A method for optimizing injection delay ejection parameters based on demolding force feature extraction

Abstract

The application relates to a kind of injection moulding delay ejection parameter optimization methods based on demoulding force feature extraction, through the synchronous acquisition of multiple cycle high-precision demoulding force signal in key time window, after zero-phase filtering, short-time fourier transform and hilbert transform, the extraction and analysis of 8-25Hz mechanical microvibration signal instantaneous phase sequence are realized, and the benchmark phase evolution fingerprint independent of amplitude accuracy is constructed.When running in real time, the topological deviation between the current and the benchmark fingerprint is quantified using the Wasserstein distance, the adaptive fine-tuning of the ejection time is realized, and the closed-loop optimization is formed in combination with the visual inspection feedback. This scheme effectively improves the robustness and adaptive ability of demoulding integrity detection and ejection time calibration, and significantly reduces the occurrence rate of molding defects.

Description

Technical Field

[0001] This invention relates to the field of signal processing and intelligent closed-loop control technology in injection molding process, and in particular to a method for optimizing injection molding delay ejection parameters based on demolding force feature extraction. Background Technology

[0002] Currently, in the field of injection molding process signal processing and intelligent closed-loop control, demolding force signals are commonly used as key feedback variables for mold cooling and ejection control. Traditional delayed ejection parameter optimization methods often rely on the amplitude, peak value, or statistical characteristics of the demolding force signal, directly relating it to the cooling state and shrinkage degree, and setting thresholds through manual experience or performing simple model fitting based on offline averages. However, the inherent mechanical vibrations, temperature fluctuations, and abnormal mold forces in the injection molding process often cause the demolding force signal to be accompanied by high-frequency noise, amplitude fluctuations, and short-term abnormal jumps. Signal preprocessing is difficult to accurately separate the real micro-vibration response from environmental noise, which easily affects the model input quality and restricts the prediction accuracy and control reliability of the delayed ejection time model.

[0003] Existing technologies mainly employ the following methods for feature analysis: one type relies on physical filters or low-pass algorithms for front-end denoising, but is prone to getting bogged down in the bottleneck of high-frequency micro-vibrations and process interference signals being mixed together; another type utilizes amplitude statistical features within a static window for parameter mapping, but due to the nonlinear dynamic evolution of the cooling and contraction process and the short-term drift of the mold state, the amplitude features lack long-term stability, making it difficult to support fine closed-loop optimization of dynamic ejection time. Although some technologies attempt to integrate higher-order algorithms such as wavelet transform and principal component analysis, they still suffer from limitations such as the model being easily dominated by noise and the loss of intrinsic signal characteristics during feature dimensionality reduction, achieving only limited improvement in scenarios with small parameter fluctuations and having a narrow range of applications.

[0004] Current technical shortcomings are mainly reflected in the following aspects: Existing feature extraction methods are overly sensitive to high-frequency noise and non-process interference, lack consistent identification capabilities for the intrinsic characteristics of micro-vibrations in shrinkage dynamics, and have unstable model input signal quality, failing to provide a reliable basis for closed-loop optimization of delayed ejection windows. Traditional denoising algorithms struggle to distinguish between real micro-vibrations and environmental noise; signal amplitude and statistical characteristics drift with process and environmental changes, requiring frequent manual resetting of feature thresholds and model parameters, increasing maintenance and downtime costs. Furthermore, existing methods generally neglect the advantages of the stability and topological consistency of the demolding force signal's temporal phase spectrum in the dynamic characterization of the cooling and shrinkage stage, failing to fully exploit the inherent robustness of the signal and hindering the realization of adaptive closed-loop control. Summary of the Invention

[0005] This application provides a method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0006] This application provides a method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction, specifically including: S1: Obtain the original demolding force signal sequence within a preset time window from the end of the cooling phase of the mold shaping stage to the ejection trigger; S2: Perform zero-phase filtering on the original demolding force signal sequence, and extract local time-frequency blocks in a preset frequency range based on the filtered original demolding force signal sequence; S3: Perform Hilbert transform based on the local time-frequency block to obtain the instantaneous phase sequence, and calculate the cumulative distribution histogram of adjacent phase differences to construct a reference phase evolution fingerprint; S4: During the real-time operation phase, continuously collect the current demolding force signal within the same preset time window and synchronously generate the current phase evolution fingerprint, and calculate the topological offset between the current phase evolution fingerprint and the reference phase evolution fingerprint; S5: If the topology offset continues to exceed the preset threshold, the ejection time fine-tuning amount is generated by looking up the table based on the monotonicity trend direction and cumulative increment of the topology offset; otherwise, the original delayed ejection time command preset by the injection molding machine control system is maintained. S6: Apply the ejection time fine-tuning amount to the triggering time of the original delayed ejection time command to perform dynamic calibration operation, and after the ejection action is completed, use the product edge continuity discrimination result as demolding integrity detection data; S7: Associate the actual ejection time after each compensation with the demolding integrity detection data, count the number of defects occurring within the continuous compensation cycle, and if the number of defects occurs reaches the preset trigger threshold, trigger the adaptive model update flag.

[0007] S8: In response to the adaptive model update flag, freeze the current reference phase evolution fingerprint and perform local retraining on the reference phase evolution fingerprint using only the latest acquired tail interval distribution data of the phase evolution fingerprint.

[0008] This application provides a method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction, which has the following beneficial effects: (1) By constructing a mold demolding state recognition mechanism based on phase evolution fingerprint, the instantaneous phase dynamic characteristics of the demolding force signal in the critical time window from the end of cooling to before ejection are utilized to extract the robust "phase evolution fingerprint" as the benchmark dynamic mode. This overcomes the shortcomings of traditional methods that are highly dependent on signal amplitude and susceptible to noise interference, and realizes the technological leap from "static numerical judgment" to "dynamic process matching", significantly improving the noise resistance and consistency of state discrimination.

[0009] (2) The topological offset between the real-time phase spectrum and the reference spectrum is measured by the Wasserstein distance, and a rolling compensation mechanism based on the offset trend is established. Fine control can be achieved by adjusting only the ejection trigger time without reconstructing the control model, which greatly reduces the control delay. At the same time, a closed-loop feedback learning system is constructed by combining the visual inspection results of the product integrity. Only the tail distribution features of the phase fingerprint are updated, which balances the model update efficiency and maintenance cost.

