Aluminum alloy forging process digital twin monitoring method and device

Abstract

The application provides an aluminum alloy forging process digital twin monitoring method and device, and relates to the technical field of intelligent manufacturing.The method comprises the following steps: collecting physical parameters of a forging process in real time; performing management, standardization processing and fusion on the physical parameters to obtain processed data; and predicting the size of a forging and performing real-time quality monitoring based on the processed data.The application can improve the real-time performance of aluminum alloy forging process monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, specifically to a digital twin monitoring method and device for aluminum alloy forging processes. Background Technology

[0002] Aluminum alloy forgings are widely used in key load-bearing components of high-end equipment such as aerospace due to their excellent specific strength and formability. To ensure assembly accuracy and service reliability, the geometric dimensional accuracy of forgings typically needs to meet strict tolerance control requirements. In existing technologies, quality control of the aluminum alloy die forging process mainly relies on post-forging dimensional inspection and statistical analysis. Offline inspection of finished forgings is performed using coordinate measuring machines, specialized gauges, or 3D scanning equipment, combined with process parameter records for post-process traceability of abnormal batches. Although some production lines have deployed pressure sensors, displacement sensors, and temperature acquisition devices, these are mostly used for equipment operation monitoring or simple parameter over-limit alarms, and a real-time mapping relationship between these sensors and the final dimensions of the forgings has not yet been established. Overall, the existing quality control system still focuses on "process acquisition, post-forging inspection, and post-adjustment," lacking the ability to dynamically model and predict the coupled processes of multiple physical fields such as heat, force, and motion.

[0003] However, the forging process itself is a dynamic process with strong coupling of multiple physical fields and high nonlinearity. Changes in equipment status, die wear, material property fluctuations, and environmental disturbances can all cause time-varying and nonlinear fluctuations in key forging process parameters, thereby affecting material flow behavior and dimensional springback characteristics. Under the current technological framework, quality judgment relies on offline inspection after forging, making it impossible to identify abnormal evolution trends within the key process window in real time during forging. When dimensional deviations are detected, the deviation often results in a batch of products already being affected, leading to increased scrap rates and greater economic losses, which cannot meet the high real-time requirements of high-end equipment manufacturing. Summary of the Invention

[0004] This invention provides a digital twin monitoring method and device for aluminum alloy forging process, which can improve the real-time monitoring of aluminum alloy forging process.

[0005] A first aspect of the present invention provides a digital twin monitoring method for an aluminum alloy forging process, the method comprising: Real-time acquisition of physical parameters during the forging process; The physical parameters are processed, standardized, and fused to obtain processed data; Based on the processed data, the dimensions of the forgings are predicted and real-time quality monitoring is performed.

[0006] Based on the above technical solutions, preferably, the step of managing, standardizing, and fusing the physical parameters to obtain processed data specifically includes: Establish a unified timeline around a single forging cycle; The pressure parameter, slider speed parameter, and billet temperature parameter in the physical parameters are mapped to the unified time axis according to the acquisition timestamp to form a time-aligned mapping relationship; Based on the time alignment mapping relationship, resampling processing is performed on physical parameters with different sampling frequencies to generate a first time-series data set; Construct a sliding time window on the first time series data set; Within each sliding time window, calculate the local mean and local fluctuation amplitude of pressure parameters, slider speed parameters, and billet temperature parameters, and mark sampling points that exceed the preset fluctuation threshold as abnormal sampling points. Neighborhood weighted interpolation is performed using valid sampling points at adjacent time locations of the abnormal sampling points to replace the abnormal sampling points, thereby obtaining a set of time-series data without abnormalities.

[0007] Based on the above technical solutions, preferably, the step of managing, standardizing, and fusing the physical parameters to obtain processed data further includes: Identify data gap segments in consecutive time segments on the anomaly-free time series dataset; Based on the adjacent valid time segments before and after the data gap segment, a linear interpolation function and a spline smoothing function are constructed, and the data gap segment is subjected to joint interpolation processing to obtain a continuous time series data set; Based on the continuous time series data set, the mean and standard deviation of the pressure parameter, the slider speed parameter, and the billet temperature parameter throughout the complete forging cycle are calculated respectively. Based on the full-cycle mean and the full-cycle standard deviation, the parameters of each time step are converted into a standardized expression with zero mean and unit variance, forming a standardized time series feature set.

[0008] Based on the above technical solutions, preferably, the step of managing, standardizing, and fusing the physical parameters to obtain processed data further includes: The pressure parameters, slider speed parameters, and billet temperature parameters within the same time step in the standardized time series feature set are vectorized according to a preset fixed feature order to form a multidimensional feature vector. The multidimensional feature vectors are stacked according to the time order of the unified time axis to construct a periodic feature matrix, thereby obtaining the processed data.

[0009] Based on the above technical solutions, preferably, the step of predicting forging dimensions and performing real-time quality monitoring based on the processed data specifically includes: The processed data is constructed into a periodic feature matrix covering a single forging cycle under a unified time axis; The periodic feature matrix is ​​input into a pre-trained temporal prediction model, and one-dimensional convolutional feature extraction is performed sequentially to obtain a local process texture feature sequence, gated recurrent unit modeling is performed to obtain a hidden state sequence, and temporal attention weighting is performed to obtain a context vector. The output layer generates key dimension prediction values ​​corresponding to the single forging cycle, including radial dimension prediction values ​​and axial dimension prediction values. The predicted key dimensions are written into the display variables that are bound to the 3D digital model; The predicted key dimensions are compared with the preset tolerance zone online to determine consistency and generate a quality status identifier. An alarm signal is generated when the quality status indicator indicates that any predicted value of a critical dimension exceeds the preset tolerance zone. A quality risk index is constructed based on the predicted key dimensions, the preset tolerance zone, the temporal attention weight, and the processed data. Based on the quality risk index and preset warning thresholds and alarm thresholds, hierarchical monitoring is performed. The state in which the quality risk index exceeds the warning threshold but does not exceed the alarm threshold is marked as a warning state, and the state in which the quality risk index exceeds the alarm threshold is marked as an alarm state.

[0010] Based on the above technical solutions, preferably, before predicting the forging dimensions based on the processed data and performing real-time quality monitoring, the method further includes: A sample set is established around historical forging samples, and the processing data corresponding to each historical forging sample is organized into a periodic feature matrix according to a unified time axis; The actual detected size value corresponding to the periodic feature matrix is ​​determined as the target size label; The sample set is partitioned using a single forging cycle as the smallest unit of division to generate a training sample subset and a validation sample subset. A network structure is constructed based on the training sample subset. The network structure sequentially includes a one-dimensional convolutional layer, a gated recurrent unit layer, a temporal attention layer, and an output layer. The one-dimensional convolutional layer is used to perform local convolution operations on the periodic feature matrix on the unified time axis to obtain a local process texture feature sequence. The gated recurrent unit layer is used to perform long-term dependency modeling on the local process texture feature sequence to generate a hidden state sequence. The temporal attention layer is used to assign attention weights to the hidden state sequence to generate a context vector. The output layer is used to map the context vector into radial dimension prediction values ​​and axial dimension prediction values. The parameters are initialized and a loss function is constructed on the training sample subset. Based on the target size label, the network parameters are iteratively updated through the backpropagation algorithm so that the radial size prediction value and the axial size prediction value gradually approach the target size label. After each training round, the validation loss is calculated using the validation sample subset and it is determined whether the preset convergence condition is met. When the preset convergence condition is met or the preset number of training rounds is reached, training is stopped and the trained time series prediction model is obtained.

[0011] Based on the above technical solutions, preferably, the step of managing, standardizing, and fusing the physical parameters to obtain processed data further includes: Based on the unified time axis, the acquisition timestamps of pressure parameters, slider speed parameters, and billet temperature parameters are uniformly converted into standard timestamps; A key process window detection mechanism is constructed based on the standard timestamp. The key process window is identified by joint threshold discrimination of the first derivative of the pressure parameter, the second derivative of the slider speed parameter, and the rate of change of the billet temperature parameter. Generate a window identifier for each of the aforementioned key process windows; Under the window identifier constraint, cross-correlation analysis is performed on the pressure parameter, the slider speed parameter, and the billet temperature parameter respectively to calculate the optimal time delay between each parameter; Based on the optimal time delay, phase compensation translation processing is performed on the pressure parameter, the slider speed parameter, and the billet temperature parameter to generate a phase correction data set; Based on the phase correction data set, a unified time axis resampling process is performed to form a second time series data set that maintains phase consistency both inside and outside the critical process window. The second time-series data set is processed to obtain the processed data.

[0012] In a second aspect of the invention, a digital twin monitoring device for an aluminum alloy forging process is provided. The device is used to execute a digital twin monitoring method for an aluminum alloy forging process as described in any of the above embodiments. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to collect physical parameters of the forging process in real time; The processing module is used to manage, standardize, and fuse the physical parameters to obtain processed data; The output module is used to predict the size of the forging based on the processed data and to perform real-time quality monitoring.

[0013] In a third aspect of the invention, an electronic device is provided, including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the preceding embodiments.

[0014] In a fourth aspect of the invention, a non-transitory computer-readable storage medium is provided, the computer-readable storage medium storing instructions that, when executed, perform the method as described in any of the preceding claims.

[0015] In summary, one or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: 1. This invention constructs a unified time axis within a single forging cycle, integrating pressure parameters, slide speed parameters, and billet temperature parameters through time alignment, phase correction, anomaly mitigation, and standardization. This forms a periodic feature matrix that can be directly input into a time-series prediction model, enabling multi-source process data to undergo structured processing immediately after acquisition and participate in dimensional prediction in real time. Simultaneously, a pre-trained time-series prediction model is used to perform online inference on the processed data, directly outputting radial and axial dimensional prediction values. These values ​​are then compared with preset tolerance zones for online consistency judgment and graded monitoring. This transforms the traditional post-forging offline detection-based post-delay judgment into in-process prediction and real-time judgment based on process data. This achieves a closed-loop chain from data acquisition and feature construction to dimensional evaluation and alarm operation during the forging process, significantly shortening the quality feedback path and improving the real-time performance of aluminum alloy forging process monitoring.

