Mold machining control system

Abstract

The invention discloses a mold processing control system, and the system comprises a multi-source data collection module which is configured to obtain the dynamic physical quantity in a mold processing process in real time; the feature extraction module is in communication connection with the multi-source data acquisition module and is used for performing dimension reduction processing on the dynamic physical quantity by adopting a multi-modal feature fusion algorithm to generate a feature vector representing a processing state, and the multi-modal feature fusion algorithm adopts a cooperative processing architecture of wavelet packet transformation and a local binary pattern; and the processing state identification module is internally provided with a hybrid classifier based on a deep belief network, and is used for receiving the feature vector and outputting a processing state identifier. According to the method, dynamic field collaborative reconstruction and cross-scale error chain suppression of multi-source sensing data are realized, and the curved surface forming precision and the optical performance of the non-metal mold are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of processing control, in particular to a mold processing control system. BACKGROUND

[0002] The complexity of the cavity surface of a precision mold, such as an injection mold (e.g., an automobile lampshade, a motorcycle plastic part shell, and a consumer electronic shell), and the requirement for optical appearance continue to rise, and the multi-physical field coupling effect of the manufacturing process becomes the core of the yield bottleneck. When the current control system analyzes the interaction between the dynamic behavior of the molten high polymer material and the mechanical response of the processing system, it faces double challenges: First, the multi-source sensing physical field is fragmented. Heterogeneous data streams such as cavity surface texture laser scanning point cloud, spindle vibration spectrum, and mold temperature field thermal imaging are difficult to reconstruct the coupled stress field of plastic flow, mold deformation, and tool chatter due to the mismatch of space-time reference, which leads to a sharp increase in the misjudgment rate of key process features (such as the resonance peak in the mold line area) in the hybrid decision model. A typical case is that the vibration caused by uneven melt filling is misidentified as tool wear, resulting in flow marks in the gate area of the automobile door panel mold.

[0003] Second, the thermal-mechanical error is chain amplified. The minute-level mold thermal expansion (slow time-varying) caused by the cooling waterway and the millisecond-level mechanical impact (fast time-varying) of the ejector pin mechanism form a strong coupling in the time-frequency domain. The traditional single-rate control architecture cannot respond to cross-scale disturbances synchronously, i.e., thermal compensation lags, causing misalignment (flash) on the parting surface of the car lamp lens mold, and vibration suppression delays, which exacerbates the molecular orientation difference at the melt front convergence (weld line). More seriously, when a key sensor (such as an in-mold pressure probe) fails, the system lacks the ability to deduce the state of multiple physical fields, causing scratches on the cavity surface of the plastic gear mold due to inaccurate compensation.

[0004] Such defects are particularly pronounced in the manufacture of transparent parts. Distortion of the mold surface is amplified after injection molding, directly leading to light scattering spots on high-end motorcycle helmet visors or automobile lenses, and the uniformity of product light transmission cannot meet the stringent optical standards. SUMMARY

[0005] To solve the above problems, embodiments of the present application provide a mold processing control system, which comprises: A multi-source data acquisition module configured to acquire dynamic physical quantities in real time during mold processing; A feature extraction module in communication connection with the multi-source data acquisition module, which adopts a multi-modal feature fusion algorithm to perform dimensionality reduction processing on the dynamic physical quantities, and generates a feature vector representing the processing state. The multi-modal feature fusion algorithm adopts a collaborative processing architecture of wavelet packet transform and local binary pattern. a processing state recognition module, which is internally provided with a hybrid classifier based on a deep belief network, receives the feature vector and outputs a processing state identifier; an adaptive compensation decision module, which calls a nonlinear compensation strategy library according to the processing state identifier, generates a compensation decision instruction, and dynamically updates a strategy weight of the nonlinear compensation strategy library through reinforcement learning; a closed-loop execution control module, which converts the compensation decision instruction into a machine tool control signal, integrates a thermal expansion real-time compensation engine, and synchronously adjusts servo drivers, spindle motors and auxiliary units of a numerical control machine tool.

