Intelligent punching control method and system for hardware stamping die

Abstract

The invention relates to the technical field of stamping control, and particularly discloses an intelligent punching control method and system for a hardware stamping die, and the method comprises the steps: carrying out the stroke feature engineering of stroke pressure time sequence data and stroke acoustic time sequence data in the hardware stamping process, so as to obtain the comprehensive state feature of the current stroke, and further, obtaining the comprehensive state feature of the current stroke; and performing anomaly detection on the current stroke comprehensive state characteristics by using the trained real-time anomaly detection model to obtain a current anomaly score, and determining whether a preliminary anomaly exists based on comparison between the anomaly score and a preset anomaly score threshold. If the initial anomaly exists, a recent anomaly score sequence and a recent high-frequency sound energy sequence are further collected, and the fault type is determined by conducting trend analysis on the two sequences. According to the method, the state of the hardware stamping die in the punching operation can be automatically monitored in real time, and potential faults are recognized and processed in time, so that stable operation of a production line is effectively guaranteed, and the product quality is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of stamping control technology, and more specifically, to an intelligent punching control method and system for metal stamping dies. Background Art

[0002] Metal stamping is an efficient and economical metal forming process in modern manufacturing, widely used in the automotive, home appliance, electronics, aerospace and other fields. Among them, punching (or punching) is a key process in metal stamping. Its quality and efficiency directly affect the performance and production cost of the final product.

[0003] When punching holes in metal stamping dies, the dies (especially the punch and die) will inevitably wear, chip, or even break under the long-term, high-frequency punching pressure. At the same time, factors such as uneven sheet material, thickness fluctuations, feeding errors, and failure to remove waste (chips) in a timely manner may also lead to poor punching, such as aperture deviation, excessive burrs, hole position offset, and even serious faults such as blockage and die jamming. These will not only lead to a decline in product quality, generate a large amount of waste, and increase production costs, but may also cause production line downtime and affect the execution of production plans. More seriously, sudden mold damage or die jamming may damage the punching equipment itself, resulting in high maintenance costs and longer downtime. Therefore, in actual production, timely identification and handling of potential faults in metal stamping dies during punching operations is of great significance for ensuring the stable operation of the production line and improving product quality.

[0004] Existing technology primarily relies on manual inspections and empirical judgment to identify faults in the metal stamping and punching process. This approach is highly subjective, lacks real-time performance, and struggles to detect subtle anomalies early on. Some solutions incorporate sensor-based monitoring systems, such as those that monitor pressure peaks or equipment vibration, to identify abnormal conditions. However, most rely solely on simple fixed-threshold alarms, making them susceptible to noise, poorly adaptable to changing operating conditions, prone to false alarms or missed alarms, and unable to provide specific information about the fault type.

[0005] Therefore, an optimized intelligent punching control method and system for metal stamping dies are expected. Summary of the Invention

[0006] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an intelligent punching control method and system for metal stamping dies, which performs stroke feature engineering on the stroke pressure time series data and stroke acoustic time series data in the metal stamping process to obtain the comprehensive state characteristics of the current stroke, and then uses the trained real-time anomaly detection model to perform anomaly detection on the current stroke comprehensive state characteristics to obtain the current anomaly score, and based on the comparison between the anomaly score and the preset anomaly score threshold, determines whether there is a preliminary anomaly. If there is a preliminary anomaly, further collect the recent anomaly score sequence and the recent high-frequency acoustic energy sequence, and perform trend analysis on the two to determine the fault type. This method can automatically and in real time monitor the status of the metal stamping die during the punching operation, identify and handle potential faults in a timely manner, thereby effectively ensuring the stable operation of the production line and improving product quality.

[0007] Accordingly, according to one aspect of the present application, there is provided an intelligent punching control method for a metal stamping die, comprising:

[0008] Acquiring stroke pressure time series data and stroke acoustic time series data collected by a punch press pressure sensor and an acoustic sensor;

[0009] performing stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of a current stroke state;

[0010] Inputting the current stroke state comprehensive feature vector into a trained real-time anomaly detection model to obtain a current anomaly score and determining a preliminary anomaly flag based on a comparison between the anomaly score and a preset anomaly score threshold;

[0011] In response to the preliminary abnormality flag being true, collecting a recent abnormality score sequence and a recent high-frequency acoustic energy sequence;

[0012] Performing trend analysis on the recent anomaly score sequence and the recent high-frequency sound energy sequence to obtain a recent anomaly score trend and a recent high-frequency sound energy trend;

[0013] A fault type is determined based on the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend.

[0014] According to another aspect of the present application, an intelligent punching control system for a metal stamping die is provided, which includes:

[0015] A stroke sensing monitoring module is used to obtain stroke pressure time series data and stroke acoustic time series data collected by the punch press pressure sensor and acoustic sensor;

[0016] a stroke feature engineering module, configured to perform stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of a current stroke state;

[0017] an anomaly detection module, configured to input the current stroke state comprehensive feature vector into a trained real-time anomaly detection model to obtain a current anomaly score and determine a preliminary anomaly flag based on a comparison between the anomaly score and a preset anomaly score threshold;

[0018] a preliminary abnormality response module, configured to collect a recent abnormality score sequence and a recent high-frequency acoustic energy sequence in response to the preliminary abnormality flag being true;

[0019] an abnormality trend analysis module, configured to perform trend analysis on the recent abnormality score sequence and the recent high-frequency sound energy sequence to obtain a recent abnormality score trend and a recent high-frequency sound energy trend;

[0020] A fault type determination module is configured to determine a fault type based on the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend.

