Stamping die wear prediction method based on multi-sensor fusion

Abstract

The invention discloses a stamping die wear prediction method based on multi-sensor fusion, and relates to the technical field of die state monitoring. Comprising the steps that S1, multi-source sensing data in the operation process of a stamping die are collected, and the multi-source sensing data comprise a vibration signal, a temperature signal, a pressure signal and an acoustic emission signal; s2, performing data preprocessing on the acquired multi-source sensing data, wherein the data preprocessing comprises denoising, dimension unification and data standardization on each path of sensing signals; through adaptive time sequence alignment, the problems of feature extraction deviation and context information loss caused by difficulty in accurate alignment of a multi-source signal time sequence in an existing method are solved; time domain statistics, frequency domain energy, time frequency analysis and deep time sequence characteristics are extracted at the same time, and the deep coupling relationship of the heterogenous data is mined through a cross-modal attention and gating mechanism of the multi-modal fusion network, so that the limitation of simple weighted fusion is avoided.

Description

Technical Field

[0001] This invention relates to the field of mold condition monitoring technology, specifically a method for predicting wear of stamping dies based on multi-sensor fusion. Background Technology

[0002] Stamping dies are key tools in manufacturing for the mass production of metal parts, and their working condition directly affects product precision and production efficiency. During continuous stamping, dies gradually wear down due to mechanical stress, friction, and fatigue, leading to decreased product quality or even scrap. Therefore, accurate prediction of die wear is of great significance. Multi-sensor fusion technology, by integrating data from various types of sensors such as vibration, temperature, pressure, and acoustic emission, can more comprehensively reflect the operating status of the die, providing a rich information foundation for wear prediction.

[0003] However, existing wear prediction methods based on multi-sensor fusion still suffer from a key drawback at the temporal feature fusion level: the sampling frequencies and response characteristics of different sensors vary significantly, making it difficult to accurately align multi-source sensor signals in the time dimension. This asynchronicity causes temporal deviations and loss of contextual information during feature extraction, thus affecting the model's ability to characterize the microscopic evolution of wear. Due to the lack of a fusion mechanism that can adaptively align multi-source sensor sequences and preserve temporal causality, existing methods often rely solely on simple weighting or post-decision fusion, failing to fully explore the deep coupling relationship between heterogeneous data such as high-frequency vibration signals and low-frequency temperature signals, thereby limiting the accuracy of wear prediction and early warning capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for predicting wear of stamping dies based on multi-sensor fusion, which solves the problems of feature extraction deviation and loss of contextual information caused by the difficulty in accurately aligning the timing of multi-source signals in existing methods.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting wear of stamping dies based on multi-sensor fusion, comprising the following steps: S1. Collect multi-source sensor data during the operation of the stamping die, including vibration signal, temperature signal, pressure signal and acoustic emission signal; S2. Perform data preprocessing on the collected multi-source sensor data. Data preprocessing includes denoising, unifying the dimensions, and standardizing the data of each sensor signal. S3. Adaptive time alignment is performed on the multi-source sensor data after data preprocessing. Adaptive time alignment is to establish a time synchronization benchmark based on the sampling frequency and response characteristics of different sensors. Through dynamic time warping algorithm, all sensor signals are mapped to a unified time dimension to eliminate time asynchrony and time deviation, and ensure that the multi-source sensor data are accurately matched in the time dimension. S4. Extract multi-dimensional temporal features from the sensor data that has completed adaptive temporal alignment. The multi-dimensional temporal features include time-domain statistical features, frequency-domain energy features, time-frequency analysis features, and deep temporal features automatically extracted by a temporal convolutional network. S5. Deep fusion is performed on the extracted multidimensional temporal features. Deep fusion utilizes a multimodal fusion network, which takes multidimensional temporal features as input, calculates the weights between features from different sensors through a cross-modal attention mechanism, and performs weighted summation to construct a fused joint feature representation. S6. Based on the joint feature representation after deep fusion, train the mold wear prediction model. The mold wear prediction model is a sequence prediction model using a long short-term memory network. S7. Using the trained die wear prediction model, predict the current wear status and future wear trend of the stamping die.

[0006] Furthermore, the vibration signal in S1 is acquired by an optical fiber vibration sensor, the temperature signal is acquired by an infrared thermometer, the pressure signal is acquired by a pressure sensor, and the acoustic emission signal is acquired by a broadband acoustic emission sensor.

[0007] Furthermore, the data preprocessing in S2 also includes outlier detection and correction, which is achieved through a sliding window mid-value filtering method to remove noise interference and improve signal quality.

[0008] Furthermore, the adaptive timing alignment in S3 specifically includes: identifying the signal with the highest sampling frequency in the multi-source sensor data as the reference signal; performing interpolation processing on the low sampling frequency signals to align them with the reference signal in the time series; and dynamically adjusting the alignment parameters by calculating the time delay and correlation between different signals to ensure high-precision synchronization.