[0010] (3) The overall solution makes full use of the phase stability characteristics in the intrinsic dynamic response of the system during demolding, breaks through the bottleneck of high-precision ejection timing determination under multi-source interference environment, and forms a state assessment framework with interpretability and good generalization ability. It shows excellent progressive adaptability when facing slow disturbances such as raw material batch differences, environmental temperature changes or mold wear, significantly reduces appearance defects such as whitening and tearing, and improves product yield and production automation level. Attached Figure Description

[0011] Figure 1 This is the main flowchart of an injection molding delayed ejection parameter optimization method based on demolding force feature extraction; Figure 2 This is a sub-flowchart of an injection molding delayed ejection parameter optimization method based on demolding force feature extraction; Figure 3 This is another sub-flowchart of an injection molding delayed ejection parameter optimization method based on demolding force feature extraction. Detailed Implementation

[0012] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0013] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0014] like Figure 1 As shown, this application provides a method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction, specifically including: Step S1: Obtain the original demolding force signal sequence within a preset time window from the end of the cooling phase of the mold shaping stage to the ejection trigger, as the initial data source for constructing the reference phase spectrum model. The preset time window is a fixed duration window from the end of the cooling phase to the end before ejection triggering, and the original demolding force signal sequence originates from data collected over multiple consecutive stable production cycles. In this embodiment, priority is given to obtaining the original demolding force signal sequence within at least 10 consecutive stable production cycles during the mold shaping stage; the preset time window is 200ms.

[0015] S2: Perform zero-phase filtering on the original demolding force signal sequence, and extract local time-frequency blocks in a preset frequency range based on the filtered original demolding force signal sequence; S3: Perform Hilbert transform based on the local time-frequency block to obtain the instantaneous phase sequence, and calculate the cumulative distribution histogram of adjacent phase differences to construct a reference phase evolution fingerprint; S4: During the real-time operation phase, continuously collect the current demolding force signal within the same preset time window and synchronously generate the current phase evolution fingerprint, and calculate the topological offset between the current phase evolution fingerprint and the reference phase evolution fingerprint; S5: If the topology offset continues to exceed the preset threshold, the ejection time fine-tuning amount is generated by looking up the table based on the monotonicity trend direction and cumulative increment of the topology offset; otherwise, the original delayed ejection time command preset by the injection molding machine control system is maintained. S6: Apply the ejection time fine-tuning amount to the triggering time of the original delayed ejection time command to perform dynamic calibration operation, and after the ejection action is completed, use the product edge continuity discrimination result as demolding integrity detection data; S7: Associate the actual ejection time after each compensation with the demolding integrity detection data, count the number of defects occurring within the continuous compensation cycle, and if the number of defects occurs reaches the preset trigger threshold, trigger the adaptive model update flag.

[0016] S8: In response to the adaptive model update flag, freeze the current reference phase evolution fingerprint and perform local retraining on the reference phase evolution fingerprint using only the latest acquired tail interval distribution data of the phase evolution fingerprint.

[0017] Step S1: Obtain the original demolding force signal sequence within a preset time window from the end of the cooling phase of the mold shaping stage to the ejection trigger, as the initial data source for constructing the reference phase spectrum model. The preset time window is a fixed duration window from the end of the cooling phase to the end before ejection triggering, and the original demolding force signal sequence originates from data collected over multiple consecutive stable production cycles. In this embodiment, priority is given to obtaining the original demolding force signal sequence within at least 10 consecutive stable production cycles during the mold shaping stage; the preset time window is 200ms. Specifically, this includes: S1.1: Real-time monitoring of the mold closing status signal and ejection command trigger signal issued by the injection molding machine control system to lock the time boundary between the end of the cooling process and the start of the ejection action, thereby generating a time window mark from the end of cooling to the ejection trigger to define the data acquisition range.

[0018] S1.1 monitors the mold closing status signal and ejection command trigger signal issued by the injection molding machine control system in real time to lock the time boundary between the end of the cooling process and the start of the ejection action, thereby generating a time window mark from the end of cooling to the ejection trigger to define the data acquisition range.

[0019] The I / O communication interface of the injection molding machine's programmable logic controller (PLC) is connected to capture digital level signals representing the mold closing state and trigger signals representing the ejection mechanism's start pulse in real time. Rising edge detection is performed on the captured digital signals to accurately identify the timestamps of the mold closing completion and ejection command issuance, establishing key timing nodes for a single production cycle. Based on preset cooling kinetics, the data acquisition termination boundary is defined as 200ms before ejection triggering, thus calculating the end time of the acquisition window. in, The timestamp at the moment the eject command is issued. This is the end time of the data collection window.

[0020] Based on the standard cooling time setting in the injection molding process parameters, the start time of the final cooling stage can be calculated: in To cover the fixed duration of the stabilization period of contraction, This is the standard cooling time setting value in the injection molding process parameters. The timestamp of the moment the mold closing is completed. This represents the start time of the final cooling phase. Based on the timestamp calculations above, a time window marker vector is constructed, containing the start time of the final cooling phase and the end time of the acquisition window. This marker vector serves as a synchronous sampling enable signal, precisely defining the effective data acquisition range of the demolding force sensor. It eliminates mechanical impact interference during the initial mold opening process and irrelevant stress fluctuations during the holding pressure stage, ensuring that subsequent phase spectrum analysis focuses only on the pure signal segment reflecting the shrinkage characteristics of the product, thus achieving high-precision spatiotemporal definition of the data acquisition range.

[0021] S1.2: Based on the time window mark from the end of cooling to the ejection trigger, the analog voltage signal output by the high-response pressure sensor installed on the mold ejector plate is synchronously sampled and controlled to obtain the instantaneous sampled data stream of demolding force covering the critical shrinkage stage within a single production cycle.

[0022] Receive the time window marker generated by S1.1 from the end of the cooling phase to the top trigger, which contains the timestamp information of the start time of the end of the cooling phase and the end time of the acquisition window.

[0023] A high-response piezoelectric pressure sensor is installed in the force concentration area of ​​the mold ejector plate, and its signal conditioning circuit is set to a sampling frequency of 10kHz to ensure that the Nyquist sampling theorem is satisfied in the 8-25Hz micro-vibration frequency band and that high-frequency noise is effectively captured.

[0024] A hardware synchronization triggering mechanism is established between the sensor's analog voltage output channel and the injection molding machine's PLC high-speed counting module. The start time of the cooling phase is used as the data acquisition enable signal, and the end time of the acquisition window is used as the acquisition stop signal.

[0025] Within the time interval between the start of the cooling phase and the end of the acquisition window, the instantaneous voltage values ​​output by the sensor are continuously read at fixed sampling intervals to form the original voltage sequence.

[0026] Using preset pressure-voltage linear conversion coefficients, a linear mapping operation is performed on the original voltage sequence, as shown in the following formula: Where F(t) is the instantaneous value of the demolding force at time t, and K calib V is the preset pressure-voltage linear conversion coefficient. raw (t) represents the original voltage sequence at time t.