[0016] 2. Establish a unified time axis around a single forging cycle, and perform time alignment, resampling, sliding time window statistics, and abnormal sampling point replacement on pressure parameters, slide speed parameters, and billet temperature parameters. This enables multi-source physical parameters to form a first time series data set with consistent structure and controlled noise under the same time reference system, thereby eliminating the interference of sampling frequency differences and instantaneous anomalies on subsequent analysis and improving the temporal consistency and stability of process data.

[0017] 3. Based on the abnormal time series data set, further identify the missing data segments and perform joint interpolation processing. At the same time, calculate the mean and standard deviation of the whole cycle and complete the standardized expression of zero mean unit variance. This ensures that all time steps within a single forging cycle have continuous and valid data, and eliminates the influence of differences in the dimensions and scales of different physical parameters on model training and inference, thereby improving the consistency and learnability of feature distribution.

[0018] 4. The standardized time series feature set is concatenated into vectors in a fixed feature order within the same time step, and stacked into a periodic feature matrix in a unified time axis time order. This enables the multi-source physical parameters to form a well-structured and semantically stable input matrix in the time and feature dimensions, thereby providing a fixed-dimensional, batch-processable data structure for the time series prediction model and improving the matching degree and operational stability of the data and model interface.

[0019] 5. Input the periodic feature matrix into the pre-trained time series prediction model to generate key dimension prediction values, and perform online consistency judgment and hierarchical monitoring of the prediction results and preset tolerance zones. At the same time, construct a quality risk index to realize a real-time closed loop from process data to dimension prediction to quality status identification and alarm signals, so that quality control is transformed from offline detection after forging to in-process prediction and real-time judgment, improving the real-time and forward-looking nature of forging process monitoring.

[0020] 6. A sample set is established based on historical forging samples, and a temporal prediction model is constructed consisting of a one-dimensional convolutional layer, a gated recurrent unit layer, a temporal attention layer, and an output layer. The model parameters are trained under supervision using target size labels, and the convergence process is controlled by a subset of validation samples. This enables the model to learn the complex nonlinear mapping relationship between multiple physical parameters and the final size under a unified time axis structure, thereby providing a size prediction model with generalization ability for the online stage and improving the size prediction accuracy and stability. Attached Figure Description

[0021] Figure 1 This is a schematic flowchart of a digital twin monitoring method for aluminum alloy forging process disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of a digital twin detection system for aluminum alloy forging process disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of a digital twin monitoring device for aluminum alloy forging process disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention.

[0022] Explanation of reference numerals in the attached drawings: 301, acquisition module; 302, processing module; 303, output module; 401, processor; 402, communication bus; 403, user interface; 404, network interface; 405, memory. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] In the description of the embodiments of the present invention, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0025] In the description of the embodiments of the present invention, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0026] The existing quality control methods for aluminum alloy die forging processes still focus on offline inspection and post-forging traceability. Although these methods can verify the accuracy of finished product dimensions, they lack the ability to perceive and predict the dynamic behavior of multi-physics coupling during the forging process in real time. They cannot establish a real-time mapping relationship between key process parameters and the final dimensions of the forging. This results in the inability to identify and intervene in time when abnormal fluctuations occur within the critical process window. Once dimensional deviations are detected, they often form batch deviations, which not only increases scrap rate and economic losses, but also makes it difficult to meet the stringent requirements of high-end equipment manufacturing such as aerospace for process real-time performance, predictability, and consistency.

[0027] This embodiment discloses a digital twin monitoring method for aluminum alloy forging processes, referring to... Figure 1 This includes the following steps S110-S130: S110 collects physical parameters of the forging process in real time.

[0028] This invention discloses a digital twin monitoring method for aluminum alloy forging processes, applied to a server. The server includes, but is not limited to, electronic devices such as mobile phones, tablets, wearable devices, and PCs (Personal Computers), and can also be a backend server running the digital twin monitoring method for aluminum alloy forging processes. The server can be a standalone server or a server cluster composed of multiple servers.

[0029] The implementation of real-time acquisition of physical parameters in the forging process can establish a traceable data link around four types of identifiers: forging equipment object identifier, acquisition point identifier, data address identifier, and time synchronization identifier. This enables pressure parameters, slide speed parameters, slide position parameters, and billet temperature parameters to form a continuous time-series data set under the same acquisition system.

[0030] First, the PLC hardware configuration is completed on the forging equipment side. The collectable process quantities of the forging equipment are fixed as global data blocks in the PLC, and a unique data address identifier is assigned to each process quantity. Simultaneously, remote PUT / GET access is enabled and optimized block access is disabled in the PLC communication properties to ensure stable readability of the data address identifier under byte offset addressing. Then, communication configuration is established on the host computer side. A PLC connection object is created based on the S7 communication protocol, and connection parameters are configured to ensure that the host computer and PLC are on the same network segment and maintain a long-term connection session. This provides a stable data channel for periodically reading pressure parameters, slider speed parameters, and slider position parameters.

[0031] After the connection is established, the global data block is periodically read based on the data address identifier. This periodic reading is driven by a timer, and upon each trigger, pressure parameters, slider speed parameters, and slider position parameters are read in a consistent batch. Simultaneously, the read results are type-converted to form numerical sampling points, and a unified acquisition timestamp is written as a time synchronization identifier when the sampling point is generated, thus forming a device-side sampling record bound to the time synchronization identifier. On the billet temperature parameter acquisition side, the output channel of the temperature acquisition device is bound as a temperature acquisition point identifier, and the sampling trigger strategy of the temperature acquisition device is aligned with the trigger cycle of the periodic reading. This ensures that the temperature sampling point is also written with a time synchronization identifier upon generation, thus forming a temperature-side sampling record that is aligned with the device-side sampling record.

[0032] Subsequently, the equipment-side sampling records and temperature-side sampling records are written into a database table structure. The database table structure includes at least the acquisition time field, pressure field, slider speed field, slider position field, and billet temperature field. The acquisition time field is used to carry a time synchronization identifier to ensure that subsequent retrieval based on the acquisition time field can restore the correspondence between pressure parameters, slider speed parameters, slider position parameters, and billet temperature parameters at the same time and position. During the database writing process, an acquisition batch identifier is attached to each written record as a forging cycle association key, enabling continuous sampling points within the same single forging cycle to be aggregated into a single forging cycle data segment. The pressure curve, slider speed curve, and slider position curve are simultaneously plotted on the upper computer visualization interface with the acquisition time field as the horizontal axis to achieve online verification of the acquisition link. If necessary, the sampling trigger cycle, data address identifier, and connection parameters are corrected for consistency to ensure that the real-time acquired physical parameter time-series data can stably support subsequent data governance, standardization processing, fusion processing, and forging size prediction and real-time quality monitoring based on the processed data.

[0033] The physical parameters are managed, standardized, and integrated to obtain processed data.

[0034] In one possible implementation, the physical parameters are managed, standardized, and fused to obtain processed data. Specifically, this includes: establishing a unified time axis around a single forging cycle; mapping the pressure parameter, slider speed parameter, and billet temperature parameter in the physical parameters to the unified time axis according to the acquisition timestamp to form a time-aligned mapping relationship; performing resampling processing on physical parameters with different sampling frequencies based on the time-aligned mapping relationship to generate a first time-series data set; constructing a sliding time window on the first time-series data set; calculating the local mean and local fluctuation amplitude of the pressure parameter, slider speed parameter, and billet temperature parameter within each sliding time window, and marking sampling points exceeding a preset fluctuation threshold as abnormal sampling points; and performing neighborhood weighted interpolation processing using valid sampling points at adjacent time positions of the abnormal sampling points to replace the abnormal sampling points, thereby obtaining a de-abnormal time-series data set.

[0035] Specifically, the start and end boundaries of a single forging cycle are determined by the forging equipment's operating state sequence, ensuring strict consistency between the unified time axis and the single forging cycle in terms of time range. The forging equipment's operating state sequence includes state bits used to mark die entry, forming holding, and return release. The start and end times of a single forging cycle are obtained by reading the flip times of these state bits, and the time interval between the start and end times is defined as the cycle time domain. The time step size of the unified time axis is determined by the target sampling frequency of subsequent resampling processing, ensuring that the unified time axis covers the cycle time domain and is discretely spaced in time. The unified time axis consists of a set of monotonically increasing time points, each serving as an anchor point for subsequent time alignment mapping relationships. The unified time axis refers to establishing a common time reference system for pressure parameters, slide speed parameters, and billet temperature parameters, enabling sampling records generated by different sampling links to be projected onto the same time coordinate for consistent processing.

[0036] The acquisition timestamps for pressure parameters, slider speed parameters, and billet temperature parameters are generated using the same time base, and each sampling record is mapped to a unified time axis through unified timestamp conversion. The acquisition timestamp refers to the time stamp corresponding to the moment the sampling record is generated. It is usually written to the record field by the PLC or host computer at a millisecond-level time resolution and serves as the primary key for subsequent alignment. The time alignment mapping relationship uses the time point on the unified time axis as the query key and the original sampling points of each physical parameter as the candidate set to establish the association between each time point and its neighboring sampling points, enabling each time point to obtain the corresponding values ​​of pressure parameters, slider speed parameters, and billet temperature parameters at that moment. To ensure that the mapping relationship can still be stably formed under different sampling frequencies, nearest neighbor index or linear interpolation index is used as the mapping rule. The nearest neighbor index is used when the sampling frequency is higher than the target frequency, and the linear interpolation index is used when the sampling frequency is lower than the target frequency or the sampling points are irregular. The time alignment mapping relationship refers to establishing a consistent value position for cross-source parameters for each time point on the unified time axis, enabling multi-source parameters to form a comparable and concatenable joint representation at the same time point.

