Mim titanium alloy injection control method and system

Abstract

The present application belongs to the technical field of metal powder injection molding, and particularly relates to a MIM titanium alloy injection molding control method and system, which comprises the following steps: obtaining a historical process parameter sequence and historical quality data, and constructing a quality prediction model and a rheological model; extracting shear rate mutation characteristics caused by a to-be-optimized sequence by using the rheological model, and combining melt rheological characteristics to generate a rheological instability index; extracting a fluctuation frequency component, combining the rheological instability index and a current injection molding stage to perform adaptive filtering, and separating state-dependent fluctuation components and high-frequency residual components; calculating the sensitivity of each state-dependent fluctuation component to a product quality index and configuring corresponding penalty weights; constructing an optimization function that fuses a product defect prediction evaluation value and a control parameter fluctuation complexity penalty term, obtaining a target control parameter sequence through optimization, and performing control. The present application can reduce warping deformation and shrinkage defects, and improve the quality stability of MIM titanium alloy injection molding.

Description

Technical Field

[0001] This invention relates to the field of metal powder injection molding technology. More specifically, this invention relates to a method and system for controlling the injection molding of MIM titanium alloys. Background Technology

[0002] In the injection molding process of MIM titanium alloys, the feedstock formed by mixing titanium alloy powder with polymer binder is a typical non-Newtonian fluid, and its melt viscosity is highly sensitive to the shear rate and shear action during the injection molding process. In the actual mold filling and holding pressure stages, the melt flow in the mold cavity is intense and the pressure state is constantly changing. The feedstock is prone to rheological instability, resulting in uneven filling density distribution in the mold cavity. This leads to stress concentration and shrinkage differences in the subsequent cooling and solidification stages, ultimately causing quality problems such as warping, surface defects, and internal shrinkage cavities in the finished product.

[0003] Currently, some technologies are attempting to incorporate machine learning, numerical calculation, and online monitoring into the optimization and control of injection molding process parameters. For example, Chinese patent application CN121551604A discloses a high-precision injection-molded titanium alloy part and its machine learning-assisted preparation method. This application obtains the original dataset through three-dimensional filling numerical calculation, and after data preprocessing, uses various machine learning algorithms to construct a prediction model for injection molding process parameters. This model obtains the optimal process parameters and guides actual production, which can shorten the process development cycle and reduce trial production costs to some extent. Another example is Chinese patent CN113043562B, which discloses a dynamic compensation system based on injection molding product quality control. By real-time acquisition of screw position and injection pressure curves, and comparing the deviation with a reference curve, a compensation command is generated to achieve real-time dynamic compensation for fluctuations in operating conditions during production, thereby improving the stability of injection molding quality. This type of feedback compensation approach has been applied and verified in the general injection molding field.

[0004] However, the aforementioned existing technologies primarily improve MIM injection molding from the perspective of offline optimization of process parameters and global control of the production process, without deeply exploring the intrinsic mechanism between the evolution of feed rheological properties and the fluctuation of control parameters. When generating active injection molding control parameter sequences such as injection speed curves and holding pressure curves, existing methods generally do not set effective constraints on the fluctuation complexity of control commands in the time dimension, resulting in parameter sequences output by optimization algorithms often exhibiting significant nonlinear oscillations and invalid fluctuations. After such high-complexity control sequences are put into actual production, on the one hand, they may exceed the dynamic response capability of the injection molding machine's servo actuator, causing command tracking deviations and increasing equipment wear and energy consumption; on the other hand, their drastic and unreasonable parameter fluctuations will exacerbate the shear mutation of titanium alloy feed melt in the runner and cavity, further inducing or even worsening the rheological instability state, causing the system to fall into a vicious cycle of control fluctuations leading to instability, and instability exacerbating product defects, ultimately making it difficult to achieve continuous and stable control of product warpage deformation and shrinkage cavity volume.

[0005] Therefore, there is an urgent need for an injection molding control mechanism that can constrain and reduce the complexity of control parameter sequence fluctuations as a whole, while suppressing undesirable control commands that are prone to inducing rheological instability, and ensuring stable and high-quality molding of MIM titanium alloy products. Summary of the Invention

[0006] To address the problem that the lack of effective constraints on the complexity of control parameter sequence fluctuations during MIM titanium alloy injection molding makes it difficult to simultaneously achieve process stability and product quality consistency, this invention provides a MIM titanium alloy injection molding control method and system.

[0007] In a first aspect, the present invention provides a method for controlling the injection molding of MIM titanium alloys, comprising: S1, acquiring historical process parameter sequences and historical quality data of the MIM titanium alloy injection molding process, and constructing a quality prediction model and a rheological model based on the historical process parameter sequences and the historical quality data; S2, during the optimization process of the control parameter sequence, using the rheological model to extract the shear rate mutation characteristics caused by the injection molding control parameter sequence to be optimized, and combining the shear rate mutation characteristics with the melt rheological properties to generate a rheological instability index; S3, extracting the fluctuation frequency components of the injection molding control parameter sequence to be optimized, and based on the rheological instability index and the corresponding components of the injection molding control parameter sequence to be optimized... In the current injection molding stage, the fluctuation frequency component is adaptively filtered to separate the state-dependent fluctuation component and the high-frequency residual component; S4, the sensitivity of the state-dependent fluctuation component to the product quality evaluation index is calculated using the quality prediction model, and a penalty weight is configured for the corresponding state-dependent fluctuation component based on the sensitivity; S5, an optimization function is constructed that includes the product defect prediction evaluation value output by the quality prediction model and the control parameter fluctuation complexity penalty term, wherein the fluctuation complexity penalty term is determined based on the state-dependent fluctuation component and its corresponding penalty weight, and the amplitude of the high-frequency residual component; the optimization function is optimized to obtain the target injection molding control parameter sequence and control is executed.

