Cold forging damage prediction method based on bidirectional LSTM

Abstract

The invention discloses a cold forging damage prediction method based on bidirectional LSTM, and relates to the field of equipment health management, and the method comprises the following steps: collecting real-time stroke data under a single stroke in real time based on an Internet of Things sensor, and forming original multi-modal time sequence data; according to the characteristics of different physical field signals, the damage sensitive characteristics of a single physical field are accurately extracted, meanwhile, based on cross-physical field triggering characteristics, the collaborative response relation between multiple physical fields when damage occurs is captured from the mechanism level, and the characteristic construction method deeply fuses multi-source heterogeneous information and highlights the coupling relation of the multi-source heterogeneous information, so that the multi-source heterogeneous information can be quickly and accurately analyzed. According to the method, the characterization capability of the feature sequence to early and weak damage is enhanced, and the model can accurately mine the damage precursor from complex multi-physics field coupling features in combination with the learning and key information focusing capability of a subsequent time sequence deep learning model, so that high-sensitivity and high-accuracy prediction can be realized in the early stage of damage initiation.

Description

Technical Field

[0001] This invention relates to the field of equipment health management, and in particular to a method for predicting cold forging damage based on bidirectional LSTM. Background Technology

[0002] Cold forging is a plastic forming process performed at or near room temperature. It is characterized by high load capacity, large deformation, extremely short deformation time, and the absence of significant temperature field changes. Unlike hot forging or casting, the die in cold forging primarily withstands high contact pressure, wear caused by strong friction, and mechanical fatigue due to cyclic loading, rather than failure mechanisms caused by temperature gradient changes, such as thermal fatigue.

[0003] Currently, existing monitoring methods mostly rely on single physical quantity signals such as extrusion pressure and displacement, which are difficult to capture the weak, cross-physical field coupling characteristics in the early stages of damage, resulting in delayed early warnings and high rates of missed and false alarms. Therefore, this paper proposes a cold forging damage prediction method based on bidirectional LSTM to solve the above problems. Summary of the Invention

[0004] The main objective of this invention is to provide a cold forging damage prediction method based on bidirectional LSTM to solve the problems mentioned in the background above.

[0005] To achieve the above objectives, the technical solution adopted by this invention is as follows: a cold forging damage prediction method based on bidirectional LSTM, comprising the following steps:

[0006] Step 1: Real-time stroke data for a single stroke is collected using IoT sensors to generate raw multimodal time-series data;

[0007] Step 2: Preprocess and extract features from the original multimodal time series data to construct a multiphysics coupling feature sequence of uniform length. The multiphysics coupling feature sequence includes cross-physics triggering features extracted from the original multimodal time series data.

[0008] Step 3: Pair the multiphysics coupling feature sequence with the workpiece damage level label to form a training dataset, and train the time series deep learning model using the training dataset;

[0009] Step 4: Based on the time-series deep learning model, collect multimodal time-series data under real-time stroke and construct a real-time multiphysics coupling feature sequence. Input the real-time multiphysics coupling feature sequence into the model and output the damage prediction result.

[0010] Step 5: Based on the prediction results, execute the control action.

[0011] Preferably, in step one, the real-time stroke data collected in a single stroke based on IoT sensors includes the extrusion pressure signal collected in real time by the force sensor deployed on the cold forging equipment, the punch displacement signal collected by the displacement sensor, the acoustic emission signal collected by the acoustic emission sensor, and the die strain signal collected by the die strain sensor. Before collection, the sampling frequency of the IoT sensors is unified, and the timestamps of all sensors are synchronized by a hardware clock.

[0012] Preferably, in step two, the original multimodal time series data is preprocessed, including denoising the acoustic emission signal using wavelet threshold denoising, dynamically filtering the extrusion pressure signal using an adaptive Kalman filter combined with a cold forging process model, smoothing the die strain signal using a 5th-order low-pass digital filter, and smoothing the punch displacement signal using a moving average filter with a window length of 10.

[0013] Simultaneously, the pre-processed extrusion pressure signal is used as a reference signal to calculate the extreme points of its second derivative of curvature, identify and divide the start and end time points of the elastic deformation, plastic deformation and holding pressure stages of the cold forging stroke, and use these time points as a reference to synchronously divide all other physical field signals into stages and perform time normalization processing. For the key stages of plastic deformation and holding pressure stages, a time resolution higher than that of elastic deformation is used for resampling.