[0037] Physical parameters with different sampling frequencies are converted into sequences of equal intervals and lengths on a unified time axis through resampling, thus forming the first time-series data set. The goal of resampling is to generate corresponding values ​​for pressure parameters, slider speed parameters, and billet temperature parameters at each time point on the unified time axis, ensuring that the three types of parameters have the same number of time steps and time intervals. For any physical parameter, let its original sampling sequence be... ,in To collect timestamps, These are sampled values, with a unified time axis. The expression for linear resampling is:

[0038] in, Indicates a point in time on a unified timeline Resampled values ​​on; and Indicates satisfaction Adjacent sampling timestamps; and Indicates and and The corresponding sampled value; This represents the k-th time point on the unified time axis; N represents the number of original sampling points; and K represents the number of time points on the unified time axis. This expression utilizes the local linear variation relationship between adjacent sampling points to fill in missing time positions on the unified time axis, ensuring the resampled sequence is completely aligned with the unified time axis. The first time series data set refers to a joint data object that simultaneously contains pressure parameter sequences, slider speed parameter sequences, and billet temperature parameter sequences under the same unified time axis, and all three have the same set of time indices.

[0039] Sliding time windows are generated consecutively on the first time-series dataset according to a unified time axis. They are used to calculate local statistics and provide local context for anomaly detection. A sliding time window is a continuous subsequence truncated from the time series. Its window length is represented by the number of time steps W, and the window step size is represented by the number of time steps S. W determines the smoothness of the local statistics, and S determines the temporal resolution of anomaly detection. For each physical parameter sequence in the first time-series dataset, a window index interval is constructed at the k-th time point on the unified time axis. or Furthermore, truncated windows or mirror fills are used at sequence boundaries to ensure window availability. A sliding time window refers to a local observation interval that shifts over time, allowing anomaly detection at each time point to be based on its neighboring time segments rather than single-point values, thereby reducing the interference of instantaneous noise on the detection.

[0040] Local mean and local fluctuation amplitude are calculated for pressure parameters, slider speed parameters, and billet temperature parameters within each sliding time window, respectively, and used to construct anomaly sampling point discrimination rules. The local mean represents the central trend of the sampled values ​​within the window, and the local fluctuation amplitude represents the dispersion of the sampled values ​​within the window relative to the local mean. Both are used together to identify abrupt changes inconsistent with the local background. For any physical parameter sequence within the window... The expressions for the local mean and local standard deviation within a given range are:

[0041] in, This represents the local mean within a sliding time window with the k-th time point as a reference. This represents the local standard deviation within the corresponding sliding time window, serving as a quantification of the local fluctuation amplitude; W represents the window length. This represents the set of indices for the k-th window; This represents the sampled value at the j-th time point within the window. A preset fluctuation threshold is used to determine whether a sampled point deviates from the local background. The fluctuation threshold is usually defined as a multiple threshold to adaptively match the local fluctuation amplitude. The expression for anomaly detection is:

[0042] in, This represents the sampled value at the k-th time point; This represents the preset fluctuation threshold coefficient, used to control the sensitivity of anomaly detection; This indicates the degree of deviation of the sampled value from the local mean. This discrimination method uses local statistics to dynamically characterize the normal fluctuation range of the sequence, avoiding over-detection or under-detection in different operating conditions when using a globally fixed threshold. Abnormal sampling points refer to sampling points that significantly deviate from the normal fluctuation range under the local statistical background. They are usually caused by instantaneous sensor jitter, communication jitter, quantization overflow, or sampling misalignment. Retaining them will lead to deviations in subsequent standardization and model prediction.

[0043] Neighborhood-weighted interpolation is used to replace outlier sampling points, ensuring that the de-outlier time-series dataset maintains consistency in temporal continuity and physical plausibility. A valid sampling point refers to the set of sampling points that are not marked as outliers and are located at time points adjacent to the outlier. The neighborhood refers to a fixed time step range extending forward and backward from the outlier, or an adaptive range until a sufficient number of valid sampling points are encountered. Neighborhood-weighted interpolation generates replacement values ​​for the outlier location by assigning weights to valid sampling points within the neighborhood and performing a weighted sum. The weights are typically inversely proportional to the time distance to emphasize closer valid sampling points. Let the outlier sampling point be located at time point... Its neighborhood valid sampling point index set is The expression for neighborhood-weighted interpolation is:

[0044] in, This represents the interpolation result used to replace abnormal sampling points; This represents the sampled value of the j-th valid sampled point within the neighborhood; This represents the weight between the outlier sampling point and the j-th valid sampling point; This indicates the time point on the unified time axis corresponding to the abnormal sampling point; This represents the time point on the unified time axis corresponding to the valid sampling point; This method prevents extremely small positive numbers with a denominator of zero. It utilizes multi-point information within the neighborhood to smoothly replace outliers, avoiding step artifacts introduced by single forward or backward filling, and ensuring the stability of the replacement result during subsequent missing interpolation, calculation of the full-cycle mean and standard deviation, and construction of the cycle feature matrix. The de-outlier time-series data set refers to the joint sequence set formed by replacing all outlier sampling points with the interpolation results while keeping the time index of the first time-series data set unchanged. This ensures that pressure parameters, slider speed parameters, and billet temperature parameters all have continuous numerical expressions on a unified time axis that can be used for subsequent processing.

[0045] In one possible implementation, the physical parameters are processed, standardized, and fused to obtain processed data. Specifically, this includes: identifying data gaps in continuous time segments on the anomaly-removed time-series data set; constructing linear interpolation functions and spline smoothing functions based on adjacent valid time segments before and after the data gaps, and performing joint interpolation processing on the data gaps to obtain a continuous time-series data set; calculating the full-cycle mean and full-cycle standard deviation of pressure parameters, slider speed parameters, and billet temperature parameters within the complete forging cycle based on the continuous time-series data set; and converting the parameters of each time step into a standardized expression with zero mean and unit variance based on the full-cycle mean and full-cycle standard deviation to form a standardized time-series feature set.

[0046] Specifically, even within a unified timeline, the anomaly removal time-series dataset may still contain intervals without valid values ​​due to communication interruptions, PLC read failures, temperature acquisition device outages, or missing database writes. Therefore, it is necessary to identify data gaps in the anomaly removal time-series dataset to clarify the scope of subsequent interpolation. Data gaps are determined by checking the validity of pressure parameters, slider speed parameters, and billet temperature parameters at each time point on the unified timeline. Validity discrimination uses a unified rule of missing value marking, non-numerical marking, or out-of-bounds placeholder value marking to ensure consistent identification of missing data from different sources. During the identification process, the validity discrimination result of each time point is first encoded into a binary sequence, and then continuous segment merging is performed on the binary sequence to obtain start and end indices, thereby aggregating consecutively missing intervals at adjacent time points into the same data gap segment. A data gap segment refers to a time interval on a unified time axis where multiple consecutive time points lack valid values ​​for at least one type of physical parameter. Its boundary is determined by the time points when the missing state switches from valid to missing and from missing to valid. The identification results are used to constrain interpolation to occur only in the missing interval without changing the original values ​​in the valid interval.

[0047] For each data gap segment, a linear interpolation function and a spline smoothing function are constructed using adjacent valid time segments before and after the data gap segment. These are then jointly interpolated to obtain a continuous time-series data set. Adjacent valid time segments refer to the set of valid values ​​located to the left and right of the data gap segment and immediately adjacent to its boundary, providing boundary conditions and local trends for interpolation. The linear interpolation function is used to create a monotonic transition consistent with the boundary values ​​within the missing interval, ensuring continuity of the interpolation result at the boundary. The spline smoothing function introduces curvature constraints that better reflect the process evolution within the missing interval, making the interpolation result not only continuous but also smooth at the first derivative level, reducing the piecewise linear artifacts generated by linear interpolation in long missing intervals. For any physical parameter in the data gap segment... Inside, let the valid point of the left boundary be... The valid points on the right boundary are For any time point within the missing interval The linear interpolation expression is:

[0048] in, This indicates the linear interpolation function at time point The interpolation value at the location; and Indicates the nearest valid time point on both sides of the data gap segment; and This indicates the parameter value at the corresponding valid time point; This represents the time point within the missing data segment. The spline smoothing function constructs a cubic spline using multiple valid points within adjacent valid time segments as control points, ensuring that the interpolated values ​​within the missing segment change continuously over time, and that the function values ​​at the control points are consistent and the derivative changes smoothly; let the spline smoothing function be... Joint interpolation obtains the final interpolated value by weighted fusion of the linear interpolation result and the spline smoothing result, ensuring that short missing intervals remain linearly stable and long missing intervals maintain reasonable curvature. The expression for joint interpolation is:

[0049] in, This represents the final interpolation value of the combined interpolation; The fusion weights are used to control the strength of spline smoothing in missing regions. It can be determined by the length of the missing segment or The normalized distance from the boundary is determined so that when it is close to the boundary, it is closer to linear interpolation to ensure boundary stability, and when it is in the middle of the segment, spline smoothing is added to ensure overall smoothness. This represents the output of the spline smoothing function; This represents the output of a linear interpolation function. A continuous time series dataset refers to a time series dataset that has no missing values ​​across the entire range of a unified time axis, and where each physical parameter maintains continuous values ​​at the boundaries of data gap segments, while maintaining a trend and local smoothness within the segments.

[0050] After the continuous time-series data set is formed, the full-cycle mean and standard deviation of pressure parameters, slide speed parameters, and billet temperature parameters are calculated separately within a complete forging cycle. This ensures that the standardized statistical benchmark covers all time points of a single forging cycle. The full-cycle mean characterizes the central level of the physical parameter within a single forging cycle, while the full-cycle standard deviation characterizes the overall fluctuation scale of the physical parameter within a single forging cycle. Together, they constitute the statistics required for zero-mean unit variance transformation. For any physical parameter sequence... Where K is the number of time points on the unified time axis, the expressions for the full-cycle mean and full-cycle standard deviation are:

[0051] in, This represents the average value of the physical parameter over the entire forging cycle; This represents the full-cycle standard deviation of the physical parameter over the entire forging cycle; Indicates the unified timeline The parameter values ​​at each time point; This indicates the number of time points on the unified time axis. The full-cycle mean and full-cycle standard deviation refer to the parameter center and scale obtained with a single forging cycle as the statistical range. They are used to ensure that the standardization process is self-consistent within the same single forging cycle and to avoid scale inconsistencies caused by batch differences introduced by cross-cycle statistics.