[0027] The DC component of the converted instantaneous demolding force value sequence is removed by subtracting the mean value within the time window to eliminate the influence of static preload, while retaining the AC component that reflects dynamic shrinkage characteristics.

[0028] The processed AC component data stream is encapsulated into a standardized data structure, and an ID tag for the current production cycle is attached to generate an instantaneous sampled data stream of demolding force covering the critical shrinkage stage.

[0029] By using the above-mentioned synchronous sampling and linear mapping processing method, the time window mark of the previous step is transformed into a high-fidelity demolding force time domain data stream, realizing the complete capture of the micro-vibration characteristics at the end of the cooling period, and providing a distortion-free initial data source for subsequent phase spectrum analysis.

[0030] S1.3: Using the production cycle counter built into the injection molding machine's programmable logic controller, the continuously generated instantaneous sampled data stream of demolding force is processed by cycle number association to filter out a set of demolding force cycle segments corresponding to at least 10 consecutive production cycles with no alarms and stable process parameters. Here, the production cycle with no alarms and stable process parameters refers to the stable production cycle.

[0031] The current production cycle count value, updated in real time, is read from the internal register of the injection molding machine's programmable logic controller (PLC) and used as a time series index key. The cycle count value is then bound to each data point in the instantaneous sampled data stream of the demolding force obtained in step S1.2 using metadata to generate an original signal record with a cycle identifier.

[0032] Call the PLC system status word interface to read the equipment alarm code field and key process parameter settings for the current cycle in parallel, including barrel temperature, injection pressure upper limit and holding time settings.

[0033] A stability criterion logic is constructed, which compares the read alarm code with a zero value. If the alarm code is non-zero, the cycle is marked as an abnormal cycle and removed.

[0034] Perform tolerance range verification on the set values ​​of key process parameters and calculate the absolute value of the deviation between the current cycle parameter value and the standard process formula baseline value.

[0035] If the absolute value of the deviation of any parameter exceeds the preset allowable fluctuation threshold, such as a temperature deviation exceeding ±2℃ or a pressure deviation exceeding ±5bar, the process state of that cycle is determined to be unstable and removed from the candidate set.

[0036] The remaining periodic data after filtering are sorted in ascending order of period count values ​​to check their continuity.

[0037] The number of consecutive, uninterrupted cycles that meet the stability conditions is counted. When the number of consecutive qualified cycles reaches 10, the demolding force data segment corresponding to these 10 cycles is extracted.

[0038] The 10 extracted demolding force data segments are encapsulated into a structured array to form a set of demolding force cycle segments.

[0039] Through the aforementioned periodic correlation and dual screening mechanism, the original data stream containing mixed noise and non-steady-state interference is transformed into a high-confidence benchmark dataset, thereby achieving quality purification of the input source for subsequent modeling.

[0040] S1.4: Based on the start timestamp of each segment in the set of demolding force cycle segments, perform time axis alignment operation to eliminate minor timing jitter during different cycles, thereby constructing a demolding force original signal sequence matrix with a unified time reference.

[0041] Extract the start timestamp of each segment from the set of demolding force cycle segments to construct an initial time index sequence. Using the start time of the first stable cycle as the global time zero point, calculate the time offset of the start time of each subsequent cycle relative to the zero point. Perform linear interpolation resampling on the demolding force data sequence of each cycle, and compensate for the micro-time shift of the data points based on the calculated time offset to eliminate the jitter at the acquisition starting point caused by mechanical transmission backlash. Truncate each cycle signal after time shift compensation to a uniform length to ensure strict alignment of all samples on the time axis. Construct a two-dimensional demolding force original signal sequence matrix, where row vectors represent independent production cycles and column vectors represent aligned discrete time sampling points. Through time axis alignment, transform the discrete cycle segments from the previous step into a demolding force original signal sequence matrix with a unified time reference, achieving precise synchronization of multi-cycle signals at the phase level, laying the data foundation for subsequent high-fidelity position evolution fingerprint extraction.

[0042] S1.5: Perform integrity verification and outlier removal processing on the original demolding force signal sequence matrix to remove non-physical jump data points caused by sensor transient interference, and finally output the original demolding force signal sequence for constructing the reference phase spectrum model.

[0043] Receive the original demolding force signal sequence matrix after time axis alignment. This matrix contains instantaneous sampling data of demolding force for no less than 10 consecutive stable production cycles, which serves as the input for integrity verification and outlier removal.

[0044] By traversing each data point in the sequence matrix, the absolute value of the first-order difference between the current sampling point and its adjacent sampling points is calculated to construct a gradient sequence that reflects the local rate of change of the signal, which is used to initially identify potential abrupt noise.

[0045] A dynamic threshold judgment condition based on physical constraints is set. This threshold is determined according to the maximum response rate of the injection molding machine hydraulic system and the upper limit of the sensor range. When the gradient value of a certain data point exceeds the dynamic threshold, the point is marked as a suspected non-physical jump anomaly point.

[0046] For data points marked as suspected anomalies, a neighborhood consistency test is performed. Five sampling points before and after the point are extracted to form a local sliding window. The mean and standard deviation of the data within the window are calculated to evaluate the statistical stability of the local signal.

[0047] If the value of a suspected outlier deviates from the local window mean by more than three times the standard deviation, it is confirmed as invalid data caused by transient interference from the sensor. Linear interpolation is then used to reconstruct the alternative value of the outlier using the values ​​of the effective neighboring points before and after it.

[0048] If the value of a suspected outlier is within the statistical range of the local window, it is determined to be a normal process fluctuation, and the original sampled value is retained to avoid mistakenly removing the true micro-vibration characteristics of the demolding force.

[0049] After completing the full sequence scan, a second smoothing check is performed. The data segments that have been repaired by interpolation are slightly smoothed with a low-pass filter to eliminate high-frequency glitches introduced by interpolation and ensure the continuity and smoothness of the signal curve.

[0050] Step S2: Perform zero-phase filtering on the original demolding force signal sequence, and extract local time-frequency blocks within a preset frequency range based on the filtered original demolding force signal sequence. The zero-phase filtering on the original demolding force signal sequence is used to eliminate group delay distortion. Extracting local time-frequency blocks within a preset frequency range based on the filtered original demolding force signal sequence includes: generating a time-frequency energy distribution map based on the filtered original demolding force signal sequence using a short-time Fourier transform algorithm, and extracting local time-frequency blocks with frequency ranges limited to the preset frequency range. These local time-frequency blocks are used as the basis data for instantaneous phase calculation. In this embodiment, the preset frequency range is preferably 8-25Hz. Of course, in other embodiments, the preset frequency range can be adaptively adjusted according to actual needs.