[0021] Compared with the prior art, the intelligent punching control method and system for metal stamping dies provided in this application performs stroke feature engineering on the stroke pressure time series data and stroke acoustic time series data in the metal stamping process to obtain the comprehensive state characteristics of the current stroke, and then uses the trained real-time anomaly detection model to perform anomaly detection on the current stroke comprehensive state characteristics to obtain the current anomaly score, and based on the comparison between the anomaly score and the preset anomaly score threshold, determines whether there is a preliminary anomaly. If there is a preliminary anomaly, further collect the recent anomaly score sequence and the recent high-frequency acoustic energy sequence, and perform trend analysis on the two to determine the fault type. This method can automatically and in real time monitor the status of the metal stamping die during the punching operation, identify and handle potential faults in a timely manner, thereby effectively ensuring the stable operation of the production line and improving product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0023] Figure 1 This is a flow chart of an intelligent punching control method for a metal stamping die according to an embodiment of the present application.

[0024] Figure 2 Schematic diagram of data flow of an intelligent punching control method for a metal stamping die according to an embodiment of the present application.

[0025] Figure 3 This is a flowchart of step S2 in the intelligent punching control method for a metal stamping die according to an embodiment of the present application.

[0026] Figure 4 This is a flowchart of step S5 in the intelligent punching control method for a metal stamping die according to an embodiment of the present application.

[0027] Figure 5 This is a block diagram of an intelligent punching control system for a metal stamping die according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] Below, an example embodiment according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiment is only a part of the embodiment of the present application, not all of the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiment described herein. It is worth noting that in the present application, all actions to obtain data are carried out in compliance with the relevant data protection laws and policies of the country where they are located, and with the authorization of the corresponding device owner.

[0029] Figure 1 This is a flow chart of an intelligent punching control method for a metal stamping die according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the intelligent punching control method for metal stamping dies according to the embodiment of the present application. Figure 1 and Figure 2 As shown, the intelligent punching control method for metal stamping dies according to the embodiment of the present application includes the following steps: S1, acquiring stroke pressure time series data and stroke acoustic time series data collected by a punch pressure sensor and an acoustic sensor; S2, performing stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of the current stroke state; S3, inputting the comprehensive feature vector of the current stroke state into a trained real-time anomaly detection model to obtain a current anomaly score and determining a preliminary anomaly flag based on a comparison between the anomaly score and a preset anomaly score threshold; S4, in response to the preliminary anomaly flag being true, collecting a recent anomaly score sequence and a recent high-frequency acoustic energy sequence; S5, performing trend analysis on the recent anomaly score sequence and the recent high-frequency acoustic energy sequence to obtain a recent anomaly score trend and a recent high-frequency acoustic energy trend; S6, determining the fault type based on the current anomaly score, the recent anomaly score trend and the recent high-frequency acoustic energy trend.

[0030] In the above-mentioned intelligent punching control method for metal stamping dies, the step S1 obtains the stroke pressure time series data and stroke acoustic time series data collected by the punch pressure sensor and the acoustic sensor. It should be understood that during the metal stamping and punching process, the interaction between the punch and the sheet metal will produce unique pressure changes and acoustic signals, which directly reflect the instantaneous state of the stamping behavior. Therefore, considering that the punch pressure and stroke acoustic data contain rich information about the stamping process and that machine learning can mine data feature patterns, this application is based on sensor technology, by integrating pressure sensors and acoustic sensors on the punch, and configuring a corresponding data acquisition system to achieve synchronous capture of the pressure and acoustic characteristics of each stroke.

[0031] In a specific embodiment of the present application, first, a sensor is installed on the key component of the punch press. The pressure sensor should have a high dynamic response frequency (for example, at least several thousand hertz to capture the rapid force changes during the punching process), high sensitivity, good linearity and overload resistance, such as a piezoelectric or strain gauge pressure sensor, which is installed on the punch holder, under the mold base, or on a force measuring component directly or indirectly connected to the punch slide, so that it can directly sense the dynamic pressure changes applied by the punch to the sheet metal throughout the stroke. The acoustic sensor can be a contact acoustic emission (AE) sensor or a non-contact microphone array. The acoustic emission sensor is usually attached or magnetically attached to the mold surface (such as near the die or punch) or the mold base to pick up structural sound waves generated by material deformation, crack expansion, friction, etc., and is particularly sensitive to high-frequency signals (usually in the range of 20kHz to 1MHz). The microphone array is arranged near the stamping area to collect the air sound generated by the stamping process. In order to shield the interference of the workshop background noise, a directional microphone or an acoustic shielding cover can be used.

[0032] Next, configure the data acquisition system (DAQ). This system includes a signal conditioning module, an analog-to-digital converter (ADC), and a data transmission interface. The signal conditioning module amplifies and filters the raw electrical signals output by the sensors (for example, low-pass filtering for pressure signals to remove high-frequency noise and band-pass filtering for acoustic signals to focus on specific frequency bands), and performs anti-aliasing processing. The ADC converts analog signals into digital signals for subsequent data processing and analysis. To ensure strict synchronization of the pressure and acoustic signals during data acquisition, these signals can be synchronized by sharing a sampling clock or using a high-precision synchronous trigger signal (for example, using the top dead center signal or slide position signal provided by the punch press crankshaft encoder as the starting trigger for each stroke data acquisition). During a stroke cycle (for example, starting from a certain position before the punch contacts the sheet material and ending when the punch completes the cut and returns to the position before the initial contact point), the DAQ system continuously acquires data from the pressure and acoustic sensors and arranges the acquired data points in chronological order, forming the pressure and acoustic time series data for that stroke. After data acquisition is complete, it is first cached in the DAQ system's memory and then transmitted in real time or near real time to a host computer or edge computing unit via Ethernet, USB, or other industrial bus interfaces for subsequent processing. The data transmission format can be a raw binary stream or pre-packaged into formats such as CSV, TDMS, HDF5, etc., with precise timestamps and stroke numbers. In this way, synchronized, high-resolution raw pressure and acoustic time series data containing rich state information can be obtained for each stamping stroke, providing high-quality raw input for subsequent anomaly detection and fault diagnosis.

