steps (3.1) to (3.2) to find sensor channel pairs with consistent breakpoints in all stages within the current cycle and form a temporary set.

[0141] (3.4) Using each sensor channel in the temporary set as a node, the consistency relationship of the stage breakpoints is satisfied. Construct an undirected graph with edges, and select the set of channels with the largest number of sensor channels in the graph, where every pair of stage breakpoints is consistent. As a reliable sensor channel set.

[0142] Reliable sensor channel set Any two sensor channels must meet the stage breakpoint consistency threshold condition, rather than a sensor channel being selected simply because it matches any one of the other channels.

[0143] ,satisfy

[0144] In the formula, This represents the set of reliable sensor channels that meets the pairwise breakpoint consistency constraint and has the largest number of sensor channels. This indicates that the set with the largest number of sensor channels is selected from the candidate channel subset; Represents any subset of candidate channels The number of sensor channels in the process. In step (3.4), by constructing an undirected graph, multiple candidate channel subsets can be obtained, and any two sensor channels in each candidate channel subset can meet the stage breakpoint consistency requirement.

[0145] After the screening is completed, if no sensor channel pairs meet the breakpoint consistency requirements in the current period, then for each sensor channel, the channel quality score is calculated based on the stage fitting residual of the single-channel dynamic programming and the candidate breakpoint interval. The sensor channel with the highest channel quality score is then taken as the reliable sensor channel and proceeds to the next step.

[0146] Among them, the current cycle number Channel quality score for each sensor channel Calculated using the following formula:

[0147]

[0148] In the formula, This indicates the total number of sampling points within the current period; Indicates the number of stages in the current cycle; Indicates the first The sensor channel in the first The sum of squared residuals within each stage, that is, the sum of squared deviations of each sampled value from the mean of that stage; Indicates the first Standardized time series of each sensor channel; Represents the population variance of the standardized time series; The time interval between adjacent stage boundaries is calculated using the following formula:

[0149]

[0150] in, Indicates the first The first sensor channel The boundaries of each stage; and let, , ;

[0151] Calculated using the following formula:

[0152]

[0153] In the formula, The value of the i-th sensor channel at time t; This represents the overall mean of the i-th sensor channel.

[0154] The first term in the above channel quality score calculation formula measures the fitting quality after dynamic programming segmentation; the second term measures the proportion of the shortest stage to the length of the entire production cycle. A higher channel quality score indicates a smaller dynamic programming fitting error and clearer breakpoint intervals for the corresponding sensor channel.

[0155] Step 4: Assemble reliable sensor channels All sensor channels are concatenated column-wise into a multivariable sequence. Then, perform dynamic programming to obtain the stage boundary of the current cycle.

[0156] For a reliable set of sensor channels, the values ​​of all sensor channels corresponding to each sampling moment constitute the sampling vector for that sampling moment. The number of stages in the current cycle is determined based on the number of candidate stage breakpoints. For each stage, the sum of the squared deviations between the sampling vector at each sampling moment and the mean vector of all sampling moments in that production stage is calculated. With the goal of minimizing the sum of the squared deviations of all production stages, the globally optimal combination of breakpoints is searched under the constraint of maintaining the time order of breakpoints, thereby obtaining the stage boundary of the current cycle.

[0157] Specifically, the sum of squared deviations (sum of squared deviations) for the k-th stage can be expressed as:

[0158]

[0159] in, Let represent the sum of squared deviations in the k-th stage, where k∈[1,K] and K is the stage number of the current cycle; Let i represent the sampling vector at the i-th sampling time in the k-th stage, where i∈[1,N] and N is the number of sampling points in the k-th stage; The mean vector for the k-th stage is represented by the following formula:

[0160]

[0161] Taking injection molding as an example, a molding cycle can be divided into four stages: injection, holding pressure, cooling, and mold opening. Since the final segmentation is determined by multiple sensor channels with consistent breakpoints in each pair of stages, it is possible to suppress isolated erroneous breakpoints caused by weak stage-related channels or noise channels such as nozzle temperature and mold opening force.