[0014] Preferably, in step two, feature extraction of the original multimodal time series data specifically includes:

[0015] S1: Perform waveform recognition on the acoustic emission signal, extract the Burst-type acoustic emission events related to the microcrack, and calculate their time-spectrum characteristics;

[0016] S2: For the extrusion force signal and the punch displacement signal, calculate the higher-order morphological features of the force and displacement curves, including the local stiffness change rate and curve fullness.

[0017] S3: For the strain signal of the mold, calculate the spatial gradient characteristics of the strain distribution.

[0018] Preferably, step two further includes calculating the cross-physical field triggering features based on the features extracted by S1, S2 and S3. This involves calculating the synergistic index of the extrusion signal drop and the acoustic emission energy mutation, as well as the hysteresis anomaly index of the mold strain change during the stable stage of the punch displacement, within a set time window. Then, the single-physical field features and cross-physical field features are spliced ​​together and standardized to the [0,1] interval to form a multi-physical field coupling feature sequence of uniform length.

[0019] Preferably, in step three, the multiphysics coupling feature sequence is paired with the workpiece damage level label to form a training dataset, as follows:

[0020] Offline acquisition of multi-physics field coupling feature sequences of one complete stroke for each of M cold forging workpieces (M≥1000), industrial CT inspection of each workpiece, quantitative calculation of internal microcrack feature parameters of the workpiece based on CT images, and mapping of the parameters to a discrete damage level label according to process standard threshold.

[0021] The multiphysics coupling feature sequence of each workpiece is paired one-to-one with its own damage level label to form a training dataset.

[0022] Preferably, in step three, the temporal deep learning model is a bidirectional LSTM model that incorporates an attention mechanism.

[0023] Preferably, in step three, the training includes: inputting the multi-physics coupled feature sequence into a bidirectional LSTM layer to learn forward and backward temporal dependencies, assigning weights to different time steps through an attention mechanism layer to focus on key time periods, using damage level labels as supervision, minimizing the classification loss function through a backpropagation algorithm, and optimizing model parameters.

[0024] Preferably, the specific steps are as follows: during the real-time stroke, a real-time multi-physics coupling feature sequence is constructed according to the methods of step one and step two. The constructed real-time multi-physics coupling feature sequence is input into the trained temporal deep learning model, and the model is calculated through its forward propagation network to output a real-time prediction result. The real-time prediction result represents the probability distribution vector of the workpiece belonging to each predefined damage level.

[0025] Preferably, in step five, the control action is a hierarchical strategy:

[0026] If the predicted damage level is a low warning level, then the warning will be triggered and the equipment will be deloaded.

[0027] If the predicted damage level is a high-level failure level, a stop command is triggered and a mold inspection is prompted.

[0028] The threshold for classifying low-level warning and high-level fault is determined by analyzing the receiver's operational characteristic curves of model prediction probability and actual damage state in historical data to maximize classification performance indicators.

[0029] The present invention has the following beneficial effects:

[0030] 1. In this invention, damage-sensitive features of a single physical field are accurately extracted based on the characteristics of different physical field signals. At the same time, based on cross-physical field triggering features, the collaborative response relationship between multiple physical fields during damage is captured from a mechanistic perspective. This feature construction method, which deeply integrates multi-source heterogeneous information and highlights their coupling relationship, enhances the ability of feature sequences to represent early and weak damage. Combined with the learning and key information focusing capabilities of subsequent time-series deep learning models, the model can accurately mine damage precursors from complex multi-physical field coupling features, thereby achieving highly sensitive and accurate prediction in the early stages of damage initiation. This effectively overcomes the problems of insufficient sensitivity and high false alarm / missed alarm rates of traditional single-signal monitoring, and significantly advances the quality control node.

[0031] 2. In this invention, by constructing a multimodal sensing network covering force, sound, displacement, and strain, and achieving strict time synchronization based on a hardware clock, highly consistent raw multimodal time-series data is obtained. This effectively solves the problem of time-series misalignment in traditional multi-sensor data, providing a precise and aligned basic data source for subsequent staged preprocessing and cross-physical field feature extraction. It avoids feature coupling deviations caused by data asynchrony and ensures the accuracy of damage monitoring from the acquisition stage.