[0006] Further, a software processing method of the multi-source data acquisition module comprises: a deformation reconstruction unit, which receives a laser displacement sensor original signal, performs a tool three-dimensional deformation calculation based on Doppler frequency shift compensation, and outputs a dynamic offset of a cutting edge; a surface enhancement analysis unit, which performs a stream processing on a high-speed industrial vision image, separates mirror reflection noise and workpiece surface feature texture through a polarization characteristic analysis algorithm; a dynamic physical field generation unit, which performs the following operations by using a space-time registration engine: establishes an affine transformation matrix of a tool coordinate system to a workpiece coordinate system; performs pixel-level fusion of the dynamic offset of the cutting edge and the feature texture; generates a dynamic stress distribution cloud diagram of a contact area of the tool and the workpiece based on a material mechanics model inversion.

[0007] Further, the feature extraction module performs wavelet packet energy spectrum reconstruction, specifically comprising: performing 8-layer wavelet packet decomposition on a vibration spectrum, and extracting normalized energy entropy of 3-5 sub-band nodes as a frequency domain resonance component amplitude ratio.

[0008] Further, the deep belief network of the processing state recognition module comprises a three-level restricted Boltzmann machine stack structure, wherein the number of input layer nodes is equal to the dimension of the feature vector, the activation function of the hidden layer node adopts a rectified linear unit, and the output layer generates a state probability distribution through a Softmax regression.

[0009] Further, the adaptive compensation decision module integrates a fuzzy logic controller, an input variable of which is a Cartesian product of a tool wear grade and a material cutting resistance abnormality index, and an output variable of which is a membership function of a feed rate attenuation coefficient.

[0010] Further, the closed-loop execution control module is internally provided with a thermal deformation compensation model, which takes a spindle temperature rise curve and an environmental temperature sensor reading as inputs, calculates a workbench thermal expansion compensation offset, and implements compensation through a linear motor micro-positioning platform.

[0011] Further, the mold processing control system is connected with a digital twin engine, which maps the actual processing state in real time based on a physical processing process simulation model, and performs residual analysis on the simulation results and the measured feature vectors to calibrate the hybrid classifier.

[0012] Further, the digital twin engine drives a virtual sensor, which predicts the missing dynamic physical quantity data stream based on a long short-term memory network when the physical sensor fails, to maintain continuous operation of the system.

[0013] A mold processing control system, the system further comprises: A vibration active suppression unit receives the amplitude of the fundamental frequency component of the vibration spectrum of the workbench, generates an inverse phase cancellation signal through an inverse model algorithm, and drives a piezoelectric ceramic actuator to apply a cancellation force.

[0014] The mold processing control system provided by the present application has the following technical effects and advantages: The present application realizes dynamic field collaborative reconstruction of multi-source sensing data and suppression of cross-scale error chains, significantly improving the surface forming precision and optical performance of non-metal molds. The present application analyzes the melt shear-tension coupling effect in real time through a dynamic stress cloud map, eliminates the flow mark defects on the high light surface, and solves the long-standing surface flaw problem in the automotive lens industry; using wavelet packet energy entropy hierarchical extraction combined with deep belief network feature decoupling improves the recognition accuracy of process state and avoids batch waste loss caused by misjudgment; using a fuzzy decision engine to drive thermal-mechanical double disturbance collaborative suppression improves the weld mark strength of precision parts; the digital twin engine realizes faultless switching through LSTM prediction, ensuring that the system can face sudden sensor failure. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A mold processing control system connection diagram in embodiment one is shown in the figure. Figure 2 A mold processing control system connection diagram in embodiment two is shown in the figure. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] Embodiment one Please refer to Figure 1 The embodiments of the present application provide a mold processing control system, which comprises: The multi-source data acquisition module is configured to acquire dynamic physical quantities in real time during mold processing. The feature extraction module is in communication connection with the multi-source data acquisition module, adopts a multi-modal feature fusion algorithm to perform dimension reduction processing on the dynamic physical quantities, and generates a feature vector representing a processing state. The processing state recognition module internally has a hybrid classifier based on a deep belief network (DBN), receives the feature vector, and outputs a processing state identifier. The adaptive compensation decision module calls a nonlinear compensation strategy library according to the processing state identifier, generates a compensation decision instruction, and dynamically updates strategy weights of the nonlinear compensation strategy library through reinforcement learning. The closed-loop execution control module converts the compensation decision instruction into a machine tool control signal, integrates a thermal expansion real-time compensation engine, and synchronously adjusts servo drivers, spindle motors, and auxiliary units of a numerical control machine tool.

[0018] The dynamic physical quantities include spindle load current, workbench vibration spectrum, tool acoustic emission signal, and workpiece surface topography image.