[0051] Step S2 specifically includes: S2.1: Perform zero-phase filtering on the original demolding force signal sequence to eliminate group delay distortion and obtain a zero-phase filtered demolding force signal, ensuring that the phase information in subsequent time-frequency analysis is not distorted due to the time lag introduced by the filter.

[0052] The system receives the original demolding force signal sequence output from S1.5. This sequence contains discrete sampled data within a 200ms time window from the end of cooling to the ejection trigger. To address the mechanical vibration and electromagnetic interference noise present in the injection molding process, a Butterworth low-pass filter with a cutoff frequency of 50Hz is designed to retain the critical micro-vibration frequency band of 8-25Hz and filter out high-frequency noise. A forward filter is used to perform an initial traversal of the original signal sequence, generating a preliminarily smoothed intermediate signal state. At this point, the signal phase lags relative to the original time point. The intermediate signal sequence obtained from the forward filter is then time-reversed to construct a reverse-order signal stream, preparing to eliminate the phase lag. The same Butterworth low-pass filter coefficients are applied again to the reverse-order signal stream for backward filtering, causing the signal phase to shift in the opposite direction to the forward filter. The signal output from the backward filter is then time-reversed again to restore the normal time sequence order. At this point, the forward and reverse phase delays cancel each other out. By using zero-phase filtering, the noisy original signal from the previous step is transformed into a zero-phase filtered demodulation force signal without group delay distortion, thereby ensuring accurate preservation of phase information in subsequent time-frequency analysis and avoiding phase spectrum distortion caused by the time lag introduced by the filter.

[0053] S2.2: Based on the zero-phase filtered demolding force signal input short-time Fourier transform algorithm, perform time-frequency decomposition operation to generate a time-frequency energy distribution map containing energy distribution information in the time and frequency dimensions, realizing the mapping and transformation from one-dimensional time domain signal to two-dimensional time-frequency domain features.

[0054] The demolding force signal sequence is received after being processed by zero-phase filtering. This sequence has eliminated group delay distortion and retained the true micro-vibration timing characteristics of the cooling and shrinkage stage.

[0055] The sliding window function of the short-time Fourier transform is set as the Hanning window, and the window length is determined based on the sampling frequency of the injection molding machine and the target frequency band resolution of 8-25Hz, so as to ensure the best balance between the time domain and the frequency domain.

[0056] Signal segments are truncated along the time axis with a fixed overlap rate. A fast Fourier transform is performed on the signal segments within each window to calculate the complex spectral coefficients.

[0057] Where X(k,n) represents the spectral coefficients of the nth time window at the kth frequency point, x is the input signal, w is the window function (specifically, the Hanning window), N is the number of FFT points, H is the step size, n is the index number of the time window, m is the sampling point index of the signal within a single window, x(m+nH) is the sampled value of the input time-domain demolding force signal at time point m+nH, N is the number of points in the Fast Fourier Transform (FFT), N is also the length of a single analysis window, and j is the imaginary unit. This is the rotation factor used to transform a time-domain signal to the frequency domain.

[0058] By taking the square of the modulus of the complex spectral coefficients, the energy spectral density values ​​at each frequency point of each time window are obtained, and a two-dimensional time-frequency energy matrix is ​​constructed.

[0059] By mapping the time axis and frequency axis to image pixel coordinates and mapping energy values ​​to grayscale or pseudo-color intensity, a time-frequency energy distribution map that intuitively reflects the evolution of signal energy over time and frequency is generated.

[0060] By using short-time Fourier transform processing, the one-dimensional time-domain signal from the previous step is transformed into time-frequency energy distribution map data containing both time and frequency dimensions. This achieves a mapping transformation from single time-domain analysis to joint time-frequency domain feature extraction, providing a visualized and quantitative data foundation for subsequent accurate extraction of the 8-25Hz key frequency band.

[0061] S2.3: Based on the inherent frequency characteristics of mechanical micro-vibration during the injection molding cooling and shrinkage stage, set frequency range limitations in the time-frequency energy distribution diagram, extract local time-frequency blocks with frequency ranges between 8 Hz and 25 Hz, and eliminate low-frequency thermal drift noise and high-frequency mechanical impact interference components.

[0062] The system receives time-frequency energy distribution map data generated by zero-phase filtering and short-time Fourier transform. This data contains a two-dimensional matrix composed of time and frequency axes, where each element represents the signal energy intensity at the corresponding spatiotemporal point.

[0063] Based on the physical mechanism of mechanical micro-vibration caused by the interaction between material volume shrinkage and mold wall constraint during the cooling and shrinkage stage of injection mold, the characteristic frequency band containing dynamic state information in the demolding force signal is determined to be between 8 Hz and 25 Hz. This frequency band effectively avoids thermal drift noise below 8 Hz and high-frequency mechanical impact interference above 25 Hz.

[0064] A filtering mask is established on the frequency dimension of the time-frequency energy distribution map. The lower limit frequency threshold is set to 8 Hz and the upper limit frequency threshold is set to 25 Hz. The frequency band data outside this range is zeroed or removed, and the energy distribution sub-matrix within the target frequency band is retained.

[0065] Extract a time-domain segment along the time axis that matches the original demolding force signal acquisition window to ensure that the extracted local time-frequency block completely covers the critical evolution interval from the end of cooling to before ejection triggering.

[0066] Boundary smoothing is performed on the truncated local time-frequency blocks, and the Hanning window function is used to weighted attenuate the spectral edges to suppress the spectral leakage effect caused by the frequency band truncation and ensure the continuity of phase calculation.

[0067] By using the above frequency band limitation and time domain truncation processing methods, the full-spectrum time-frequency energy distribution map is transformed into local time-frequency block data focusing on the dynamic characteristics of cooling contraction. This achieves effective isolation of low-frequency drift and high-frequency noise, providing a high signal-to-noise ratio input basis for subsequent high-precision instantaneous phase extraction.

[0068] For example, the sampling rate is set to 1000Hz, the short-time Fourier transform window length is 64 points, and the overlap rate is 50%. In the generated time-frequency graph, a band-shaped region with energy mainly concentrated around 12Hz is identified. When performing frequency band truncation, the frequency rows corresponding to the index from 8Hz to 25Hz are retained, and the data in the remaining rows are discarded. For a time window of 200ms, the corresponding 10 time frames are truncated to form an 18-row, 10-column local time-frequency block matrix. Hanning window coefficients are applied to the edges of this matrix to make the edge energy smoothly transition to zero. After this processing, the signal-to-noise ratio is significantly improved, effectively removing interference components below 5Hz caused by low-frequency pulsation of the hydraulic system and high-frequency noise above 30Hz caused by high-speed movement of the ejector pin, ensuring the purity of the subsequent Hilbert transform input.