[0162] Step 5, as follows Figure 2 As shown in (c), the multi-channel sensor trajectory of the current cycle is synchronously extracted into multiple stages according to the stage boundary, and multiple stage data blocks are formed according to the process sequence.

[0163] Each stage data block contains the time series of all sensor channels corresponding to that stage within the current cycle. For the m-th cycle and the n-th stage, its sensor trajectory (time series) is denoted as... Dimension is Where S is the number of sensor channels, This represents the number of sampling points in the nth stage.

[0164] Step 6: Repeat steps 2-5 until all production cycles have been traversed, and obtain the stage data blocks corresponding to all production cycles arranged in process order.

[0165] Step 7: Divide the production cycle data collected in Step 1 into training set, validation set and test set according to process parameters, and ensure that repeated production cycles of the same parameter group are always located in the same data subset.

[0166] First, the constructed quality prediction network is trained and optimized using the preset process parameters of all production cycles in the training set, the corresponding measured quality, and the corresponding stage data blocks obtained in step 6. The quality prediction network includes a historical cycle encoder, a process memory library construction module, a parameter query decoder, a quality prediction head, and an auxiliary quality prediction head used only during the training phase.

[0167] Specifically, the process of training and optimizing the quality prediction network using the training set includes:

[0168] 7.1 For each production cycle in the training set, using its preset process parameter vector and corresponding stage data block as input, the historical cycle encoder outputs the corresponding process memory vector. The specific process is as follows:

[0169] Each stage data block within any production cycle is input into a one-dimensional convolutional network to extract the stage sensor feature matrix; a preset process parameter vector is projected through a fully connected layer to obtain a query vector; using the query vector as the query and the stage sensor feature matrix as the key and value, cross-modal attention is used to calculate the stage fusion features; then, time-adaptive attention is used to calculate the attention weight of each stage, and the fusion features of all stages are weighted and summed to obtain the process memory vector of that production cycle.

[0170] The process involves using a stage-level multimodal encoder to input stage data blocks into a two-layer one-dimensional convolutional network. Each layer includes one-dimensional convolution, batch normalization, and ReLU activation to extract local temporal features from each sensor channel. Global average pooling and global max pooling are then performed on the local temporal features of each sensor channel. The results of these two pooling operations are concatenated and projected onto a D-dimensional array to obtain the stage sensor feature matrix. .

[0171] At the same time, the preset process parameter vector of the m-th production cycle will be... Input a fully connected projection layer to obtain the process condition query vector. .

[0172]

[0173] In the formula, and represents the trainable weight matrix and bias vector; P represents the number of process parameters; ReLU represents the nonlinear activation function.

[0174] The formula for calculating the stage fusion feature is:

[0175]

[0176] In the formula, As a query Q; The feature matrix of the stage sensor in the m-th period and n-th stage serves as both the key K and the value V; MHA represents multi-head cross-attention; Dropout represents random deactivation; LayerNorm represents layer normalization; residual terms... Used for stable training; output The parameter-related fusion features (stage fusion features) are for the m-th period and n-th stage.

[0177] The above design enables the model to adaptively select the more important sensor responses at different stages based on process parameters. For example, during the injection stage, injection speed and injection pressure typically better reflect the filling state; during the holding stage, injection pressure and screw position better reflect the shrinkage state; and during the cooling stage, temperature and thermal history are more sensitive to the transmittance and haze of transparent products.

[0178] The specific process of calculating the contribution weight of each stage to the final quality (attention weight) using time-adaptive attention and then summing them in a weighted manner to form a periodic process memory vector is as follows:

[0179] After obtaining N stage fusion features, the nth stage fusion feature is first projected onto the attention latent space, and then... Calculate the latent space vector, and then use the context vector. Calculate the normalized attention weights for each stage The process memory vector for the m-th production cycle is obtained by weighted summation of the stage fusion features. The calculation formulas are as follows:

[0180]

[0181]

[0182] In the formula, and These are the trainable weight matrix and bias vector for time-adaptive attention, respectively. For time-adaptive attention, the trainable context vector is exp; exp represents the exponential function; the denominator is summed over all N stages, therefore all All are non-negative and their sum is 1.