[0052] Based on the full-cycle mean and standard deviation, parameters at each time step are converted into a standardized expression with zero mean and unit variance, thus forming a standardized time-series feature set. This makes pressure parameters, slider speed parameters, and billet temperature parameters comparable on a numerical scale and meets the requirements of the prediction model for input distribution stability. The zero-mean, unit-variance standardized expression uses Z-score transformation, which applies to any physical parameter at any time point. The standardized value at point is defined as:

[0053] in, A standardized representation of the k-th time point on a unified timeline; This indicates the original parameter values ​​at that point in time, taken from a continuous time series dataset. This represents the average value of the physical parameter over the entire forging cycle; This represents the full-cycle standard deviation of the physical parameter over the entire forging cycle; Indicates prevention The smallest positive number that causes a division-by-zero error when the parameter is approximately constant within a certain period is the one that causes the division-by-zero error. Maintaining numerical stability. The standardized temporal feature set refers to a set of joint sequences consisting of standardized sequences of pressure parameters, slider speed parameters, and billet temperature parameters on a unified time axis. Each sequence is normalized with zero mean and unit variance, ensuring that different physical parameters do not dominate feature contributions due to differences in numerical magnitude during subsequent vector concatenation and periodic feature matrix construction. This also enables the one-dimensional convolutional layer, gated recurrent unit layer, and temporal attention layer of the temporal prediction model to learn transferable temporal correlation structures under a stable input distribution.

[0054] In one possible implementation, the physical parameters are managed, standardized, and fused to obtain processed data. Specifically, this includes: concatenating pressure parameters, slider speed parameters, and billet temperature parameters within the same time step in the standardized time-series feature set into vectors according to a preset fixed feature order to form a multi-dimensional feature vector; stacking the multi-dimensional feature vectors according to the time order of a unified time axis to construct a periodic feature matrix to obtain processed data.

[0055] Specifically, at each time step on the unified time axis, the pressure parameter, slider speed parameter, and billet temperature parameter have already been standardized to zero mean and unit variance. Therefore, within the same time step, these three standardized expressions need to be combined into a multi-dimensional feature vector that can be directly processed by the subsequent time series prediction model according to a preset fixed feature order. The preset fixed feature order remains consistent during the training and online inference phases, and consistency is achieved through immutable field position constraints, such as a fixed arrangement of pressure parameter, slider speed parameter, and billet temperature parameter. This ensures that each feature dimension corresponds stably in semantics and physical origin, avoiding the model misinterpreting the same feature dimension as different physical parameters due to changes in order. For the k-th time step on the unified time axis, let the standardized value of the pressure parameter be... The standardized value of the slider speed parameter is The standardized value of the billet temperature parameter is Then the expression for the multidimensional feature vector is:

[0056] in, This represents the multidimensional feature vector at the k-th time step; This represents the standardized value of the pressure parameter at the k-th time step; This represents the standardized value of the slider velocity parameter at the k-th time step; This represents the standardized value of the billet temperature parameter at the k-th time step. This vector concatenation method achieves cross-parameter fusion at the time step granularity, enabling the pressure parameter, slider speed parameter, and billet temperature parameter to form a joint state representation at the same time position. This allows the local process texture features extracted by the subsequent one-dimensional convolutional layer to simultaneously reflect the coupled changes of the three types of parameters, and establishes the modeling of long-term dependencies of the gated recurrent unit layer on a unified joint state sequence.

[0057] After generating the multidimensional feature vectors for each time step, the multidimensional feature vectors from each time step are stacked according to the time order of a unified time axis to form a periodic feature matrix covering a single forging cycle. This periodic feature matrix maintains consistent time order in the row direction and consistent feature order in the column direction, thus serving as the input tensor structure for the processed data output and direct integration with the time series prediction model. Let the number of time steps on the unified time axis be . Then the expression for the periodic characteristic matrix is:

[0058] Where X represents the periodic characteristic matrix; This represents the transpose of the multidimensional eigenvector at the k-th time step, used to write the vector row-wise into the matrix; K represents the number of time steps covered by a single forging cycle, determined by a unified time axis; the matrix has 3 columns, corresponding to the fixed characteristic order of pressure parameters, slide speed parameters, and billet temperature parameters. By... As the output of the processing data, the periodic feature matrix in the sample pair set during the training phase is structurally identical to the periodic feature matrix generated in a single forging cycle during the online phase, avoiding input dimension drift and ensuring that subsequent one-dimensional convolutional feature extraction, gated recurrent unit modeling, and temporal attention weighting based on the periodic feature matrix can all run stably on a consistent data structure.

[0059] Furthermore, in the forging process of key components, critical process windows such as the moment of die closing, the moment of punch contact, the rapid pressure increase section, and the friction state abrupt change section are often extremely short in duration, but they have a decisive impact on the material flow path, contact pressure distribution, and final dimensional springback. In a multi-source data acquisition system, pressure parameters are usually obtained by press-side sensors through periodic scanning by a PLC, slider speed parameters may be output by a displacement encoder through a high-speed acquisition module, and billet temperature parameters are inherently lagging due to the infrared temperature measurement response time and filtering algorithm. When there are different levels of buffer delay, network jitter, scanning cycle differences, or sensor dynamic response lag in each sampling link, even if time alignment is achieved in form through a unified time axis and resampling processing, a phase deviation of milliseconds to tens of milliseconds may still remain within the critical process window. Since the critical process window itself is extremely short in duration, this phase deviation will cause the peak pressure parameter, the abrupt change point of slider speed, and the change in billet temperature gradient to be out of sync in time, resulting in the multidimensional feature vectors formed within the same time step not being a true physical synchronous state, but a spliced ​​state across time segments. In this context, the local process texture features extracted by the one-dimensional convolutional layer will form distortion patterns based on the misaligned signal structure. When modeling long-term dependencies, the gated recurrent unit layer may misjudge pseudo-synchronization relationships caused by phase misalignment as true causal relationships, thereby reducing prediction stability and reliability during the critical process window, the time period most sensitive to size. In high-energy, rapid prototyping equipment environments, the changes in physical quantities per unit time are greater and the time scale is shorter, further amplifying the relative impact of phase deviations. This causes small errors in multi-source time alignment to evolve into significant sources of deviation in size prediction.

[0060] In one possible implementation, the physical parameters are managed, standardized, and fused to obtain processed data. Specifically, this includes: converting the acquisition timestamps of pressure parameters, slider speed parameters, and billet temperature parameters to standard timestamps based on a unified time axis; constructing a key process window detection mechanism based on the standard timestamps, identifying key process windows by performing joint threshold discrimination on the first derivative of the pressure parameter, the second derivative of the slider speed parameter, and the rate of change of the billet temperature parameter; generating a window identifier for each key process window; performing cross-correlation analysis on the pressure parameter, slider speed parameter, and billet temperature parameter respectively under the constraints of the window identifier to calculate the optimal time delay between each parameter; performing phase compensation translation processing on the pressure parameter, slider speed parameter, and billet temperature parameter based on the optimal time delay to generate a phase-corrected data set; performing unified time axis resampling processing on the phase-corrected data set to form a second time-series data set that maintains phase consistency inside and outside the key process window; and processing the second time-series data set to obtain processed data.

[0061] Specifically, the unified time axis has defined the time range and target time step of a single forging cycle. However, the timestamps generated by different sampling links may originate from the PLC scanning time, the host computer reading time, or the internal clock of the temperature acquisition device. Therefore, the acquired timestamps are first uniformly converted into standard timestamps to eliminate systematic offsets caused by inconsistent time bases. The standard timestamp uses the start time of the unified time axis as the zero point and the time step of the unified time axis as the discrete scale. Any acquired timestamp is first mapped to a relative time offset, and then quantized to the time grid of the unified time axis, so that the pressure parameters, slide speed parameters, and billet temperature parameters have the same reference system at the time mark level. Let the start time of the unified time axis be... The time step is Any collection timestamp is Then the discrete index of the standard timestamp can be written as:

[0062] in, This represents a time index on a unified timeline. This represents the rounding mapping rule, which ensures a one-to-one correspondence between standard timestamps and a unified timeline, and can be used for alignment operations in subsequent window discrimination and cross-correlation analysis.

[0063] The critical process window detection mechanism operates within the standard timestamp domain, ensuring that the detected object is unaffected by sampling link clock differences. It captures short-duration, highly nonlinear segments by using a joint threshold to determine the first derivative of the pressure parameter, the second derivative of the slider speed parameter, and the rate of change of the billet temperature parameter. The first derivative of the pressure parameter characterizes abrupt changes in the pressurization rate, the second derivative of the slider speed parameter characterizes abrupt changes in acceleration or impact response, and the rate of change of the billet temperature parameter characterizes the temperature gradient changes caused by heat input and contact heat transfer. These three parameters often exhibit synchronous abrupt changes at the moment of mold closing, the moment of contact, and the rapid pressurization segment; therefore, using a joint threshold can avoid false triggering by a single signal. (Standard timestamp index) At this point, the derivative is calculated using discrete difference calculus, assuming the pressure parameter is... The slider speed parameter is The billet temperature parameters are as follows: ,but:

[0064] in, This represents a discrete approximation of the first derivative of the pressure parameter. This represents a discrete approximation of the second derivative of the slider velocity parameter. A discrete approximation representing the rate of change of billet temperature parameters. To unify the time axis step size, the joint threshold discrimination uses the simultaneous satisfaction of three types of derivatives or the satisfaction of preset combinational logic as the trigger condition for the critical process window. For example, it adopts...