[0069] S2.4: Using the local time-frequency block as the basic data source for instantaneous phase calculation, perform data formatting and encapsulation operations to generate a standardized local time-frequency block dataset, ensuring that the output data format meets the input dimension requirements of the subsequent Hilbert transform.

[0070] S2.5: Execute validity verification logic based on the local time-frequency block dataset to confirm the signal integrity within the main energy peak bandwidth, and output the final instance of the verified local time-frequency block as the sole input basis for constructing a reference phase evolution fingerprint that does not depend on the absolute accuracy of the amplitude.

[0071] Receive the normalized local time-frequency block dataset output by S2.4, which contains complex energy distribution matrices in the time dimension and the frequency dimension of 8-25Hz.

[0072] Energy integration is performed on the spectral vector corresponding to each time slice within the local time-frequency block to calculate the total energy value within the main frequency band, forming an energy envelope sequence that varies with time.

[0073] The ratio of the standard deviation to the mean, i.e., the coefficient of variation, is calculated based on the energy envelope sequence and used to quantify the energy fluctuation stability of the signal during the cooling and contraction phase.

[0074] Set an energy integrity threshold. When the coefficient of variation exceeds the preset upper limit, it is determined that there is a non-periodic mechanical shock or sensor transient failure within that period, and it is marked as an invalid data segment.

[0075] Perform a purge operation on local time-frequency blocks marked as invalid, retaining only qualified time-frequency block instances with a coefficient of variation below a threshold.

[0076] Further check the continuity of the spectral peaks of qualified local time-frequency blocks to ensure that there is a single and stable main energy peak in the 8-25Hz range, and eliminate the risk of phase ambiguity caused by multi-peak interference.

[0077] Through the above validity verification logic, the original time-frequency data from the previous step is transformed into a high-quality local time-frequency block final instance after noise suppression and anomaly removal, thus realizing a high signal-to-noise ratio input basis for constructing a reference phase evolution fingerprint that does not depend on the absolute accuracy of the amplitude.

[0078] like Figure 2 As shown, step S3: Based on the local time-frequency block, perform a Hilbert transform to obtain the instantaneous phase sequence, and calculate the cumulative distribution histogram of adjacent phase differences to construct a reference phase evolution fingerprint. Specifically, the cumulative distribution histogram of adjacent phase differences is calculated by sliding with a preset fixed step size; the reference phase evolution fingerprint does not depend on the absolute accuracy of the amplitude. This includes: S3.1: Perform Hilbert transform on the complex analytic signal within the local time-frequency block to reconstruct the analytic representation of the signal, thereby obtaining a complex analytic signal sequence containing instantaneous amplitude and instantaneous phase information, which serves as the direct input for instantaneous phase calculation.

[0079] The system receives local time-frequency block data that has been intercepted and verified in step S2. This data contains a complex time-frequency component matrix within a frequency range of 8-25Hz, spanning the time window from the end of the cooling phase to the top trigger. The local time-frequency block is reconstructed into a real-time signal using an inverse short-time Fourier transform, and then a Hilbert transform is performed on this real-time signal to obtain the real part signal.

[0080] Perform a Hilbert transform operation on the real part of the signal in the local time-frequency block to construct the corresponding orthogonal imaginary part signal components, thereby establishing the basis for the analytical representation of the signal.

[0081] By using the real part signal and the generated imaginary part signal to construct a complex analytic signal sequence, it is ensured that the trajectory of the signal on the complex plane can fully reflect the instantaneous state of the micro-vibration of the demolding force.

[0082] Calculate the complex analytic signal sequence using the following formula: Where z(t) is a complex analytic signal, x(t) is the real part of the signal in the local time-frequency block, H[·] represents the Hilbert transform operator, and j is the imaginary unit.

[0083] Instantaneous amplitude information is extracted from complex analytic signal sequences to assist in subsequent verification of signal energy stability and to eliminate phase jump anomalies caused by drastic amplitude fluctuations.

[0084] The phase angle information in the complex analytic signal sequence is preserved to form an instantaneous phase raw dataset containing a time dimension index, which serves as the direct input object for solving the instantaneous phase sequence.

[0085] By using Hilbert transform, the local time-frequency block from the previous step is transformed into a complex analytic signal sequence containing instantaneous amplitude and phase information, realizing the mapping from time-frequency domain energy distribution to instantaneous dynamic phase state, and providing a high-fidelity data foundation for constructing phase evolution fingerprints that do not depend on amplitude accuracy.

[0086] For example, a Hilbert transform is performed on a local time-frequency block of data with a sampling rate of 1000Hz and a duration of 200ms. The FFT points are set to 256, and a Hanning window is used for windowing to suppress spectral leakage. The real part of the signal x(t) is transformed to generate the imaginary part H[x(t)], constructing a complex sequence z(t). At t=50ms, if the real part is 0.8V and the imaginary part is 0.6V, the instantaneous amplitude is calculated to be 1.0V, and the instantaneous phase angle is 0.6435 radians. This operation is performed on all 200 sampling points, generating a complex analytic signal sequence of length 200. This sequence effectively preserves the micro-vibration phase characteristics within the 8-25Hz frequency band, eliminates the interference of amplitude fluctuations on phase calculation, significantly improves the signal-to-noise ratio of phase information, and ensures the accuracy of the subsequent phase difference cumulative distribution histogram construction.

[0087] S3.2: Based on the ratio of the real part to the imaginary part of the complex analytic signal sequence, perform arctangent operation to calculate the angle value at each time sampling point, thereby obtaining an instantaneous phase sequence that changes continuously with time, which serves as the basic data characterizing the temporal evolution law of the demolding force micro-vibration.

[0088] The system receives a complex analytic signal sequence generated by the Hilbert transform and extracts the real and imaginary signal components corresponding to each sampling time. The real part corresponds to the original zero-phase filtered demolding force time-domain waveform, and the imaginary part corresponds to its orthogonal components after the Hilbert transform.

[0089] A division operation is performed on the imaginary signal component and the real signal at the same sampling time to construct the tangent ratio of the instantaneous phase angle. This ratio eliminates the influence of signal amplitude fluctuations on angle calculation, retaining only phase evolution information.

[0090] The four-quadrant arctangent function is applied to perform a nonlinear mapping solution on the above tangent ratio to obtain the principal phase angle in the interval [-π, π]. The quadrant in which the phase lies is determined by judging the combination of the signs of the real and imaginary parts, avoiding the numerical jump error of the traditional single-quadrant arctangent function when crossing quadrants.

[0091] Phase unwrapping is performed on the calculated principal phase angle sequence to detect the absolute value of the phase difference between adjacent sampling points. When the phase difference exceeds the π threshold, a multiple of 2π is added or subtracted according to the direction of change to eliminate phase entanglement and generate an unwrapped instantaneous phase sequence that monotonically increases or continuously changes with time.