[0008] This invention first constructs a quality prediction model and a rheological model based on historical process and quality data, providing dual-model computational support for subsequent optimization of control parameter sequences. During the optimization process, a rheological instability index is generated by combining shear rate mutation characteristics and melt rheological properties. Then, based on the injection molding stage and the degree of instability, adaptive filtering separates the fluctuation components. Corresponding penalty weights are configured based on the sensitivity of each component to quality indicators. Finally, defect prediction and fluctuation complexity penalties are jointly incorporated into the objective function to complete the optimization. This scheme can simultaneously constrain product defects and the degree of fluctuation in the control sequence, suppress rheological instability, break the vicious cycle of control fluctuations exacerbating quality deterioration, and improve the stability and batch consistency of molding quality.

[0009] Preferably, the historical process parameter sequence includes state parameters, active injection molding control parameters, and passive state responses; the construction of a quality prediction model and a rheological model based on the historical process parameter sequence and the historical quality data includes: constructing a quality prediction model using a long short-term memory network, taking the historical process parameter sequence as input, and using the measured warpage deformation and equivalent shrinkage volume of the product as the historical quality data for supervised training; An improved Cross-WLF equation incorporating a metal powder volume fraction term was used to construct a rheological model for titanium alloy feedstock. The equation coefficients were determined by fitting test data under different temperatures, pressures, and shear rates.

[0010] By constructing a quality prediction model using a long short-term memory network, the model can be fully adapted to the temporal characteristics of process parameters and accurately map the correlation between the process and product defects. By introducing an improved Cross-WLF equation with a powder volume fraction term, a rheological model is constructed, which fits the non-Newtonian fluid characteristics of titanium alloy feed. After fitting with measured data, it can accurately reflect the melt rheological behavior under different working conditions, providing a reliable model basis for subsequent instability assessment and parameter analysis.

[0011] Preferably, the step of extracting the shear rate mutation feature caused by the injection control parameter sequence to be optimized, and combining the shear rate mutation feature with the melt rheological properties to generate a rheological instability index, includes: using the rheological model to calculate the second derivative of the feed melt shear rate with respect to time caused by the injection control parameter sequence to be optimized to generate a Jerk value; taking the absolute value of the Jerk value to construct a Jerk amplitude sequence; and extracting a characteristic peak value from the Jerk amplitude sequence as the shear rate mutation feature; multiplying the characteristic peak value by the absolute value of the partial derivative of the feed melt viscosity with respect to the shear rate in the rheological model to obtain the rheological instability index.

[0012] The Jerk value is obtained by calculating the second derivative of the shear rate with respect to time. The peak amplitude characterizes the severity of the shear shock. The rheological instability index is then calculated by combining the partial derivative of the melt viscosity with respect to the shear rate. This allows for the integration of shear impact intensity and the melt's own rheological sensitivity, comprehensively assessing the risk of rheological instability in the melt. This provides a direct basis for the adaptive adjustment of the subsequent filter bandwidth, enabling the early identification of instability risks.

[0013] Preferably, extracting characteristic peaks from the Jerk amplitude sequence as the shear rate abrupt change feature includes: setting a sliding time window of fixed length and calculating the mean and standard deviation of the Jerk amplitude sequence within the current window; A constant coefficient is determined based on the difference between the predicted temperature of the current feed melt and the preset ideal injection temperature. The constant coefficient is negatively correlated with the difference, that is, the lower the predicted temperature is relative to the ideal injection temperature, the smaller the value of the constant coefficient. The equivalent probability quantile threshold is the product of the current window mean plus the constant coefficient and the current window standard deviation. Traverse the Jerk amplitude sequence within the current window and extract the local maxima of Jerk amplitudes that are greater than the equivalent probability quantile threshold as the feature peaks.

[0014] Preferably, the fluctuation frequency component is subjected to adaptive filtering to separate the state-dependent fluctuation component and the high-frequency residual component, including: The injection control parameter sequence to be optimized is obtained by continuous wavelet transform to obtain the time-frequency domain spectrum and the corresponding wavelet coefficients, and an asymmetric Gaussian frequency domain filter window is constructed. The basic bandwidth of the high-frequency side of the filter window is limited by the current injection stage. The basic bandwidth of the pressure rise stage is set to be greater than the basic bandwidth of the pressure holding stage, and the final bandwidth of the filter window is positively correlated with the reciprocal of the rheological instability index. The time-frequency domain spectrum is filtered using the filter window configured with bandwidth, and the state-dependent fluctuation component is reconstructed by inverse transformation. The difference between the wavelet coefficients and the filtered coefficients is extracted as the high-frequency residual component.

[0015] Preferably, the product quality evaluation index includes warpage deformation and shrinkage volume; the sensitivity of the state-dependent fluctuation component to the product quality evaluation index is calculated using the quality prediction model, including: applying a positive small perturbation to the amplitude corresponding to the extracted state-dependent fluctuation component, and replacing it in the corresponding control channel of the injection molding control parameter sequence to be optimized after time-domain reconstruction; The relevant state response channels are updated by the rheological model or process response mapping model, and then input into the quality prediction model to obtain the predicted increments of warpage deformation and shrinkage volume. The partial derivative in the finite difference sense is calculated using the ratio of the predicted increment after dimensionless processing based on the preset upper limit of quality tolerance to the positive small perturbation, and is used as the sensitivity.

[0016] Preferably, the penalty weight is configured for the corresponding state-dependent fluctuation component according to the sensitivity, including: concatenating the sensitivity of the warp deformation amount corresponding to all state-dependent fluctuation components with the sensitivity of the cavity volume to form a two-dimensional sensitivity matrix; For the sensitivity in the sensitivity matrix, when its absolute value exceeds a preset dimensionless safety threshold, the corresponding penalty weight is non-linearly increased based on an exponential function.