[0009] Furthermore, S4 extracts multidimensional temporal features, including time-domain statistical features such as mean, variance, kurtosis, skewness, kurtosis, and root mean square value; frequency-domain energy features such as dominant frequency, energy entropy, and frequency band energy ratio; time-frequency analysis features such as wavelet packet energy spectrum and Hilbert-Huang transform marginal spectrum; and deep temporal features automatically extracted by the temporal convolutional network are high-level abstract features related to mold wear state obtained by automatically learning from the preprocessed and aligned temporal signals input into the temporal convolutional network.

[0010] Furthermore, the deep fusion in S5 specifically includes: using the extracted time-domain, frequency-domain, and deep temporal features as the input layer of the multimodal fusion network; the multimodal fusion network includes multiple convolutional layers, pooling layers, and fully connected layers to learn the representation within each feature modality; the multimodal fusion network includes cross-modal attention mechanisms and gating mechanisms to adaptively weight and fuse features from different modalities to enhance the contribution of key features to wear prediction and suppress redundant information.

[0011] Furthermore, before training, the mold wear prediction model in S6 divides the deeply fused time-series feature dataset into training, validation, and test sets. During model training, mean squared error is used as the loss function. The Adam optimizer is used for iterative parameter updates, and learning rate decay strategy and early stopping mechanism are set to prevent model overfitting.

[0012] Furthermore, the die wear prediction model in S7 outputs the wear level or remaining service life of the stamping die.

[0013] Furthermore, after training, the mold wear prediction model in S7 adjusts the model's internal parameters through the backpropagation algorithm to minimize the error between the predicted output and the actual wear label; the model also includes a validation phase to evaluate the model's generalization ability and prediction accuracy on independent datasets.

[0014] Furthermore, the die wear prediction model in S7 can receive process parameters representing different materials, stamping speeds, and lubrication conditions as auxiliary inputs, and thus be applied to predict the wear evolution of stamping dies under different process parameters.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the issues of feature extraction bias and contextual information loss caused by the difficulty in accurately aligning the timing of multi-source signals in existing methods through adaptive temporal alignment. Simultaneously, it extracts temporal statistics, frequency energy, time-frequency analysis, and deep temporal features. Then, a multimodal fusion network with cross-modal attention and gating mechanisms is used to uncover deep coupling relationships between heterogeneous data, avoiding the limitations of simple weighted fusion. Subsequently, an LSTM prediction model is constructed, combined with process parameters as auxiliary input, which not only improves the accuracy of die wear prediction and early warning capabilities but also adapts to the wear evolution patterns under different process conditions, effectively ensuring the precision and production efficiency of stamping dies. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1 This invention provides a method for predicting wear of stamping dies based on multi-sensor fusion, comprising the following steps: S1. Collect multi-source sensor data during the operation of the stamping die, including vibration signal, temperature signal, pressure signal and acoustic emission signal; S2. Perform data preprocessing on the collected multi-source sensor data. Data preprocessing includes denoising, unifying the dimensions, and standardizing the data of each sensor signal. S3. Adaptive time alignment is performed on the multi-source sensor data after data preprocessing. Adaptive time alignment is to establish a time synchronization benchmark based on the sampling frequency and response characteristics of different sensors. Through dynamic time warping algorithm, all sensor signals are mapped to a unified time dimension to eliminate time asynchrony and time deviation, and ensure that the multi-source sensor data are accurately matched in the time dimension. S4. Extract multi-dimensional temporal features from the sensor data that has completed adaptive temporal alignment. The multi-dimensional temporal features include time-domain statistical features, frequency-domain energy features, time-frequency analysis features, and deep temporal features automatically extracted by a temporal convolutional network. S5. Deep fusion is performed on the extracted multidimensional temporal features. Deep fusion utilizes a multimodal fusion network, which takes multidimensional temporal features as input, calculates the weights between features from different sensors through a cross-modal attention mechanism, and performs weighted summation to construct a fused joint feature representation. S6. Based on the joint feature representation after deep fusion, train the mold wear prediction model. The mold wear prediction model is a sequence prediction model using a long short-term memory network. S7. Using the trained die wear prediction model, predict the current wear status and future wear trend of the stamping die.

[0019] Specifically, sensors are deployed at key locations on the stamping die to collect multi-dimensional operational data. For example, vibration sensors are deployed at the guide post connection of the upper die holder to collect vibration signals during the die opening and closing process; temperature sensors are deployed on the cavity sidewalls to monitor local temperature changes caused by stamping friction; pressure sensors are deployed at the contact end between the stamping head and the die to record real-time pressure during the stamping process; and acoustic emission sensors are deployed at the support part of the lower die holder to capture acoustic emission signals generated by material deformation and friction. The acquisition parameters can be set to a vibration signal sampling frequency of 1kHz, a temperature signal sampling frequency of 1Hz, a pressure signal sampling frequency of 500Hz, and an acoustic emission signal sampling frequency of 2kHz. The acquisition duration is the complete cycle of 1000 consecutive stamped workpieces to ensure coverage of typical stages of die wear.