[0033] In the above-mentioned intelligent punching control method for metal stamping dies, the step S2 performs stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of the current stroke state. It should be understood that the original stroke pressure time series data and the stroke acoustic time series data have high data dimensions and information redundancy, and it is difficult to directly and efficiently reflect the key state changes of the stamping process, such as the degree of wear of the mold, the hardness change of the material, or whether blockage occurs. Therefore, in order to extract more representative current stroke health status information from the original time series data and integrate the complementary advantages of multimodal sensing data, this application performs feature extraction in the time domain, frequency domain, etc. on the stroke pressure time series data and the stroke acoustic time series data, and integrates the extracted multi-dimensional heterogeneous features to achieve a comprehensive characterization of the current stroke state.

[0034] Figure 3 FIG is a flow chart of step S2 in the intelligent punching control method for metal stamping dies according to an embodiment of the present application. Figure 3As shown, the step S2 includes: S21, performing pressure data feature extraction on the stroke pressure time series data to obtain stroke pressure time series feature parameters, and the stroke pressure time series feature parameters include peak pressure, shear slope, breakthrough point pressure, stroke area and waveform similarity; S22, performing acoustic data feature extraction on the stroke acoustic time series data to obtain stroke acoustic time series feature parameters, and the stroke acoustic time series feature parameters include total energy, high-frequency sound energy, time domain peak and spectrum characteristics; S23, performing multimodal feature integration on the stroke pressure time series feature parameters and the stroke acoustic time series feature parameters to obtain the comprehensive feature vector of the current stroke state.

[0035] Specifically, in step S21, the stroke pressure time series data is subjected to pressure data feature extraction to obtain stroke pressure time series characteristic parameters, wherein the stroke pressure time series characteristic parameters include peak pressure, punching slope, breaking point pressure, stroke area and waveform similarity. It should be understood that the peak pressure represents the maximum pressure value applied by the punch to the sheet during the punching process, reflects the maximum impact force required for the punching process, and is closely related to the material thickness, strength, mold gap, sharpness, etc. The punching slope refers to the slope of the main linear segment in the rising stage of the pressure curve, that is, when the punch begins to shear the material but has not yet completely broken it, reflecting the shear resistance characteristics of the material and the sharpness of the mold edge, and can reveal the stability and efficiency of the stamping process. The breaking point pressure is the pressure value at the moment the punch penetrates the sheet, which is usually the inflection point where the pressure begins to drop sharply from the peak or near-peak area, and is related to factors such as the thickness, hardness and mold design of the material. The stroke area represents the integral of the pressure variation over time—that is, the area enclosed by the pressure-time curve and the time axis during the blanking phase—and reflects the work performed by the punch on the sheet metal throughout the entire stroke. Waveform similarity evaluates the consistency and stability of the stamping process by calculating the degree of similarity between the current stroke pressure waveform and a pre-stored standard waveform for a normal stroke. These stroke pressure time-series characteristic parameters reflect the interaction between the die and material and the health of the stamping process from different perspectives, providing critical information for subsequent fault diagnosis. For example, an abnormal increase in peak pressure may indicate increased die edge wear or an abnormal increase in material hardness; a decrease in the shear slope may indicate a blunting of the die edge, leading to an unstable stamping process; abnormal fluctuations in the breakpoint pressure may be related to improper die clearance or uneven material hardness; a decrease in stroke area may indicate reduced stamping efficiency, possibly due to increased energy loss caused by die wear or increased material hardness; and a decrease in waveform similarity may indicate a decrease in stamping process consistency, possibly due to a mold or equipment malfunction. Based on this, this application extracts the above-mentioned stroke pressure timing characteristic parameters to achieve sensitive capture and accurate description of key state changes in the stamping process.

[0036] In a specific embodiment of the present application, for peak pressure, the maximum value of the pressure value is extracted by traversing the stroke pressure time series data to obtain the peak pressure; for the shear slope, the stroke pressure time series data is linearly fitted using the least squares method to obtain the slope of the fitting line as the shear slope; for the breakthrough point pressure, the second-order derivative of the pressure curve is calculated to find the inflection point where the second-order derivative turns from positive to negative, and the pressure value corresponding to the inflection point is the breakthrough point pressure; for the stroke area, the stroke pressure time series data is integrated to obtain the stroke area; for waveform similarity, the dynamic time warping (DTW) algorithm can be used to calculate the minimum distance between the current stroke pressure waveform and the pre-stored normal stroke standard waveform, or the cosine similarity can be calculated after the waveforms are normalized to the same length.

