Deep learning-based multi-process cutting machining quality prediction method

Abstract

The present application relates to the technical field of precision machining quality prediction, in particular to a multi-process cutting machining quality prediction method based on deep learning, comprising: collecting production parameters and equipment state data of the whole process multi-process, performing time sequence alignment and missing value interpolation to form a unified process data sequence. Extracting multi-scale process features to generate an initial feature vector, inputting a multi-modal fusion machining knowledge network to mine cross-process deep coupling correlation, and generating potential quality indicators. Combine the potential indicators with real-time monitoring data to construct a dynamic prediction input, use the prediction model to deduce the final machining quality parameters, and perform uncertainty quantization to output the prediction results and confidence. According to the prediction results and confidence, a reverse traceability report is generated to identify key process links, and a closed-loop model is formed by combining the previous machining data to adaptively optimize the instructions. The method realizes high-precision, interpretable quality prediction and intelligent process optimization.

Description

Technical Field

[0001] This invention relates to the field of precision machining quality prediction technology, and in particular to a method for predicting the quality of multi-process cutting based on deep learning. Background Technology

[0002] Existing cutting process quality prediction technologies mostly rely on monitoring data or statistical models based on a single process for quality assessment. These methods typically focus on local process parameters, such as cutting line speed, tension, and coolant flow rate, establishing predictive relationships through linear regression or shallow machine learning algorithms. They fail to adequately consider the temporal dependencies and nonlinear interactions between multiple processes throughout the entire process, resulting in fragmented data and limited feature representation capabilities. In terms of feature engineering, existing technologies depend on manually designed statistical indicators, making it difficult to capture multi-scale dynamic changes and implicit states during processing. The prediction models are also susceptible to interference from inter-process coupling effects, leading to decreased accuracy.

[0003] Existing prediction systems often only output point estimates of quality parameters, lacking a quantitative assessment of prediction reliability, making it impossible for operators to determine the confidence level of the results. Furthermore, these systems lack effective back-analysis mechanisms, failing to identify specific process steps that significantly impact the final quality, resulting in a disconnect between prediction results and process control. When fluctuations occur in the processing or changes in equipment status, the model cannot adaptively adjust, requiring manual parameter recalibration based on experience, leading to optimization lag and inefficiency. Therefore, how to automatically mine deep correlations from cross-process data to achieve accurate dynamic prediction and establish a closed-loop optimization mechanism based on prediction feedback is a key technical challenge that needs to be addressed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a multi-process cutting quality prediction method based on deep learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a multi-process cutting quality prediction method based on deep learning, comprising: Collect production parameters and equipment status data from multiple processes throughout the entire cutting and processing workflow, construct a cross-process processing data set, and perform time-series alignment and missing value imputation to form a unified process data sequence; Extract multi-scale process features related to the final cutting quality from the unified process data sequence to generate an initial feature vector; The initial feature vector is input into a multimodal fusion processing knowledge network to uncover deep coupling and correlation patterns across processes. Based on the output of the processing knowledge network, potential quality indicators for characterizing the state of intermediate processes are generated. The potential quality indicators are combined with real-time monitoring data of the current process to construct a dynamic prediction input for final quality prediction. The dynamic prediction input is processed using a prediction model to deduce the values ​​of the final processing quality parameters. The uncertainty of the deduced final processing quality parameters is quantified, and the quality prediction results and their confidence levels are output. Based on the quality prediction results and their confidence levels, a reverse traceability report is generated to identify key process steps that significantly affect the quality results. The information of the key process steps is combined with the preceding processing data to form a closed-loop model adaptive optimization instruction.

[0006] As a further aspect of the present invention, the step of performing time alignment and missing value imputation to form a unified process data sequence includes: The time when the workpiece enters the processing flow is taken as the absolute time starting point, and a relative timestamp is marked for the data of each process. Based on the process logic sequence between each process, a time delay compensation model between processes is established, and the relative timestamps are compensated and aligned. The missing positions of the aligned data on the time axis are detected, and the data is filled using a spline interpolation algorithm based on the data change trend of adjacent processes before and after the missing data. The overall data stream after interpolation and filling is smoothed and filtered to remove high-frequency noise points caused by abnormal equipment acquisition. The processed data from each process are spliced ​​together in chronological order to form a continuous, unified process data sequence with a consistent time base.

[0007] As a further aspect of the present invention, the step of extracting multi-scale process features related to the final cutting quality from the unified process data sequence to generate an initial feature vector includes: At the process level, the statistics of the core parameters within each process window in the unified process data sequence are calculated, including the mean, variance, and extreme values. At the event level, transient fluctuation events exceeding a preset threshold are identified in the unified process data sequence, and the duration, amplitude, and shape of the transient fluctuation events are extracted. At the trend level, the key parameters in the unified process data sequence are piecewise linearly fitted, and the slope and intercept changes of each fitted line segment are extracted. All features extracted from the process-level, event-level, and trend-level scales will be spliced ​​and dimensionality-reduced fused. The fused features are standardized and scaled to conform to a preset numerical distribution range, thereby generating the initial feature vector.