[0020] A wavelet thresholding denoising algorithm is used to denoise each signal. For example, the db4 wavelet basis is selected, the decomposition layer is 5 layers, and the wavelet coefficients are processed by a soft thresholding function to eliminate mechanical vibration interference and electromagnetic noise. Due to the differences in the units of data from different sensors, such as vibration being measured in g and temperature in °C, Z-score standardization is used to unify the units. The formula is as follows: ; in, For raw sensor data, This is the average value of the signal from that channel. The standard deviation of this signal is used to standardize all data, ensuring consistent dimensions, allowing direct use in subsequent fusion. Outlier detection and correction are achieved through a sliding window midpoint filter; for example, setting the window size to 5 and filtering for each data point... Calculate the data within the window the median of ,like If the standard deviation of the data within the window is used, then the judgment is made. It is an outlier, and is used Replacement is used to avoid the impact of sudden interference, such as instantaneous voltage fluctuations, on data quality.

[0021] To address the timing asynchrony issue caused by differences in sampling frequencies between different sensors, such as temperature at 1Hz and vibration at 1kHz, alignment is achieved using a reference signal selection combined with Dynamic Time Warping (DTW). The sampling frequencies of all sensing signals are calculated, and the signal with the highest sampling frequency is selected as the reference signal. For example, in this embodiment, the acoustic emission signal has the highest sampling frequency of 2kHz, which is used as the timing reference. The DTW algorithm is used to map low sampling frequency signals, such as temperature, pressure, and vibration, to the time dimension of the reference signal. The core of DTW is to calculate the minimum cumulative distance between the two signals, as shown in the following formula: Distance matrix elements: ; in, For the reference signal Data points, For the signal to be aligned The data points have been standardized through step S2. and The units are consistent, and the distance can be calculated directly; Dynamic programming recurrence formula: ; in, From arrive The minimum cumulative distance, with the boundary condition being: ; Calculate the cross-correlation coefficients between each signal and the reference signal after alignment. , ); like If necessary, adjust the distance weights of the DTW, for example, by increasing the distance weights of high-frequency signals, and re-perform alignment until the cross-correlation coefficient between all signals and the reference signal is ≥0.8, ensuring that the multi-source data are accurately matched in the time dimension and eliminating phase deviation.

[0022] Four-dimensional features—time-domain, frequency-frequency-time-frequency, and deep-level—are extracted from the aligned signal to comprehensively characterize the mold wear state: Time-domain statistical features are calculated based on the aligned time-domain signal. The mean reflects the overall signal level, the variance reflects the signal fluctuation degree, the kurtosis reflects the signal peak characteristics (kurtosis increases with increased wear), and the root mean square (RMS) reflects the signal energy intensity. Other features include skewness and kurtosis, resulting in a total of 12 time-domain features. Frequency-domain energy features are obtained by performing a Fast Fourier Transform (FFT) on the aligned signal, extracting 8 frequency-domain features including the dominant frequency, energy entropy, and bandwidth energy ratio. Time-frequency analysis features are extracted using wavelet packet decomposition and Hilbert-Huang Transform (HHT). The db6 wavelet basis is selected for the wavelet packet energy spectrum, with a decomposition level of 4 layers. The energy of 16 wavelet packet subbands is calculated, forming a 16-dimensional energy spectrum feature. The HHT marginal spectrum is obtained by performing Empirical Mode Decomposition (EMD) on the signal to obtain the Intrinsic Mode Function (IMF). A Hilbert Transform is performed on each IMF to obtain the instantaneous frequency and amplitude, and integration yields the marginal spectrum. In the formula, For instantaneous amplitude, six features, including peak value and mean value, are extracted from the marginal spectrum. For deep temporal features, the aligned original signal is input into the temporal convolutional network (TCN) for automatic extraction. The TCN structure includes six causal convolutional layers, layer normalization, and ReLU activation function. Finally, global average pooling is used to output 24-dimensional deep features, which can automatically capture the implicit temporal patterns related to wear in the signal, such as the vibration period changes caused by wear.

[0023] Deep feature fusion is achieved using a multimodal fusion network. The network input layer concatenates the 12-dimensional temporal, 8-dimensional frequency, 22-dimensional temporal, and 24-dimensional deep features extracted in step S4 to obtain a 66-dimensional multimodal feature vector. Intramodal feature learning is performed by learning local feature representations within each modality through two convolutional layers and a max-pooling layer. Finally, two fully connected layers are used to further map the features to a unified-dimensional modality. Cross-modal attention fusion introduces a cross-modal attention mechanism to calculate the weights of each modality, as shown in the following formula: Similarity calculation: ,in For the first Feature vectors of each modality , These are learnable parameters; Weight normalization: , For the first The weights of each modality are set to ensure that the sum of the weights is 1. Joint feature construction: , The fused joint feature vector has a dimension of 64. The gating mechanism is optimized by using the sigmoid gating function. The sigmoid function controls the proportion of features to suppress redundant features, such as repeated time-domain and frequency-domain energy information, and strengthens key features, such as the contribution of deep time-series features to wear prediction.