[0037] Specifically, step S22 performs acoustic data feature extraction on the stroke acoustic time series data to obtain stroke acoustic time series characteristic parameters, which include total energy, high-frequency acoustic energy, time-domain peak value, and spectral characteristics. It should be understood that total energy is the sum of the energy of the acoustic signal throughout the entire stroke time, reflecting the overall intensity of the acoustic signal during the stroke and related to the energy release and vibration of the stamping process. High-frequency acoustic energy is acoustic energy in a specific high-frequency band, focusing on the high-frequency components in the acoustic signal. It is usually related to events such as material fracture and mold impact. Therefore, high-frequency acoustic energy can be used as an important indicator for evaluating the impact characteristics of the stamping process. The time-domain peak value represents the maximum amplitude in the acoustic signal, reflecting the strongest instantaneous impact of the acoustic event generated during the stamping process, and is related to the impact force between the mold and the material and the sound of the material fracture. The spectral characteristics are the frequency distribution characteristics of the acoustic signal obtained through spectral analysis methods such as Fourier transform. They reflect the spectral components and changes of the acoustic signal during the stamping process and can provide important clues for analyzing the dynamic characteristics of the stamping process. For example, an increase in a specific frequency component within a spectral signature may indicate a specific type of vibration within the mold or a specific fracture mode within the material. Based on this, this application further enriches the description of the stamping process and improves the accuracy and reliability of fault diagnosis by extracting and analyzing the aforementioned stroke acoustic timing characteristic parameters.

[0038] In a specific embodiment of the present application, for total energy, the root mean square of the sum of squares of the stroke acoustic time series data is calculated to obtain the total energy; for high-frequency sound energy, the stroke acoustic time series data is first converted from the time domain to the frequency domain signal by fast Fourier transform (FFT), and then the high frequency band of 5kHz-20kHz is selected to calculate the total energy value in the frequency band to obtain the high-frequency sound energy; for the time domain peak, the maximum amplitude of the acoustic signal is extracted by traversing the stroke acoustic time series data to obtain the time domain peak; for spectral characteristics, the stroke acoustic time series data is converted from the time domain to the frequency domain using Fourier transform to obtain the spectral distribution of the acoustic signal, and then the spectral characteristics are analyzed to calculate the main frequency of the spectrum (the frequency point with the highest energy), the center of mass of the spectrum (the center frequency of the spectral energy distribution), the spectral bandwidth (the range of the spectral energy distribution) and the flatness of the spectrum (the uniformity of the spectral energy distribution) and other parameters to comprehensively describe the spectral characteristics of the acoustic signal.

[0039] Specifically, in step S23, the stroke pressure timing characteristic parameters and the stroke acoustic timing characteristic parameters are subjected to multimodal feature integration to obtain the comprehensive feature vector of the current stroke state. It should be understood that the stroke pressure timing characteristic parameters and the stroke acoustic timing characteristic parameters respectively reflect the state of the stamping process from different aspects. Relying solely on a single type of characteristic parameter cannot fully and comprehensively describe the operation of the mold, and it is difficult to accurately judge complex fault types. Therefore, in order to more accurately reflect the true state of the stamping process and improve the accuracy and reliability of fault diagnosis, the present application constructs a comprehensive feature representation of the current stroke state by conducting multimodal feature integration on the stroke pressure timing characteristic parameters and the stroke acoustic timing characteristic parameters.

[0040] In a specific embodiment of the present application, step S23 includes: arranging the stroke pressure time series characteristic parameters and the stroke acoustic time series characteristic parameters in a predetermined order to obtain the original characteristic vector of the current stroke state; and normalizing the original characteristic vector of the current stroke state to obtain the comprehensive characteristic vector of the current stroke state. Specifically, first, the stroke pressure time series characteristic parameters (peak pressure, shear slope, breakthrough point pressure, stroke area and waveform similarity) and the stroke acoustic time series characteristic parameters (total energy, high-frequency sound energy, time domain peak and spectrum characteristics) are arranged in a predetermined and fixed order to form a fixed-length original characteristic vector of the current stroke state. Furthermore, in order to eliminate the influence of the differences in dimensions and numerical ranges between different characteristic parameters on subsequent analysis, the Z-score normalization method is used to normalize each characteristic parameter in the original characteristic vector of the current stroke state, thereby obtaining a standardized comprehensive characteristic vector of the current stroke state. Here, the mean and standard deviation parameters used in the Z-score normalization method are obtained based on a large amount of healthy stroke data statistics.

[0041] In the above-mentioned intelligent punching control method for metal stamping dies, in step S3, the current stroke state comprehensive feature vector is input into the trained real-time anomaly detection model to obtain the current anomaly score and determine the preliminary anomaly flag based on the comparison between the anomaly score and the preset anomaly score threshold. It should be understood that, considering that the alarm method based on a fixed threshold in the prior art has poor adaptability to changes in working conditions. Therefore, in order to improve the adaptability of the intelligent punching control method to changes in working conditions, the present application uses machine learning technology to train a real-time anomaly detection model to dynamically determine whether the stamping process is abnormal. More specifically, the training data of the trained real-time anomaly detection model are stroke pressure time series data and stroke acoustic time series data marked as healthy strokes. That is, considering that in the training process of the real-time anomaly detection model, stroke state abnormality samples are scarce and diverse and difficult to obtain, while normal stroke state samples are relatively abundant. Therefore, based on the idea of ​​one-class learning in machine learning, this application trains the real-time anomaly detection model using only stroke pressure time series data and stroke acoustic time series data labeled as healthy strokes, so that the model learns the normal characteristic patterns of acoustic signals and pressure signals in healthy strokes. After the model training is completed, the real-time anomaly detection model evaluates the input current stroke state comprehensive feature vector in real time based on the learned normal stamping pattern, thereby calculating the degree of deviation between the current stroke state and the normal stroke state and obtaining the anomaly score of the current stroke.