[0008] As a further aspect of the present invention, the step of inputting the initial feature vector into a multimodal fusion processing knowledge network to mine deep coupling and correlation patterns across processes includes: The multimodal fusion processing knowledge network includes a shared encoder and multiple process-specific decoders; The shared encoder receives the initial feature vector and maps it to a common feature space through multiple nonlinear transformations. In the common feature space, the correlation weights between feature segments of different processes are calculated through a self-attention mechanism; Based on the association weights, the information in the common feature space is dynamically weighted, fused, and enhanced. The enhanced common features are input into the decoders specific to each process, and the decoders reconstruct the enhanced process feature representations. The enhanced process feature representations, which aggregate all process-specific decoder outputs, are presented as the output of the processing knowledge network, which contains deep coupling and association patterns between processes.

[0009] As a further aspect of the present invention, the step of generating potential quality indicators for characterizing the intermediate process status based on the output of the processing knowledge network includes: The output of the processing knowledge network is fed into a latent space mapping layer; The latent space mapping layer projects the output of the processing knowledge network into a lower-dimensional latent quality space; In the potential quality space, each coordinate axis represents an independent abstract process factor that affects the final quality; Calculate the projection coefficients of the output of the processing knowledge network onto each abstract process factor axis in the potential quality space; The projection coefficients are weighted and combined with the preset benchmark values ​​of the corresponding abstract process factors to generate a set of quantified values, which are the potential quality indicators.

[0010] As a further aspect of the present invention, the step of combining the potential quality indicators with real-time monitoring data of the current process to construct a dynamic prediction input for final quality prediction includes: Continuously acquire real-time sensor monitoring data streams for the currently executing process; The sensor's real-time monitoring data stream is subjected to feature extraction at the same scale to generate real-time features for the current process; The real-time characteristics of the current process are combined with the potential quality indicators corresponding to the historical processes. Add a position encoding vector that identifies the current process's position in the overall process before the concatenated feature vector; The concatenated features with added positional encoding vectors are normalized to form the dynamic prediction input.

[0011] As a further aspect of the present invention, the step of processing the dynamic prediction input using a prediction model to deduce the values ​​of the final processing quality parameters includes: The iterative prediction model contains a long short-term memory network unit and a feedforward neural network. At each processing time step, the Long Short-Term Memory network unit receives the current dynamic prediction input and the hidden state of the previous time step; The Long Short-Term Memory (LSTM) network unit updates its internal memory unit state based on the current input and the previous state, and outputs the hidden state of the current time step. The hidden state at the current time step is passed to the feedforward neural network; The feedforward neural network maps the hidden state at the current time step to a preliminary quality prediction value. This process is carried out cyclically within the model, simulating the progress of the process, until a specific time step representing the end of the process is reached. At this point, the output of the feedforward neural network is the value of the final processing quality parameter obtained through deduction.

[0012] As a further aspect of the present invention, the uncertainty quantification of the final processing quality parameters obtained from the deduction, and the output of the quality prediction result and its confidence level, includes: The Monte Carlo random dropout method is used to perform multiple forward propagations during the inference phase of the feedforward neural network of the cyclically iterative prediction model; In each forward propagation, the connections of some nodes in the feedforward neural network are randomly discarded; Record the values ​​of the final processing quality parameters output by the prediction model in the iterative cycle after each random discard of a connection; Calculate the mean of all recorded values ​​and use the mean as the final quality prediction result; Calculate the standard deviation of all recorded values, and look up the corresponding confidence score in a preset mapping table based on the standard deviation. Use the confidence score as the confidence level for this prediction.

[0013] As a further aspect of the present invention, the step of generating a reverse traceability report based on the quality prediction results and their confidence levels, and identifying key process steps that significantly affect the quality results, includes: In the common feature space of the processed knowledge network, the gradient of the feature representation of the final time step with respect to the feature representations of each previous time step is calculated. The magnitude of the gradient characterizes the degree of influence of historical process characteristics on the final quality prediction result; Filter out all historical time steps whose absolute gradient value exceeds a preset importance threshold; Map the historical time steps back to their corresponding original process steps; The frequency of each original process step in the historical time step is counted, and one or more original process steps with the highest frequency are identified as the key process steps, and this mapping relationship is written into the reverse traceability report.

[0014] As a further aspect of the present invention, the step of combining the key process information with the preceding processing data to form a closed-loop model adaptive optimization instruction includes: Retrieve the unified process data sequence corresponding to all processing processes that include the key process steps from the historical database; Detailed process parameter data of the time window in which the key process steps are located are extracted from the unified process data sequence. The detailed process parameter data is paired with the final actual quality data of the corresponding processing process to form an optimized sample pair; Using the newly added optimized sample pairs, a round of fine-tuning training is performed on the shared encoder parameters in the multimodal fusion processing knowledge network; Save the fine-tuned shared encoder parameters and generate the model adaptive optimization instructions containing parameter version identifiers.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By constructing a multimodal fusion-based processing knowledge network, deep nonlinear correlation patterns are automatically mined from cross-process time-series data, generating a series of potential quality indicators characterizing the implicit states of intermediate processes. These potential indicators are dynamically combined with real-time sensor data to construct predictive inputs containing process knowledge. This enables the predictive model to not only rely on the explicit parameters of the current process but also capture process states that cannot be directly measured, such as bow dynamics and material removal consistency. Prediction accuracy is thus improved, early warnings of quality fluctuations are provided, and the model's adaptability to complex coupling effects between processes is enhanced.