[0065] in, Indicates the trigger strength indicator. , , These represent the threshold values ​​for the first derivative of the pressure parameter, the second derivative of the slider speed parameter, and the rate of change of the billet temperature parameter, respectively. As an indicator function, when the trigger intensity indicator meets the preset minimum trigger intensity, the continuous interval near the time index is identified as the critical process window to ensure that the identification result is sensitive to short-term mutations and insensitive to isolated noise.

[0066] The critical process window is identified as one or more consecutive intervals on a standard timestamp index. A window identifier needs to be generated for each consecutive interval to ensure that phase correction is subsequently confined within that interval while maintaining the original alignment strategy outside the interval. The window identifier consists of a window number and a window boundary index. The window boundary index includes a window start index and a window end index, allowing any time index to be determined as belonging to a specific window. The window identifier is generated using a continuous segment aggregation method, merging time-adjacent indices with true trigger intensity indicators into the same window. A window number is written for each window, and the window start index and window end index are recorded, ensuring that the window identifier is strictly bound to a unified time axis and can be directly used to constrain the sampling interval for cross-correlation analysis.

[0067] Cross-correlation analysis was performed on pressure parameters, slider speed parameters, and billet temperature parameters under window constraints. This was used to quantify the relative time delays of different parameters within the critical process window and calculate the optimal time delay, enabling subsequent phase compensation to correct phase misalignments within the critical process window without introducing unnecessary full-cycle translation. The cross-correlation analysis used window boundary indices to extract discrete sequences within the window and calculated cross-correlation values ​​within a preset hysteresis search range. The peak position of the cross-correlation value corresponded to the optimal time delay. The hysteresis search range was determined by the maximum buffer delay that the sampling link might generate and the sensor dynamic response hysteresis, ensuring the search covered the true time delay while avoiding the introduction of spurious peaks over an excessively large range. Let the pressure parameter sequence within a certain window be... The slider speed parameter sequence is as follows The billet temperature parameter sequence is as follows When the pressure parameter is used as the reference sequence, the cross-correlation between the slider speed parameter and the billet temperature parameter relative to the pressure parameter can be written as:

[0068] in, Represents the lag, with values ​​ranging from Within the integer range, and These represent the starting and ending indices of the window, respectively. During cross-correlation calculations, out-of-bounds terms are truncated or zero-padded to ensure valid summation. The optimal time delay is determined by the peak value of the cross-correlation, aiming for phase compensation with maximum correlation alignment.

[0069] in, and These represent the optimal time delays of the slider speed parameter and the billet temperature parameter relative to the pressure parameter, respectively. If other parameters are selected as the reference sequence, the corresponding optimal time delay set is calculated according to the same rule and saved separately in the window identifier dimension.

[0070] Based on the optimal time delay, phase compensation translation is performed on the pressure parameter, slider speed parameter, and billet temperature parameter to ensure that the multiple source values ​​at the same standard timestamp index within the critical process window correspond to the same physical event time as much as possible, thereby reducing the joint state splicing error caused by phase misalignment. Phase compensation translation is performed separately for each window, and each parameter is indexed and translated according to its optimal time delay relative to the reference parameter. The translation direction is determined by the sign of the optimal time delay, and the translation magnitude is determined by the absolute value of the optimal time delay. To avoid new gaps at the window boundaries caused by translation, a boundary preservation strategy is adopted. After translation within the window, the vacated index positions are preserved using the effective values ​​adjacent to the window boundary or filled using short-distance interpolation, ensuring that the sequence within the window maintains the same length and continuous processability. Assuming the pressure parameter is the reference sequence and is not translated, while the slider speed parameter and billet temperature parameter are translated according to the optimal time delay, then for index k within the window:

[0071] in, and This represents the phase-compensated slider speed parameters and billet temperature parameters. When the index exceeds the limit, the corresponding value is generated according to the boundary preservation strategy. The phase correction dataset consists of the phase-compensated pressure parameters, slider speed parameters, and billet temperature parameters within each window, and the corresponding uncompensated sequence outside the window. The compensated and uncompensated segments are concatenated into a full-cycle sequence using the window identifier as an index, so that subsequent resampling and anomaly handling are still completed within a unified time axis framework.

[0072] Based on the phase-corrected dataset, a unified time-axis resampling process is performed. This ensures that the phase-compensated data within and outside the window form a structurally consistent second time-series dataset at the same target sampling frequency, maintaining the phase consistency constraint both inside and outside the window after resampling. Resampling uses the discrete-time index of the unified time axis as the output index. Within the window, the phase-compensated sequence is used as the resampling input; outside the window, the original aligned sequence is used, achieving segmented consistency on the input side. The resampling method employs the same interpolation or nearest-neighbor rules as the previous sequence, ensuring the output sequence has the same length and time step across the entire cycle. Consistent boundary handling at window boundaries prevents artificial step jumps. The second time-series dataset includes pressure parameter sequences, slider speed parameter sequences, and billet temperature parameter sequences. All three parameters have values ​​at each index on the unified time axis. Phase consistency is ensured within the critical process window by a phase compensation mechanism, and temporal consistency is ensured outside the critical process window by a conventional alignment mechanism, providing a more reliable joint time-series foundation for subsequent standardization and fusion.

[0073] When processing the second time series dataset to obtain the processed data, the established processing chain of data governance, missing data imputation, full-cycle statistics and standardization, vector concatenation and periodic feature matrix construction is used, and phase compensation information is retained in the window identifier dimension to support quality traceability. The second time-series dataset first enters a sliding time window for statistical analysis to identify outlier sampling points and performs neighborhood-weighted interpolation replacement. Then, it identifies data gaps and performs joint interpolation using linear interpolation and spline smoothing functions to form a continuous time-series dataset. On this continuous time-series dataset, the full-cycle mean and full-cycle standard deviation of pressure parameters, slider speed parameters, and billet temperature parameters are calculated, generating a standardized time-series feature set with zero mean and unit variance. At each time step on a unified time axis, the pressure parameters, slider speed parameters, and billet temperature parameters are concatenated into a multi-dimensional feature vector according to a fixed feature order, and stacked in time order on a unified time axis to form a periodic feature matrix as the processed data output. This ensures that the subsequent one-dimensional convolutional feature extraction, gated cyclic unit modeling, and time-series attention weighting of the time-series prediction model are based on a joint state sequence with consistent phase inside and outside the critical process window, thereby suppressing spurious correlations within short-term strong nonlinear windows and improving the stability and usability of critical dimension prediction values.

[0074] Forging dimensions are predicted based on processed data, and real-time quality monitoring is performed.

[0075] In one possible implementation, before predicting forging dimensions based on processed data and performing real-time quality monitoring, the method further includes: establishing a sample set around historical forging samples, organizing the processed data corresponding to each historical forging sample into a periodic feature matrix according to a unified time axis; determining the actual detected size value corresponding to the periodic feature matrix as the target size label; performing sample partitioning processing on the sample set with a single forging cycle as the smallest partitioning unit to generate a training sample subset and a validation sample subset; constructing a network structure based on the training sample subset, the network structure sequentially including a one-dimensional convolutional layer, a gated recurrent unit layer, a temporal attention layer, and an output layer, wherein the one-dimensional convolutional layer is used to perform local convolution operations on the periodic feature matrix on a unified time axis to obtain local process texture features. The sequence, gated recurrent unit layer is used to perform long-term dependency modeling on local process texture feature sequences to generate hidden state sequences, temporal attention layer is used to assign attention weights to the hidden state sequences to generate context vectors, and output layer is used to map the context vectors to radial and axial dimension prediction values. Parameter initialization and loss function are performed on a subset of training samples. Based on the target size label, the network parameters are iteratively updated through backpropagation algorithm so that the radial and axial dimension prediction values ​​gradually approach the target size label. After each training round, the validation loss is calculated using a subset of validation samples to determine whether the preset convergence condition is met. When the preset convergence condition is met or the preset number of training rounds is reached, training stops and the trained temporal prediction model is obtained.

[0076] Specifically, a sample set of historical forging samples is established using a single forging cycle as the basic recording unit. Each record simultaneously contains the processing data of that single forging cycle and the subsequent traceable dimensional inspection results, and its structure is consistent with the processing data in the online stage. The sample set is constructed using a unified time axis as the alignment benchmark. The processing data of each historical forging sample is reorganized into a periodic feature matrix according to the time index order of the unified time axis. The rows of the periodic feature matrix correspond to the time steps of the unified time axis, and the columns correspond to the fixed feature order of pressure parameters, slider speed parameters, and billet temperature parameters. When there is a difference between the original sampling length of the historical forging sample and the number of time steps of the unified time axis, a resampling rule consistent with the online stage is used to generate a sequence of equal length on the unified time axis, thereby ensuring that the periodic feature matrices of all historical forging samples are completely consistent in dimension and can be directly stacked into a training tensor. To ensure the auditability of the sample set, a forging batch identifier, mold identifier, and equipment identifier are written to each historical forging sample, so that the sample source can be traced back and the source of domain difference can be located when anomalies are verified later.

[0077] The actual detected size values ​​and the periodic feature matrix are bound to the same single forging cycle identifier in the sample set, creating a semantic one-to-one correspondence between the supervision signal and the process input and supporting end-to-end learning. The actual detected size values ​​come from the post-forging inspection stage, including at least radial and axial size detection values, and undergo metrological consistency processing upon warehousing. This ensures that inspection results from different inspection equipment, measurement benchmarks, and temperature compensation strategies can be mapped to the same size definition space. Metrological consistency processing is achieved by fixing the measurement benchmark, compensation strategy, and value rules, minimizing the differences in actual detected size values ​​obtained for the same forging under different measurement conditions. Target size labels are stored in vector form and are consistent with the sample index of the periodic feature matrix, ensuring that each periodic feature matrix corresponds to a unique supervision target during training. Let the first... The target size label for each historical forging sample is: ,but:

[0078] in, This indicates the radial dimension measurement value. The axial dimension detection value is determined by the detection record of the same forging and is bound to the same single forging cycle identifier with the periodic feature matrix of the nth historical forging sample.