[0042] In a specific embodiment of the present application, a large amount of stroke pressure time series data and stroke acoustic time series data marked as healthy strokes are used as training data, and the training data is processed based on the above-mentioned stroke feature engineering, the training sample stroke state comprehensive feature vector is extracted, and a training sample set is constructed. Then, the autoencoder model in deep learning is used for training. The autoencoder consists of two parts, an encoder and a decoder. The encoder compresses the input training sample stroke state comprehensive feature vector into a low-dimensional code, and the decoder reconstructs the low-dimensional code into the original feature vector. During the training process, by minimizing the mean square error (MSE) between the input training sample stroke state comprehensive feature vector and the reconstructed feature vector, the model is forced to learn the characteristic distribution law of healthy stroke data. After the training is completed, the current stroke state comprehensive feature vector is input into the trained autoencoder, and the autoencoder encodes and reconstructs it according to the learned healthy stroke data characteristic distribution law. The reconstruction error, that is, the mean square error (MSE) between the current stroke state comprehensive feature vector and the reconstructed feature vector output by the autoencoder, is calculated as the abnormality score of the current stroke. Here, the abnormality score reflects the degree of deviation between the current stroke state and the normal stroke state. The higher the score, the greater the difference between the current stroke state and the normal stroke state, that is, the greater the possibility of abnormality. Furthermore, a preset abnormality score threshold is set. In one example of the present application, the preset abnormality score threshold is the 99th percentile of the abnormality score distribution obtained after the model training is completed, based on a large amount of healthy stroke state data. The abnormality score of the current stroke is compared with the preset abnormality score threshold. If the current abnormality score is greater than or equal to the preset abnormality score threshold, the preliminary abnormality flag is determined to be true, indicating that the current stroke state may have a fault or abnormality; if the current abnormality score is less than the preset abnormality score threshold, the preliminary abnormality flag is determined to be false, indicating that the current stroke state is in a healthy or normal state. In this way, real-time abnormality detection can be achieved during the intelligent punching process of metal stamping dies, improving the accuracy and timeliness of fault diagnosis, providing strong support for mold maintenance and servicing, and further ensuring production efficiency and product quality.

[0043] In the aforementioned intelligent punching control method for metal stamping dies, step S4, in response to the initial abnormality flag being true, collects recent abnormality score sequences and recent high-frequency acoustic energy sequences. It should be understood that after the initial abnormality is determined, it is difficult to accurately determine the fault type and development trend based solely on the current abnormality score. Therefore, in order to deeply analyze the abnormal situation and accurately determine the fault type, this application further collects historical information that can reflect the evolution trend of the abnormal state. By backtracking and collecting recent abnormality score sequences and high-frequency acoustic energy sequences, a richer contextual basis is provided for subsequent fault type diagnosis. Here, the recent abnormality score sequence can reflect the changing trend between the current abnormal state and the previous state, helping to identify the duration and development trend of the abnormality. The recent high-frequency acoustic energy sequence can reflect the changes in the acoustic characteristics produced by the mold under abnormal conditions, which are often associated with specific fault types. By collecting recent abnormality score sequences and recent high-frequency acoustic energy sequences, key time window data reflecting the evolution of the stroke abnormality level and the physical response of the potential fault can be quickly obtained, providing the necessary multi-dimensional input with temporal context for subsequent fault mode identification.

[0044] In one embodiment of the present application, a fixed-length first-in, first-out (FIFO) buffer is maintained within the system to continuously store the anomaly scores for the most recent L strokes and the high-frequency acoustic energy values ​​calculated using the aforementioned method. When a preliminary anomaly is triggered, the anomaly scores and high-frequency acoustic energy values ​​corresponding to the most recent N strokes (L>N, where N is typically set to 20 strokes) are extracted from this buffer to construct a recent anomaly score sequence and a recent high-frequency acoustic energy sequence.

[0045] In the above-mentioned intelligent punching control method for metal stamping dies, the step S5 performs trend analysis on the recent abnormal score sequence and the recent high-frequency acoustic energy sequence to obtain the recent abnormal score trend and the recent high-frequency acoustic energy trend. It should be understood that different fault modes (such as mold wear, cracks, material jamming, insufficient lubrication, and foreign body intrusion) will show different statistical characteristics and change patterns in the evolution trend of abnormal scores and high-frequency acoustic energy. Therefore, in order to accurately distinguish the fault types, this application is based on the principle of pattern recognition, and extracts key trend indicators by performing statistical trend analysis on the recent abnormal score sequence and the recent high-frequency acoustic energy sequence, thereby providing an effective basis for subsequent fault mode identification.

[0046] In a specific embodiment of the present application, the recent anomaly score trend includes the anomaly score mean, anomaly score standard deviation, and anomaly score slope. The anomaly score mean reflects the average level of recent anomaly conditions and is calculated by calculating the arithmetic mean of the recent anomaly score sequence. A higher mean indicates a persistent and significant anomaly. The anomaly score standard deviation measures the degree of fluctuation in the anomaly score and is calculated by calculating the standard deviation of the recent anomaly score sequence. A larger standard deviation indicates more dramatic fluctuations in the anomaly condition, potentially indicating a rapidly developing fault. The anomaly score slope reflects the rate of change of the anomaly score over time and is calculated by performing a least squares linear fit on the recent anomaly score sequence. A larger slope indicates a rapidly worsening anomaly. The recent high-frequency acoustic energy trend includes the high-frequency acoustic energy mean and the high-frequency acoustic energy standard deviation. The high-frequency acoustic energy mean reflects the average intensity of the acoustic signature generated by the mold during an abnormal condition. A higher mean may indicate that the fault type is closely related to the mold's vibration or impact characteristics. The high-frequency acoustic energy standard deviation measures the degree of fluctuation in the high-frequency acoustic energy. A larger standard deviation may indicate that the acoustic signature generated by the fault exhibits significant variations in intensity and frequency, helping to further distinguish different fault modes. Through comprehensive analysis of the above-mentioned recent abnormal score trend indicators and recent high-frequency sound energy trend indicators, the essential characteristics and development laws of the fault can be grasped more accurately, providing strong support for the accurate identification of subsequent fault modes.