[0024] A prediction model is constructed using a Long Short-Term Memory (LSTM) network, and the fused joint features are used. The actual wear of the corresponding mold is measured using a laser thickness gauge, for example, once every 100 stamped workpieces, forming a sample set, which is divided into a training set, a validation set, and a test set in a ratio of 7:2:1; the LSTM model includes an input layer, two hidden layers, and an output layer; The mean squared error (MSE) is used to measure the difference between the predicted and actual values. The formula is as follows: ; in, The number of samples in the training set. This represents the actual wear and tear. To predict wear and tear; The optimizer used is the Adam optimizer, with an initial learning rate of 0.001 and a momentum parameter. , Numerical stability parameters Overfitting is suppressed by using a learning rate decay strategy, and a dropout layer is introduced into the LSTM hidden layer with a dropout probability of 0.2. During the training process, the model is trained iteratively for 50 rounds. After each round of training, the model performance is evaluated using a validation set. If the MSE on the validation set does not decrease for 5 consecutive rounds, training is stopped and the optimal model parameters are saved.

[0025] The joint features collected in real time and processed through steps S1 to S5 Input the trained LSTM model and output the current wear level of the mold. Based on 10 consecutive real-time prediction results, linear fitting is used to predict the wear trend in the next 100 stamping cycles.

[0026] In this embodiment, the vibration signal in S1 is acquired by an optical fiber vibration sensor, the temperature signal is acquired by an infrared thermometer, the pressure signal is acquired by a pressure sensor, and the acoustic emission signal is acquired by a broadband acoustic emission sensor.

[0027] Specifically, the vibration signal uses an optical fiber vibration sensor, such as the FBG-VS100 optical fiber vibration sensor. This sensor has a measurement range of 0-5kHz, a sensitivity of 1pm / g, and a resolution of 0.01g. It is installed at the connection between the guide post and the template on the upper mold base. Vibration is transmitted most directly at this location, which can reflect the vibration state of the mold opening and closing. Data is transmitted through optical fiber to avoid electromagnetic interference. Temperature signals are measured using an infrared thermometer, such as the FLIRA655sc model, with a measurement range of -40℃ to 150℃, an accuracy of ±0.03℃, and a sampling frequency of 1Hz. It is installed on the side wall of the mold cavity, 5mm from the cavity surface, using non-contact measurement to avoid damage to the sensor from high temperatures. Pressure signals are measured using a piezoelectric pressure sensor, such as the Kistler9311B model, with a range of 0-200MPa, an accuracy of ±0.1%FS, and a sampling frequency of 500Hz. It is installed in the center hole of the stamping head, directly contacting the upper mold die, to capture changes in stamping pressure in real time. Acoustic emission signals are measured using a broadband acoustic emission sensor, such as the PACR15α model, with a frequency response range of 100kHz-1MHz and a sensitivity of -65dB. It is installed at the support rib of the lower mold base, where background noise is low, effectively capturing acoustic emission signals generated by material friction.

[0028] In this embodiment, the data preprocessing in S2 also includes outlier detection and correction, which is achieved by a sliding window mid-value filtering method to remove noise interference and improve signal quality.

[0029] Specifically, outlier detection and correction in data preprocessing are achieved through sliding window median filtering. The window size is determined based on the sampling frequency of the sensor signal. For example, for vibration signals with a sampling frequency of 1kHz, a window size of 5 is set, corresponding to a time window of 5ms, balancing real-time performance and filtering effect; for temperature signals with a sampling frequency of 1Hz, a window size of 3 is set, corresponding to a time window of 3s. The median is calculated for the signal sequence... For each data point ( and ), retrieve data from the window After sorting, take the middle value as the median. Outlier identification and correction; calculation of the standard deviation of data within the window. ,like Then determine For outliers, such as sudden changes in pressure sensor data due to instantaneous impact, Replace with ;like or For the first and last data of the sequence, the original data is directly retained to avoid boundary distortion.

[0030] In this embodiment, the adaptive timing alignment in S3 specifically includes: identifying the signal with the highest sampling frequency in the multi-source sensor data as the reference signal; performing interpolation processing on the low sampling frequency signals to align them with the reference signal in the time series; and dynamically adjusting the alignment parameters by calculating the time delay and correlation between different signals to ensure high-precision synchronization.