[0079] The sample partitioning is performed using a single forging cycle as the smallest unit, ensuring that the training and validation sample subsets are independent at the data level. This avoids the possibility of different segments from the same single forging cycle being included in different subsets, which could lead to an underestimation of the validation loss. The sample partitioning process first removes duplicates from the sample set based on the single forging cycle identifier, and then divides it into training and validation sample subsets according to a preset ratio. Hierarchical constraints can be further added during the partitioning process to ensure that identifiers from different equipment, molds, or material batches are approximately consistently distributed across the two subsets. This allows the validation sample subset to accurately reflect the generalization error across different working conditions. The training sample subset is used for network parameter updates, while the validation sample subset is used only for calculating the validation loss and convergence determination. Furthermore, the validation sample subset does not participate in backpropagation updates throughout the training process, ensuring that the validation loss serves as a stable measure of generalization ability.

[0080] The network structure determines the input interface based on the dimension of the periodic feature matrix of a subset of training samples, and constructs an end-to-end mapping in the order of one-dimensional convolutional layers, gated recurrent unit layers, temporal attention layers, and output layers. This allows the model to extract local texture along the time axis, learn long-term dependencies, and focus on key process windows before outputting radial and axial dimension predictions. The one-dimensional convolutional layer performs local convolution operations on the periodic feature matrix along a unified time axis. The convolution kernel slides along the time axis and weights and converges multiple feature channels, thus forming a sequence of local process texture features. Let the periodic feature matrix be... The convolutional kernel length is L, and the weight of the m-th convolutional kernel is... , bias is Then the convolution output at the k-th time position can be written as:

[0081] in, This indicates the time position of the m-th convolutional kernel. The output; Indicates the convolution kernel at the th... Weight vectors at each time offset; This represents the three-dimensional eigenvector of the periodic feature matrix at time position ki; Indicates the inner product; This represents a nonlinear activation function used to enhance nonlinear representation capabilities. The gated recurrent unit layer receives a local process texture feature sequence and recursively generates a hidden state sequence along a unified time axis, enabling the hidden state sequence to encode the cumulative influence of historical time steps on the current state. Let the input of the gated recurrent unit layer be... Hidden state is Then the expressions for updating the gate, resetting the gate, and the candidate state are:

[0082] in, This represents the update gate vector, used to control the intensity of historical information writing; This represents the reset gate vector, used to control the intensity of historical information participating in candidate state calculation; Indicates the candidate hidden state; Represents the Sigmoid function; This indicates element-wise multiplication; and This represents a trainable weight matrix; This represents a trainable bias vector. The temporal attention layer assigns attention weights to the hidden state sequence and generates a context vector, enabling the model to assign higher weights to the hidden states of critical process windows during the output stage; let the hidden state sequence be... Then the expressions for attention scoring, normalization, and context vector are:

[0083] in, This represents the attention score at the k-th time step; represents the attention weights, which are normalized over all time steps; s represents the context vector. and Let represent the trainable parameters of the attention layer; u represents the trainable vector used for projection scoring. The output layer maps the context vector to radial and axial dimension predictions. The output layer uses linear mapping or multilayer perceptron mapping to adapt to different complexity requirements. Let the output layer weights be . , bias is ,in If the context vector dimension is used, then:

[0084] in, This represents the dimension prediction vector output by the model, with the first dimension being the radial dimension prediction value and the second dimension being the axial dimension prediction value.

[0085] Parameter initialization is performed after the network structure is determined, ensuring that the weights of each layer are in an optimizable state at the start of training and avoiding training failures caused by vanishing or exploding gradients. Parameter initialization employs a layered, consistent random initialization strategy, ensuring that the weight matrices of the one-dimensional convolutional layers, gated recurrent unit layers, temporal attention layers, and output layers meet preset variance scale constraints. The bias vectors are initialized with zero or small values ​​to achieve a stable start. The loss function is constructed using the target size label as the supervision target, simultaneously constraining the fitting errors of the radial and axial size predictions, enabling the model to learn a joint mapping that considers both size indices. For the nth sample, the size prediction vector is... The target size label is The weighted mean squared error loss can be written as:

[0086] in, This represents the training loss; N represents the number of samples in the training sample subset. and These represent the predicted radial dimension and the predicted axial dimension, respectively. and These represent the radial dimension measurement value and the axial dimension measurement value, respectively. and The weighting coefficients represent the two types of size error terms, used to balance the training objective under different accuracy requirements for the two size indices. The backpropagation algorithm calculates the gradient of the network parameters based on the loss in each training epoch and uses a gradient descent-type optimizer to iteratively update the parameters. This allows the output layer error to propagate layer by layer through the temporal attention layer, gated recurrent unit layer, and one-dimensional convolutional layer, driving the entire network to converge. Consequently, the radial and axial size predictions gradually approach the target size label.

[0087] After each training epoch, forward inference is performed on a subset of validation samples, and the validation loss is calculated. This ensures that convergence is determined based on the data distribution that did not participate in parameter updates, avoiding a situation where the training loss decreases but the generalization error increases. The validation loss uses the same loss function as the training loss, allowing for direct comparison and determining whether the model is overfitting. Preset convergence conditions can be defined by rules such as the validation loss decreasing below a threshold for several consecutive epochs, the validation loss failing to improve after reaching its historical best, or the difference between the validation loss and the training loss exceeding an upper limit. Training is stopped when any of these convergence conditions are met. Training also stops when a preset number of training epochs is reached to limit training costs. The network parameters corresponding to the epoch with the minimum validation loss are then fixed as the trained temporal prediction model. This allows subsequent online loading of the model to use the same unified time axis structure and periodic feature matrix structure for inference, outputting radial and axial dimension predictions.

[0088] Furthermore, based on the processed data, the forging dimensions are predicted and real-time quality monitoring is performed. Specifically, this includes: constructing a periodic feature matrix covering a single forging cycle from the processed data under a unified time axis; inputting the periodic feature matrix into a pre-trained temporal prediction model, sequentially performing one-dimensional convolutional feature extraction to obtain local process texture feature sequences, gated recurrent unit modeling to obtain hidden state sequences, and temporal attention weighting to obtain context vectors; generating key dimension prediction values ​​corresponding to a single forging cycle through the output layer, including radial and axial dimension prediction values; writing the key dimension prediction values ​​into display variables bound to the three-dimensional digital model; performing online consistency judgment between the key dimension prediction values ​​and the preset tolerance zone to generate a quality status identifier; generating an alarm signal when the quality status identifier indicates that any key dimension prediction value exceeds the preset tolerance zone; constructing a quality risk index based on the key dimension prediction values, the preset tolerance zone, the temporal attention weights, and the processed data; performing hierarchical monitoring based on the quality risk index, preset warning thresholds, and alarm thresholds, marking the state where the quality risk index exceeds the warning threshold but does not exceed the alarm threshold as a warning state, and marking the state where the quality risk index exceeds the alarm threshold as an alarm state.

[0089] Specifically, the data processing has already undergone governance, standardization, and fusion on a unified timeline in the preceding steps. Therefore, the processed data is organized into a periodic feature matrix using the unified timeline as the unique index. This ensures that the periodic feature matrix fully covers a single forging cycle, and each time step contains standardized expressions of pressure parameters, slide speed parameters, and billet temperature parameters at the same time position. The construction of the periodic feature matrix first determines the start and end indices of a single forging cycle. Within this index range, the multidimensional feature vectors corresponding to the processed data are read in chronological order along the unified timeline. These multidimensional feature vectors are then stacked row-wise to obtain the matrix structure. When there are a small number of boundary padding values ​​due to window boundary processing, these padding values ​​are kept in place in the matrix to maintain a constant time step length, ensuring stable input dimensions for subsequent models. Let the number of time steps on the unified timeline be K, and the multidimensional feature vector of the k-th time step be... Then the periodic characteristic matrix can be written as:

[0090] Where X is a periodic feature matrix, the row index corresponds one-to-one with the time index of the unified time axis, and the column index corresponds one-to-one with the fixed feature order, ensuring that subsequent feature extraction is performed on the same semantic structure.

[0091] After the periodic feature matrix is ​​input into the pre-trained temporal prediction model, a one-dimensional convolutional layer first performs local convolution operations along a unified time axis to extract short-term coupled textures. Then, a gated recurrent unit layer recursively updates the convolutional output sequence to encode long-term dependencies. Subsequently, a temporal attention layer calculates attention weights on the hidden state sequence and weights them to generate a context vector, highlighting the model's contribution to key process windows in the output stage. Let the periodic feature matrix be... The local process texture feature sequence output by the one-dimensional convolutional layer is The hidden state sequence output by the gated recurrent unit layer is The attention weights output by the temporal attention layer are: If the context vector is s, then the attention-weighted convergence can be written as:

[0092] in, Normalization is satisfied for all time steps, so that s can be numerically interpreted as a weighted combination of hidden state sequences; local process texture feature sequences are used to capture the local morphology at the moment of mold closing and the rapid pressurization stage; hidden state sequences are used to accumulate the influence of historical process evolution on the current forming state; and context vectors are used to form a compact representation for size regression at the full cycle scale.

[0093] The output layer receives the context vector and generates key dimension predictions, ensuring that the radial and axial dimension predictions for the same forging cycle are output synchronously in a single forward inference, thus avoiding inconsistencies introduced by separate modeling. Let the output layer parameters be... and ,in If the context vector dimension is used, then the key size prediction value is written as:

[0094] in, This is the predicted radial dimension. These are the predicted axial dimensions, and both are linked to this single forging cycle and enter the subsequent consistency judgment and risk assessment chain.

[0095] When writing critical dimension prediction values ​​into display variables bound to the 3D digital model, display variables corresponding one-to-one with the radial and axial dimension prediction values ​​are first defined within the 3D digital model. Reference relationships are then established between these display variables and interface controls, dimension annotation objects, or curve display objects, enabling updates to the display variables to drive the visualization refresh of the 3D digital model. The writing process uses a single forging cycle identifier as an index, binding the critical dimension prediction value to the corresponding 3D digital model instance for that cycle. The critical dimension prediction value is cached as a traceable display record, allowing the interface to simultaneously display the current prediction value and historical prediction trajectories. When the critical process window detection mechanism outputs a window identifier, the corresponding time window can be highlighted synchronously in the 3D digital model to associate prediction fluctuations with the process window display.