[0019] The feature vector includes time domain energy entropy, frequency domain resonance component amplitude ratio, image texture gradient amplitude, stress gradient amplitude entropy for non-metallic brittle materials, sub-band energy ratio for low-frequency vibration of an automobile mold, and polarization texture features for transparent plastic surface detection.

[0020] The processing state identifier includes tool wear grade, material cutting resistance anomaly index, surface micro-defect probability, and variable temperature cooling liquid injection timing for precision injection molding of plastic products.

[0021] The compensation decision instruction includes spindle speed correction amount, feed rate attenuation coefficient, and cooling liquid injection mode switching instruction.

[0022] The software processing flow of the multi-source data acquisition module includes: The deformation reconstruction unit receives original signals of a laser displacement sensor, performs three-dimensional deformation calculation of a tool based on Doppler frequency shift compensation, and outputs a dynamic offset of a cutting edge; The surface enhancement analysis unit performs streaming processing on high-speed industrial vision images, separates mirror reflection noise and workpiece surface feature texture through a polarization characteristic analysis algorithm; The dynamic physical field generation unit adopts a space-time registration engine to perform the following operations: An affine transformation matrix of a tool coordinate system to a workpiece coordinate system is established; The dynamic offset of the cutting edge is fused with the feature texture at the pixel level; Based on the material mechanics model inversion to generate the dynamic stress distribution nephogram of the contact area between the tool and the workpiece; wherein the dynamic stress distribution nephogram serves as the input source of the feature extraction module, and stress gradient data thereof directly participate in the calculation of time domain energy entropy and image texture gradient amplitude.

[0023] The multi-source data acquisition module realizes physical field reconstruction through an algorithm chain, including: When the deformation reconstruction unit processes the laser displacement signal, a Doppler frequency shift compensation algorithm is embedded. When the system detects that the spindle speed exceeds the safety threshold (for example, reaches 12000 rpm when machining titanium alloy), the algorithm automatically calculates the dynamic drift amount of the laser reflection wavelength, and real-time corrects the measurement error caused by the centrifugal effect. The three-dimensional deformation offset output by this processing still maintains sub-micron level precision under high-speed working conditions, providing a reliable basis for subsequent analysis.

[0024] The surface enhancement analysis unit develops a polarization characteristic analysis algorithm for visual images, dynamically constructs a virtual polarization filter according to the cooling liquid jet angle, effectively separates the specular reflection noise points and real texture features of the metal surface, and in the actual measurement of die steel machining, this method significantly improves the imaging contrast of surface micro-cracks. The sub-pixel level topological data extracted can directly represent the degree of material plastic deformation.

[0025] The dynamic physical field generation unit performs multi-source fusion through a space-time registration engine, including: A rigid transformation relationship between the tool coordinate system and the workpiece coordinate system is established, the deformation offset and the surface texture are aligned at the pixel level, and the stress distribution in the contact area is calculated by inversion based on the material constitutive equation; In the output dynamic stress nephogram, the tool rake face presents a high stress concentration area (typical value 150-180MPa), and the machined surface of the workpiece shows a gradient change of the residual stress field.

[0026] The dynamic stress nephogram as the core carrier drives the operation of the subsequent modules, including: In the feature extraction module: The time domain energy entropy calculation directly uses the stress fluctuation frequency component as the input, and the image texture gradient amplitude is taken from the edge mutation feature of the stress concentration area. The abnormal detection index constructed by the two cooperates, for example, when machining IN718 high-temperature alloy: The nephogram detects a sudden 200MPa stress peak on the rake face, the feature extraction module generates the abnormal frequency domain resonance component amplitude ratio, and triggers the adaptive compensation module to switch the cooling liquid injection mode.

[0027] The digital twin engine calibrates the model using the dynamic stress nephogram, including: When the stress distribution shown by the dynamic stress nephogram deviates from the simulation prediction by more than the tolerance (such as 15%), the system automatically updates the friction coefficient boundary strip in the finite element analysis.

[0028] The feature extraction module performs wavelet packet energy spectrum reconstruction, specifically including: performing 8-level wavelet packet decomposition on the vibration spectrum, and extracting the normalized energy entropy of the 3rd to 5th sub-band nodes as the amplitude ratio of the frequency domain resonance component.