[0092] By using arctangent operation and phase expansion processing, the complex analytical signal from the previous step is transformed into continuous and unambiguous instantaneous phase sequence data, thereby achieving accurate quantitative characterization of the evolution law of micro-vibration of demolding force and providing high-fidelity input for subsequent phase difference statistics.

[0093] For example, the sampling frequency is set to 1000Hz, and the local time-frequency block contains 200 sampling points. For the k-th sampling point, the real part Re(k) is 0.5V, and the imaginary part Im(k) is 0.866V. The ratio Im(k) / Re(k) is calculated to be 1.732. The principal phase π / 3 rad is obtained by calling the atan2(0.866, 0.5) function. If the phase at the previous moment is -3.1 rad and the calculated value at the current moment is 3.1 rad, and a jump amplitude greater than π is detected, then 2π is subtracted from the current phase to correct it to -3.18 rad, ensuring the continuity of the phase sequence. Finally, a smooth instantaneous phase sequence of 200 points is output, effectively eliminating amplitude noise interference.

[0094] S3.3: The instantaneous phase sequence is subjected to sliding window truncation processing according to a preset fixed step size to generate multiple overlapping short-time phase segments, and then the phase difference between adjacent sampling points in each short-time phase segment is calculated to obtain an instantaneous phase difference sequence that reflects the phase evolution rate.

[0095] Receive the instantaneous phase sequence output by S3.2, set the sliding window length parameter W and the step size parameter S, where the value of W covers at least 3 complete cycles of micro-vibration during the injection molding cooling and shrinkage stage, and the value of S is 1 / 4 of W to ensure sufficient overlap between segments.

[0096] A sliding window truncation operation is performed on the instantaneous phase sequence. Starting from the beginning of the sequence, a subsequence of length W is extracted every S sampling points to generate a set of short-time phase segments with temporal continuity.

[0097] For each short-time phase segment, the phase difference between adjacent sampling points is calculated, and the phase accumulation error is eliminated by discrete difference operation to obtain the instantaneous phase difference vector that reflects the local phase change rate.

[0098] The phase difference value of the i-th sampling point within the k-th short-time phase segment is calculated using the following formula: in, The phase difference value. The instantaneous phase of the next sampling point, It represents the instantaneous phase of the current sampling point.

[0099] The instantaneous phase difference vectors of all short-time phase segments are spliced ​​together to form a global instantaneous phase difference sequence covering the entire time window from the end of the cooling process to the top trigger.

[0100] Phase expansion correction is performed on the global instantaneous phase difference sequence to identify and correct the 2π jump point caused by arctangent operation, ensuring the continuity and physical authenticity of the phase evolution rate data.

[0101] By using a sliding window differential processing method, the instantaneous phase sequence from the previous step is transformed into an instantaneous phase difference sequence that reflects the evolution rate of the micro-vibration dynamics of the demolding force. This achieves high sensitivity in capturing the phase change trend and provides robust basic data for the subsequent construction of a phase evolution fingerprint that does not depend on amplitude accuracy.

[0102] S3.4: Perform statistical frequency accumulation processing based on the instantaneous phase difference sequence to map continuous phase difference values ​​to discretized bins intervals and count the probability of occurrence of each interval, thereby obtaining the original phase difference cumulative distribution histogram data describing the distribution characteristics of phase change.

[0103] The system receives the instantaneous phase difference sequence generated in the previous steps. This sequence contains continuous values ​​reflecting the phase evolution rate of the micro-vibration of the demolding force. A discretized set of statistical intervals is constructed, dividing the range of phase difference values ​​into a predetermined number of bins, each representing a specific phase change rate level. Each phase difference data point in the instantaneous phase difference sequence is traversed, determining which bin it falls into. The frequency of phase difference values ​​falling into the same bin is accumulated and counted, recording the total number of occurrences of data points within each interval. The frequency count results of all bins are arranged in interval order to form a raw phase difference cumulative distribution histogram data describing the distribution characteristics of the phase change pattern. Through the above statistical frequency accumulation processing method, continuous phase difference values ​​are mapped to a discrete probability distribution, achieving a structured characterization of the dynamic characteristics of the demolding force. This provides fundamental data support for subsequently constructing a reference phase evolution fingerprint independent of amplitude accuracy.

[0104] S3.5: Normalize and integrate the original phase difference cumulative distribution histogram data to eliminate the influence of sample number differences and form a probability density function curve, thereby constructing a benchmark phase evolution fingerprint that does not depend on the absolute accuracy of the amplitude and has unique identification, serving as a standardized reference model for subsequent topological consistency comparison.

[0105] Receive the raw phase difference cumulative distribution histogram data output by S3.4. The raw phase difference cumulative distribution histogram data contains discretized bins intervals and their corresponding frequency count values.

[0106] Perform a summation operation on the frequency count values ​​corresponding to each bin interval in the original phase difference cumulative distribution histogram to obtain the total number N of valid phase difference samples in the current sampling window.

[0107] Based on the total number of effective phase difference samples, a division operation is performed on the frequency count value of each bin interval to convert the absolute frequency into a relative frequency value, thereby obtaining a relative frequency distribution sequence and eliminating the influence of sample number differences caused by fluctuations in sampling duration or sampling rate.

[0108] The trapezoidal numerical integration method is used to accumulate the normalized relative frequency distribution sequence to construct a monotonically increasing cumulative distribution function curve.

[0109] The cumulative probability value of the k-th bins interval is calculated using the following formula: Where, p i Let P be the normalized relative frequency of the i-th bin interval, k be the index of the currently traversed bin, and P be the relative frequency of the bin interval. k This represents the cumulative probability value of the k-th bins interval.

[0110] The generated cumulative distribution function curve is smoothed by a smoothing filter to suppress the step-like quantization noise caused by the limited samples, resulting in a continuous and smooth probability density feature curve.

[0111] The probability density feature curve is mapped to a standardized vector space representation and given a unique topological structure identifier, which serves as a reference phase evolution fingerprint.

[0112] By using a normalized integral processing method, the discrete histogram data from the previous step is transformed into a standardized benchmark phase evolution fingerprint with probability conservation properties, thereby achieving a dynamic state characterization that does not depend on the absolute accuracy of the amplitude and providing a consistent comparison benchmark for subsequent Wasserstein distance measurement.

[0113] like Figure 3 As shown, step S4: During the real-time operation phase, continuously collect the current demolding force signal within the same preset time window and synchronously generate the current phase evolution fingerprint, calculating the topological offset between the current phase evolution fingerprint and the reference phase evolution fingerprint. Specifically, this includes: S4.1: Obtain the original signal sequence of the current demolding force within the same preset time window from the end of the cooling phase to the ejection phase during the real-time operation phase, and use it as the initial data source for real-time dynamic response analysis.