[0096] Online consistency assessment generates a quality status identifier by comparing the predicted values ​​of key dimensions with preset tolerance zones in real time. This identifier directly indicates whether the dimensional requirements of a single forging cycle are met. The preset tolerance zones are configured with upper and lower tolerance thresholds for radial and axial dimensions, respectively. During consistency assessment, both types of predicted dimensional values ​​are checked simultaneously to ensure they fall within their respective tolerance ranges. If any predicted dimensional value exceeds its corresponding tolerance range, the quality status identifier is set to an unqualified state; otherwise, it is set to a qualified state. Let the radial dimension tolerance zone be... The axial dimension tolerance zone is Then the consistency criterion can be written as:

[0097] in, For consistency judgment results, a value of 1 indicates pass, and a value of 0 indicates fail; For indicator functions; and These are the lower and upper tolerance thresholds for the radial dimension, respectively. and These are the lower and upper tolerance thresholds for the axial dimension, respectively. The quality status identifier is mapped from the consistency judgment result and written into the event log along with the single forging cycle identifier for traceability.

[0098] An alarm signal is generated when the quality status indicator indicates that any predicted value of a critical dimension exceeds the preset tolerance zone. The alarm signal is triggered in an event-driven manner and is linked to the alarm display variables of the 3D digital model, so that the alarm is displayed synchronously on the interface and peripheral devices. When the alarm signal is generated, the trigger time index, trigger dimension type, trigger value and corresponding tolerance boundary are recorded, and the window identifier and attention weight distribution are written into the alarm record, so that the process window that the model is concerned with when the alarm occurs can be located later. The alarm signal can further trigger audio-visual equipment or a pop-up window on the host computer, and the alarm status is maintained until manual confirmation or automatic reset when the next forging cycle begins, so as to meet the on-site handling procedures.

[0099] The quality risk index is used to assess the risks of approach deviation and process drift within the preset tolerance zone. It quantifies the critical dimension predictions, preset tolerance zones, temporal attention weights, and processed data by linking them along the same unified time axis. The quality risk index first calculates the dimensional approximation degree, converting the radial and axial dimension predictions into normalized deviations relative to the tolerance zone center, making different dimensional indicators comparable on the same risk scale. Next, it calculates the process drift degree, using temporal attention weights to weight and aggregate the time step deviations of the processed data, making drift within the critical process window contribute more significantly to the risk. Finally, it weights and fuses the two risk components and maps them to the 0-1 interval using a monotonic compression function. Let the radial dimension tolerance zone center be... The center of the axial dimension tolerance zone is Half bandwidth is , The multidimensional feature vector of the processed data at time step k is The full-cycle reference mean vector is The quality risk index is then written as:

[0100] Wherein, R is the quality risk index, with a value ranging from 0 to 1; It is a monotonic compression function used to map the fusion value to 0 to 1 and enhance the differentiation of high-risk values; , , These are weighting coefficients used to balance the contributions of radial dimension approximation, axial dimension approximation, and process drift to risk; For temporal attention weights, normalization is satisfied for the entire cycle; The Euclidean norm is used to measure the performance of processed data at time steps. The deviation from the full-cycle reference mean vector is considered. This construction ensures that risk is determined not only by the degree of dimensional approximation but also by the degree of process drift within the critical process window, thus exposing risk trends in advance before dimensions exceed tolerances.

[0101] The tiered monitoring system uses a quality risk index as a continuous risk quantity and preset early warning and alarm thresholds as discrete tier boundaries, enabling the system to adopt different response strategies at different risk levels. Let the early warning threshold be... The alarm threshold is And satisfy The classification judgment is then written as:

[0102] The system is divided into three levels: Normal status, Warning status, and Alarm status. Normal status indicates a low level of risk and maintains a consistent display. Warning status indicates significant drift or size approaching but has not yet reached the alarm threshold. Alarm status indicates the risk has reached a level requiring intervention and can be triggered in parallel with the out-of-tolerance alarm of the quality status indicator, forming a monitoring closed loop with both out-of-tolerance and high-risk alarms. The classification results, along with the single forging cycle identifier, window identifier, critical dimension prediction value, and attention weight statistics, are written into the event log, enabling consistent tracing of the triggering reasons for warning and alarm statuses and supporting iterative updates to threshold calibration.

[0103] Reference Figure 2 The present invention discloses a digital twin detection system for aluminum alloy forging process, comprising three continuous links. The first link is used to constrain the stable formation of traceable time-series records of multi-source physical parameters in the same time reference system within a single forging cycle. The second link is used to convert the time-series records into processing data that can be directly consumed by the model. The third link is used to send the processing data into the time-series prediction model to obtain the predicted value of the key dimension, and after performing online consistency judgment between the predicted value and the preset tolerance zone, it is diverted to different quality handling paths.

[0104] The first row corresponds to the acquisition-side link driven by a unified time axis. A single forging cycle is first determined as a periodic time domain with start and end boundaries. Within this periodic time domain, the unified time axis is discretized into a set of time indices at fixed time steps, so that all subsequent data are organized using this time index as the primary key. The acquisition timestamps of pressure parameters, slider speed parameters, and billet temperature parameters are uniformly converted into standard timestamps. The standard timestamps correspond one-to-one with the time indices of the unified time axis, thereby eliminating the systematic offset caused by clock differences and buffer delays in different sampling links. Subsequently, joint threshold discrimination is performed on the first derivative of the pressure parameter, the second derivative of the slider speed parameter, and the rate of change of the billet temperature parameter within the standard timestamp domain to identify short-term strong nonlinear segments such as the instant of mold closing, the instant of contact, and the rapid pressure increase section. A window identifier is generated for each short-term strong nonlinear segment, so that subsequent phase correction only occurs within the interval covered by the window identifier. Under window identifier constraints, cross-correlation analysis is performed on pressure parameters, slider speed parameters, and billet temperature parameters to obtain the optimal time delay between each parameter. The optimal time delay is used to characterize the relative arrival time difference of the same physical event on different sampling links. Subsequently, phase compensation translation processing is performed on the corresponding parameters based on the optimal time delay to generate a phase correction data set, so that the values ​​of the three types of parameters at the same time index within the critical process window correspond to the same physical moment as much as possible, providing a synchronized joint state for subsequent feature splicing. Based on the phase correction data set, unified time axis resampling processing is performed to form a second time series data set that maintains phase consistency inside and outside the critical process window, thereby simultaneously solidifying the phase consistency inside the window and the time consistency outside the window into the same time series structure.

[0105] The second row corresponds to the data governance, standardization, and fusion process. The second time-series data set is first statistically analyzed within a sliding time window on a unified time axis. The sliding time window continuously slides along the time index, calculating the local mean and local fluctuation amplitude of pressure parameters, slider speed parameters, and billet temperature parameters within the window. Sampling points deviating from the local background are marked with a preset fluctuation threshold, thus obtaining an abnormal sampling point set. The abnormal sampling point set triggers neighborhood-weighted interpolation replacement. The replacement value is generated by weighting the effective sampling points before and after the abnormal sampling point according to the time distance, ensuring that the abnormal replacement remains stable in terms of boundary continuity and local trend consistency. Even after removing the abnormal time-series data set, there may still be continuous missing segments. Therefore, missing data segments are identified, and linear interpolation functions and spline smoothing functions are constructed based on the effective time segments before and after the missing segments. Joint interpolation is performed on the missing segments to obtain a continuous time-series data set, ensuring that each time index within the entire range of the unified time axis has effective values ​​for the three types of parameters. The full-cycle mean and standard deviation of pressure, slider speed, and billet temperature parameters are calculated on a continuous time-series dataset. A zero-mean, unit variance transformation is performed on each time index to obtain a standardized time-series feature set, ensuring that the three types of parameters are comparable in numerical scale and consistent with the input distribution during training. Subsequently, within the same time step, the standardized value vectors of pressure, slider speed, and billet temperature parameters are concatenated into a multi-dimensional feature vector according to a fixed feature order. These vectors are then stacked in a time-ordered manner along a unified time axis to form a periodic feature matrix. The rows of the periodic feature matrix strictly correspond to the time indices, and the columns strictly correspond to the fixed feature order. This periodic feature matrix serves as the unique input structure for the processed data on the model side.

[0106] The third row corresponds to the intelligent prediction and quality monitoring link. After the periodic feature matrix is ​​input into the pre-trained time-series prediction model, a one-dimensional convolutional layer first extracts local process texture feature sequences along the time axis, explicitly representing the coupling pattern within a short time window. Then, a gated recurrent unit layer performs long-term dependency modeling on the local process texture feature sequences to generate hidden state sequences, encoding the evolution history of the entire cycle into the recursive state. Subsequently, a time-series attention layer assigns attention weights to the hidden state sequences and weights them to form a context vector, highlighting the contribution of key process windows to the output. The output layer receives the context vector and generates key dimension prediction values, including radial and axial dimension prediction values. After the key dimension prediction values ​​are written into the display variables bound to the 3D digital model, the 3D digital model synchronously presents the prediction results on the interface and maintains binding with the corresponding single forging cycle identifier, so that the prediction trajectory and process curve are displayed together on the same time axis. Then, the system enters the discrimination node, where online consistency discrimination is performed between the predicted value of the critical dimension and the preset tolerance zone. The preset tolerance zone is configured with upper and lower tolerance thresholds for radial and axial dimensions, respectively. The result of the consistency discrimination is solidified as a quality status indicator. When the quality status indicator indicates that any predicted value of the critical dimension exceeds the preset tolerance zone, an alarm signal is generated and the system enters the alarm handling branch on the right. When the quality status indicator indicates that the predicted value of the critical dimension does not exceed the preset tolerance zone, the system enters the non-alarm branch on the right and maintains normal monitoring.