[0183] This is the process memory vector for the m-th production cycle, used to store the stage sensor responses and their stage importance information after being filtered by process parameters.

[0184] 7.2. Take any untraversed production cycle as the current cycle, and construct the current process memory library using the process memory library construction module based on the process memory vectors corresponding to the other production cycles besides the current cycle.

[0185] Specifically, M process memory vectors (the number of production cycles in the training set is M+1) are used. Arranged in rows, forming The matrix is ​​then used to apply a shared fully connected linear transformation and ReLU nonlinear mapping to each row, resulting in the process memory bank, calculated as follows:

[0186]

[0187] in, Represents the procedural memory; This indicates that the M process memory vectors are stacked row by row; The weight matrix is ​​trainable. is a trainable bias vector, broadcast over M rows; ReLU represents a non-linear activation function.

[0188] 7.3. Using the preset process parameter vector of the current cycle and the current process memory as input, the context vector is obtained by using the parameter query decoder.

[0189] Specifically, let the preset process parameter vector for the current cycle be... First, the query vector is obtained by querying the projection matrix. ; then with As a query, based on the current process memory Simultaneously, multi-head cross-attention is performed as both key and value to obtain the context vector of the historical process. Throughout the entire process, it is not necessary to input the current cycle sensor trajectory or the current cycle quality detection value.

[0190]

[0191] In the formula, Learnable soft retrieval weights are assigned to the M historical process memory vectors based on the current parameter conditions. MHA represents multi-head cross-attention; LayerNorm represents layer normalization; Dropout represents random deactivation; residual terms... Used for stable training.

[0192] 7.4. Using the context vector as input, obtain the main quality prediction vector for the current cycle through the quality prediction head, and calculate the main loss based on the measured quality of the current cycle; at the same time, input the process memory vector of each production cycle that builds the current process memory into the auxiliary quality prediction head to obtain the corresponding auxiliary quality prediction vector, and calculate the auxiliary loss based on the measured quality of each production cycle that builds the current process memory; and then calculate the total loss based on the main loss and auxiliary loss.

[0193] In this process, auxiliary quality prediction supervision is applied to the process memory vectors of each historical production cycle, enabling the process memory to not only describe sensor trajectories but also discriminate against final quality differences. The auxiliary prediction head only participates in auxiliary loss calculation during the training phase and is removed during the inference phase.

[0194] The formula for calculating the main loss is as follows:

[0195]

[0196] Where L represents the number of quality indicators, ; Indicates the first [number]th ... The measured quality of each quality indicator; Represents the current period principal mass prediction vector The One portion, ;

[0197] Principal mass prediction vector The following formula is used to calculate:

[0198]

[0199] In the formula, and These represent the trainable weight matrices of the quality prediction head; and represents the trainable bias vector of the quality prediction head, respectively; ReLU represents the nonlinear activation function; Dropout represents random deactivation.

[0200] The formula for calculating auxiliary loss is:

[0201]

[0202] Where L represents the number of quality indicators, ; This indicates the number of production cycles in the current process memory. ; This indicates the m-th production cycle. The measured quality of each quality indicator; Represents the auxiliary quality prediction vector for the m-th production cycle. The One portion, ;

[0203] Auxiliary quality prediction vector The following formula is used to calculate:

[0204]

[0205] In the formula, and These represent the trainable weight matrices of the auxiliary quality prediction head; and These represent the trainable bias vectors of the auxiliary quality prediction head, respectively. Represents the process memory vector for the m-th production cycle; ReLU represents the nonlinear activation function.

[0206] The total loss is obtained by the following formula:

[0207]

[0208] in, Indicates the total loss; The auxiliary supervision weights are used to balance the main loss and auxiliary loss, and are preferably determined based on the validation set. Specifically, λ is set as a hyperparameter before model training, and is then determined by training the model on the candidate value set separately and comparing the performance on the validation set. This setting ensures that short-term stages highly correlated with defects or quality fluctuations are not overwhelmed by stages with large values ​​but weak quality correlation.