[0029] The feature extraction module enhances fault features using wavelet packet energy spectrum reconstruction technology, and its technical process is as follows: The physical field signal input, namely the dynamic stress distribution cloud map (from the multi-source data acquisition module), provides the vibration spectrum source signal. The key input feature is the high-frequency vibration energy release mode in the contact area between the tool and the workpiece. The vibration spectrum was decomposed into 8-level wavelet packets to generate 256 frequency bands. The focus was on the 3rd to 5th sub-band nodes (corresponding to the mid-to-high frequency characteristic bands), which contain the core fault spectrum characteristics such as material plastic deformation and tool chatter. The signal energy value of the selected sub-band is calculated and normalized. A normalized energy entropy index is constructed based on information entropy theory, and its calculation conforms to the time-frequency analysis specifications in IEEE Standard 1241-2010, including: ; In the formula, Representing the The proportion of subband energy to total energy The entropy value quantifies the dispersion of the spectral energy distribution. Under abnormal operating conditions, the entropy value increases significantly.

[0030] Examples of tool wear monitoring: Normal processing stage: The energy entropy of the 3rd to 5th subbands remains in a stable range (e.g., 0.18-0.25), while the time-domain energy entropy shows a smooth fluctuation. Abnormal triggering phase: Wavelet packet analysis shows that the energy entropy of the fourth subband continues to rise; When the entropy value exceeds the entropy increase threshold (example value 0.6), the system determines that the back face wear is aggravated, and the decision response chain includes: An abnormal index for the amplitude ratio of the frequency domain resonance component is generated, and the adaptive compensation module initiates spindle speed adjustment and enhanced coolant injection.

[0031] When the tool enters a worn state: The entropy value of the fourth sub-band jumps into the red warning zone, and the third and fifth sub-bands show a synergistic enhancement effect, with the time-domain energy entropy curve rising sharply in sync.

[0032] The deep belief network of the processing state recognition module includes a three-level restricted Boltzmann machine (RBM) stack structure, wherein the number of input layer nodes is equal to the dimension of the feature vector, the hidden layer node activation function adopts a rectified linear unit (ReLU), and the output layer generates a state probability distribution through Softmax regression.

[0033] The processing state recognition module realizes accurate classification of fault states through a three-level restricted Boltzmann machine stack architecture, and forms a cascade processing chain with the feature extraction module, including: The number of input nodes in the input layer is matched in real time with the vector dimension output by the feature fusion module, and directly receives the mixed features generated by wavelet packet energy spectrum reconstruction (including frequency domain energy entropy and resonance amplitude ratio) For example, in the milling tool chatter monitoring scene, the input layer is automatically configured with 5 nodes to receive the following features: 3rd sub-band energy entropy, 4th sub-band energy entropy, 5th sub-band energy entropy, time domain energy entropy, and frequency domain resonance amplitude ratio.

[0034] The three-level restricted Boltzmann machine (RBM) stack structure includes three levels of feature abstraction: The first level RBM extracts frequency band coupling features and analyzes the frequency energy correlation patterns between input features, for example, when the 4th sub-band entropy value increases sharply and the 3rd sub-band entropy value increases slowly, a primary feature of tool eccentric moment enhancement is generated.

[0035] The second level RBM constructs a time sequence dependence model, correlates the feature change trend of consecutive time windows, and identifies wear evolution patterns such as "energy entropy continuously rising + amplitude ratio step transition".

[0036] The third level RBM generates a state abstraction vector, integrates the abstraction results of the previous two levels, and forms high-order features representing the processing state, such as a 128-dimensional feature vector representing "tool flank wear - mid-term".

[0037] The three-level restricted Boltzmann machine (RBM) stack structure includes an activation and decision mechanism: All hidden layer nodes use a rectified linear unit activation function, which has the characteristics of outputting zero for negative input and maintaining linear increase for positive input, and the rectified linear unit activation function enhances the sensitivity to tool progressive wear while suppressing vibration noise interference.

[0038] The output layer generates a four-class state probability distribution through Softmax regression, including normal cutting, initial wear, mid-term wear, and severe failure.