[0032] 3. In this invention, differentiated actions are performed based on the damage prediction level: a load reduction warning is triggered when a low-level warning is issued, which can intervene in the early stage of damage to delay damage to the mold and workpiece and reduce unnecessary downtime; when a high-level fault occurs, the machine is stopped immediately and the mold is prompted for inspection, which can avoid equipment damage and safety accidents. Moreover, the level threshold is determined by the receiver's operating characteristic curve, making the control strategy more precise, realizing a dynamic balance between production safety and production efficiency, and improving the intelligent management and control level of cold forging production. Attached Figure Description

[0033] Figure 1 This is a flowchart of a cold forging damage prediction method based on bidirectional LSTM according to the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Please see Figure 1 This invention provides a technical solution: a cold forging damage prediction method based on bidirectional LSTM, comprising the following steps:

[0036] Step 1: Real-time stroke data for a single stroke is collected using IoT sensors to generate raw multimodal time-series data;

[0037] Step 2: Preprocess and extract features from the original multimodal time series data to construct a multiphysics coupling feature sequence of uniform length. The multiphysics coupling feature sequence contains cross-physics triggering features extracted from the original multimodal time series data.

[0038] Step 3: Pair the multiphysics coupling feature sequence with the workpiece damage level label to form a training dataset, and train the time series deep learning model using the training dataset;

[0039] Step 4: Based on the time-series deep learning model, collect multimodal time-series data under real-time stroke and construct a real-time multiphysics coupling feature sequence. Input the real-time multiphysics coupling feature sequence into the model and output the damage prediction result.

[0040] Step 5: Based on the prediction results, execute the control actions.

[0041] In step one, the real-time stroke data for a single stroke is collected in real time based on IoT sensors, including the extrusion pressure signal collected in real time by the force sensor deployed on the cold forging equipment, the punch displacement signal collected by the displacement sensor, the acoustic emission signal collected by the acoustic emission sensor, and the die strain signal collected by the die strain sensor. Before the collection, the sampling frequency of the IoT sensors is unified, and the timestamps of all sensors are synchronized by a hardware clock.

[0042] Specifically: the force sensor is installed on the punch drive end, the displacement sensor is coaxially installed on the punch guide shaft, the acoustic emission sensor is attached to the side wall of the mold cavity (≤50mm from the workpiece), the strain sensor array is arranged in the stress concentration area of ​​the mold, and the sampling frequency of all sensors is uniformly set to 1kHz-5kHz.

[0043] Through the above technical solutions, multimodal signals reflect the mechanical behavior, material deformation, microscopic damage, and mold state during the cold forging process from different perspectives, providing rich and complementary information for subsequent damage prediction. At the same time, by unifying the sampling frequency of IoT sensors and synchronizing the timestamps of all sensors using a hardware clock, the high consistency and precise alignment of all collected multimodal time-series data in the time dimension are ensured, avoiding analysis errors caused by data misalignment. This allows the coupling relationship between different physical fields to be accurately captured, thereby significantly improving the quality and reliability of the original multimodal time-series data. This lays a solid foundation for the subsequent construction of multiphysics coupling feature sequences and training of time-series deep learning models, thereby improving the accuracy and robustness of cold forging damage prediction.

[0044] In step two, the original multimodal time series data are preprocessed, including denoising the acoustic emission signal using wavelet threshold denoising, dynamically filtering the extrusion pressure signal using an adaptive Kalman filter combined with the cold forging process model, smoothing the die strain signal using a 5th-order low-pass digital filter, and smoothing the punch displacement signal using a moving average filter with a window length of 10.

[0045] Simultaneously, the pre-processed extrusion pressure signal is used as a reference signal to calculate the extreme points of its second derivative of curvature, identify and divide the start and end time points of the elastic deformation, plastic deformation and holding pressure stages of the cold forging stroke, and use these time points as a reference to synchronously divide all other physical field signals into stages and perform time normalization processing. For the key stages of plastic deformation and holding pressure stages, a time resolution higher than that of elastic deformation is used for resampling.

[0046] For example: the time resolution of the elastic deformation stage is set to 50ms (one set of data is collected every 50ms); the time resolution of the plastic deformation and pressure holding stages is set to 10ms (one set of data is collected every 10ms).

[0047] Time normalization processing: The sensor data sequences corresponding to different cold forging workpieces are uniformly mapped to a fixed time length (such as 200 time steps), and linear interpolation is used to fill the sequence to ensure consistent sequence length.