[0031] Specifically, the sampling frequencies of all sensing signals are iterated, such as vibration 1kHz, temperature 1Hz, pressure 500Hz, and acoustic emission 2kHz, and the acoustic emission signal with the highest sampling frequency is selected as the reference signal. Its time series is , The reference signal data length; for signals with low sampling frequencies, such as temperature signals. Its sampling frequency is 1Hz, and the time series... The time series is mapped to the time dimension of the reference signal using linear interpolation. The linear interpolation formula is as follows: ; in, For the reference signal At a certain point in time, , , The temperature signal values ​​are located at adjacent sampling points. The vibration and pressure signals are interpolated using the same method to obtain a preliminary aligned signal with the time dimension of the reference signal. The cross-correlation function between each signal and the reference signal after preliminary alignment is calculated. ; in , Find the mean values ​​of the reference signal and the alignment signal, respectively. The maximum value corresponding to This is the time delay; if If the time delay exceeds two sampling points of the reference signal, the interpolation step size is adjusted, such as changing linear interpolation to cubic spline interpolation. The interpolation and cross-correlation calculations are then re-executed until the time delay is ≤0.001s to ensure high-precision synchronization.

[0032] In this embodiment, the extraction of multidimensional temporal features in S4 includes time-domain statistical features such as mean, variance, kurtosis, skewness, kurtosis, and root mean square value; frequency-domain energy features such as dominant frequency, energy entropy, and frequency band energy ratio; time-frequency analysis features such as wavelet packet energy spectrum and Hilbert-Huang transform marginal spectrum; and deep temporal features automatically extracted by the temporal convolutional network are high-level abstract features related to the wear state of the mold obtained by automatically learning from the preprocessed and aligned temporal signals by inputting them into the temporal convolutional network.

[0033] Specifically, the formulas and physical meanings of each feature are as follows: Mean: This reflects the overall amplitude level of the signal. When mold wear intensifies, the average value of vibration and acoustic emission signals will increase. variance: This reflects the degree of dispersion of the signal. When wear causes the mold to operate unstably, the variance will increase. Skewness: This reflects the asymmetry of signal distribution; in the early stages of wear, the skewness will shift in the positive direction. kurtosis: This reflects the peak characteristics of the signal; when the mold experiences localized wear, the kurtosis will increase significantly. Root mean square value: RMS reflects the energy intensity of the signal; when wear leads to increased friction, RMS will rise.

[0034] Frequency domain signal based on FFT transform: , in For time-domain signals, Frequency calculation: Clock speed: ,in As the power spectral density is used, when mold wear leads to a decrease in stiffness, the dominant frequency will shift towards lower frequencies. Energy entropy: ,in , For the first Energy of each frequency band For total energy, The frequency band is the number of frequency bands, for example, divided into 32 frequency bands. The energy entropy reflects the uniformity of the energy distribution in the frequency domain. The energy entropy will increase when the frequency band is worn. Bandwidth energy ratio: ,in The energy source is in the low-frequency band, and the proportion of low-frequency energy will increase in the early stages of wear.

[0035] The wavelet packet energy spectrum was calculated by decomposing the signal into four layers using the db6 wavelet basis, resulting in 16 wavelet packet subbands. The energy of each subband was then calculated. ,in For the first The reconstructed signals of each sub-band, 16 The wavelet packet energy spectrum characteristics are constituted; The Hilbert-Huang Transform (HHT) marginal spectrum is obtained by first performing EMD decomposition on the signal. IMF components Perform a Hilbert transform on each IMF: ; Where PV is the Cauchy principal value, the analytic signal is obtained: , in For instantaneous amplitude, Instantaneous phase, instantaneous frequency Marginal spectrum is ,in, for In frequency The projection at the point is used to extract three features of the marginal spectrum: peak frequency, mean frequency, and energy centroid.

[0036] Automatic extraction is performed using a Temporal Convolutional Network (TCN). The input to the TCN is the aligned original temporal signal with dimensions [missing information]. , Given the data length, the network structure includes: Six causal convolutional layers with a kernel size of 3 and dilation coefficients of 1, 2, 4, 8, 16, and 32 respectively. Each convolutional layer is followed by a normalization layer and a ReLU activation function. 1st Global Average Pooling Layer: Pools the feature map output by the convolutional layer into a 1×64 feature vector; One fully connected layer: maps 64-dimensional features to 24-dimensional deep temporal features. This feature can automatically capture the implicit patterns related to wear in the signal, such as the periodic changes of vibration signals and the pulse interval changes of acoustic emission signals.

[0037] In this embodiment, the deep fusion in S5 specifically includes: using the extracted time-domain, frequency-domain, and deep temporal features as the input layer of the multimodal fusion network; the multimodal fusion network includes multiple convolutional layers, pooling layers, and fully connected layers to learn the representation within each feature modality; the multimodal fusion network includes cross-modal attention mechanisms and gating mechanisms to adaptively weight and fuse features from different modalities to enhance the contribution of key features to wear prediction and suppress redundant information.