[0039] For example, when the "mid-term wear" probability is greater than 85%, the strong cooling strategy of the adaptive compensation module is triggered, and the die steel processing scene is continued. The feature extraction module detects that the 4th sub-band entropy value breaks through 0.6 (entropy rise threshold), generates a 5-dimensional input vector, and transmits it to the processing state recognition module. The three-level RBM recognizes that the entropy value mutation is accompanied by a sustained amplitude ratio higher than 1.8, and the Softmax output "medium wear" probability reaches 92%. The spindle speed regulation (reduced to the original value of 75%) and the cooling liquid flow rate are increased to the baseline value of 300%, so that the processing vibration energy falls below the safety threshold within 3 seconds Compared with traditional machine learning, the three-level RBM stack deeply couples the frequency domain features and the time sequence features, reduces the misjudgment rate of "pseudo-normal" state, avoids gradient disappearance with ReLU activation function, ensures continuous iteration of the model in industrial environment, and is compatible with different numerical control system alarm threshold settings with Softmax probability output.

[0040] The frequency band entropy index of the feature extraction module is mathematically isomorphic to the RBM abstract layer of the recognition module. When the wavelet packet analysis adds new feature dimensions, the number of input layer nodes can be expanded adaptively.

[0041] The adaptive compensation decision module integrates a fuzzy logic controller. The input variables are the Cartesian product of the tool wear grade and the material cutting resistance anomaly index, the output variable is the membership function of the feed rate attenuation coefficient, and the rule base contains 49 fuzzy inference rules in the form of IF-THEN.

[0042] The specific implementation process of the adaptive compensation decision module includes: After receiving the tool wear grade output by the processing state recognition module, the system performs Cartesian product operation (i.e. set intersection combination) on the tool wear grade and the material cutting resistance anomaly index generated by the feature extraction module to form a two-dimensional decision matrix. Here, the wear grade is divided into three levels: level I represents early wear (Softmax probability output 50%-70%), level II represents medium wear (Softmax probability output 70%-90%), and level III represents severe failure (Softmax probability output >90%). The material cutting resistance anomaly index is dynamically calculated by time energy entropy. When the index is greater than 1.8, it indicates that the workpiece has hard spots or built-up edges. For example, in automobile die machining, if the system detects level II wear (probability 85%) and the resistance index reaches 2.0, the Cartesian product generates a specific working condition point (II high resistance).

[0043] Based on the product result, the fuzzy logic controller calls the preset rule base for decision making. The rule base defines the fuzzy set of the feed rate attenuation coefficient (i.e. the proportion coefficient of the actual feed rate and the set value), including three types of fuzzy variables: "fine tuning", "medium attenuation", and "strong speed reduction".

[0044] Each variable quantifies its action range through a triangular membership function, for example, the membership degree of "medium attenuation" under the condition of (II, high resistance) is 0.7, indicating that the feed rate should be reduced to 60%-70% of the original speed; the controller uses the barycentric method to de-fuzzify (weights the membership degree of each fuzzy variable and its typical value) to output precise control instructions. Continuing the previous example, the system calculates the attenuation coefficient as 0.65, which means the control servo unit will reduce the feed speed to 65%.

[0045] Exemplary: During the machining process of a certain hardened steel mold, the feature extraction module detects that the energy entropy value of the 4th sub-band breaks through 0.8 (entropy rise threshold), the recognition module determines that the probability of grade III wear reaches 92%, and the material resistance index jumps to 2.3 due to local carbide segregation. The fuzzy controller performs the following actions: The Cartesian product generates a (grade III wear, high resistance) working condition point, and the matching rule library gets a "strong speed reduction" membership degree of 0.8, outputting an attenuation coefficient of 0.48 (i.e., the feed rate is reduced to 48%). The spindle speed reduction and cooling and pressure increase are executed in linkage, effectively preventing tool chipping.

[0046] The material cutting resistance anomaly index is constructed based on the time domain differential operation of the wavelet packet sub-band energy entropy, which is used to quantify the degree of material property mutation.

[0047] The 9 kinds of working condition combinations generated by Cartesian product (3 wear grades x 3 resistance grades) cover the main abnormal scenarios of mold machining. When the resistance index fluctuates instantaneously for more than 0.3 seconds, it triggers a recalculation to avoid false actions. When the attenuation coefficient is less than 0.4, the third level response (emergency stop + spindle unloading + chip flushing) is activated. The membership function parameter library supports automatic calibration based on historical machining data, improving the adaptability of different mold materials.

[0048] The closed-loop execution control module has a built-in thermal expansion real-time compensation engine. The thermal expansion real-time compensation engine takes the spindle temperature rise curve and the ambient temperature sensor readings as inputs, calculates the worktable thermal expansion compensation offset, and implements compensation through the linear motor micro-positioning platform.