[0017] By concatenating the sensitivity of two types of quality indicators into a two-dimensional sensitivity matrix, the impact data of multiple frequency components on different defects can be managed in a unified manner. When the sensitivity exceeds the safety threshold, an exponential function is used to increase the penalty weight, thereby imposing stronger constraints on the fluctuation components that are prone to causing defects. This guides the optimization process to actively avoid high-risk fluctuations and improves the targeting of defect suppression.

[0018] Preferably, the nonlinear increase of the corresponding penalty weight based on the exponential function specifically involves: calculating the difference between the absolute value of the sensitivity and the dimensionless safety threshold, and multiplying the difference by a preset sensitivity adjustment factor to obtain the exponential term; The power function of the exponential term is calculated with the natural constant as the base, and the power function is multiplied by the preset base weight to obtain the final penalty weight.

[0019] By calculating the exponential term based on the difference between the sensitivity exceeding the threshold, and combining it with the power function of the natural constant to amplify the penalty intensity, the more the sensitivity exceeds the threshold, the more significant the increase in penalty weight. Unlike the weak constraint effect of linear penalty, it can form a stronger constraint on the fluctuation component of strong defects, further enhancing the avoidance effect of the optimization process on high-risk control trajectories.

[0020] Preferably, the optimization function is optimized to obtain a target injection molding control parameter sequence, including: the optimization function includes the product of a first preset weight and a dimensionless warp deformation prediction value, the product of a second preset weight and a dimensionless shrinkage volume prediction value, and the sum of the product of a third preset weight and a dimensionless control parameter fluctuation complexity penalty term; The particle swarm optimization algorithm is used to minimize the optimization function. The particle positions are iteratively updated within the set injection speed boundary and screw position constraint boundary. The globally optimal particle positions when the convergence condition is met are decoded into the target injection molding control parameter sequence.

[0021] Secondly, the present invention provides a MIM titanium alloy injection molding control system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned MIM titanium alloy injection molding control method is implemented.

[0022] By adopting the above technical solution, a computer program is generated from the above-mentioned MIM titanium alloy injection molding control method and stored in the memory so that it can be loaded and executed by the processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0023] The technical solution of the present invention has the following beneficial technical effects: This invention constructs an injection molding quality prediction model and a rheological model. During the optimization of the control parameter sequence, it extracts shear rate mutation features and generates a rheological instability index. Based on this, the control parameter sequence to be optimized is adaptively filtered to separate state-dependent fluctuation components and high-frequency residual components. Then, penalty weights are configured based on the sensitivity of each component to product quality indicators. Finally, an optimization function integrating defect prediction values ​​and fluctuation complexity penalty terms is constructed for parameter optimization. This approach, starting from the time-frequency domain fluctuation characteristics of the control parameter sequence, simultaneously constrains the expected level of product defects and the fluctuation complexity of the parameter sequence during the optimization process. This effectively suppresses nonlinear invalid fluctuations in control commands that easily induce rheological instability, reduces the following deviation and wear of the injection molding machine's actuator, and solves the problems of existing technologies that fail to correlate rheological characteristics with control fluctuations and insufficient sequence complexity constraints. It alleviates the vicious cycle of mutual deterioration between the two, ensuring stable control of product warpage and shrinkage volume, and guaranteeing the quality consistency of MIM titanium alloy injection molding. Attached Figure Description

[0024] Figure 1 A flowchart of a method for controlling the injection molding of MIM titanium alloys; Figure 2 A schematic diagram showing the temperature setting range for each section of the barrel; Figure 3 This is a schematic diagram comparing the warping deformation. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] This invention discloses a method for controlling the injection molding of MIM titanium alloys, referring to... Figure 1 This includes the following steps: S1. Obtain process data and construct quality prediction and rheological models.

[0028] This step obtains the historical process parameter sequence and historical quality data of the MIM titanium alloy injection molding process, and constructs a quality prediction model and a rheological model based on the historical process parameter sequence and historical quality data.

[0029] The historical process parameter sequence is obtained as follows: Injection molding machine sensors collect barrel temperature and melt temperature at corresponding sampling periods as state parameters, screw position and injection speed as active injection control parameters, and mold cavity pressure as a passive state response, collectively forming time-series data. For implementations including active adjustment during the holding pressure stage, the holding pressure setpoint can be further collected as an extended active injection control parameter. Data acquisition from each sensor is achieved via an industrial Ethernet bus. For rapidly changing channels such as mold cavity pressure, screw position, and injection speed, the sampling period is set to 2ms to 5ms. For channels with slower thermal responses, such as barrel temperature and melt temperature, a lower sampling frequency is used based on the sensor response time, and the data is aligned with the rapidly changing channels using timestamp linear interpolation. After timestamp alignment, the multi-dimensional sensor data is combined into a sliding time window feature tensor with a step size of 100 to 500, constructing continuous time-series data as the historical process parameter sequence.

[0030] Barrel temperature sensors are arranged in four sections: the feeding section, the compression section, the homogenization section, and the nozzle. The temperature parameters collected range from 160℃ to 170℃, 175℃ to 185℃, 190℃ to 200℃, and 200℃ to 210℃, respectively. In the active injection control parameters, the screw position stroke range is set from 0mm to 80mm, and the injection speed range is controlled from 10mm / s to 150mm / s. In embodiments that include active adjustment during the holding pressure stage, the holding pressure setpoint range is controlled from 20MPa to 120MPa. The holding pressure stage parameter sequence follows the injection stage parameter sequence and is uniformly spliced ​​into a complete control parameter sequence according to the injection sequence. The mold cavity pressure sensor is installed at the end of the mold cavity or in the runner, and its data collection range is from 0MPa to 150MPa.