[0016] While outputting the final quality parameter point predictions, the system rigorously quantifies the uncertainty of the prediction results and automatically performs reverse traceability analysis based on the quantification results to identify the specific process steps and their parameters that have the most significant impact on the current prediction results and confidence levels. This key process information is combined with preceding actual processing data to form structured model adaptive optimization instructions. This system achieves transparent assessment of prediction reliability and can accurately locate weak links in the process, providing a clear direction for adjusting control parameters. This forms a closed loop from quality prediction back to process optimization, reducing reliance on manual experience in model maintenance and improving the response speed and targeting of process optimization. Attached Figure Description

[0017] Figure 1 This is a flowchart of the deep learning-based multi-process cutting quality prediction method described in this invention; Figure 2 A flowchart for creating a unified sequence of process data; Figure 3 A flowchart illustrating the operation of a multimodal fusion-based processing knowledge network; Figure 4 This is a composite analysis diagram of multi-process parameters and quality influence coefficients; Figure 5 A graph showing the prediction of final processing quality across multiple processes and its confidence level analysis. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1The process involves collecting production parameters and equipment status data from multiple steps throughout the cutting and processing workflow. This constructs a cross-step processing dataset, which undergoes temporal alignment and missing value imputation to form a unified process data sequence. From this unified process data sequence, multi-scale process features related to the final cutting quality are extracted and combined to generate an initial feature vector. This initial feature vector is input into a multi-modal fusion processing knowledge network, which is used to uncover deep coupling patterns across processes. Based on the output of the processing knowledge network, a set of potential quality indicators characterizing the intermediate process status is generated. This set of potential quality indicators is combined with real-time monitoring data from the current process to construct a dynamic prediction input for final quality prediction. A prediction model processes this dynamic prediction input to deduce the values ​​of the final processing quality parameters and quantifies the uncertainty of these values, thus outputting the quality prediction result and its confidence level. Based on the quality prediction result and its confidence level, a reverse traceability report is generated to identify key process steps that significantly impact the quality result. Information on these key process steps is combined with preceding processing data to form a closed-loop model adaptive optimization instruction.

[0021] See Figure 2 In one embodiment of the present invention, taking the actual processing flow of a cut silicon rod as an example, this process includes three main steps: wire feeding, cutting, and cleaning. In specific implementation, the wire feeding step records tension and linear velocity, the cutting step records cutting current and coolant flow rate, and the cleaning step records cleaning pressure and time. The collected raw data set is discrete and unaligned in the time dimension. Taking the time when the silicon rod enters the wire feeding machine, 2024-05-10 08:00:00, as the absolute time starting point, the relative timestamp for the wire feeding step data is 0 seconds, the relative timestamp for the cutting step data starting at 08:00:30 is 30 seconds, and the relative timestamp for the cleaning step data starting at 08:05:00 is 300 seconds. According to the process logic, the cutting step must start after the wire feeding step begins, and the cleaning step must start after the cutting step ends. Based on this logic, a time delay compensation model between the steps is established. In practice, the time delay compensation model identifies that the sensors in the cutting process have an inherent response delay of 2 seconds, and the start of the cleaning process is affected by the handling robot arm with a random delay of 5 seconds. Based on these delay parameters, the relative timestamps of all data points in the cutting and cleaning processes are compensated and aligned. The timestamp of the cutting process is corrected to start at 32 seconds, and the timestamp of the cleaning process is corrected to start at 305 seconds.

[0022] In practice, gaps in the aligned data on the timeline are detected. For example, gaps in the tension sensor data recording occur between the 15th and 18th seconds after the start of the wire laying process. Based on the data change trends of adjacent processes before and after the gaps—specifically, examining the tension values ​​recorded at the 14th and 19th seconds, and the changes in linear velocity during the same period—a spline interpolation algorithm is used to fill the data. Specifically, the spline interpolation algorithm is applied to the cutting current data stream. The algorithm fits a smooth spline curve based on the known current data points before and after the gap window, and uses this curve to calculate the interpolated current values ​​for the 15th, 16th, 17th, and 18th seconds, which are 152.3A, 151.8A, 151.5A, and 152.1A, respectively. The overall data stream after interpolation is then smoothed and filtered to remove high-frequency noise points caused by equipment acquisition anomalies. In practice, a high-frequency noise point manifests as a sudden spike in the cutting current from 150A to 200A within a millisecond, followed by a drop back. A moving average filter is used for smoothing to remove this abnormal peak, ensuring a smooth transition between the current value at that point and its values ​​in the preceding and following time periods. The processed data from each process are then concatenated in chronological order to form a continuous process data sequence with a unified time base. In practice, this unified process data sequence is a multi-dimensional time series, with each row of data corresponding to a millisecond-level time point, and the columns containing all production parameters and equipment status data from each process: wire feeding, cutting, and cleaning.