[0047] In particular, considering that the abnormality score and the recent abnormality score trend are both determined by abnormality score detection based on the stroke pressure characteristic data and the stroke acoustic characteristic data, the fault type decision based on the current high-frequency sound energy, the current abnormality score, the recent abnormality score trend and the recent high-frequency sound energy trend is essentially a multi-physical field signal and coupling trend decision, and is therefore also affected by the non-stationary characteristics of the multi-physical field signal coupling, thereby affecting the decision accuracy.

[0048] Therefore, the recent abnormal score sequence and the recent high-frequency sound energy sequence for trend analysis are set to be a1, a2, ..., a n and b1,b2,…,b n , where a n and b n That is, the current anomaly score and the current high-frequency sound energy, then it is expected that for the anomaly score sequence a1, a2, ..., a n and the high frequency sound energy sequence b1, b2, ..., b nJoint spatial embedding is performed to compensate for the local trend coupling irregularities of the signal sequence. Based on this, in a preferred embodiment of the present application, step S5 includes: first, performing multi-physics signal coupling stationarity optimization on the recent anomaly score sequence and the recent high-frequency acoustic energy sequence to obtain an optimized recent anomaly score sequence and an optimized recent high-frequency acoustic energy sequence; then, based on the optimized recent anomaly score sequence and the optimized recent high-frequency acoustic energy sequence, calculating the recent anomaly score trend and the recent high-frequency acoustic energy trend.

[0049] Specifically, first, the recent anomaly score sequence and the recent high-frequency sound energy sequence are mapped based on the anomaly score-high-frequency sound energy correlation state response trajectory to obtain the anomaly score correlation state response trajectory mapping vector and the high-frequency sound energy correlation state response trajectory mapping vector. More specifically, first, the recent anomaly score sequence a1, a2, ..., a n The anomaly score vector A=(a1,a2,…,a n ) and the recent high-frequency sound energy sequence b1, b2, ..., b n The high-frequency sound energy vector B=(b1,b2,…,b n ) in,(·) T represents the transpose of a vector, Represents vector multiplication operation. Then calculate the anomaly score associated state response trajectory mapping vector and high-frequency sound energy associated state response trajectory mapping vector Among them, M -1 Represents the inverse matrix of matrix M. In this way, the anomaly score and high energy vector and their time series trend representation are jointly dimensionally mapped while considering the possible loss of local sequence trend coupling caused by abnormal correlation state, so as to represent the structured response characteristics of the system correlation state under the predetermined time series length.

[0050] Then, based on the anomaly score associated state response trajectory mapping vector and the high-frequency sound energy associated state response trajectory mapping vector, the anomaly score associated state response trajectory curvature vector and the high-frequency sound energy associated state response trajectory curvature vector are calculated. That is, based on the spatial canonicalization mapping, the square of the gradient pattern is used to represent the curvature correspondence of each trajectory in the joint space, which is expressed as:

[0051]

[0052] in, represents the difference by position, ⊙ represents the multiplication by position, (·) ⊙-1 Indicates taking the reciprocal of each position, (·) ⊙2represents the position-by-position square calculation, A” represents the curvature vector of the state response trajectory associated with the anomaly score, and B” represents the curvature vector of the state response trajectory associated with the high-frequency sound energy.

[0053] In this way, based on the transition probability between the curvature vector of the state response trajectory associated with the anomaly score and the curvature vector of the state response trajectory associated with the high-frequency sound energy, the recent anomaly score sequence and the recent high-frequency sound energy sequence can be mapped priori to obtain the optimized recent anomaly score sequence and the optimized recent high-frequency sound energy sequence. That is, the prior mapping is performed using the sequence transition probability as the joint likelihood distribution, which is expressed as:

[0054]

[0055] Among them, M c Represents the transition probability matrix between A' and B', A c represents the optimized recent anomaly score sequence, B c Represents the optimized recent high-frequency acoustic energy sequence.

[0056] Thus, the recent abnormal score sequence a1, a2, ..., a n and the recent high-frequency sound energy sequence b1, b2, ..., b n The joint spatial embedding optimization representation A c and B c , thereby improving the accuracy of the current high-frequency sound energy, the current anomaly score, the recent anomaly score trend, and the recent high-frequency sound energy trend.

[0057] In the above-mentioned intelligent punching control method for metal stamping dies, the step S6 determines the fault type based on the current abnormality score, the recent abnormality score trend and the recent high-frequency sound energy trend. That is, by comprehensively utilizing the instantaneous abnormality information of the current stroke and the recent dynamic trend information, the statistical characteristics and change laws corresponding to different fault modes are identified to infer the specific fault type that may occur, thereby improving the accuracy and guidance of fault diagnosis. Figure 4 As shown, the step S6 includes: S61, extracting the current high-frequency sound energy from the stroke acoustic timing characteristic parameters; S62, inputting the current high-frequency sound energy, the current abnormality score, the recent abnormality score trend and the recent high-frequency sound energy trend into a decision tree model to obtain the fault type.