[0031] The quality prediction model is constructed as follows: a Long Short-Term Memory (LSTM) network is used, with a feature tensor composed of historical process parameter time-series data as input, and the measured warpage deformation and shrinkage volume of the product as historical quality data for supervised training. The LSM network structure includes an input layer, a hidden layer, and an output layer. The basic units of the hidden layer include a forget gate, an input gate, a cell state update, and an output gate. The input feature tensor includes at least the temperature of each section of the barrel, melt temperature, mold cavity pressure, screw position, and injection speed; for implementations that include active adjustment during the holding pressure stage, a holding pressure setpoint may be further included. During model training, the supervised label of the measured output uses the warpage deformation obtained by a 3D optical scanner and the shrinkage volume measured by an industrial CT scan; when the industrial CT output is a shrinkage volume fraction, the shrinkage volume fraction is multiplied by the nominal volume of the product to obtain the equivalent shrinkage volume. For example, the historical time-series input length is set to 200 time steps, and the hidden layer contains 2 to 3 layers, each containing 64 to 128 neuron nodes.

[0032] The rheological model was constructed as follows: a modified Cross-WLF equation incorporating a metal powder volume fraction term was used to build the titanium alloy feed rheological model. This modified equation adds a Krieger-Dougherty factor term to the original zero-shear viscosity calculation term to represent the concentration effect of the metal powder volume fraction on the melt viscosity. The equation coefficients were determined by conducting orthogonal tests using a dual-capillary rheometer at different temperature gradients and shear rates. The Levenberg-Marquardt nonlinear least squares method was used to perform multivariate regression fitting on the test rheological data to determine the core coefficients in the equation. For example, the test temperatures were selected as 170℃, 190℃, and 210℃, the extrusion pressure range was 10MPa to 50MPa, and the shear rate range was 10²s. - ¹ to 10 4 s -¹; The preferred range for the non-Newtonian fluid index determined by the fitting is 0.35 to 0.55, and the preferred range for the maximum powder loading is 0.62 to 0.68.

[0033] S2. Extract shear mutation features and generate rheological instability index.

[0034] This step is performed during the optimization of the control parameter sequence. A rheological model is used to extract the shear rate abrupt change characteristics caused by the injection control parameter sequence to be optimized, and these shear rate abrupt change characteristics are combined with the melt rheological properties to generate a rheological instability index. The injection control parameter sequence to be optimized is the relationship curve between injection speed and screw position over time; in embodiments that include active adjustment during the holding pressure stage, it also includes the relationship curve between the holding pressure setpoint and time.

[0035] The extraction method for shear rate mutation features is as follows: The second derivative of the melt shear rate with respect to time caused by the injection control parameter sequence to be optimized is calculated using a rheological model to generate Jerk values. The absolute values ​​of the Jerk values ​​are then used to construct a Jerk amplitude sequence, and characteristic peak values ​​are extracted from this sequence as shear rate mutation features. When calculating the shear rate, based on the screw cross-sectional area, nozzle orifice diameter, equivalent hydraulic radius of the flow channel, and equivalent flow cross-section of the cavity, the injection speed and screw position changes are first converted into volumetric flow rates at each time step. Then, the equivalent shear rate at each time step is calculated using the rheological model and process response mapping relationship. Subsequently, the first derivative of the shear rate sequence with time is obtained to get the shear acceleration, and the first derivative of the shear acceleration with time is then obtained to get the second derivative, i.e., the Jerk value.

[0036] The process response mapping relationship is established by simplifying the flow channel into an equivalent circular pipe model, using volumetric flow rate... and apparent viscosity of melt For input, the Poiseuille flow formula for a circular tube is used. Calculate the change in mold cavity pressure, where The equivalent length of the flow channel. The equivalent hydraulic radius of the flow channel is given; the viscous heat generation energy conservation equation is adopted. Calculate the change in melt temperature, where The density of the melt. For isobaric specific heat capacity, Let be the thermal conductivity. The geometric and physical property parameters in the above equations are determined through rheological testing and mold design parameters. For complex flow channel structures, a simplified empirical transfer function based on historical production data can be used instead.

[0037] The rheological instability index is generated as follows: the extracted characteristic peak values ​​are multiplied by the absolute values ​​of the partial derivatives of the feed melt viscosity with respect to the shear rate in the rheological model to obtain an instability product sequence; the root mean square value of this instability product sequence is taken to obtain the scalar rheological instability index within the current sliding window, which is used to adjust the filtering bandwidth in subsequent steps. Each sliding window outputs a rheological instability index, corresponding to all sampling times within the window, and is updated sequentially as the window slides.

[0038] The specific implementation of extracting characteristic peaks from the Jerk amplitude sequence includes the following processing: First, a fixed-length sliding time window is set, and the window is updated point by point, moving forward one sampling period at a time. The mean and standard deviation of the Jerk amplitude sequence within the current window are calculated. The length of the sliding time window is set to cover 50 to 200 consecutive sampling periods, for example, taking 100 discrete sampling points, corresponding to a time span of 200ms.

[0039] Secondly, a constant coefficient is determined based on the difference between the predicted temperature of the current feed melt and the preset ideal injection temperature. The constant coefficient is negatively correlated with this difference; that is, the lower the predicted temperature is relative to the ideal injection temperature, the smaller the constant coefficient. The ideal injection temperature is preset in the range of 190℃ to 200℃ based on the material characteristics of the titanium alloy feed, for example, set to 195℃. The predicted temperature is calculated by the thermodynamic model of the injection molding machine controller. The baseline value of the constant coefficient is set to 3, corresponding to the 99.7% high quantile level under the normal distribution, and the temperature sensitivity adjustment gain coefficient is set to 0.1 / ℃. When the predicted temperature is 185℃, which is 10℃ lower than the ideal temperature, the constant coefficient is lowered to 2.0, corresponding to a reduction in the quantile threshold to detect more potential abnormal high-frequency fluctuations.

[0040] Then, the product of the current window mean plus the constant coefficient and the current window standard deviation is used as the equivalent probability quantile threshold. This threshold is constructed based on the local normal approximation model and is used to characterize the high quantile abnormal amplitude level within the current window.

[0041] Finally, the Jerk amplitude sequence within the current window is traversed using a non-maximum suppression algorithm to identify all local maxima, and local maxima greater than the equivalent probability quantile threshold are extracted as feature peaks.