[0023] In practical implementation, the process of extracting multi-scale process features from a unified process data sequence and generating an initial feature vector is as follows: At the process level, the statistics of core parameters within each process window in the unified process data sequence are calculated. For example, the mean current of the cutting process is calculated with a window of 10 seconds. The current reading sequence within a window is [150.1, 151.3, 149.8, 152.2, 150.5, 151.0, 149.5, 152.8, 150.0, 151.2], with a mean of 150.84, a variance of 1.12, and extreme values ​​of a maximum of 152.8 and a minimum of 149.5. At the event level, transient fluctuation events exceeding a preset threshold in the unified process data sequence are identified. For example, with a preset tension threshold upper limit of 25N, an event lasting 3 seconds during the wire laying process, in which the tension rises from 22N to 28N and then falls back, is identified. This transient fluctuation event is extracted to have a duration of 3000 milliseconds, an amplitude of 6N, and a morphology of a single-peak pulse. At the trend level, key parameters in a unified process data sequence are piecewise linearly fitted. For example, a 60-second coolant flow rate data sequence is piecewise linearly fitted, divided into two segments. The slope of the first segment is extracted to be 0.02L / s, and the intercept is 10.5L; the slope of the second segment is extracted to be -0.01L / s, and the intercept is 12.8L. All features extracted from the process-level, event-level, and trend-level scales are then concatenated and dimensionality-reduced. For example, process-level statistical features have 20 dimensions, event-level features have 15 dimensions, and trend-level features have 10 dimensions. After concatenation, a 45-dimensional feature vector is obtained, which is then dimensionality-reduced and fused to a 25-dimensional feature vector using principal component analysis. The fused features are standardized and scaled to conform to a preset numerical distribution range, thereby generating an initial feature vector.

[0024] See Figure 3 In one embodiment of the present invention, the processed initial feature vector is used as input. The initial feature vector is a numerical vector containing 25 dimensions, representing multi-scale features of the wire feeding, cutting, and cleaning processes. In a specific implementation, the multimodal fusion processing knowledge network includes a shared encoder and multiple process-specific decoders. The shared encoder consists of a three-layer fully connected neural network, with each layer followed by a ReLU activation function and a batch normalization operation. The shared encoder receives the initial feature vector and maps it to a common feature space through multiple nonlinear transformations. The common feature space is a 16-dimensional vector space. In the common feature space, the correlation weights between feature segments of different processes are calculated through a self-attention mechanism. In a specific implementation, the 8 feature dimensions representing the wire feeding process, the 10 feature dimensions representing the cutting process, and the 7 feature dimensions representing the cleaning process in the initial feature vector are considered as three feature segments. The self-attention mechanism calculates the correlation weight matrix between these three segments, and the correlation weight matrix has a dimension of 3x3.

[0025] Based on the correlation weight matrix, information in the common feature space is dynamically weighted and fused, and information enhancement is performed. In some embodiments, the calculation formula for dynamic weighted fusion is: in: It is an enhanced representation of common features. This is the total number of process segments (3 in this case). It is a process segment calculated by a self-attention mechanism. Process segments Attention weights It is a trainable parameter matrix. It is the original common feature space corresponding to the process segment The eigenvectors. In the calculation, The matrix has a dimension of 16x16. The enhanced common features are input into the process-specific decoders, each of which is a two-layer fully connected network. Each decoder reconstructs an enhanced process feature representation; for example, the decoder for the wire feeding process outputs an 8-dimensional vector, the decoder for the cutting process outputs a 10-dimensional vector, and the decoder for the cleaning process outputs a 7-dimensional vector. The enhanced process feature representations from all the process-specific decoders are aggregated and used as the output of the processing knowledge network. This output is a recombined 25-dimensional vector, which contains deep coupling patterns between processes. For example, this vector might reflect a strong correlation between the stability of the cutting current and the specific fluctuation pattern of the wire feeding tension.

[0026] In practical implementation, the output of the machining knowledge network generates potential quality indicators to characterize the intermediate process state. The output of the machining knowledge network is then fed into a latent space mapping layer. This latent space mapping layer is a linear transformation layer without an activation function, projecting the 25-dimensional output of the machining knowledge network into a lower-dimensional latent quality space, which is set to 5 dimensions. In the latent quality space, each coordinate axis represents an independent abstract process factor affecting the final quality. For example, the first axis represents the "line stability factor," the second the "thermal load factor," the third the "chip removal efficiency factor," the fourth the "surface micro-damage factor," and the fifth the "stress distribution factor." The projection coefficients of the machining knowledge network output onto each abstract process factor axis in the latent quality space are calculated using the weight matrix of the latent space mapping layer, resulting in five projection coefficient values, for example, [0.85, -0.23, 0.41, 0.12, -0.05]. These projection coefficients are then weighted and combined with preset benchmark values ​​for the corresponding abstract process factors to generate a set of quantified values. In some embodiments, the preset baseline values ​​are typical values ​​of each factor obtained from offline statistical analysis in qualified products. The weighted combination adopts linear weighted summation. For example, for the "line bow stability factor", its baseline value is 1.0 and the projection coefficient is 0.85, then the combined value is 1.0 * 0.85 = 0.85. The final set of 5 quantified values ​​is the potential quality index, such as [0.85, -0.11, 0.32, 0.04, 0.01]. This set of values ​​collectively characterizes the potential impact of the composite state of intermediate processes on the final quality.