[0048] Specifically, the steps for calculating the extreme points of the second derivative of curvature are as follows:

[0049] The preprocessed extrusion pressure signal is fitted with a cubic spline to obtain a continuous function:

[0050] ;

[0051] In the formula, Represents a continuous function. , , and Representing the The cubic spline coefficients for each interpolation interval are obtained by solving for boundary conditions, such as the first derivative being 0 at both endpoints. Representing the The start and end times of each interpolation interval;

[0052] The curvature and its second derivative are calculated as follows:

[0053] First derivative of the extrusion pressure curve: ;

[0054] The second derivative of the extrusion pressure curve: ;

[0055] Curvature of the extrusion pressure curve: ;

[0056] The second derivative of curvature: ;

[0057] Extreme point extraction: for The sequence is used to extract points that satisfy the condition of "maximum / minimum value in neighborhood" and whose magnitude meets the standard as extreme points;

[0058] The elastic deformation, plastic deformation, and pressure holding stages are divided by the distribution of extreme points:

[0059] Elastic deformation: start and end time (T1-T2), where T1 is the stroke start time and T2 is the end time. The time corresponding to the first positive extreme point;

[0060] Plastic deformation stage: start and end time (T2-T3), T3 as The time corresponding to the last negative extreme point;

[0061] The start and end times of the pressure holding phase (T3-T4), where T4 is the stroke end time;

[0062] By performing refined preprocessing and stage division on the original multimodal time series data as described above, this application can effectively solve the problems of high noise in the original data, inconsistent timing of different physical field signals, and insufficient capture of key damage stage information. Based on the preprocessed extrusion pressure signal, by calculating the extreme points of its second derivative of curvature, the start and end times of the elastic deformation, plastic deformation, and holding pressure stages of the cold forging stroke can be accurately identified and divided. Based on this, all other physical field signals are synchronously divided into stages and time-normalized, ensuring high alignment and comparability of multi-physical field data in the time dimension. In particular, the time resolution of plastic deformation and holding pressure, which are prone to damage, is used for resampling, which is higher than that of elastic deformation, so that the model can capture the weak signals and dynamic changes related to damage evolution in these stages more meticulously.

[0063] Step two involves feature extraction from the original multimodal time series data, specifically including:

[0064] S1: Perform waveform recognition on the acoustic emission signal, extract the Burst-type acoustic emission events related to the microcrack, and calculate their time-spectrum characteristics;

[0065] Specifically: Set an amplitude threshold, which is 3 times the amplitude of the background noise. When the signal amplitude is greater than or equal to the amplitude threshold and the duration is greater than or equal to 5 milliseconds, it is determined to be a Burst type event.

[0066] For a single Burst-type event signal segment, the time-spectral characteristics are calculated using the Short Time Fourier Transform (STFT).

[0067] And extract core features from the time spectrum, including:

[0068] Center frequency: ,in, Represents the spectral characteristics during short-time Fourier transform calculation;

[0069] Peak frequency: ;

[0070] Bandwidth: ;in, and These represent the frequency values ​​corresponding to 10% and 90% of the cumulative energy in the acoustic emission signal's spectrum, respectively.

[0071] S2: Calculate the higher-order morphological features of the force and displacement curves for the extrusion pressure signal and the punch displacement signal. The higher-order morphological features include the local stiffness change rate and the curve fullness.

[0072] The formula for calculating the rate of change of local stiffness is as follows:

[0073] ;

[0074] In the formula, Represents the rate of change of local stiffness. Represents a time window, The average stiffness over the entire stroke, and the local stiffness. , Represents extrusion pressure. Represents the displacement of the punch;

[0075] The formula for calculating curve fullness is:

[0076] ;

[0077] In the formula, Represents the fullness of the curve. This represents the area under the compressive stress versus displacement curve. Represents the maximum compressive force. This represents the maximum displacement of the punch.

[0078] S3: For the mold strain signal, calculate the spatial gradient characteristics of the strain distribution;

[0079] The formula for spatial gradient features is:

[0080] ;

[0081] In the formula, Representing the The sensor and the first The strain space gradient of +1 sensor, Representing the The sensor and the first +1 physical spacing between sensors, The first in the mold strain sensor array The strain values ​​collected by each sensor are then extracted. The maximum and mean gradient values ​​are used as features.