[0049] Based on the real-time collection of spindle temperature data by the multi-source sensor network, the system dynamically offsets the machine tool thermal deformation error through the thermal expansion real-time compensation engine. The thermal expansion real-time compensation engine takes the spindle temperature rise curve (a temperature change rate function generated by the feature extraction module) as the core input, combined with the ambient temperature sensor readings installed on the machine bed (used to correct the influence of seasonal temperature difference), to calculate the thermal expansion compensation offset of the worktable in the X / Y / Z axis direction. For example, when the spindle temperature rise curve shows that the temperature rises by more than 120% of the set temperature rise threshold within 30 minutes during the continuous machining of a certain automobile mold, the system predicts the trend of the worktable offset caused by thermal expansion.

[0050] The calculation of the compensation offset is based on the thermal mechanical properties of the machine tool structure, including: The temperature rise and expansion mapping relationship, that is, for each unit temperature rise of the spindle, a specific proportion of expansion of the worktable base body is caused by heat conduction.

[0051] Exemplary: In the processing of a certain large injection mold, the slope of the spindle temperature rise curve suddenly changes (the temperature rise exceeds 80% of the set threshold within 30 minutes), and the model maps the worktable Y axis to expand 0.01 mm according to historical data.

[0052] The ambient temperature correction coefficient, that is, when the ambient temperature is relatively high in summer (the sensor reading is greater than the set temperature rise threshold), the compensation amount is reduced in proportion (because the thermal expansion allowance is reduced).

[0053] Exemplary: When the ambient temperature is relatively low in winter, the spindle temperature rise set temperature rise threshold is 100%, and the model automatically increases the compensation amount by 15% to offset the low-temperature shrinkage effect; The compensation offset generated by the calculation is implemented in real time through the linear motor micro-positioning platform, including: The worktable receives micron-level position compensation instructions at each time of feeding interruption, and the compensation action is synchronized with the feeding rate adjustment.

[0054] Exemplary: During the finishing of a certain automobile cover mold, the system monitors that: The spindle temperature rise curve continuously exceeds 90% of the set temperature rise threshold within 2 hours (feature extraction), and the ambient temperature sensor shows a sudden drop of 15% of the set temperature rise threshold (low-temperature shrinkage warning), The thermal expansion real-time compensation engine outputs: X-axis compensation offset: +0.008 mm; Z-axis compensation offset: -0.005 mm (considering the coupling effect of column thermal deformation).

[0055] The spindle temperature rise curve inherits the temperature sampling data of the multi-source sensor network, and the temperature change rate function curve generated by time domain feature extraction is used to reflect the cumulative trend of spindle bearing friction heat.

[0056] The compensation offset is a three-dimensional space position correction value output by the thermal expansion real-time compensation engine.

[0057] The calculation logic includes: offset = temperature rise contribution × material expansion coefficient + ambient temperature correction term.

[0058] The system is connected to a digital twin engine, which maps the actual processing state in real time based on a physical processing process simulation model, and performs residual analysis on the simulation results and the measured feature vectors to calibrate the hybrid classifier.

[0059] The system mirrors the actual machining process in the virtual space through a physical simulation model (based on the dynamic characteristics of the machine tool and the material cutting mechanism), which receives real-time data collected by a multi-source sensing network (such as spindle power, vibration spectrum, cutting temperature), generates a simulation feature vector (including virtual vibration energy entropy, theoretical thermal deformation, etc.) that is completely synchronized with the physical machine tool. For example, during the milling of an automotive mold cavity, the actual sensor measures a cutting force fluctuation of 120% of the set fluctuation threshold, and the physical simulation model calculates that the theoretical fluctuation should be 105% of the set threshold based on the current tool wear state and material properties.

[0060] The residual analysis calibration process includes: The extracted measured feature vector (such as wavelet packet 4th sub-band energy entropy) is compared with the theoretical feature vector output by the simulation model item by item, for example: Measured: finishing section energy entropy = 0.85; Simulation: theoretical energy entropy = 0.72; Residual value = |0.85-0.72| = 0.13.

[0061] When the residual continuously exceeds the set fluctuation threshold (such as residual > 0.1 for 5 consecutive sampling periods): Determine that the hybrid classifier (DBN-SVM model) has recognition bias, activate the model parameter calibration program to calibrate the model.