[0027] In one embodiment of the invention, the generated potential quality index and real-time monitoring data from the production line are used. The potential quality index is a set of 5-dimensional vectors, such as [0.85, -0.11, 0.32, 0.04, 0.01], representing an abstraction of the process state up to the end of the previous process. Constructing a dynamic prediction input for the final quality prediction requires continuously acquiring the real-time monitoring data stream from the sensors of the currently executing process. When the current process is a precision cutting process, the real-time monitoring data stream includes the cutting line speed, cutting tension, and coolant pressure sampled per second. Multi-scale feature extraction is performed on the real-time monitoring data stream of cutting line speed, cutting tension, and cooling hydraulic pressure. This extraction method is consistent with the feature extraction scale. In specific implementation, for a real-time data window with a length of 10 seconds, the mean value of the cutting tension is calculated to be 24.5N and the variance is 0.8. A pulse event of cooling fluid pressure lasting 2 seconds with an amplitude of 0.2MPa is identified. Piecewise linear fitting is performed on the cutting line speed to obtain a slope of -0.05m / s². These process-level, event-level, and trend-level features extracted from the real-time data are concatenated to generate an 8-dimensional real-time feature vector of the current process.

[0028] In specific implementation, the real-time feature vector of the current process is concatenated with the potential quality indicators corresponding to the historical processes. The real-time feature vector of the current process is 8-dimensional, and the historical potential quality indicators are 5-dimensional, forming a 13-dimensional combined feature vector. A position encoding vector is added before the concatenated 13-dimensional combined feature vector to identify the current process's position in the overall process. The overall process includes 5 processes, and the current precision cutting process is the 3rd process. The position encoding vector is a 5-dimensional one-hot vector [0,0,1,0,0]. The concatenated feature vector with the added position encoding vector is normalized. The normalization operation scales all feature values ​​to the [0,1] interval. For example, the original average cutting tension of 24.5N becomes 0.61 after normalization, thus forming the dynamic prediction input, which is an 18-dimensional vector. In some embodiments, the normalization operation uses the max-min normalization method, and its calculation formula is: in: These are the normalized eigenvalues. These are the original eigenvalues. It is the minimum value of this feature in the training set. It is the maximum value of this feature in the training set.

[0029] In practice, a prediction model is used to process the dynamic prediction input. The iterative prediction model internally comprises a Long Short-Term Memory (LSTM) network unit with 128 hidden units and a two-layer feedforward neural network. At each processing time step, the LSM unit receives the current 18-dimensional dynamic prediction input and the 128-dimensional hidden state from the previous time step, initially set as a zero vector. The LSM unit updates its internal memory unit state based on the current input and the previous state, and outputs the 128-dimensional hidden state for the current time step. This 128-dimensional hidden state is then passed to the feedforward neural network. The first layer of the feedforward neural network maps the 128-dimensional input to 64 dimensions, and the second layer maps the 64-dimensional input to 1 dimension. The feedforward neural network maps the hidden state of the current time step to a preliminary quality prediction value; for example, at a certain time step, it outputs a preliminary line mark depth prediction of 12.5 micrometers. This process is carried out cyclically within the model, simulating the progress of the process. The dynamic prediction input is updated at each time step. For example, every second, new real-time features are concatenated with historical potential quality indicators, positional codes are added, and normalized before being input into the model, until a specific time step representing the end of the process is reached. At this point, the output of the feedforward neural network is the value of the final processing quality parameter obtained through deduction. For example, the final output line mark depth prediction value is 13.2 micrometers.

[0030] In one embodiment of the present invention, the final processing quality parameter value obtained from the aforementioned prediction model is used as input. For example, the model predicts a line mark depth of 13.2 micrometers for a certain cutting batch. The uncertainty of the final processing quality parameter value is quantified using a Monte Carlo random discarding method. Multiple forward propagations are performed during the feedforward neural network inference phase of the iterative prediction model, with the number of forward propagations set to 100. In each forward propagation, connections of some nodes in the feedforward neural network are randomly discarded. Specifically, before the activation functions of the first layer (64 neurons) and the second layer (1 neuron) of the feedforward neural network, the output of the neuron is randomly set to zero with a preset discarding probability (e.g., 20%). The final processing quality parameter value output by the iterative prediction model after each random discarding of connections is recorded, resulting in 100 predicted line mark depth values. The mean of all recorded values ​​is calculated and used as the final quality prediction result. The standard deviation of all recorded values ​​is calculated using the following formula: in: It is the standard deviation. It is the total number of forward propagations. It is the first The predicted value obtained from the second forward propagation This is the arithmetic mean of 100 predicted values. The calculated standard deviation is used to look up the corresponding confidence score in a pre-defined mapping table. In some embodiments, the pre-defined mapping table defines the correspondence between standard deviation ranges and confidence scores; for example, a standard deviation less than 0.5 micrometers corresponds to a confidence score of 95%, a standard deviation between 0.5 and 1.0 micrometers corresponds to a confidence score of 85%, and a standard deviation greater than 1.0 micrometer corresponds to a confidence score of 70%. The found confidence score is used as the output confidence score for this prediction. For example, if the calculated standard deviation is 0.42 micrometers, the output quality prediction result is 13.2 micrometers, with a confidence score of 95%. Refer to Table 1, which shows example recorded values ​​from 10 forward propagations.