[0209] 7.5. Perform backpropagation with the goal of minimizing total loss, and update the network parameters of the historical cycle encoder, process memory library construction module, parameter query decoder, quality prediction head and auxiliary quality prediction head.

[0210] Among them, stage feature extraction, cross-modal fusion, and time-adaptive stage aggregation belong to the historical cycle encoder; process memory vector stacking and nonlinear mapping belong to the process memory library construction module; memory retrieval based on current process parameters belongs to the parameter query decoder; the main quality prediction is completed by the quality prediction head; and the auxiliary quality prediction head only provides supervision during the training phase.

[0211] Model training comprises both a primary prediction task and an encoder-assisted supervision task. The primary loss constrains the parameter query decoder and quality prediction head for the final prediction of the current cycle quality, while the auxiliary loss constrains the process memory vectors generated by the historical cycle encoder to recover their known quality results. Both types of losses are backpropagated together to jointly learn the network parameters of the historical cycle encoder, process memory library construction module, parameter query decoder, quality prediction head, and auxiliary quality prediction head. The stacking, linear mapping, and activation of process memory vectors belong to the process memory library construction module; auxiliary quality supervision and memory library construction are completed in the same forward pass of training and the parameters are jointly updated through the total loss.

[0212] 7.6. Take each production cycle in the training set sequentially or in batches as the current cycle, repeat steps 7.1 to 7.5, and update the network parameters according to the total loss until the preset training termination condition is met, thus obtaining the optimized network parameters. After each training iteration, update the network parameters once, and the next training will be based on the network parameters updated in the previous iteration.

[0213] In each training iteration, the current cycle only provides a preset process parameter vector as a query, and the remaining training cycles are used to form the current process memory. The network can process multiple current cycles in parallel using a mini-batch approach, and is not limited to updating parameters only once per production cycle.

[0214] 7.7. Re-encode all production cycles in the training set using the optimized historical cycle encoder to obtain the corresponding process memory vectors, and construct the final process memory library using the optimized process memory library construction module.

[0215] 7.8. The final process memory is used together with the optimized parameter query decoder and quality prediction head for quality prediction of the period to be predicted.

[0216] In this step, when training and optimizing the quality prediction network using the training set, for each training target sample (current cycle), only its preset process parameters are used as queries for the parameter query decoder; the preset process parameters and sensor trajectories of the remaining training samples are used by the historical cycle encoder and process memory library construction module to form the current process memory library, and the measured quality is used to calculate the auxiliary loss. For each current cycle in the mini-batch, the process memory vector corresponding to that production cycle is excluded from the memory library candidate to avoid the target sample retrieving itself; this leave-one-out mechanism can be used simultaneously with mini-batch parallel training, and the network parameters are jointly updated based on the total loss.

[0217] After training, the optimized quality prediction network is validated and tested using validation and test sets.

[0218] During the validation phase, a frozen final process memory built from all training set samples is used. Validation samples are only queried using preset process parameters and are not written to the memory. Validation results are used for hyperparameter selection, model selection, and early stopping. During the testing phase, the model with the best validation performance is loaded. The final process memory built from all training set samples is still used. Test samples are only input with preset process parameters, do not participate in parameter updates, and are not written to the memory. This avoids direct sample leakage and process memory leakage.

[0219] Step 8: Using the preset process parameters of the cycle to be predicted as input, the optimized quality prediction network outputs the corresponding quality prediction value.

[0220] The trained quality prediction system (network) consists of a final process memory, a parameter query decoder, and a two-layer fully connected quality prediction head. The historical cycle encoder and process memory construction module are used to build the final process memory after training and do not receive current cycle data during inference in a new cycle. The input for quality prediction is only the process parameter to be predicted or a candidate process parameter, and the output is an L-dimensional quality prediction value.

[0221] Furthermore, the trained quality prediction network can be embedded as a surrogate model into multi-objective optimization algorithms such as NSGA-II, with the goal of minimizing the deviation between the predicted quality and the target quality, to obtain recommended process parameters.