[0062] Correct the hybrid classifier weights through the backpropagation algorithm, including: Residual > set threshold: increase vibration feature weight; Residual < set threshold negative value: increase temperature feature weight.

[0063] For example: In the processing of a certain injection mold, the measured vibration energy entropy is continuously higher than the simulation value by 15% of the set vibration threshold, and the digital twin engine executes: The decision weight of the vibration feature in the hybrid classifier is increased by 20% of the set vibration threshold, and the temperature feature weight is simultaneously reduced by 10% of the set threshold.

[0064] The physical simulation model is the core carrier of the physical-based machining process simulation, where the input is the execution parameters (speed and feed), sensing data, and wear state, and the output is the theoretical vibration spectrum and cutting force curve, etc. comparable parameters.

[0065] The digital twin engine drives the virtual sensor, and when the physical sensor fails, it predicts the missing dynamic physical quantity data stream based on the long short-term memory network (LSTM), maintaining continuous operation of the system.

[0066] When the physical sensors (such as vibration accelerometers, spindle power meters) of the multi-source sensing network fail, the digital twin engine immediately activates a virtual sensing channel based on a long short-term memory network (LSTM). The long short-term memory network takes the machine tool control parameters (spindle speed and feed rate) and the remaining valid sensor data (such as environmental temperature) as input, and real-time predicts the missing dynamic physical quantity data stream (including vibration spectrum, cutting force, etc. Key parameters), for example, in continuous processing of automobile molds, if the vibration sensor suddenly breaks down: The LSTM receives 80% of the current spindle speed set threshold, combines the valid temperature measurement point data (75% of the bearing temperature set threshold), and outputs the predicted vibration energy entropy = 0.78 (with an error of <5% of the set vibration threshold from the actual historical working condition).

[0067] The basis for failure determination is that the feature extraction module detects sensor signals. If the value continues to be set to zero or exceeds 120% of the physical quantity range set threshold, it is determined to be a failure.

[0068] Example explanation: When the current is 0, the LSTM predicts the current to be 42A (with a deviation of 3% from the average before failure), the predicted value automatically replaces the failed sensor channel, the feature extraction module continues to calculate the wavelet packet energy entropy, and the hybrid classifier maintains tool wear recognition. During the failure of the vibration sensor, the system performs: Based on the LSTM predicted data, determine the Ⅱ wear (probability 80%), trigger the feed attenuation coefficient 0.7. Embodiment two As Figure 2 shown, this embodiment is further improved based on embodiment one. The difference is that in the actual operation of embodiment one, it is found that the workbench vibration fundamental component amplitude continuously exceeds the limit when the tool wears or the material resistance is abnormal, and the response of the existing adaptive compensation decision module to vibration suppression has a lag, resulting in a decrease in workpiece surface topography precision and an increase in tool chatter, which cannot real-time offset the vibration energy to maintain the stability of processing. Based on this, a mold processing control system further comprises: A vibration active suppression unit receives the fundamental component amplitude of the workbench vibration spectrum, generates an inverse compensation signal through an inverse model algorithm, and drives a piezoelectric ceramic actuator to apply a compensation force.

[0069] The data source is the workbench vibration spectrum (dynamic physical quantity) directly received from the multi-source data acquisition module output. The fundamental component amplitude in the spectrum, i.e. the main frequency band amplitude of the workbench vibration energy, is extracted. This feature has been quantified by the wavelet packet energy spectrum reconstruction technology in embodiment one (such as the 3-5 sub-band energy entropy).

[0070] The inverse model algorithm constructs a dynamic inverse model based on the machine tool structure's frequency response function (obtained through offline impact testing calibration), including: Establish a mathematical model for the vibration transmission path, including: ; In the formula, The transfer function characterizes the transfer from excitation force to vibration response. For complex frequencies, The input excitation force frequency domain function is derived from the cutting force excitation during the machining process. The frequency domain function of the vibration response of the workbench is directly input from the measured values ​​of the vibration sensors in the multi-source data acquisition module and the amplitude of the fundamental frequency component quantized by the feature extraction module.

[0071] Construct the inverse transfer function based on the mathematical model of the vibration transmission path. The fundamental frequency amplitude will be collected in real time. Converted to an inverting control signal, including: ; In the formula, based on Inverse of the zero-pole configuration, The direct input source is for feature extraction Take the wavelet packet energy conversion of the module. The actual conversion of the piezoelectric ceramic actuator drive signal into the canceling force requires multiplication by the actuator gain coefficient.