[0082] Based on the characteristics of different physical field signals during cold forging, waveform recognition of acoustic emission signals and extraction of time-spectrum features of Burst-type events related to microcracks can accurately capture the initiation and propagation information of micro-damage inside the material. By calculating the higher-order morphological features of extrusion pressure signals and punch displacement signals, such as the rate of change of local stiffness and curve fullness, abnormal mechanical responses during macroscopic plastic deformation can be sensitively reflected. The spatial gradient features of strain distribution of die strain signals can be calculated to effectively monitor local stress concentration and potential fatigue damage in the die. These refined feature extraction methods designed for the characteristics of different physical field signals can more comprehensively and accurately characterize the damage precursor information in the cold forging process, and significantly improve the ability of multi-physics coupled feature sequences to characterize the damage state of workpieces.

[0083] Step two also includes calculating cross-physical field triggering features based on the features extracted from S1, S2 and S3. Within a set time window, the synergistic index of the extrusion signal drop and the acoustic emission energy mutation, as well as the hysteresis anomaly index of the mold strain change during the stable stage of the punch displacement, are calculated. Then, the single-physical field features and cross-physical field features are spliced ​​together and standardized to the [0,1] interval to form a multi-physical field coupling feature sequence of uniform length.

[0084] The formula for calculating the synergistic index between the pressure signal drop and the acoustic emission energy mutation is as follows:

[0085] ;

[0086] In the formula, Representing time The synergy index, represent The constant pressure Represents the average extrusion pressure. represent Constant sound emission of energy Represents the average emitted energy;

[0087] The formula for calculating the hysteresis anomaly index of die strain change during the stable phase of punch displacement is:

[0088] ;

[0089] In the formula, represent The time lag anomaly index, Represents a lag time window. Represents the mean strain. The normal strain variation threshold value represents the process calibration.

[0090] By introducing cross-physics triggering features, the complex synergistic effects and causal relationships between different physical phenomena during cold forging are effectively captured. For example, the synergistic index of extrusion pressure signal drop and acoustic emission energy mutation can accurately indicate the instantaneous correlation between mechanical load change and energy release when damage occurs, while the hysteresis anomaly index of die strain change during the stable stage of punch displacement reveals the abnormal response of material or die under continuous load. By splicing and standardizing these cross-physics features with strong damage indicativeness with single-physics features, a more information-rich and comprehensive multi-physics coupling feature sequence is constructed.

[0091] In step three, the multiphysics coupling feature sequences are paired with workpiece damage level labels to form a training dataset, as detailed below:

[0092] Offline acquisition of multi-physics field coupling feature sequences of one complete stroke for M cold forging workpieces (M≥1000), industrial CT inspection of each workpiece, quantitative calculation of internal microcrack feature parameters of the workpiece based on CT images, and mapping of parameters to a discrete damage level label according to process standard threshold.

[0093] Specifically, the formula for calculating the characteristic parameters of internal microcracks is as follows:

[0094] ;

[0095] In the formula, This represents the total number of microcracks with a length ≥50μm in the CT image. The representative is the first The area of ​​the microcrack. This represents the area of ​​the effective forming region of the workpiece. This represents the characteristic parameters of internal microcracks, namely the proportion of microcrack area.

[0096] The damage level label is as follows:

[0097] Level 0: N is less than or equal to 2, and the maximum length of the microcrack is less than or equal to 50 μm. Less than or equal to 0.5%;

[0098] Level 1: N is greater than 2 and less than or equal to 5, and the maximum length of the microcrack is greater than 50 μm and less than or equal to 300 μm. Less than or equal to 2% and greater than or equal to 0.5%;

[0099] Level 2: N is greater than 5, and the maximum length of the microcrack is greater than 300 μm. Greater than 2%;

[0100] The multi-physics coupling feature sequence of each workpiece is paired one-to-one with its own damage level label to form a training dataset.

[0101] The training dataset is divided into a training set (for model parameter optimization), a validation set (for tuning during training), and a test set (for model performance evaluation) in a 7:1:2 ratio.

[0102] Mapping these parameters to discrete damage level labels based on process standard thresholds ensures the objectivity and consistency of the labels, avoids errors caused by subjective judgment, not only provides a sufficient sample size, but also significantly improves the accuracy and reliability of damage level labels, providing high-quality supervision signals for time-series deep learning models. As a result, the models can more effectively learn the complex mapping relationship between multi-physics field features and microcrack damage inside the workpiece, improving the accuracy and robustness of cold forging damage prediction.