[0042] S3, filtering separates state fluctuations and high-frequency residual components.

[0043] This step extracts the fluctuation frequency components from the injection control parameter sequence to be optimized. Based on the rheological instability index and the current injection stage corresponding to the injection control parameter sequence to be optimized, adaptive filtering is performed on the fluctuation frequency components to separate state-dependent fluctuation components and high-frequency residual components. State-dependent fluctuation components refer to the effective fluctuation components that are directly related to the injection conditions and rheological state, and are distinct from irregular high-frequency chattering.

[0044] The method for extracting the fluctuation frequency components is as follows: A complex Morlet wavelet basis function is used to perform a continuous wavelet transform on the injection molding control parameter sequence to be optimized, calculating the time-frequency domain spectrum in two-dimensional complex matrix form and the corresponding wavelet coefficients. The center frequency of the wavelet basis is 1Hz, the scale is discretized into 32 layers, corresponding to a sampling frequency of 500Hz, and the frequency coverage range is from 0.5Hz to 250Hz. A continuous wavelet transform is performed on each control channel in the active injection molding control parameter sequence, and the corresponding state-dependent fluctuation frequency components are extracted. These components are then concatenated according to the injection stage and control channel sequence to form a frequency component set.

[0045] The implementation of adaptive filtering includes constructing an asymmetric Gaussian frequency domain filter window, determining the filter window bandwidth, performing filtering, and separating the components.

[0046] The asymmetric Gaussian frequency domain filter window is constructed as follows: the filter window function is a piecewise Gaussian function, with the dominant frequency of the current active injection molding control parameter sequence as the center frequency. Low-frequency side ( Using a fixed bandwidth B_low Gaussian attenuation, the high-frequency side ( A Gaussian attenuation with adjustable bandwidth B_high is employed, resulting in an asymmetric structure with steep low-frequency sides and adjustable high-frequency sides. The highest extreme point of the energy spectrum of the current active injection molding control parameter sequence is identified, and the frequency corresponding to this extreme point is taken as the center frequency. When multiple local extreme frequencies exist in the current spectrum, the local extreme frequencies exceeding a preset energy threshold are taken as candidate center frequencies, and filtering windows are constructed one by one to complete the filtering reconstruction, resulting in multiple state-dependent fluctuation frequency components. When only one dominant frequency exists, the reconstruction result corresponding to that dominant frequency is taken as a single state-dependent fluctuation frequency component.

[0047] The high-frequency base bandwidth of the filter window is limited by the current injection stage, where the base bandwidth of the pressure rise stage is set to be greater than that of the holding stage. The current injection stage is determined by monitoring the gradient of the cavity pressure with respect to time: when the pressure derivative is greater than a positive threshold of 5 MPa / s, it is determined to be the pressure rise stage; when the cavity pressure is at a high pressure level and the absolute value of the pressure derivative is less than a preset stable threshold or the pressure derivative is less than zero and in a slow-decreasing state, it is determined to be the holding stage or the holding pressure slow-decreasing stage.

[0048] The final bandwidth of the filter window is positively correlated with the reciprocal of the rheological instability exponent. The bandwidth calculation formula is as follows: ; in, This represents the actual bandwidth of the high-frequency side of the filter window, in Hz. This refers to the basic bandwidth on the high-frequency side, measured in Hz. This is a preset gain coefficient, dimensionless, and its value is set according to the typical range of the rheological instability index. In this embodiment, based on the titanium alloy feed at a shear rate... to The rheological property test data are within the range, with the typical value of the rheological instability index ranging from 50 to 500. To ensure that the bandwidth adjustment has appropriate sensitivity when changing, the following settings are made: At this point, when the rheological instability index increases from 50 to 500, From the agreement Reduced to approximately It covers the entire adjustment range; This is the rheological instability index generated in S2. This mapping relationship ensures that the larger the rheological instability index, the narrower the filter window bandwidth, thereby achieving adaptive attenuation of the high-frequency chattering signal induced by rheological instability.

[0049] The filtering and separation process is as follows: the time-frequency domain spectrum is processed by frequency domain point-by-point multiplication filtering using a filter window with configured bandwidth; the state-dependent wavelet component is reconstructed by inverse continuous wavelet transform, and the difference between the original wavelet coefficients and the filtered coefficients is extracted as the high-frequency residual component.

[0050] For example, the low-frequency side fixed bandwidth of the filter window is set to 2Hz; the main frequency of the center frequency is typically distributed in the range of 5Hz to 50Hz. The high-frequency side basic bandwidth corresponding to the pressure rise phase is set to 15Hz; the high-frequency side basic bandwidth is reduced to 5Hz during the pressure holding phase or the pressure holding and pressure reduction phase. When the system is in the pressure rise phase, the basic bandwidth is 15Hz, and the rheological instability index reaches 200, the high-frequency side bandwidth is calculated to be 5Hz according to the formula.

[0051] S4. Calculate the sensitivity and configure the penalty weights.

[0052] This step uses a quality prediction model to calculate the sensitivity of state-dependent fluctuation components to product quality evaluation indicators, and then assigns penalty weights to the corresponding state-dependent fluctuation components based on this sensitivity. The product quality evaluation indicators include warpage and shrinkage cavity volume.