[0114] During the real-time operation of the injection molding machine, the programmable logic controller monitors the mold closing signal and the status of the cooling timer to lock the time node when the cooling process ends.

[0115] Based on the locked time node, the start and end times of the current demolding force data acquisition are determined by tracing back a preset fixed time window length.

[0116] A synchronous sampling trigger command is sent to the high-response pressure sensor mounted on the mold ejector plate to start the analog-to-digital conversion module to perform high-frequency acquisition of the analog voltage signal.

[0117] The timestamp of each sampling point is recorded with microsecond-level precision, and the collected discrete voltage values ​​are converted into a sequence of instantaneous numerical values ​​of demolding force in physical dimensions.

[0118] Perform time axis alignment verification on the generated instantaneous numerical sequence of demolding force, and remove non-equal interval sampling points caused by clock jitter to ensure the temporal uniformity of the data sequence.

[0119] Through the above processing method, the baseline model construction requirements of the previous step are transformed into the initial data source for real-time dynamic response analysis, realizing the accurate extraction and standardized output of the demolding force signal in the time domain.

[0120] S4.2: Perform zero-phase filtering on the original current demolding force signal sequence to eliminate group delay distortion, and input the filtered current demolding force signal sequence into the short-time Fourier transform algorithm to generate the current time-frequency energy distribution map.

[0121] The system receives the original demolding force signal sequence acquired by S4.1. This sequence contains real-time sampled data within the time window from the end of cooling to before ejection triggering. Zero-phase digital filtering is performed on the original demolding force signal sequence, using a Butterworth low-pass filter as the basic filter core. A cutoff frequency of 50Hz is set to retain the key frequency band of 8-25Hz and filter out high-frequency mechanical noise. Forward and backward filtering eliminate the group delay distortion introduced by traditional causal filters, ensuring that the phase information of the filtered signal is strictly aligned with the original signal on the time axis, thus obtaining a zero-phase filtered demolding force signal. The zero-phase filtered demolding force signal is input into a short-time Fourier transform algorithm, selecting a Hanning window function as the window function, setting the window length to 64 sampling points, and the overlap rate to 50%, performing time-frequency decomposition. The discrete Fourier transform coefficients within each time window are calculated to generate a complex time-frequency matrix containing energy distribution information in both time and frequency dimensions. The squared modulus of the complex time-frequency matrix is ​​extracted to construct the current time-frequency energy distribution map, realizing the mapping transformation from a one-dimensional time-domain signal to two-dimensional time-frequency domain features. By using zero-phase filtering and time-frequency transformation processing, the result of the previous step is transformed into current time-frequency energy distribution map data with high time resolution and no phase distortion, thus achieving the expected technical effect of providing high-quality input for subsequent local time-frequency block truncation.

[0122] For example, for the current demolding force signal with a sampling rate of 1kHz, a 4th-order Butterworth low-pass filter is used with a cutoff frequency of 50Hz, and a FLTFFET bidirectional filter is performed to eliminate group delay. The STFT Hanning window length is set to 64 points, the frame shift is 32 points, and a 64-point FFT is calculated per frame. A time-frequency energy distribution map with a time resolution of 32ms and a frequency resolution of 15.6Hz is generated. This processing effectively preserves the micro-vibration energy peaks in the 10-20Hz range, while suppressing high-frequency interference above 50Hz, reducing the phase deviation to within 0.1ms, and significantly improving the fidelity of the time-frequency characteristics.

[0123] S4.3: Based on the current time-frequency energy distribution map, extract the current local time-frequency block with a frequency range limited to 8 to 25 Hz, and perform Hilbert transform on the current local time-frequency block to obtain the current instantaneous phase sequence.

[0124] Receive the current time-frequency energy distribution map generated by the short-time Fourier transform. This map contains the time axis, the frequency axis, and the corresponding energy amplitude matrix.

[0125] Based on the physical characteristics of mechanical micro-vibration during the cooling and shrinkage stage of injection mold, a bandpass filter range of 8 Hz to 25 Hz is set in the frequency axis direction to eliminate low-frequency noise of thermal drift below 8 Hz and high-frequency mechanical impact interference above 25 Hz.

[0126] Complex signal sub-matrices of corresponding frequency intervals are extracted from the time-frequency energy distribution map to form the current local time-frequency block, preserving the phase continuity information of the signal in the key frequency band.

[0127] Perform a Hilbert transform on the real part of the signal in the current local time-frequency block to construct the orthogonal components of the imaginary part of the analytic signal, thereby realizing the mapping transformation from the real signal to the complex analytic signal.

[0128] By combining the original real part of the signal with the imaginary part generated by the Hilbert transform, a complex analytic signal sequence is constructed, the mathematical expression of which is: ,in The real part of the signal within the local time-frequency block. This represents the Hilbert transform operator.

[0129] Extract the real and imaginary parts of each sampling point in the complex analytic signal sequence, calculate their ratio, and use it as the basis for instantaneous phase calculation.

[0130] Perform a four-quadrant arctangent operation on the ratio to obtain the instantaneous phase angle that changes continuously with time, thereby eliminating phase ambiguity and ensuring the continuity of the phase value.

[0131] A current instantaneous phase sequence covering the entire time window is generated, which accurately characterizes the dynamic evolution trajectory of the demolding force micro-vibration in the final stage of cooling.

[0132] By using Hilbert transform and arctangent solution processing, the time-frequency local block from the previous step is transformed into the current instantaneous phase sequence data reflecting the intrinsic dynamic state of the system. This enables phase feature extraction that is insensitive to amplitude noise, providing high-fidelity input for subsequent construction of phase evolution fingerprints.

[0133] S4.4: Based on the current instantaneous phase sequence, calculate the cumulative distribution histogram of adjacent phase differences by sliding with a fixed step size, thereby constructing a current phase evolution fingerprint that does not depend on the absolute accuracy of the amplitude.

[0134] Receive the current instantaneous phase sequence output by S4.3, set the sliding window length and step size parameters, and perform segmentation and truncation processing on the phase sequence.

[0135] By iterating through the phase data points within each sliding window, the phase difference between adjacent sampling points is calculated, and a differential sequence reflecting the local phase evolution rate is constructed.

[0136] The calculated phase difference values ​​are mapped to preset discretization intervals bins, and the frequency of occurrence of phase difference values ​​in each interval is counted to generate the original phase difference histogram.

[0137] Normalization is performed on the original histogram to convert the frequency into a probability density distribution, eliminating dimensional differences caused by fluctuations in the number of sampling points.