[0038] Specifically, deep fusion is based on a multimodal fusion network, the structure and process of which are as follows: Input layer and modality splitting: The 66-dimensional multimodal features extracted in step S4 are split into 4 modality branches: 12-dimensional time-domain feature branch, 8-dimensional frequency-domain feature branch, 22-dimensional time-frequency feature branch, and 24-dimensional deep temporal feature branch. Each branch is input separately into the intramodal feature learning module. The intra-modal feature learning module, each branch of which contains a structure of "convolutional layer + pooling layer + fully connected layer": The convolutional layer uses 1D convolution with a kernel size of 3×1. The number of output channels is: 32 in the time domain, 32 in the frequency domain, 64 in the time-frequency branch, and 64 in the deep branch. The activation function is ReLU. The pooling layer uses 1D max pooling with a 2×1 kernel size and a stride of 2, which is used for dimensionality reduction and preservation of key features. Fully connected layers: Each layer contains 64 neurons, mapping the features of each branch to a unified 64-dimensional modal feature vector. Each corresponds to one of the four modes; The cross-modal attention fusion module calculates the weights of features from each modality to achieve adaptive fusion. Similarity calculation for each modality feature The similarity between it and the fusion target is calculated through linear transformation: ,in The weight matrix is ​​64×1. These are bias terms, all of which are learnable parameters; Weight normalization is performed by using the softmax function to normalize the similarity, resulting in the weights for each modality: ,make sure ; Weighted fusion involves summing the features of each modality according to their respective weights to obtain a joint feature vector. , The dimension is 64; The gating mechanism optimization module introduces a sigmoid gating function to suppress redundant features: Gating vector calculation: ,in The weight matrix is ​​64×64. For 64-dimensional bias terms, For the sigmoid function, Each element takes a value in the range [0,1], representing the pass rate of the corresponding feature dimension; Gated post-joint features: By using a gating function, redundant features with low relevance to wear prediction are assigned low weights, such as the weight of repeated energy information which is close to 0, while key features are assigned high weights, such as the weight of wear cycle information in deep time series features which is close to 1.

[0039] In this embodiment, before training, the mold wear prediction model in S6 divides the deeply fused time-series feature dataset into training set, validation set and test set; during model training, mean squared error is used as the loss function; the Adam optimizer is used for parameter iterative updates, and learning rate decay strategy and early stopping mechanism are set to prevent model overfitting.

[0040] Specifically, the combined features after deep integration The sample pairs are composed of the actual wear amount of the corresponding mold. ,in For the first The actual wear amount of each sample was measured, for example, by a laser thickness gauge with a measurement accuracy of ±0.001mm. Measurements were taken once for every 100 stamped workpieces, for a total of 1000 samples. The sample sets were randomly divided in a ratio of 7:2:1: 700 samples for training the model parameters, 200 samples for validation the hyperparameters and overfitting monitoring, and 100 samples for testing the model's generalization ability. The division ensured that the wear amount distribution of each set was consistent and covered the typical wear range of 0-0.5mm.

[0041] The loss function uses the mean squared error (MSE), and the formula is as follows: ; in, This specifies the number of samples in the current training batch; for example, the batch size can be set to 32. This represents the actual wear and tear. The model outputs the predicted wear amount. MSE effectively reflects the overall deviation between the predicted and actual values, making it suitable for predicting wear amounts for continuous values. The Adam optimizer is used, with parameters set as follows: initial learning rate... Momentum parameters , , To estimate the attenuation coefficient for the first moment, To estimate the attenuation coefficient for the second moment, numerical stability parameters The Adam optimizer can adaptively adjust the learning rate, balancing training speed and stability.

[0042] The learning rate decay strategy uses a step-wise decay, where the learning rate is multiplied by a decay factor of 0.8 every 15 training epochs. To avoid parameter oscillations caused by excessively high learning rates in the later stages of training, in addition to learning rate decay, a dropout layer is introduced between the two hidden layers of the LSTM, with a dropout probability set to 0.2, which means randomly discarding 20% ​​of the neurons. At the same time, an early stopping strategy is adopted: if the MSE of the validation set does not decrease for 5 consecutive rounds, training is stopped, the current optimal model parameters are saved, and the model is prevented from overfitting on the training set.

[0043] In this embodiment, the die wear prediction model in S7 outputs the wear level or remaining service life of the stamping die.

[0044] Specifically, the output layer of the mold wear prediction model can output two forms of prediction results depending on the application scenario: Wear level output The mold wear is divided into multiple levels, and the model outputs the current wear level, for example: Grade 0 Slight Wear: Wear Amount At this point, the mold precision meets the requirements and requires no maintenance. Level 1 moderate wear: At this point, the mold precision has slightly decreased, and monitoring needs to be strengthened. Level 2 severe wear: At this point, the mold precision exceeds the allowable range, and the machine must be stopped immediately for maintenance; The model output layer uses the softmax activation function and outputs three probability values, corresponding to level 0, level 1, and level 2, respectively. The level with the highest probability is the predicted wear level. For example, if the output probability is [0.02, 0.95, 0.03], then the predicted wear level is level 1, moderate wear.

[0045] Remaining lifespan RUL output Remaining service life (RUL) refers to the number of stamping operations required for a die to go from its current state to reaching a severe wear level. The model calculates RUL using the following steps: Based on 10 consecutive real-time wear prediction results Corresponding number of stampings The wear rate was obtained by linear fitting. The fitting formula is: ; in, For the number of stamping cycles, For wear rate, The intercept is obtained by solving the least squares method. and ,make Minimum; Calculate remaining useful life: ,in, The severe wear threshold, For the current predicted wear amount, for example , Next, then Maintenance is required after 150 stamping cycles remaining.