[0053] The sensitivity is calculated as follows: First, a 0.5% positive small perturbation is applied to the amplitude of the filtered wavelet coefficients corresponding to the extracted state-dependent fluctuation components. After being reconstructed into time-domain fluctuation components by inverse continuous wavelet transform, these components are superimposed onto the baseline control parameter sequence to be optimized, and then replaced in the corresponding active control channel in the process parameter feature tensor. Simultaneously, the rheological model, combined with the process response mapping relationship, calculates the corresponding shear rate, mold cavity pressure, and melt temperature response based on the perturbed active control channel, updating the relevant state response channels to form a perturbed process parameter sequence consistent with the input dimension of the training phase. This sequence is input into the quality prediction model to obtain the predicted increments of warpage deformation and cavity volume. Then, the predicted increments and perturbation are dimensionless: the warpage prediction increment uses the maximum allowable warpage deformation of 0.2 mm as the normalized denominator, and the cavity prediction increment uses the maximum allowable cavity volume as the normalized denominator; the frequency domain positive small perturbation is dimensionless by dividing by the global maximum allowable amplitude of the corresponding control parameter. The partial derivative in the finite difference sense is calculated by using the ratio of the dimensionless prediction increment to the dimensionless disturbance, and is used as the sensitivity.

[0054] The penalty weights are configured as follows: First, the sensitivity of warpage deformation corresponding to all state-dependent fluctuation components is concatenated with the sensitivity of cavity volume to form a two-dimensional sensitivity matrix. Specifically, the partial derivatives of each component corresponding to the dimensionless warpage deformation are arranged as the first column, and the partial derivatives corresponding to the dimensionless cavity volume are arranged as the second column, concatenating them to form a two-dimensional sensitivity matrix with a dimension of 2 times the number of features. Then, for the sensitivity in the sensitivity matrix, when its absolute value exceeds a preset dimensionless safety threshold, the corresponding penalty weight is non-linearly increased based on an exponential function.

[0055] The formula for calculating penalty weights is: ; in, This is the final penalty weight; The preset base weight is initialized to 1, serving as the normalization benchmark for the penalty system; This is a preset sensitivity adjustment factor used to control the rate of exponential growth; This represents the absolute value of the sensitivity corresponding to the fluctuation component that the current state depends on. This is a dimensionless safety threshold. In this embodiment, Set to 0.5, this threshold corresponds to the critical sensitivity when the influence of the state-dependent fluctuation component on the product quality index reaches 50% of the upper limit of the quality tolerance. When the sensitivity is below this threshold, the contribution of the fluctuation component to product defects is considered acceptable, and no additional penalty is required. When the sensitivity exceeds this threshold, it indicates that the fluctuation component has a significant impact on product defects, and additional constraints need to be imposed. Sensitivity Adjustment Factor The value is 3, and its selection is based on the typical range of sensitivity s (0.1 to 3). Time penalty weight amplification times, when Time magnification This multiplier can effectively constrain highly sensitive fluctuation components while avoiding optimization oscillations caused by excessively rapid penalty growth. At that time, the penalty weight remains the base weight. This formula uses the magnitude by which the absolute value of sensitivity exceeds the safety threshold as the input of the exponential term. Through the non-linear growth characteristics of the natural exponential function, the fluctuation component whose impact on the product defect exceeds the safety range receives a rapidly increasing penalty weight, thereby guiding the subsequent optimization process to actively avoid such highly sensitive fluctuation frequency bands.

[0056] For example, if the absolute value of the cavity volume sensitivity corresponding to a certain frequency component reaches 0.8, the difference between it and the threshold of 0.5 is 0.3, and the exponent term is 0.9, then the penalty weight will be amplified to the power of the natural constant, approximately 2.46 times.

[0057] S5. Construct optimization functions and optimize execution control.

[0058] This step involves constructing an optimization function that includes the product defect prediction evaluation value based on the output of the quality prediction model and the control parameter fluctuation complexity penalty term. Then, the optimization function is optimized to obtain the target injection molding control parameter sequence and the control is executed.

[0059] The method for obtaining the product defect prediction and evaluation value is as follows: divide the product warpage deformation and shrinkage volume output by the quality prediction model by the corresponding upper limit value of the preset quality index to obtain the dimensionless product defect prediction and evaluation value. Among them, the warpage deformation uses the upper limit of the length tolerance as the normalized denominator, and the shrinkage volume uses the upper limit of the volume tolerance as the normalized denominator; when the shrinkage detection uses volume fraction, the upper limit of the shrinkage volume fraction is used as the normalized denominator.

[0060] The method for determining the control parameter fluctuation complexity penalty term is as follows: Frequency domain amplitude statistics are performed on each state-dependent fluctuation component to obtain a dimensionless absolute amplitude sequence; the high-frequency residual component obtained in step S3 is statistically analyzed and used as a supplementary fluctuation component in the calculation, with the penalty weight of the high-frequency residual component using the basic weight. The dimensionless absolute amplitude sequences of each state-dependent fluctuation component and the dimensionless absolute amplitude sequences of the high-frequency residual component are multiplied element-wise with the corresponding penalty weight sequence. The trapezoidal integral rule is used to numerically integrate the product sequence over the frequency band, and the integration result represents the weighted accumulation of the control parameter fluctuation amplitude. The integration result is divided by the maximum permissible control fluctuation amplitude integral reference constant for dimensionless processing to obtain the fluctuation penalty value for a single channel; the processing results for all control channels and all frequency components are summed to obtain the final control parameter fluctuation complexity penalty term. The maximum permissible control fluctuation amplitude integral reference constant is the frequency domain amplitude integral benchmark value corresponding to when the control parameter reaches full-scale fluctuation; for example, the reference constant for the injection speed channel is [value missing]. It has the same dimensions as the weighted amplitude integral result and is used to convert the penalty term into a dimensionless quantity that can be directly added to the dimensionless defect prediction value.

[0061] The optimized function expression is: ; in, To optimize the objective value of the function, the optimization process aims to minimize... For direction; The dimensionless warpage deformation prediction value is obtained from the output of the mass prediction model and after tolerance normalization. This is the dimensionless predicted value for the shrinkage cavity volume. This is a penalty term for the fluctuation complexity of dimensionless control parameters; , , The first preset weight, the second preset weight, and the third preset weight are respectively, and satisfy the following conditions: The weighting coefficients are determined using either the analytic hierarchy process (AHP) or an experimental tuning method based on process experience. For applications where dimensional accuracy is the primary requirement, the weighting coefficients can be increased. The value of can be increased; for applications where internal density is the primary requirement, it can be increased. The value of can be increased; for application scenarios where the stable operation of the actuator is the primary requirement, the value of can be increased. The values ​​of are determined as follows. In this embodiment, considering both product quality and control stability, the three weighting coefficients are configured as 0.4, 0.4, and 0.2, respectively.