[0072] For example: Precision milling of hardened steel mold (material resistance index = 2.1), input is: =18μm (fundamental frequency 250Hz), then =1.1 N / μm, =1.1×(-18)=-19.8V, the counteracting force generated by the piezoelectric ceramic actuator is =0.1×-19.8=-1.98N.

[0073] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations of this invention fall within the scope of the claims of this invention... The present invention also intends to include these modifications and variations within the scope of the art and its equivalents.

[0074] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. A mold processing control system, characterized by a system The system comprises: a multi-source data acquisition module configured to acquire dynamic physical quantities in real time during mold processing; a feature extraction module in communication with the multi-source data acquisition module, which uses a multi-modal feature fusion algorithm to reduce the dimensionality of the dynamic physical quantities, generating a feature vector representing the processing state, and the multi-modal feature fusion algorithm uses a collaborative processing architecture of wavelet packet transform and local binary pattern; a processing state recognition module with a hybrid classifier based on a deep belief network, which receives the feature vector and outputs a processing state identifier; an adaptive compensation decision module that calls a nonlinear compensation strategy library based on the processing state identifier to generate a compensation decision instruction, and the nonlinear compensation strategy library dynamically updates the strategy weight through reinforcement learning; a closed-loop execution control module that converts the compensation decision instruction into machine tool control signals and integrates a thermal expansion real-time compensation engine to synchronize the servo drive, spindle motor, and auxiliary units of the numerical control machine tool.

2. The mold processing control system of claim 1, wherein, The software processing method of the multi-source data acquisition module includes: a deformation reconstruction unit that receives laser displacement sensor raw signals, performs three-dimensional tool deformation calculation based on Doppler frequency shift compensation, and outputs dynamic blade offset; a surface enhancement analysis unit that performs stream processing on high-speed industrial vision images, separates mirror reflection noise from workpiece surface feature textures through polarization characteristic analysis algorithms; a dynamic physical field generation unit that uses a space-time registration engine to perform the following operations: establishes an affine transformation matrix from the tool coordinate system to the workpiece coordinate system; performs pixel-level fusion of the dynamic blade offset and the feature texture; generates a dynamic stress distribution cloud map of the tool and workpiece contact area based on a material mechanics model.

3. The mold processing control system of claim 1, wherein The feature extraction module performs wavelet packet energy spectrum reconstruction, specifically including: 8-layer wavelet packet decomposition of the vibration frequency spectrum, extraction of the normalized energy entropy of the 3-5 sub-band nodes as the frequency domain resonance component amplitude ratio.

4. The mold processing control system of claim 3, wherein The deep belief network of the processing state recognition module contains a three-level restricted Boltzmann machine stack structure, wherein the number of input layer nodes is equal to the dimension of the feature vector, the activation function of the hidden layer nodes uses a rectified linear unit, and the output layer generates a state probability distribution through Softmax regression.

5. The mold processing control system of claim 1, wherein, The adaptive compensation decision module integrates a fuzzy logic controller, with the Cartesian product of the tool wear grade and the material cutting resistance anomaly index as the input variable, and the membership function of the feed rate attenuation coefficient as the output variable.

6. The mold processing control system of claim 1, wherein, The closed-loop execution control module has a thermal deformation compensation model, which takes the spindle temperature rise curve and the environmental temperature sensor reading as input, calculates the workbench thermal expansion compensation offset, and implements compensation through a linear motor micro-positioning platform.

7. The mold processing control system of claim 1, wherein, The mold processing control system is connected to a digital twin engine, which maps the actual processing state in real time based on a physics-based processing simulation model, and performs residual analysis on the simulation results and the measured feature vector to calibrate the hybrid classifier.

8. The mold processing control system of claim 7, wherein, The digital twin engine drives virtual sensors, and when physical sensors fail, it predicts missing dynamic physical quantity data streams based on a long short-term memory network to maintain continuous system operation.

9. The mold processing control system of claim 1, wherein, The system further comprises: A vibration active suppression unit receives the amplitude of the fundamental component of the worktable vibration spectrum, generates an anti-phase cancellation signal through an inverse model algorithm, and drives a piezoelectric ceramic actuator to apply a cancellation force.

Images