[0031] Table 1: Example Record Table of Monte Carlo Random Dropout Forward Propagation Method In practice, a reverse traceability report is generated based on the quality prediction results and their confidence levels. In the common feature space of the processing knowledge network, the gradient of the feature representation of the final time step with respect to the feature representations of each previous time step is calculated. Taking a processing procedure with 20 processing time steps as an example, the feature representation of the final time step (time step 20) is a 16-dimensional vector. The gradient of this vector with respect to the 16-dimensional feature representations of time steps 1 to 19 needs to be calculated. The magnitude of the gradient characterizes the degree of influence of historical process features on the final quality prediction result; a larger absolute gradient value indicates a greater contribution or influence of the corresponding historical time step's features on the final prediction. All historical time steps whose absolute gradient values ​​exceed a preset importance threshold are selected. The preset importance threshold is set to 1.5 times the average absolute gradient values ​​of all time steps. For example, calculations show that the absolute gradient values ​​of time steps 5, 8, 12, and 15 exceed the threshold. These historical time steps are mapped back to their corresponding original process steps, based on a time step-to-process correspondence table. For example, time steps 1-7 correspond to the "layout" process, time steps 8-14 correspond to the "main cutting" process, and time steps 15-20 correspond to the "fine grinding" process. Therefore, time step 5 belongs to the "layout" process, time steps 8 and 12 belong to the "main cutting" process, and time step 15 belongs to the "fine grinding" process. The frequency of each original process step in the selected historical time steps is counted: the "layout" process appears once, the "main cutting" process appears twice, and the "fine grinding" process appears once. It can be understood that one or more original process steps with the highest frequency are identified as critical process steps. In this example, the "main cutting" process appears most frequently and is identified as a critical process step. The mapping relationship that "the main cutting process (time steps 8 and 12) has a significant impact on the final line mark depth prediction result" is written into the reverse traceability report.

[0032] See Figure 4 This is a composite analysis chart showing the relationship between process parameters and quality influence coefficients across multiple processes in the cutting process. It illustrates the impact of process parameters at different stages on the final quality. The chart visually reveals the correlation between "process parameter fluctuations" and "quality influence." The main cutting process exhibits the largest tension variance and the highest quality influence coefficient, indicating that parameter stability in this process is crucial for controlling final quality. This provides data support for tracing back to key processes and optimizing process parameters. This chart is a key analytical output of the cutting quality prediction system, primarily used to help process engineers quickly identify core processes affecting quality; provide visual guidance for subsequent process parameter optimization; and validate the analysis results of the multimodal fusion processing knowledge network.

[0033] In one embodiment of the present invention, the key process information identified in the aforementioned reverse traceability report is used as input. For example, the report identifies "main cutting process" as a key process that significantly affects the final trace depth. The key process information is combined with preceding processing data to retrieve unified process data sequences corresponding to all processing steps containing key process steps from the historical processing database. The historical processing database stores unified process data sequences for all silicon rod cutting tasks completed in the past six months. The search criteria are that the process tag contains "main cutting," and a total of 125 unified process data sequence records matching the criteria are retrieved. From these retrieved unified process data sequences, detailed process parameter data for the time window of the key process step is extracted. It is understood that the start and end timestamps of the "main cutting" process are known in each unified process data sequence. All parameter data sampled at 100-millisecond intervals within this time window are extracted, including four types of parameters: cutting line speed, tension, cutting current, and coolant flow rate. Each sequence can yield a parameter matrix consisting of approximately 2000 data points.

[0034] In practice, the extracted detailed process parameter data is paired with the final actual quality data of the corresponding processing process to form optimized sample pairs. The final actual quality data comes from the offline measurement report after the silicon rod cutting of that batch is completed, for example, the measured value of the line mark depth is 13.8 micrometers. Therefore, in a complete historical processing record, the "main cutting" process parameter matrix extracted from the unified process data sequence and the corresponding final actual line mark depth value of 13.8 micrometers constitute an optimized sample pair. Using these newly added optimized sample pairs, a round of fine-tuning training is performed on the shared encoder parameters in the multimodal fusion processing knowledge network. In the fine-tuning training, the shared encoder receives the initial feature vector calculated from the "main cutting" process parameter matrix, and its output, together with the subsequent network parts, calculates a predicted line mark depth value. In some embodiments, the loss function used in the fine-tuning training is the mean squared error, calculated as follows: in: It is the loss value. This is the number of optimized sample pairs used for this fine-tuning. It is the first The actual line mark depth value in each sample pair It is the network to the first The predicted line depth values ​​for each sample are used. The gradient of the loss value with respect to the shared encoder parameters is calculated using the backpropagation algorithm, and the parameters of the shared encoder are updated using the optimizer. The training process continues until the loss value converges.