[0222] When the quality prediction network is used as a surrogate model for NSGA-II, the historical cycle encoder and process memory are pre-trained and constructed based on training data and remain fixed during optimization. For each candidate process parameter vector generated by NSGA-II, the candidate parameter is input into the parameter query decoder, and then processed by the quality prediction head to obtain the predicted values ​​of each quality index. The obtained predicted values ​​are used to calculate the deviation between the predicted value and the target quality and form a multi-objective fitness. The surrogate model outputs the corresponding multi-index quality prediction values. The final output of NSGA-II is a set of Pareto candidate parameters that satisfy the process constraints and the recommended process parameters selected from them.

[0223] Specifically, mold temperature, injection speed, injection pressure, barrel temperature, and hot runner temperature are set as decision variables and limited to within allowable process ranges. A pre-trained historical process memory is fixed before optimization begins. In each generation of NSGA-II, each candidate parameter vector undergoes only one forward computation via a parameter query decoder and two layers of quality prediction heads, yielding predicted values ​​for L quality metrics. For the ... For each quality indicator, the objective function can be defined as follows: ,in, Indicates the first Predicted values ​​for each quality indicator Indicates the first The target quality of each quality index. Multiple quality deviations together constitute the multi-objective fitness. The NSGA-II algorithm obtains Pareto candidate parameters through non-dominated sorting and crowding selection, and then gives recommended process parameters according to the principle of minimizing comprehensive deviation. This optimization is a further application of the quality prediction method of this invention. It eliminates the need for real-world testing of each candidate parameter, significantly reducing material consumption and testing time.

[0224] In this embodiment, injection molding data from both PC and PMMA materials are used, with product weight and transmittance as representative quality indicators. Data from 300 production cycles for each material are collected, divided into 100 parameter groups, with each group repeated three times to ensure that repeated production cycles for the same parameter group always fall within the same data subset. The dataset is divided into 70 training parameter groups, 15 validation parameter groups, and 15 test parameter groups. (Auxiliary supervisory weights are also included.) The value is determined to be 0.5. Test results show that the above method can still achieve ex-ante quality prediction by utilizing staged process evidence from historical sensor trajectories without using current period sensor data.

[0225] On the PC dataset, the RMSE of the method in this embodiment for transmittance and weight are 0.0792±0.0076 and 0.4212±0.0460, respectively. In the comparison method using the current cycle process sensor sequence as input, the corresponding RMSEs of PDA-LSTM (an LSTM quality prediction model employing stage modeling and dual attention mechanisms) are 0.1244±0.0040 and 0.5405±0.1512, respectively. Compared to PDA-LSTM, the RMSE of the method in this embodiment is reduced by 36.3% and 22.1%, respectively. On the PMMA dataset, the RMSE of the method in this embodiment for transmittance and weight are 0.1910 and 0.5304, respectively, lower than PDA-LSTM's 0.2106 and 0.6043. These results indicate that under the same equipment and mold conditions, when the material changes from PC to PMMA, the method in this embodiment maintains consistent improvement, demonstrating its applicability to different materials under these conditions.

[0226] In summary, this invention achieves parameter-only pre-production quality prediction for multi-stage manufacturing processes through channel consistency stage segmentation, parameter-guided cross-modal attention, time-adaptive stage aggregation, and quality-supervised process memory construction. It can be used for pre-production risk identification, process parameter recommendation, and intelligent manufacturing quality decision-making.