[0062] The parameter optimization method is as follows: The particle swarm optimization algorithm is used to minimize the optimization function. The injection stage is evenly divided into 15 time nodes, each corresponding to an injection velocity value. A continuous velocity curve is obtained through linear interpolation between nodes. The screw position is obtained by integrating the velocity curve over time, thus naturally satisfying the kinematic consistency constraint between injection velocity and screw position. The pressure holding stage is evenly divided into 10 time nodes, each corresponding to a pressure holding setpoint. The values ​​of all nodes are concatenated in time sequence into a one-dimensional real vector, encoded as the particle position vector. The particle position is iteratively updated within the set injection velocity boundary and screw position constraint boundary. The inertia weight adopts a decay strategy that smoothly and linearly decreases from 0.9 to 0.4 with the increase of the iteration number. Both the individual cognitive learning factor and the social experience learning factor are configured to be 2. The optimal particle objective function value is achieved when the relative fluctuation change is less than 1×10⁻⁶ after reaching the preset maximum number of iterations or 15 consecutive iterations. -4 When the convergence accuracy threshold is reached, convergence is determined and the iteration is terminated. The globally optimal particle position that satisfies the convergence condition is decoded into the target injection molding control parameter sequence.

[0063] The target injection molding control parameter sequence is converted into low-level register instructions and sent to the PLC controller via industrial Ethernet. The PLC controller then drives the injection molding machine's servo motors, servo valves, and other actuators to perform closed-loop injection molding control.

[0064] For example, the population size of the particle swarm optimization algorithm is set to 50 active particles; the maximum number of iterations is limited to 200 generations; the search range for injection speed is set to 10 mm / s to 150 mm / s, the search range for screw position is 0 mm to 80 mm, and the search range for holding pressure is 20 MPa to 120 MPa.

[0065] This invention also discloses a MIM titanium alloy injection molding control system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a MIM titanium alloy injection molding control method according to the present invention.

[0066] To verify the technical effectiveness of the present invention, a batch production test was conducted using an industrial-grade servo injection molding machine of the same model and a titanium alloy feedstock. The experiment was divided into a control group and an experimental group. The control group used a long short-term memory network prediction combined with a particle swarm optimization algorithm for injection control, with optimization objectives including only two quality indicators: warpage deformation and shrinkage volume, without applying fluctuation complexity constraints. The experimental group adopted the complete scheme of the above-described embodiments of the present invention. Both groups were set with an injection speed range of 10 mm / s to 150 mm / s, a holding pressure range of 20 MPa to 120 MPa, and barrel temperatures controlled between 160°C and 210°C. Each group continuously executed 500 injection molding cycles.

[0067] According to the statistical analysis of three-dimensional optical scanning and industrial CT, the average warpage of titanium alloy products produced in the control group was 0.18 mm, the average shrinkage volume fraction was 1.25%, and the average operating current fluctuation of the injection molding machine servo motor was as high as 12.4% due to control command jitter. The average warpage of products produced in the experimental group was reduced to 0.07 mm, the average shrinkage volume fraction was reduced to 0.42%, and the average operating current fluctuation of the servo motor was reduced to 3.5%. The dispersion of various quality indicators was significantly reduced.

[0068] The experimental results above show that by extracting the peak value of Jerk amplitude and combining it with asymmetric Gaussian frequency domain filtering, the present invention can suppress high-frequency chattering interference caused by rheological instability, making the extraction of state-dependent fluctuation components purer. By constructing a sensitivity matrix and applying exponential penalty weights, the optimization algorithm is guided to avoid control frequency bands that are prone to inducing shape and size deviations in multi-objective search, thereby improving the dimensional accuracy and density of titanium alloy products while reducing the operating load of the actuator.

[0069] like Figure 2 As shown in the figure, the reference set temperature of each temperature zone is displayed in the form of bars, and the error range reflects the reasonable fluctuation range of the temperature in each segment. The overall distribution shows a gradient distribution that gradually increases from the feed end to the nozzle end, which is suitable for the process requirements of gradual melting and plasticization of titanium alloy feed, ensuring uniform feeding plasticization and stable binder performance.

[0070] like Figure 3 As shown in the figure, the difference in average warpage deformation of the product under the two injection molding control schemes is intuitively presented. The warpage deformation of the product in the experimental group using the control method of this invention is significantly lower than that in the control group, which shows that the method has a good suppression effect on warpage defects and can effectively improve the dimensional accuracy and molding quality of MIM titanium alloy products.