[0103] In step three, the temporal deep learning model is a bidirectional LSTM model that incorporates an attention mechanism.

[0104] In step three, the training includes: inputting the multi-physics coupled feature sequence into the bidirectional LSTM layer to learn the forward and backward temporal dependencies, assigning weights to different time steps through the attention mechanism layer to focus on key time periods, using damage level labels as supervision, minimizing the classification loss function through the backpropagation algorithm, and optimizing the model parameters.

[0105] The classification function uses cross-entropy loss, and the optimizer is Adam (initial learning rate = 0.001, decaying by 10% every 10 rounds). During training, the weighted F1 score is used as the core evaluation metric (taking into account class imbalance). Training is stopped when the weighted F1 score on the validation set fails to improve for 5 consecutive rounds to avoid overfitting.

[0106] Specifically, by inputting the multi-physics coupled feature sequence into the bidirectional LSTM layer, the model can simultaneously capture the forward and backward temporal dependencies of data in the cold forging stroke, thereby gaining a more comprehensive understanding of the dynamic process of damage evolution. The attention mechanism layer can intelligently assign weights to different time steps in the sequence, enabling the model to automatically focus on the stroke stage or abnormal event that is most critical for damage prediction, effectively avoiding interference from irrelevant or secondary information. Combined with the damage level label as a supervision signal, and by minimizing the classification loss function through the backpropagation algorithm to optimize the model parameters, the model can efficiently and accurately learn the mapping relationship between features and damage levels.

[0107] The specific steps are as follows: During the real-time stroke, a real-time multi-physics coupling feature sequence is constructed according to the methods in Step 1 and Step 2. The constructed real-time multi-physics coupling feature sequence is input into the trained temporal deep learning model, and the model is calculated through its forward propagation network to output the real-time prediction result. The real-time prediction result represents the probability distribution vector of the workpiece belonging to each predefined damage level.

[0108] For example: Layer 1 bidirectional LSTM: performs forward + backward temporal dependency learning on the input feature sequence. The forward LSTM extracts features from the beginning to the end of the sequence, and the backward LSTM extracts features from the end to the beginning of the sequence. The output dimension is "time step × 128" (64-dimensional features from both the forward and backward directions are concatenated).

[0109] The second layer, a bidirectional LSTM, performs further temporal correlation learning on the output features of the first layer, with the output dimension still being "time step × 128".

[0110] The attention mechanism layer performs forward computation on the output features of the second-layer bidirectional LSTM, and calculates real-time weights through an additive attention mechanism.

[0111] Perform a linear transformation on the output features to obtain the real-time query vector Q, key vector K, and value vector V (each with a dimension of "time step × 64").

[0112] Calculate the matching score between Q and K, and obtain the real-time time step weight after Softmax normalization (focusing on the key stage of the current stroke, such as the peak moment of plastic deformation).

[0113] We use weighted summation on V to obtain the attention-weighted real-time feature sequence (dimension "1×128", compression time step of 1).

[0114] Output layer computation and output of attention-weighted features. Input to fully connected output layer:

[0115] The fully connected layer maps 128-dimensional features to an output dimension of "predefined number of damage levels" (e.g., if there are 3 damage levels, then the output dimension is 3).

[0116] The output is converted into a probability distribution vector by using the Softmax activation function. Each element in the vector corresponds to a probability of a damage level, and the level corresponding to the maximum probability is the real-time damage level of the current workpiece.

[0117] In real-time cold forging, the sensor data collected in real time is transformed into feature sequences that can be recognized by the model. The trained temporal deep learning model is then used for efficient and real-time damage prediction. The prediction results are represented as probability distribution vectors of the workpiece belonging to various predefined damage levels, which greatly improves the practicality and decision-making value of the prediction results. It not only provides a quantitative assessment of the workpiece damage state, but also reflects the model's confidence in the prediction results. This enables operators or automated control systems to formulate more refined and flexible control strategies based on the probability of different damage levels, thereby improving the intelligence level and reliability of cold forging production.

[0118] In step five, the control action is a hierarchical strategy:

[0119] If the predicted damage level is a low warning level, then the warning will be triggered and the equipment will be deloaded.

[0120] If the predicted damage level is a high-level failure level, a stop command is triggered and a mold inspection is prompted.