[0046] In this embodiment, after the mold wear prediction model in S7 is trained, the internal parameters of the model are adjusted through the backpropagation algorithm to minimize the error between the predicted output and the actual wear label; the model also includes a validation phase to evaluate the model's generalization ability and prediction accuracy on an independent dataset.

[0047] Specifically, during model training, the gradient of the loss function with respect to each parameter is calculated using the backpropagation algorithm, and then the parameters are updated. The specific steps are as follows: Forward propagation: The training set samples are input into the model and passed sequentially through the input layer, hidden layer, and output layer of the LSTM to obtain the predicted values. ; Loss Calculation: Calculate the loss according to the MSE formula. ; Backpropagation: Starting from the output layer, calculate the loss using the chain rule. For each parameter, the weights of the LSTM Bias The weights of the fully connected layer Bias gradient: Output layer gradient: ; LSTM hidden layer gradient: Based on LSTM-based gating structures such as input gate, forget gate, and output gate, the gradient is calculated through the backpropagation algorithm, taking into account the dependencies between time series data. Parameter update: Update parameters using gradients according to the update rules of the Adam optimizer. ; ; ; ; ; in, , These are the estimates of the first and second moments, respectively. , This is the moment estimate after bias correction. The current parameter weights, This is the loss gradient for the current batch. This represents the number of iteration steps.

[0048] Model validation phase After the model is trained, its generalization ability and prediction accuracy are evaluated on independent validation and test sets. Evaluation metrics include: Mean Absolute Error (MAE): This reflects the average absolute deviation between the predicted value and the actual value; Coefficient of determination : ,in The mean of the true values. The closer it is to 1, the stronger the explanatory power of the model; During the verification process, if the MAE of the test set is less than 0.01 mm and If the model meets the requirements, then the model is deemed to meet the usage requirements; if not, return to step S6 to adjust the model structure or retrain until the performance requirements are met.

[0049] In this embodiment, the die wear prediction model in S7 can receive process parameters representing different materials, stamping speeds and lubrication conditions as auxiliary inputs, and thus be used to predict the wear evolution of stamping dies under different process parameters.

[0050] Specifically, to enable the model to adapt to wear prediction under different process conditions, process parameters representing the working environment of the mold are used as auxiliary inputs and incorporated into the model training process: Selection of process parameters and pretreatment: Select three types of process parameters that significantly affect mold wear, for example: Material parameters: The hardness of the metal material being stamped, such as the hardness of low carbon steel Q235 being 150HB and the hardness of stainless steel 304 being 200HB. Stamping speed: The downward speed of the stamping head, such as 50mm / s, 100mm / s, 150mm / s, 200mm / s; Lubrication conditions are represented by a classification code, such as 0 for dry friction, 1 for oily lubrication, and 2 for watery lubrication. Preprocessing of process parameters: Numerical parameters such as hardness and stamping speed are standardized using Z-score to eliminate dimensional differences. Classified parameters such as lubrication conditions are converted into 3-dimensional vectors using unique thermal encoding, such as [1,0,0] for dry friction and [0,1,0] for oily lubrication; the pre-processed process parameters are 5-dimensional in total, with 1 dimension for hardness, 1 dimension for stamping speed, and 3 dimensions for lubrication conditions.

[0051] The preprocessed 5-dimensional process parameters are used as auxiliary features, and combined with the 64-dimensional joint features obtained in step five. Concatenating the features yields a 69-dimensional extended feature vector. ,in For process parameter characteristics; As the input layer of the LSTM model for mold wear prediction, the input layer dimension was adjusted from 64 to 69, while the hidden layers and output layers of the rest of the network structure remained unchanged.

[0052] Model training under multiple process conditions Collect sample data under different process conditions, such as covering 2 materials, 4 stamping speeds, and 3 lubrication conditions, for a total of 2×4×3=24 process combinations. Collect 50 samples for each combination, for a total of 1200 samples. Divide the data into training set, validation set, and test set in a 7:2:1 ratio, and train the model using the method in step S6.

[0053] In summary, this invention solves the problems of feature extraction deviation and contextual information loss caused by the difficulty in accurately aligning the timing of multi-source signals in existing methods through adaptive timing alignment. Simultaneously, it extracts time-domain statistics, frequency-domain energy, time-frequency analysis, and deep timing features. Then, it uses a multi-modal fusion network with cross-modal attention and gating mechanisms to mine the deep coupling relationships of heterogeneous data, avoiding the limitations of simple weighted fusion. Subsequently, it constructs a prediction model using LSTM, combined with process parameters as auxiliary input, which not only improves the accuracy of die wear prediction and early warning capabilities but also adapts to the wear evolution patterns under different process conditions, effectively ensuring the precision and production efficiency of stamping dies.