[0058] Specifically, first, the current high-frequency sound energy is extracted from the stroke acoustic time series feature parameters, and then the current high-frequency sound energy, the current anomaly score, the recent anomaly score trend (anomaly score average, anomaly score standard deviation and anomaly score slope) and the recent high-frequency sound energy trend (high-frequency sound energy average and high-frequency sound energy standard deviation) are used as input features of the decision tree model. It should be understood that the decision tree is a common machine learning classification model that divides data into different categories by learning a series of splitting rules. In the present application, the decision tree model is constructed by training a large amount of sample data with fault type labels, wherein each sample data contains the above 7 features and the corresponding actual fault type (such as "potential blockage", "mold wear", "cutting edge cracking", "waste not discharged", etc., these labels are manually recorded after the fault occurs). During the training process, information gain is used as the metric for feature selection, and the feature with the largest information gain is selected as the node for splitting, and the decision tree structure is gradually constructed. When the trained decision tree model is applied, it will start from the root node according to the value of the input feature, and make judgments along the corresponding branches according to the size of the feature value, and finally reach the leaf node. The label corresponding to the leaf node is the predicted fault type. For example, the root node may be "Is the current anomaly score greater than X?" If so, go to the left subtree and then judge "Is the current high-frequency sound energy greater than Y?", and so on, until it reaches the leaf node. In a specific example of the present application, if the current anomaly score is greater than the medium threshold and less than the extremely high threshold, and the slope of the anomaly score is less than a certain positive threshold and the current high-frequency sound energy is less than the chipping feature threshold, the fault type is determined to be potential blockage. Here, the medium threshold, the extremely high threshold, the positive threshold and the chipping feature threshold are automatically learned and determined by the decision tree training algorithm in the process of node splitting and constructing the tree. In practical applications, in order to improve the generalization ability and accuracy of the decision tree model, the cross-validation method is used to evaluate and optimize the model and adjust the parameters of the decision tree, such as the maximum depth, the minimum number of sample splits, etc. At the same time, training data is regularly updated, and the model is retrained based on new production situations and failure cases to ensure that the model can adapt to different working conditions and accurately judge various failure types. This method can quickly and accurately determine the type of failure, providing clear guidance for timely and effective repair and adjustment measures, which in turn helps significantly improve the stability and production efficiency of the metal stamping die punching process and effectively reduce production losses caused by failures.

[0059] In summary, according to the embodiment of the present application, the intelligent punching control method for metal stamping dies is explained, which performs stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data in the metal stamping process to obtain the comprehensive state characteristics of the current stroke, and then uses the trained real-time anomaly detection model to perform anomaly detection on the current stroke comprehensive state characteristics to obtain the current anomaly score, and based on the comparison between the anomaly score and the preset anomaly score threshold, determines whether there is a preliminary anomaly. If there is a preliminary anomaly, further collect the recent anomaly score sequence and the recent high-frequency acoustic energy sequence, and perform trend analysis on the two to determine the fault type. This method can automatically and in real time monitor the status of the metal stamping die during the punching operation, identify and handle potential faults in a timely manner, thereby effectively ensuring the stable operation of the production line and improving product quality.

[0060] Furthermore, the present application also provides an intelligent punching control system for metal stamping dies.

[0061] Figure 5 FIG is a block diagram of an intelligent punching control system for a metal stamping die according to an embodiment of the present application. Figure 5 As shown, the intelligent punching control system 100 for metal stamping dies according to the embodiment of the present application includes: a stroke sensing monitoring module 110, which is used to obtain the stroke pressure time series data and the stroke acoustic time series data collected by the punch pressure sensor and the acoustic sensor; a stroke feature engineering module 120, which is used to perform stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a current stroke state comprehensive feature vector; an anomaly detection module 130, which is used to input the current stroke state comprehensive feature vector into the trained real-time anomaly detection model to obtain the current anomaly. The system comprises a first abnormality response module 140 for collecting a recent abnormality score sequence and a recent high-frequency acoustic energy sequence in response to the preliminary abnormality flag being true; an abnormality trend analysis module 150 for performing trend analysis on the recent abnormality score sequence and the recent high-frequency acoustic energy sequence to obtain a recent abnormality score trend and a recent high-frequency acoustic energy trend; and a fault type determination module 160 for determining a fault type based on the current abnormality score, the recent abnormality score trend, and the recent high-frequency acoustic energy trend.

[0062] Here, those skilled in the art will appreciate that the specific operations of each module in the intelligent punching control system for metal stamping dies have been described in the above description. Figures 1 to 4 The invention has been introduced in detail in the description of the intelligent punching control method for metal stamping dies, and therefore, its repeated description will be omitted.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent punching control method for a metal stamping die, characterized in that: include: Acquiring stroke pressure time series data and stroke acoustic time series data collected by a punch press pressure sensor and an acoustic sensor; performing stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of a current stroke state; Inputting the current stroke state comprehensive feature vector into a trained real-time anomaly detection model to obtain a current anomaly score and determining a preliminary anomaly flag based on a comparison between the anomaly score and a preset anomaly score threshold; In response to the preliminary abnormality flag being true, collecting a recent abnormality score sequence and a recent high-frequency acoustic energy sequence; Performing trend analysis on the recent anomaly score sequence and the recent high-frequency sound energy sequence to obtain a recent anomaly score trend and a recent high-frequency sound energy trend; A fault type is determined based on the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend.