[0107] To enable the discrimination node to identify not only out-of-tolerance situations but also near-out-of-tolerance situations and process drift, a quality risk index is constructed in parallel within the same link and hierarchical monitoring is performed. The quality risk index integrates the normalized deviation of the predicted critical dimension value relative to the preset tolerance zone center with the deviation of the processed data under attention weighting into a single risk quantity. This risk quantity is then mapped to a warning or alarm state through preset warning and alarm thresholds. Based on the comparison results between the risk quantity and the preset warning and alarm thresholds, a warning state is marked when the risk quantity exceeds the warning threshold but does not exceed the alarm threshold, and an alarm state is marked when the risk quantity exceeds the alarm threshold. This allows for early triggering of handling actions even in non-out-of-tolerance scenarios, and together with the out-of-tolerance alarm from the discrimination node, it forms a closed-loop link of prediction, discrimination, and handling.

[0108] This embodiment also discloses a digital twin monitoring device for aluminum alloy forging processes, referring to... Figure 3 The device includes an acquisition module 301, a processing module 302, and an output module 303. It is used to execute any of the above-described digital twin monitoring methods for aluminum alloy forging processes, wherein: The acquisition module 301 is used to collect physical parameters of the forging process in real time; Processing module 302 is used to manage, standardize, and fuse physical parameters to obtain processed data; Output module 303 is used to predict forging dimensions and perform real-time quality monitoring based on processed data.

[0109] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0110] This embodiment also discloses an electronic device, as shown in the reference. Figure 4 The electronic device may include: at least one processor 401, at least one communication bus 402, user interface 403, network interface 404, and at least one memory 405.

[0111] The communication bus 402 is used to enable communication between these components.

[0112] The user interface 403 may include a display screen and a camera. Optionally, the user interface 403 may also include a standard wired interface and a wireless interface.

[0113] The network interface 404 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0114] The processor 401 may include one or more processing cores. The processor 401 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 405, and by calling data stored in memory 405. Optionally, the processor 401 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 401 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 401.

[0115] The memory 405 may include random access memory (RAM) or read-only memory. Optionally, the memory may include a non-transitory computer-readable storage medium. The memory 405 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 405 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 405 may also be at least one storage device located remotely from the aforementioned processor 401. As a computer storage medium, the memory 405 may include an operating system, a network communication module, a user interface 403 module, and an application program for a digital twin monitoring method for aluminum alloy forging processes.

[0116] exist Figure 4In the electronic device shown, the user interface 403 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 401 can be used to call the application program of a digital twin monitoring method for aluminum alloy forging process stored in the memory 405. When executed by one or more processors 401, the electronic device performs one or more methods as described in the above embodiments.

[0117] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0118] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0119] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0122] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 405 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned memory 405 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0123] The present invention also discloses a non-transitory computer-readable storage medium storing instructions. When executed by one or more processors 401, these instructions cause an electronic device to perform one or more methods as described in the above embodiments.

[0124] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truths. This invention is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A digital twin monitoring method for aluminum alloy forging process, characterized in that, The method includes: Real-time acquisition of physical parameters during the forging process; The physical parameters are processed, standardized, and fused to obtain processed data; Based on the processed data, the dimensions of the forgings are predicted and real-time quality monitoring is performed.

2. The digital twin monitoring method for aluminum alloy forging process according to claim 1, characterized in that, The process of managing, standardizing, and fusing the physical parameters to obtain processed data specifically includes: Establish a unified timeline around a single forging cycle; The pressure parameter, slider speed parameter, and billet temperature parameter in the physical parameters are mapped to the unified time axis according to the acquisition timestamp to form a time-aligned mapping relationship; Based on the time alignment mapping relationship, resampling processing is performed on physical parameters with different sampling frequencies to generate a first time-series data set; Construct a sliding time window on the first time series data set; Within each sliding time window, calculate the local mean and local fluctuation amplitude of pressure parameters, slider speed parameters, and billet temperature parameters, and mark sampling points that exceed the preset fluctuation threshold as abnormal sampling points. Neighborhood weighted interpolation is performed using valid sampling points at adjacent time locations of the abnormal sampling points to replace the abnormal sampling points, thereby obtaining a set of time-series data without abnormalities.

3. The digital twin monitoring method for aluminum alloy forging process according to claim 2, characterized in that, The process of managing, standardizing, and fusing the physical parameters to obtain processed data specifically includes: Identify data gap segments in consecutive time segments on the anomaly-free time series data set; Based on the adjacent valid time segments before and after the data gap segment, a linear interpolation function and a spline smoothing function are constructed, and the data gap segment is subjected to joint interpolation processing to obtain a continuous time series data set; Based on the continuous time series data set, the mean and standard deviation of the pressure parameter, the slider speed parameter, and the billet temperature parameter throughout the complete forging cycle are calculated respectively. Based on the full-cycle mean and the full-cycle standard deviation, the parameters of each time step are converted into a standardized expression with zero mean and unit variance, forming a standardized time series feature set.

4. The digital twin monitoring method for aluminum alloy forging process according to claim 3, characterized in that, The process of managing, standardizing, and fusing the physical parameters to obtain processed data specifically includes: The pressure parameters, slider speed parameters, and billet temperature parameters within the same time step in the standardized time series feature set are vectorized according to a preset fixed feature order to form a multidimensional feature vector. The multidimensional feature vectors are stacked according to the time order of the unified time axis to construct a periodic feature matrix, thereby obtaining the processed data.

5. The digital twin monitoring method for aluminum alloy forging process according to claim 1, characterized in that, The process of predicting forging dimensions and performing real-time quality monitoring based on the processed data specifically includes: The processed data is constructed into a periodic feature matrix covering a single forging cycle under a unified time axis; The periodic feature matrix is ​​input into a pre-trained temporal prediction model, and one-dimensional convolutional feature extraction is performed sequentially to obtain a local process texture feature sequence, gated recurrent unit modeling is performed to obtain a hidden state sequence, and temporal attention weighting is performed to obtain a context vector. The output layer generates key dimension prediction values ​​corresponding to the single forging cycle, including radial dimension prediction values ​​and axial dimension prediction values. The predicted key dimensions are written into the display variables that are bound to the 3D digital model; The predicted key dimensions are compared with the preset tolerance zone online to determine consistency and generate a quality status identifier. An alarm signal is generated when the quality status indicator indicates that any predicted value of a critical dimension exceeds the preset tolerance zone. A quality risk index is constructed based on the predicted key dimensions, the preset tolerance zone, the temporal attention weight, and the processed data. Based on the quality risk index and preset warning thresholds and alarm thresholds, hierarchical monitoring is performed. The state in which the quality risk index exceeds the warning threshold but does not exceed the alarm threshold is marked as a warning state, and the state in which the quality risk index exceeds the alarm threshold is marked as an alarm state.

6. The digital twin monitoring method for aluminum alloy forging process according to claim 5, characterized in that, Before predicting forging dimensions based on the processed data and performing real-time quality monitoring, the method further includes: A sample set is established around historical forging samples, and the processing data corresponding to each historical forging sample is organized into a periodic feature matrix according to a unified time axis; The actual detected size value corresponding to the periodic feature matrix is ​​determined as the target size label; The sample set is partitioned using a single forging cycle as the smallest unit of division to generate a training sample subset and a validation sample subset. A network structure is constructed based on the training sample subset. The network structure sequentially includes a one-dimensional convolutional layer, a gated recurrent unit layer, a temporal attention layer, and an output layer. The one-dimensional convolutional layer is used to perform local convolution operations on the periodic feature matrix on the unified time axis to obtain a local process texture feature sequence. The gated recurrent unit layer is used to perform long-term dependency modeling on the local process texture feature sequence to generate a hidden state sequence. The temporal attention layer is used to assign attention weights to the hidden state sequence to generate a context vector. The output layer is used to map the context vector into radial dimension prediction values ​​and axial dimension prediction values. The parameters are initialized and a loss function is constructed on the training sample subset. Based on the target size label, the network parameters are iteratively updated through the backpropagation algorithm so that the radial size prediction value and the axial size prediction value gradually approach the target size label. After each training round, the validation loss is calculated using the validation sample subset and it is determined whether the preset convergence condition is met. When the preset convergence condition is met or the preset number of training rounds is reached, training is stopped and the trained time series prediction model is obtained.

7. The digital twin monitoring method for aluminum alloy forging process according to claim 2, characterized in that, The process of managing, standardizing, and fusing the physical parameters to obtain processed data specifically includes: Based on the unified time axis, the acquisition timestamps of pressure parameters, slider speed parameters, and billet temperature parameters are uniformly converted into standard timestamps; A key process window detection mechanism is constructed based on the standard timestamp. The key process window is identified by joint threshold discrimination of the first derivative of the pressure parameter, the second derivative of the slider speed parameter, and the rate of change of the billet temperature parameter. Generate a window identifier for each of the aforementioned key process windows; Under the window identifier constraint, cross-correlation analysis is performed on the pressure parameter, the slider speed parameter, and the billet temperature parameter respectively to calculate the optimal time delay between each parameter; Based on the optimal time delay, phase compensation translation processing is performed on the pressure parameter, the slider speed parameter, and the billet temperature parameter to generate a phase correction data set; Based on the phase correction data set, a unified time axis resampling process is performed to form a second time series data set that maintains phase consistency both inside and outside the critical process window. The second time-series data set is processed to obtain the processed data.

8. A digital twin monitoring device for aluminum alloy forging process, characterized in that, The device is used to execute a digital twin monitoring method for aluminum alloy forging process as described in any one of claims 1-7. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to collect physical parameters of the forging process in real time; The processing module is used to manage, standardize, and fuse the physical parameters to obtain processed data; The output module is used to predict the size of the forging based on the processed data and to perform real-time quality monitoring.

9. An electronic device, characterized in that, The device includes a processor, a communication bus, a user interface, a network interface, and a memory. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The communication bus is used to enable communication between the components within the electronic device. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images