[0054] 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 process, method, article, or apparatus.

[0055] 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 alterations 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 press die wear prediction method based on multi-sensor fusion, characterized by, The method comprises the following steps: S1, collecting multi-source sensing data in the running process of the stamping die, the multi-source sensing data comprising vibration signals, temperature signals, pressure signals and acoustic emission signals; S2, performing data preprocessing on the collected multi-source sensing data, the data preprocessing comprising denoising, dimension unification and data standardization of each sensing signal; S3, performing adaptive time sequence alignment on the multi-source sensing data after data preprocessing, the adaptive time sequence alignment being to establish a time sequence synchronization benchmark according to the sampling frequency and response characteristics of different sensors, and to map all sensing signals to a unified time dimension through a dynamic time warping algorithm to eliminate time sequence asynchrony and time sequence deviation and ensure accurate matching of the multi-source sensing data in the time dimension; S4, extracting multi-dimensional time sequence features from the sensing data after adaptive time sequence alignment, the multi-dimensional time sequence features comprising time domain statistical features, frequency domain energy features, time-frequency analysis features and deep time sequence features automatically extracted through a time sequence convolution network; S5, performing deep fusion on the extracted multi-dimensional time sequence features, the deep fusion being to use a multi-modal fusion network, input the multi-dimensional time sequence features into the network, calculate the weight between different sensor features through a cross-modal attention mechanism, and perform weighted summation to construct a fused joint feature representation; S6, training a die wear prediction model based on the deep fused joint feature representation, the die wear prediction model being a sequence prediction model using a long short-term memory network; S7, predicting the current wear state and future wear trend of the stamping die by using the trained die wear prediction model.

2. The multi-sensor fusion based stamping die wear prediction method according to claim 1, wherein, The vibration signals in S1 are collected by a fiber optic vibration sensor, the temperature signals are collected by an infrared thermometer, the pressure signals are collected by a pressure sensor, and the acoustic emission signals are collected by a broadband acoustic emission sensor.

3. The multi-sensor fusion based stamping die wear prediction method according to claim 1, wherein, The data preprocessing in S2 further comprises anomaly value detection and correction, which is achieved by a sliding window median filtering method to remove noise interference and improve signal quality.

4. The multi-sensor fusion based stamping die wear prediction method of claim 1, wherein, The adaptive time sequence alignment in S3 specifically comprises: identifying the signal with the highest sampling frequency in the multi-source sensing data as a reference signal; performing interpolation processing on the signals with low sampling frequency to align them with the reference signal in time sequence; and dynamically adjusting the alignment parameters by calculating the time delay and correlation between different signals to ensure high-precision synchronization.

5. The multi-sensor fusion based stamping die wear prediction method of claim 1, wherein, The extraction of multi-dimensional time sequence features in S4 comprises time domain statistical features including mean, variance, kurtosis, skewness, kurtosis and root mean square value; frequency domain energy features including dominant frequency, energy entropy and frequency band energy ratio; time-frequency analysis features including wavelet packet energy spectrum and Hilbert-Huang transform marginal spectrum; and deep time sequence features automatically extracted through a time sequence convolution network by inputting the preprocessed and aligned time sequence signals into the time sequence convolution network to automatically learn high-level abstract features related to the die wear state.

6. The multi-sensor fusion based stamping die wear prediction method of claim 1, wherein, The deep fusion in S5 specifically includes: taking the extracted time domain, frequency domain and deep temporal features as the input layer of the multi-modal fusion network; the multi-modal fusion network includes multiple convolutional layers, pooling layers and fully connected layers, which are used to learn the representation within each feature modality; the multi-modal fusion network includes cross-modal attention mechanism and gating mechanism, which are used to adaptively weight and fuse the features of different modalities to strengthen the contribution of key features to wear prediction and suppress redundant information.

7. The multi-sensor fusion based stamping die wear prediction method of claim 1, wherein, The die wear prediction model in S6 divides the deep fused time series feature dataset into training set, validation set and test set before training; during model training, mean square error is used as the loss function; Adam optimizer is used for parameter iteration update, and learning rate decay strategy and early stopping mechanism are set to prevent model overfitting.

8. The multi-sensor fusion based stamping die wear prediction method of claim 1, wherein, The die wear prediction model in S7 has an output layer for outputting the wear grade or remaining service life of the stamping die. 9.The multi-sensor fusion based stamping die wear prediction method according to claim 1, wherein, After the training of the die wear prediction model in S7 is completed, the internal parameters of the model are adjusted through the back propagation algorithm to minimize the error between the predicted output and the actual wear label; the model also includes a validation phase to evaluate the generalization ability and prediction accuracy of the model on independent datasets.

10. The multi-sensor fusion based stamping die wear prediction method of claim 1, wherein, The die wear prediction model in S7 can receive process parameters representing different materials, stamping speeds and lubrication conditions as auxiliary inputs, so as to be applied to predict the wear evolution law of the stamping die under different process parameters.

Images