[0035] Save the fine-tuned shared encoder parameters and store the updated parameter matrix in a new file. Generate a model adaptive optimization instruction file containing parameter version identifiers; the optimization instruction is a structured text file. In some embodiments, the model adaptive optimization instruction file includes the following fields: the target process to be optimized, the number of historical samples used for optimization, the version number of the fine-tuned shared encoder parameters, the storage path of the parameter file, and the timestamp of the instruction's generation. Optionally, the model adaptive optimization instruction file may also include a suggestion that, when encountering processing tasks involving the "main cutting" process in the future, the prediction system should prioritize loading the shared encoder parameters of version Encoder_V2.1 for inference. This process allows the model to self-adjust based on newly discovered key process knowledge, thus forming a closed loop from problem diagnosis to model update.

[0036] See Figure 5 This is a chart analyzing the final processing quality prediction and confidence level across multiple processes. It simultaneously displays the trends of actual quality values, predicted quality values, and prediction confidence levels as samples change. The high degree of agreement between predicted and actual values ​​verifies the effectiveness of the multimodal fusion processing knowledge network; while the dynamic changes in confidence level provide a quantitative basis for assessing the reliability of the prediction results. Stages with lower confidence levels can be used as a focus for reverse tracing to identify key process steps that lead to quality fluctuations. This chart is primarily used for model validation and performance evaluation of the segmentation quality prediction system. It helps engineers intuitively assess the accuracy and stability of the prediction model; identify risky samples with low model confidence levels to trigger further process optimization; and, combined with the reverse tracing mechanism, pinpoint the key processes that cause quality fluctuations.

[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting the quality of multi-process cutting based on deep learning, characterized in that, The method includes: Collect production parameters and equipment status data from multiple processes throughout the entire cutting and processing workflow, construct a cross-process processing data set, and perform time-series alignment and missing value imputation to form a unified process data sequence; Extract multi-scale process features related to the final cutting quality from the unified process data sequence to generate an initial feature vector; The initial feature vector is input into a multimodal fusion processing knowledge network to uncover deep coupling and correlation patterns across processes. Based on the output of the processing knowledge network, potential quality indicators for characterizing the state of intermediate processes are generated. The potential quality indicators are combined with real-time monitoring data of the current process to construct a dynamic prediction input for final quality prediction. The dynamic prediction input is processed using a prediction model to deduce the values ​​of the final processing quality parameters. The uncertainty of the deduced final processing quality parameters is quantified, and the quality prediction results and their confidence levels are output. Based on the quality prediction results and their confidence levels, a reverse traceability report is generated to identify key process steps that significantly affect the quality results. The information of the key process steps is combined with the preceding processing data to form a closed-loop model adaptive optimization instruction.

2. The method for predicting the quality of multi-process cutting based on deep learning according to claim 1, characterized in that, The step of performing time-series alignment and missing value imputation to form a unified process data sequence includes: The time when the workpiece enters the processing flow is taken as the absolute time starting point, and a relative timestamp is marked for the data of each process. Based on the process logic sequence between each process, a time delay compensation model between processes is established, and the relative timestamps are compensated and aligned. The missing positions of the aligned data on the time axis are detected, and the data is filled using a spline interpolation algorithm based on the data change trend of adjacent processes before and after the missing data. The overall data stream after interpolation and filling is smoothed and filtered to remove high-frequency noise points caused by abnormal equipment acquisition. The processed data from each process are spliced ​​together in chronological order to form a continuous, unified process data sequence with a consistent time base.

3. The method for predicting the quality of multi-process cutting based on deep learning according to claim 2, characterized in that, The step of extracting multi-scale process features related to the final cutting quality from the unified process data sequence to generate an initial feature vector includes: At the process level, the statistics of the core parameters within each process window in the unified process data sequence are calculated, including the mean, variance, and extreme values. At the event level, transient fluctuation events exceeding a preset threshold are identified in the unified process data sequence, and the duration, amplitude, and shape of the transient fluctuation events are extracted. At the trend level, the key parameters in the unified process data sequence are piecewise linearly fitted, and the slope and intercept changes of each fitted line segment are extracted. All features extracted from the process-level, event-level, and trend-level scales will be spliced ​​and dimensionality-reduced fused. The fused features are standardized and scaled to conform to a preset numerical distribution range, thereby generating the initial feature vector.