[0071] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for controlling the injection molding of MIM titanium alloys, characterized in that, include: S1. Obtain the historical process parameter sequence and historical quality data of the MIM titanium alloy injection molding process, and construct a quality prediction model and a rheological model based on the historical process parameter sequence and the historical quality data; S2. During the optimization process of the control parameter sequence, use the rheological model to extract the shear rate mutation characteristics caused by the injection control parameter sequence to be optimized, and combine the shear rate mutation characteristics with the melt rheological properties to generate a rheological instability index; S3. Extract the fluctuation frequency component of the injection control parameter sequence to be optimized, and based on the rheological instability index and the current injection stage corresponding to the injection control parameter sequence to be optimized, perform adaptive filtering on the fluctuation frequency component to separate the state-dependent fluctuation component and the high-frequency residual component; S4. Use the quality prediction model to calculate the sensitivity of the state-dependent fluctuation component to the product quality evaluation index, and configure a penalty weight for the corresponding state-dependent fluctuation component according to the sensitivity; S5. Construct an optimization function that includes the product defect prediction evaluation value and the control parameter fluctuation complexity penalty term based on the output of the quality prediction model. The fluctuation complexity penalty term is determined based on the state-dependent fluctuation component and its corresponding penalty weight, and the amplitude of the high-frequency residual component. Perform parameter optimization on the optimization function to obtain the target injection molding control parameter sequence and execute the control. The historical process parameter sequence includes state parameters, active injection molding control parameters, and passive state responses; The quality prediction model and rheological model are constructed based on the historical process parameter sequence and the historical quality data, including: constructing a quality prediction model using a long short-term memory network, taking the historical process parameter sequence as input, and using the measured warpage deformation and equivalent shrinkage volume of the product as the historical quality data for supervised training. An improved Cross-WLF equation incorporating a metal powder volume fraction term was used to construct a rheological model for titanium alloy feedstock. The equation coefficients were determined by fitting test data under different temperatures, pressures, and shear rates. The step of extracting the shear rate mutation feature caused by the injection control parameter sequence to be optimized, and combining the shear rate mutation feature with the melt rheological properties to generate a rheological instability index, includes: using the rheological model to calculate the second derivative of the feed melt shear rate with respect to time caused by the injection control parameter sequence to be optimized to generate a Jerk value; taking the absolute value of the Jerk value to construct a Jerk amplitude sequence; and extracting the characteristic peak value from the Jerk amplitude sequence as the shear rate mutation feature; multiplying the characteristic peak value by the absolute value of the partial derivative of the feed melt viscosity with respect to the shear rate in the rheological model to obtain the rheological instability index.

2. The method for controlling the injection molding of MIM titanium alloys according to claim 1, characterized in that, Extracting characteristic peaks from the Jerk amplitude sequence as the shear rate mutation feature includes: setting a sliding time window of fixed length and calculating the mean and standard deviation of the Jerk amplitude sequence within the current window; A constant coefficient is determined based on the difference between the predicted temperature of the current feed melt and the preset ideal injection temperature. The constant coefficient is negatively correlated with the difference, that is, the lower the predicted temperature is relative to the ideal injection temperature, the smaller the value of the constant coefficient. The equivalent probability quantile threshold is the product of the current window mean plus the constant coefficient and the current window standard deviation. Traverse the Jerk amplitude sequence within the current window and extract the local maxima of Jerk amplitudes that are greater than the equivalent probability quantile threshold as the feature peaks.

3. The method for controlling the injection molding of MIM titanium alloys according to claim 1, characterized in that, Adaptive filtering is performed on the fluctuation frequency components to separate the state-dependent fluctuation components and high-frequency residual components, including: The injection control parameter sequence to be optimized is obtained by continuous wavelet transform to obtain the time-frequency domain spectrum and the corresponding wavelet coefficients, and an asymmetric Gaussian frequency domain filter window is constructed. The basic bandwidth of the high-frequency side of the filter window is limited by the current injection stage. The basic bandwidth of the pressure rise stage is set to be greater than the basic bandwidth of the pressure holding stage, and the final bandwidth of the filter window is positively correlated with the reciprocal of the rheological instability index. The time-frequency domain spectrum is filtered using the filter window configured with bandwidth, and the state-dependent fluctuation component is reconstructed by inverse transformation. The difference between the wavelet coefficients and the filtered coefficients is extracted as the high-frequency residual component.

4. The method for controlling the injection molding of MIM titanium alloys according to claim 1, characterized in that, The product quality evaluation indicators include warpage deformation and shrinkage volume; the sensitivity of the state-dependent fluctuation component to the product quality evaluation indicators is calculated using the quality prediction model, including: applying a positive small perturbation to the amplitude corresponding to the extracted state-dependent fluctuation component, and replacing it in the corresponding control channel of the injection molding control parameter sequence to be optimized after time-domain reconstruction; The relevant state response channels are updated by the rheological model or process response mapping model, and then input into the quality prediction model to obtain the predicted increments of warpage deformation and shrinkage volume. The partial derivative in the finite difference sense is calculated using the ratio of the predicted increment after dimensionless processing based on the preset upper limit of quality tolerance to the positive small perturbation, and is used as the sensitivity.

5. The method for controlling the injection molding of MIM titanium alloys according to claim 4, characterized in that, According to the sensitivity, the corresponding state-dependent fluctuation component is assigned a penalty weight, including: concatenating the sensitivity of the warp deformation amount corresponding to all state-dependent fluctuation components with the sensitivity of the cavity volume to form a two-dimensional sensitivity matrix. For the sensitivity in the sensitivity matrix, when its absolute value exceeds a preset dimensionless safety threshold, the corresponding penalty weight is non-linearly increased based on an exponential function.

6. The method for controlling the injection molding of MIM titanium alloys according to claim 5, characterized in that, The nonlinear increase of the corresponding penalty weight based on the exponential function is specifically as follows: calculate the difference between the absolute value of the sensitivity and the dimensionless safety threshold, and multiply the preset sensitivity adjustment factor by the difference to obtain the exponential term; The power function of the exponential term is calculated with the natural constant as the base, and the power function is multiplied by the preset base weight to obtain the final penalty weight.

7. The method for controlling the injection molding of MIM titanium alloys according to claim 1, characterized in that, The optimization function is optimized to obtain the target injection molding control parameter sequence, including: the optimization function includes the product of the first preset weight and the dimensionless warp deformation prediction value, the product of the second preset weight and the dimensionless shrinkage volume prediction value, and the sum of the product of the third preset weight and the dimensionless control parameter fluctuation complexity penalty term. The particle swarm optimization algorithm is used to minimize the optimization function. Within the set injection speed boundary and screw position constraint boundary, the particle position is iteratively updated, and the global optimal particle position when the convergence condition is met is decoded into the target injection molding control parameter sequence.

8. A MIM titanium alloy injection molding control system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the MIM titanium alloy injection molding control method according to any one of claims 1-7.

Images