2. The intelligent punching control method for metal stamping dies according to claim 1, characterized in that: Performing stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of the current stroke state, including: performing pressure data feature extraction on the stroke pressure time series data to obtain stroke pressure time series characteristic parameters, wherein the stroke pressure time series characteristic parameters include peak pressure, punching slope, breaking point pressure, stroke area, and waveform similarity; Performing acoustic data feature extraction on the stroke acoustic time series data to obtain stroke acoustic time series feature parameters, wherein the stroke acoustic time series feature parameters include total energy, high-frequency sound energy, time domain peak value and spectrum characteristics; Multimodal feature integration is performed on the stroke pressure time series feature parameters and the stroke acoustic time series feature parameters to obtain the current stroke state comprehensive feature vector.

3. The intelligent punching control method for metal stamping dies according to claim 2, characterized in that: Performing multimodal feature integration on the stroke pressure time series feature parameters and the stroke acoustic time series feature parameters to obtain the current stroke state comprehensive feature vector, including: Arranging the stroke pressure time series characteristic parameters and the stroke acoustic time series characteristic parameters in a predetermined order to obtain a current stroke state original characteristic vector; The original feature vector of the current stroke state is normalized to obtain the comprehensive feature vector of the current stroke state.

4. The intelligent punching control method for a metal stamping die according to claim 1, characterized in that: The training data of the trained real-time anomaly detection model are stroke pressure time series data and stroke acoustic time series data marked as healthy strokes.

5. The intelligent punching control method for metal stamping dies according to claim 1, characterized in that: The recent anomaly score trend includes an anomaly score average, an anomaly score standard deviation, and an anomaly score slope; the recent high-frequency sound energy trend includes an high-frequency sound energy average and a high-frequency sound energy standard deviation.

6. The intelligent punching control method for metal stamping dies according to claim 1, characterized in that: Performing trend analysis on the recent anomaly score sequence and the recent high-frequency sound energy sequence to obtain a recent anomaly score trend and a recent high-frequency sound energy trend, including: Performing multi-physics signal coupling stationarity optimization on the recent anomaly score sequence and the recent high-frequency acoustic energy sequence to obtain an optimized recent anomaly score sequence and an optimized recent high-frequency acoustic energy sequence; Based on the optimized recent anomaly score sequence and the optimized recent high-frequency acoustic energy sequence, a recent anomaly score trend and a recent high-frequency acoustic energy trend are calculated.

7. The intelligent punching control method for a metal stamping die according to claim 6, characterized in that: Performing multi-physics signal coupling stationarity optimization on the recent anomaly score sequence and the recent high-frequency acoustic energy sequence to obtain an optimized recent anomaly score sequence and an optimized recent high-frequency acoustic energy sequence, including: Performing response trajectory mapping based on anomaly score-high-frequency sound energy correlation state on the recent anomaly score sequence and the recent high-frequency sound energy sequence to obtain an anomaly score correlation state response trajectory mapping vector and a high-frequency sound energy correlation state response trajectory mapping vector; Calculating an anomaly score associated state response trajectory curvature vector and a high-frequency sound energy associated state response trajectory curvature vector based on the anomaly score associated state response trajectory mapping vector and the high-frequency sound energy associated state response trajectory mapping vector; Based on the transition probability between the anomaly score-associated state response trajectory curvature vector and the high-frequency acoustic energy-associated state response trajectory curvature vector, the recent anomaly score sequence and the recent high-frequency acoustic energy sequence are respectively priori mapped to obtain the optimized recent anomaly score sequence and the optimized recent high-frequency acoustic energy sequence.

8. The intelligent punching control method for a metal stamping die according to claim 5, characterized in that: Determining a fault type based on the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend includes: extracting current high-frequency sound energy from the stroke acoustic time series characteristic parameters; The current high-frequency acoustic energy, the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend are input into a decision tree model to obtain the fault type.

9. The intelligent punching control method for a metal stamping die according to claim 8, characterized in that: Inputting the current high-frequency acoustic energy, the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend into a decision tree model to obtain the fault type includes: If the current anomaly score is greater than a medium threshold and less than a very high threshold, the anomaly score slope is less than a positive threshold, and the current high-frequency acoustic energy is less than a chipping characteristic threshold, the fault type is determined to be potential material blockage.

10. An intelligent punching control system for metal stamping dies, characterized in that: include: A stroke sensing monitoring module is used to obtain stroke pressure time series data and stroke acoustic time series data collected by the punch press pressure sensor and acoustic sensor; a stroke feature engineering module, configured to perform stroke feature engineering on the stroke pressure time series data and the stroke acoustic time series data to obtain a comprehensive feature vector of a current stroke state; an anomaly detection module, configured to input the current stroke state comprehensive feature vector into a trained real-time anomaly detection model to obtain a current anomaly score and determine a preliminary anomaly flag based on a comparison between the anomaly score and a preset anomaly score threshold; a preliminary abnormality response module, configured to collect a recent abnormality score sequence and a recent high-frequency acoustic energy sequence in response to the preliminary abnormality flag being true; an abnormality trend analysis module, configured to perform trend analysis on the recent abnormality score sequence and the recent high-frequency sound energy sequence to obtain a recent abnormality score trend and a recent high-frequency sound energy trend; A fault type determination module is configured to determine a fault type based on the current anomaly score, the recent anomaly score trend, and the recent high-frequency acoustic energy trend.

Images