[0121] The threshold for classifying low-level warning and high-level fault is determined by analyzing the receiver's operational characteristic curves of model prediction probability and actual damage state in historical data to maximize classification performance indicators.

[0122] By introducing a tiered control strategy, the limitations of traditional single-threshold control strategies are avoided. When a low-level warning is predicted, a timely warning is triggered and equipment load is reduced, allowing for intervention in the early stages of damage, effectively delaying damage development, extending mold life, reducing unnecessary downtime, and thus improving production efficiency. Conversely, when a high-level fault is predicted, a shutdown command is decisively triggered and mold inspection is prompted, minimizing equipment damage and safety accidents and ensuring production safety. Furthermore, by analyzing receiver operating characteristic curves in historical data to determine the classification threshold, the boundaries of the tiered strategy are made more scientific and precise, maximizing classification performance indicators and ensuring the reliability and effectiveness of control actions. This achieves more refined and intelligent management of the cold forging process.

[0123] In this invention, the wavelet threshold denoising method adopts the general threshold wavelet denoising method proposed by Donoho et al., which is suitable for non-stationary and non-Gaussian acoustic emission signals, can effectively suppress background noise, and at the same time preserve the time-frequency characteristics of sudden damage events.

[0124] The adaptive Kalman filter method can dynamically track the nonlinear changes of the extrusion pressure signal during the plastic deformation stage and effectively filter out vibration interference.

[0125] A 5th-order low-pass digital filter is used to smooth the mold strain signal. The cutoff frequency is set to 1 / 5 of the system's natural frequency to preserve the strain trend and suppress high-frequency noise.

[0126] Moving average filtering is suitable for the initial smoothing of punch displacement signals. The window length was determined experimentally and can suppress small fluctuations while preserving the displacement step characteristics.

[0127] The extreme point detection of the second derivative of curvature: the extreme point of the second derivative of curvature of the extrusion pressure curve corresponds to the turning point of the deformation stage, which can objectively divide the elastic, plastic and pressure holding stages.

[0128] Linear interpolation time normalization maps sequences of different lengths to a fixed time step, ensuring that the length of the input feature sequences is consistent, which is convenient for subsequent deep learning model processing.

[0129] Acoustic emission spectral characteristics: the center frequency and peak frequency reflect the frequency domain energy distribution of the damage event, and the bandwidth characterizes the frequency dispersion of the damage event;

[0130] Local stiffness change rate and curve fullness, based on the morphological derivatives of the force-displacement curve, reflect abnormal material hardening behavior and energy absorption efficiency, both of which are sensitive to early plastic damage.

[0131] The strain spatial gradient, based on the spatial difference calculation of the strain field, reflects the local non-uniformity of the strain distribution on the mold surface and can be used to detect stress concentration and microcrack initiation locations.

[0132] Synergy index, a statistical correlation index based on covariance normalization, resolves the temporal synergy anomalies between pressure and acoustic emission energy and is sensitive to cross-field coupling response in the damage initiation stage.

[0133] The hysteresis index, a derived index based on time hysteresis cross-correlation analysis, detects the abnormal delayed response of the die strain after the punch displacement stabilizes, reflecting the relaxation anomaly of the material or structure.

[0134] Internal microcrack characteristic parameters, damage quantification method based on industrial CT images, and damage level classification by combining process experience thresholds;

[0135] The bidirectional LSTM with attention mechanism is a classic structure for temporal modeling. The attention mechanism originates from the field of neural machine translation. The bidirectional LSTM can capture the sequential dependence of damage evolution. The attention mechanism can focus on critical stroke stages and improve feature selection ability.

[0136] The cross-entropy loss function and the Adam optimizer are the standard loss function and adaptive moment estimator optimizer for classification tasks, respectively applicable to multi-class tasks and non-stationary objective function convergence.

[0137] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cold forging damage prediction method based on bidirectional LSTM, characterized in that, Includes the following steps: Step 1: Real-time stroke data for a single stroke is collected using IoT sensors to generate raw multimodal time-series data; Step 2: Preprocess and extract features from the original multimodal time series data to construct a multiphysics coupling feature sequence of uniform length. The multiphysics coupling feature sequence includes cross-physics triggering features extracted from the original multimodal time series data. Step 3: Pair the multiphysics coupling feature sequence with the workpiece damage level label to form a training dataset, and train the time series deep learning model using the training dataset; Step 4: Based on the time-series deep learning model, collect multimodal time-series data under real-time stroke and construct a real-time multiphysics coupling feature sequence. Input the real-time multiphysics coupling feature sequence into the model and output the damage prediction result. Step 5: Based on the prediction results, execute the control action.

2. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step one, the real-time stroke data collected in a single stroke based on IoT sensors includes the extrusion pressure signal collected in real time by the force sensor deployed on the cold forging equipment, the punch displacement signal collected by the displacement sensor, the acoustic emission signal collected by the acoustic emission sensor, and the die strain signal collected by the die strain sensor. Before collection, the sampling frequency of the IoT sensors is unified, and the timestamps of all sensors are synchronized by a hardware clock.

3. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step two, the original multimodal time series data is preprocessed, including denoising the acoustic emission signal using wavelet threshold denoising, dynamically filtering the extrusion pressure signal using an adaptive Kalman filter combined with a cold forging process model, smoothing the die strain signal using a 5th-order low-pass digital filter, and smoothing the punch displacement signal using a moving average filter with a window length of 10. Simultaneously, the pre-processed extrusion pressure signal is used as a reference signal to calculate the extreme points of its second derivative of curvature, identify and divide the start and end time points of the elastic deformation, plastic deformation and holding pressure stages of the cold forging stroke, and use these time points as a reference to synchronously divide all other physical field signals into stages and perform time normalization processing. For the key stages of plastic deformation and holding pressure stages, a time resolution higher than that of elastic deformation is used for resampling.

4. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step two, feature extraction is performed on the original multimodal time-series data, specifically including: S1: Perform waveform recognition on the acoustic emission signal, extract the Burst-type acoustic emission events related to the microcrack, and calculate their time-spectrum characteristics; S2: For the extrusion force signal and the punch displacement signal, calculate the higher-order morphological features of the force and displacement curves, including the local stiffness change rate and curve fullness. S3: For the strain signal of the mold, calculate the spatial gradient characteristics of the strain distribution.

5. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, Step two also includes calculating the cross-physical field triggering features based on the features extracted by S1, S2 and S3. Within a set time window, the synergistic index of the extrusion signal drop and the acoustic emission energy mutation, as well as the hysteresis anomaly index of the mold strain change during the stable stage of the punch displacement, are calculated. Then, the single-physical field features and cross-physical field features are spliced ​​together and standardized to the [0,1] interval to form a multi-physical field coupling feature sequence of uniform length.

6. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step three, the multiphysics coupling feature sequence is paired with the workpiece damage level label to form a training dataset, as detailed below: Offline acquisition of multi-physics field coupling feature sequences of one complete stroke for each of M cold forging workpieces (M≥1000), industrial CT inspection of each workpiece, quantitative calculation of internal microcrack feature parameters of the workpiece based on CT images, and mapping of the parameters to a discrete damage level label according to process standard threshold. The multiphysics coupling feature sequence of each workpiece is paired one-to-one with its own damage level label to form a training dataset.

7. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step three, the temporal deep learning model is a bidirectional LSTM model that incorporates an attention mechanism.

8. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step three, the training includes: inputting the multi-physics coupled feature sequence into a bidirectional LSTM layer to learn the forward and backward temporal dependencies, assigning weights to different time steps through an attention mechanism layer to focus on key time periods, using damage level labels as supervision, minimizing the classification loss function through a backpropagation algorithm, and optimizing model parameters.

9. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, The specific steps of step four are as follows: During the real-time stroke, a real-time multi-physics coupling feature sequence is constructed according to the methods of step one and step two. The constructed real-time multi-physics coupling feature sequence is input into the trained temporal deep learning model, and the model is calculated through its forward propagation network to output the real-time prediction result. The real-time prediction result represents the probability distribution vector of the workpiece belonging to each predefined damage level.

10. The cold forging damage prediction method based on bidirectional LSTM according to claim 1, characterized in that, In step five, the control action is a hierarchical strategy: If the predicted damage level is a low warning level, then the warning will be triggered and the equipment will be deloaded. If the predicted damage level is a high-level failure level, a stop command is triggered and a mold inspection is prompted. The threshold for classifying low-level warning and high-level fault is determined by analyzing the receiver's operational characteristic curves of model prediction probability and actual damage state in historical data to maximize classification performance indicators.

Images