4. The method for predicting the quality of multi-process cutting based on deep learning according to claim 3, characterized in that, The step of inputting the initial feature vector into a multimodal fusion processing knowledge network to mine deep coupling and correlation patterns across processes includes: The multimodal fusion processing knowledge network includes a shared encoder and multiple process-specific decoders; The shared encoder receives the initial feature vector and maps it to a common feature space through multiple nonlinear transformations. In the common feature space, the correlation weights between feature segments of different processes are calculated through a self-attention mechanism; Based on the association weights, the information in the common feature space is dynamically weighted, fused, and enhanced. The enhanced common features are input into the decoders specific to each process, and the decoders reconstruct the enhanced process feature representations. The enhanced process feature representations, which aggregate all process-specific decoder outputs, are presented as the output of the processing knowledge network, which contains deep coupling and association patterns between processes.

5. The method for predicting the quality of multi-process cutting based on deep learning according to claim 4, characterized in that, The generation of potential quality indicators to characterize the intermediate process status based on the output of the processing knowledge network includes: The output of the processing knowledge network is fed into a latent space mapping layer; The latent space mapping layer projects the output of the processing knowledge network into a lower-dimensional latent quality space; In the potential quality space, each coordinate axis represents an independent abstract process factor that affects the final quality; Calculate the projection coefficients of the output of the processing knowledge network onto each abstract process factor axis in the potential quality space; The projection coefficients are weighted and combined with the preset benchmark values ​​of the corresponding abstract process factors to generate a set of quantified values, which are the potential quality indicators.

6. The method for predicting the quality of multi-process cutting based on deep learning according to claim 5, characterized in that, The step of combining the potential quality indicators with real-time monitoring data of the current process to construct a dynamic prediction input for final quality prediction includes: Continuously acquire real-time sensor monitoring data streams for the currently executing process; The sensor's real-time monitoring data stream is subjected to feature extraction at the same scale to generate real-time features for the current process; The real-time characteristics of the current process are combined with the potential quality indicators corresponding to the historical processes. Add a position encoding vector that identifies the current process's position in the overall process before the concatenated feature vector; The concatenated features with added positional encoding vectors are normalized to form the dynamic prediction input.

7. The method for predicting the quality of multi-process cutting based on deep learning according to claim 6, characterized in that, The process of using a prediction model to process the dynamic prediction input and deduce the values ​​of the final processing quality parameters includes: The iterative prediction model contains a long short-term memory network unit and a feedforward neural network. At each processing time step, the Long Short-Term Memory network unit receives the current dynamic prediction input and the hidden state of the previous time step; The Long Short-Term Memory (LSTM) network unit updates its internal memory unit state based on the current input and the previous state, and outputs the hidden state of the current time step. The hidden state at the current time step is passed to the feedforward neural network; The feedforward neural network maps the hidden state at the current time step to a preliminary quality prediction value. This process is carried out cyclically within the model, simulating the progress of the process, until a specific time step representing the end of the process is reached. At this point, the output of the feedforward neural network is the value of the final processing quality parameter obtained through deduction.

8. The method for predicting the quality of multi-process cutting based on deep learning according to claim 7, characterized in that, The uncertainty quantification of the final processing quality parameters obtained from the deduction, and the output of the quality prediction results and their confidence levels, include: The Monte Carlo random dropout method is used to perform multiple forward propagations during the inference phase of the feedforward neural network of the cyclically iterative prediction model; In each forward propagation, the connections of some nodes in the feedforward neural network are randomly discarded; Record the values ​​of the final processing quality parameters output by the prediction model in the iterative cycle after each random discard of a connection; Calculate the mean of all recorded values ​​and use the mean as the final quality prediction result; Calculate the standard deviation of all recorded values, and look up the corresponding confidence score in a preset mapping table based on the standard deviation. Use the confidence score as the confidence level for this prediction.

9. The method for predicting the quality of multi-process cutting based on deep learning according to claim 8, characterized in that, Based on the quality prediction results and their confidence levels, a reverse traceability report is generated to identify key process steps that significantly impact the quality results, including: In the common feature space of the processed knowledge network, the gradient of the feature representation of the final time step with respect to the feature representations of each previous time step is calculated. The magnitude of the gradient characterizes the degree of influence of historical process characteristics on the final quality prediction result; Filter out all historical time steps whose absolute gradient value exceeds a preset importance threshold; Map the historical time steps back to their corresponding original process steps; The frequency of each original process step in the historical time step is counted, and one or more original process steps with the highest frequency are identified as the key process steps, and this mapping relationship is written into the reverse traceability report.

10. The method for predicting the quality of multi-process cutting based on deep learning according to claim 9, characterized in that, The step of combining the key process information with the preceding processing data to form a closed-loop model adaptive optimization instruction includes: Retrieve the unified process data sequence corresponding to all processing processes that include the key process steps from the historical database; Detailed process parameter data of the time window in which the key process steps are located are extracted from the unified process data sequence. The detailed process parameter data is paired with the final actual quality data of the corresponding processing process to form an optimized sample pair; Using the newly added optimized sample pairs, a round of fine-tuning training is performed on the shared encoder parameters in the multimodal fusion processing knowledge network; Save the fine-tuned shared encoder parameters and generate the model adaptive optimization instructions containing parameter version identifiers.

Images