Claims

1. A multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory, characterized in that, Includes the following steps: (1) Collect multiple production cycle data of multi-stage manufacturing. Each cycle data includes preset process parameters, multi-channel process sensor trajectories and measured product quality. (2) Take any untraversed production cycle as the current cycle, and smooth and standardize each sensor trajectory within it to obtain the corresponding standardized time series; (3) Perform single-channel dynamic programming on the standardized time series corresponding to each sensor channel to obtain the corresponding candidate stage breakpoint set; (4) Determine the consistency between different channel breakpoints based on the candidate stage breakpoint set, and select a reliable sensor channel set; (5) All sensor channels in the reliable sensor channel set are concatenated into a multivariable sequence by column and dynamic programming is performed to obtain the stage boundary of the current cycle; then, according to the stage boundary, the multi-channel sensor trajectory of the current cycle is synchronously extracted into multiple stages and formed into multiple stage data blocks arranged in the process order. (6) Repeat steps (2) to (5) until all production cycles have been traversed, and obtain the stage data blocks corresponding to all production cycles arranged in the process sequence; (7) The quality prediction network is trained and optimized using the collected preset process parameters, measured quality and stage data blocks obtained in step (6); (8) Using the preset process parameters of the cycle to be predicted as input, the corresponding quality prediction value is output by the optimized quality prediction network.

2. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 1, characterized in that, In step (4), the specific steps for obtaining the reliable sensor channel set are as follows: (4.1) For any two sensor channels, calculate the difference between each pair of candidate breakpoints with the same index in the candidate stage breakpoint set for each of them to obtain multiple corresponding time differences, and take the maximum value as the maximum time difference; (4.2) Determine the magnitude of the maximum time difference relative to the set threshold: If the maximum time difference is less than or equal to the set threshold, the two sensor channels are considered to have the same stage breakpoint and are consistent. Otherwise, the breakpoints of the two stages are inconsistent; (4.3) Repeat steps (4.1) to (4.2) to find all sensor channel pairs with consistent breakpoints in the current cycle and form a temporary set; (4.4) Using each sensor channel in the temporary set as a node and the consistency relationship of stage breakpoints as an edge, an undirected graph is constructed. The subset of channels with the largest number of sensor channels in the graph and whose stage breakpoints are consistent with each other is selected as the reliable sensor channel set.

3. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 2, characterized in that, If no sensor channel meets the breakpoint consistency requirement in the current period, then for each sensor channel, calculate the corresponding channel quality score based on the stage fitting residual of the single-channel dynamic programming and the candidate breakpoint interval, and take the sensor channel with the highest channel quality score as the reliable sensor channel and proceed to the next step. Among them, the current cycle number Channel quality score for each sensor channel Calculated using the following formula: ； In the formula, This indicates the total number of sampling points within the current period; Indicates the number of stages in the current cycle; Indicates the first The sensor channel in the first Sum of squared residuals within each stage; Indicates the first Standardized time series of each sensor channel; Represents the population variance of the standardized time series; The time interval between adjacent stage boundaries is calculated using the following formula: ； in, Indicates the first The first sensor channel The boundaries of each stage; and let, , .

4. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 1, characterized in that, In step (7), the quality prediction network includes a historical cycle encoder, a process memory bank construction module, a parameter query decoder, a quality prediction head, and an auxiliary quality prediction head used only during the training phase; the specific process for training and optimizing it is as follows: (7.1) For each production cycle, the corresponding process memory vector is output through the historical cycle encoder, with its preset process parameters and corresponding stage data blocks as input. (7.2) Take any untraversed production cycle as the current cycle, and construct the current process memory library using the process memory library construction module based on the process memory vectors corresponding to the other production cycles besides the current cycle. (7.3) Using the preset process parameters of the current cycle and the current process memory as input, the context vector is obtained by using the parameter query decoder; (7.4) Using the context vector as input, obtain the main quality prediction vector for the current cycle through the quality prediction head, and calculate the main loss based on the measured quality of the current cycle; at the same time, input the process memory vector of each production cycle that constructs the current process memory into the auxiliary quality prediction head to obtain the corresponding auxiliary quality prediction vector, and calculate the auxiliary loss based on the measured quality of each production cycle that constructs the current process memory; then calculate the total loss based on the main loss and the auxiliary loss. (7.5) Perform backpropagation with the goal of minimizing total loss, and update the network parameters of the historical cycle encoder, process memory library construction module, parameter query decoder, quality prediction head and auxiliary quality prediction head; (7.6) Take each production cycle sequentially or in batches as the current cycle, and repeat steps (7.1) to (7.5) until the preset training termination condition is met, and obtain the optimized network parameters; (7.7) The entire production cycle is re-encoded using the optimized historical cycle encoder to obtain the corresponding process memory vector, and the final process memory library is constructed using the optimized process memory library construction module. (7.8) The final process memory is used together with the optimized parameter query decoder and quality prediction head for quality prediction of the period to be predicted.

5. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 4, characterized in that, The specific operation for obtaining a process memory vector for a production cycle in step (7.1) is as follows: Perform one-dimensional convolution on each stage data block within the production cycle to extract the stage sensor feature matrix; The preset process parameters are projected through a fully connected layer to obtain the query vector; Using the query vector as the query and the stage sensor feature matrix as the key and value, cross-modal attention is used to calculate stage fusion features; Then, time-adaptive attention is used to calculate the attention weights of each stage, and the fused features of each stage are weighted and summed to obtain the process memory vector of the production cycle.

6. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 5, characterized in that, The formula for calculating the stage fusion features using cross-modal attention is as follows: ； in, This represents the stage fusion feature of the m-th cycle and the n-th stage; The query vector serves as the query Q; Let K be the stage sensor feature matrix for the m-th period and n-th stage, and V be the key and value, respectively; MHA represents multi-head cross-attention; Dropout represents random deactivation; LayerNorm represents layer normalization; residual term Used for stable training; The formula for calculating the process memory vector is: ； ； ； in, The process memory vector represents the m-th production cycle; n∈[1,N], j∈[1,N], where N represents the total number of stages in the m-th production cycle; This represents the normalized attention weight for the nth stage; This represents the stage fusion feature of the nth stage; and These are the trainable weight matrix and bias vector for time-adaptive attention, respectively. For time-adaptive attention, is a trainable context vector; exp represents the exponential function. This represents the hyperbolic tangent activation function.

7. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 4, characterized in that, In step (7.2), the formula for constructing the process memory is: ； in, Represents the procedural memory; This represents the memory vector of M processes. Stacked by row; The weight matrix is ​​trainable. is a trainable bias vector, broadcast over M rows; ReLU represents a non-linear activation function.

8. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 4, characterized in that, In step (7.3), the formula for obtaining the context vector using the parameter query decoder is: ； in, This represents the obtained context vector; The query vector is obtained by mapping the preset process parameters of the current cycle through a fully connected layer. The input process memory; MHA represents multi-head cross-attention; Dropout represents random deactivation; LayerNorm represents layer normalization; residual terms Used for stable training.

9. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 4, characterized in that, In step (7.4), the formula for calculating the main loss is: ； Where L represents the number of quality indicators, ; Indicates the first [number]th ... The measured quality of each quality indicator; Represents the current period principal mass prediction vector The One portion, ; Principal mass prediction vector The following formula is used to calculate: ； In the formula, and These represent the trainable weight matrices of the quality prediction head; and , respectively, represent the trainable bias vector of the quality prediction head; ReLU represents the nonlinear activation function; Dropout represents random deactivation; The formula for calculating auxiliary loss is: ； Where L represents the number of quality indicators, ; This indicates the number of production cycles in the current process memory. ; This indicates the m-th production cycle. The measured quality of each quality indicator; The m-th auxiliary quality prediction vector represents the first... One portion, ; Auxiliary quality prediction vector The following formula is used to calculate: ； In the formula, and These represent the trainable weight matrices of the auxiliary quality prediction head; and These represent the trainable bias vectors of the auxiliary quality prediction head, respectively. Represents the process memory vector for the m-th production cycle; ReLU represents the nonlinear activation function; The total loss is obtained by the following formula: ； in, Indicates the total loss; This represents the auxiliary supervision weight, used to balance the main loss and auxiliary loss.

10. The multi-stage manufacturing process quality prediction method based on cross-modal stage perception process memory according to claim 1, characterized in that, The multi-stage manufacturing process includes one of injection molding, die casting, and hot pressing.

Images