[0138] The probability density distribution is integrally calculated using the cumulative distribution function to construct a monotonically increasing current phase evolution fingerprint curve.

[0139] By using the above processing method, transient phase fluctuations are transformed into topologically stable distributed characteristic data, achieving robust state characterization that does not depend on amplitude accuracy.

[0140] S4.5: Calculate the topological offset between the current phase evolution fingerprint and the reference phase evolution fingerprint using the Wasserstein distance as a metric to quantify the degree of difference between the real-time cooling state and the historical reference state.

[0141] The current phase evolution fingerprint histogram data generated in S4.4 and the reference phase evolution fingerprint stored in S3.5 are received as input objects for topology offset calculation.

[0142] The baseline phase evolution fingerprint and the current phase evolution fingerprint are regarded as two discrete probability mass distributions, and their support set is defined as the phase difference bins interval index set.

[0143] Using the Wasserstein distance as a metric, we calculate the minimum "work" required to transform the current distribution into the baseline distribution. This work is defined as the integral of the probability mass multiplied by the distance moved.

[0144] It should be noted that, in general, the Wasserstein distance between two probability distributions is defined as the minimum transportation cost required to transform one distribution into another. For a one-dimensional probability distribution, the Wasserstein distance has a concise analytical form: it is equal to the L1 norm (i.e., the absolute value of the integral) of the difference between the two cumulative distribution functions. In this step, both the reference phase evolution fingerprint and the current phase evolution fingerprint are one-dimensional discrete probability distributions (characterized by the cumulative distribution function curve), and their support sets are one-dimensional ordered sets of phase difference bins. Therefore, the calculation of the Wasserstein distance can be simplified to the sum of the absolute values ​​of the difference between the cumulative distribution function values, without needing to solve a complex linear programming problem. The above simplified formula significantly reduces computational complexity while maintaining the physical meaning of the metric, making it suitable for real-time closed-loop control of the injection molding process. The specific calculation formula is as follows: Where W is the topological offset, N is the total number of histogram bins, and F base (i) represents the cumulative distribution function value of the baseline phase evolution fingerprint at the i-th bin, F curr (i) represents the cumulative distribution function value of the current phase evolution fingerprint at the i-th bin.

[0145] Iterate through all bins indices, sum the absolute values ​​of the difference between the baseline cumulative distribution and the current cumulative distribution, and obtain the scalarized topology offset value.

[0146] This value quantifies the overall deviation of the real-time cooling dynamics phase spectrum from the historical stable baseline, eliminating the interference of local noise on a single bin value.

[0147] By using the Wasserstein distance metric, the phase distribution characteristics of the previous step are transformed into a topological offset index that reflects the consistency of the system state. This achieves highly sensitive quantification of minute phase drifts in the demolding force signal, providing a robust decision-making basis for subsequent dynamic calibration of the ejection time.

[0148] Step S5: If the topology offset continuously exceeds a preset threshold, a fine-tuning amount for the ejection time is generated based on the monotonicity trend direction and cumulative increment of the topology offset, using a lookup table. Otherwise, the original delayed ejection time command preset by the injection molding machine control system is maintained. In this embodiment, it is determined whether the topology offset continuously exceeds the preset threshold for three sampling periods. Specifically, this includes: S5.1: Perform sliding window continuity judgment processing on the topological offsets calculated in real time to filter out valid offset events that continuously exceed the preset threshold for three sampling periods, thereby obtaining a statistically significant abnormal state label sequence.

[0149] Receive the current production cycle topology offset value output from step S4.5 and store it in a sliding window buffer of length N, where N is set to 3 to match the determination requirement of continuous sampling cycle.

[0150] Perform a time-series index alignment operation on the historical offset data within the sliding window buffer to ensure that the latest acquired offset is located at the end of the window, forming an offset sequence vector arranged in ascending order of time.

[0151] Iterate through each element in the offset sequence vector, compare the value of each element with the preset stability threshold point by point, and generate the corresponding Boolean state flag bit.

[0152] When a certain offset value is greater than a preset stability threshold, the state at that position is marked as a logical true value, indicating that there is a significant phase spectrum topology deviation in the current period.

[0153] When a certain offset value is less than or equal to a preset stability threshold, the state at that position is marked as a logical false value, indicating that the current cycle is in a normal stable state.

[0154] Perform a logical AND operation on the generated Boolean state flag bit sequence to determine whether the state flag bits of all three consecutive cycles within the window are logically true values.

[0155] If the result of the logical AND operation is true, it is determined that a valid offset event that has continuously exceeded the preset threshold for three sampling periods has been detected, and a high-priority abnormal state trigger signal is generated.

[0156] If the result of the logical AND operation is false, it is determined that the continuity condition is not met, the current abnormal state flag is cleared, and the system is kept in normal monitoring mode.

[0157] By using a sliding window continuity determination process, discrete periodic topological offsets are transformed into a statistically significant sequence of anomaly state markers, enabling effective filtering of transient noise interference and accurate capture of real process drift.

[0158] S5.2: Perform monotonic trend analysis processing based on the topological offset corresponding to the effective offset event marker sequence to identify the cumulative increment direction and rate of change of the topological offset over time, thereby generating a trend feature vector characterizing the degree of deviation in cooling dynamics evolution.

[0159] The topological offsets corresponding to the valid offset event marker sequences selected in S5.1 are extracted, and a time-series vector containing offset values ​​for three consecutive sampling periods is constructed. First-order differencing is performed on this time-series vector to calculate the difference in offset changes between adjacent sampling points, quantifying the instantaneous rate of change of topological deviation. The time-series vector is linearly fitted using the least squares method, and the slope parameter of the fitted line is solved. This slope characterizes the long-term trend direction of the degree of deviation in cooling kinetic evolution. The root mean square value of the fitted residuals is calculated to evaluate the stability of the fluctuations in offset changes and eliminate spurious trend interference caused by random noise. The deviation direction is determined based on the sign of the slope parameter: a positive value indicates an increasing trend in demolding resistance, and a negative value indicates a decreasing trend. The absolute value of the slope is combined with the root mean square value to generate a trend feature vector containing a trend direction identifier, the magnitude of the rate of change, and a fluctuation stability coefficient. Through monotonic trend analysis, the discrete offsets from the previous step are transformed into a trend feature vector characterizing the evolution of the cooling state, achieving the technical effect of moving from static deviation measurement to dynamic trend prediction.

[0160] S5.3: Use the preset offset to time compensation mapping table to perform a lookup operation on the trend feature vector to quantify the abstract phase spectrum topology deviation into a specific physical time adjustment value, thereby obtaining the initial top-out time fine-tuning candidate value.

[0161] The offset-to-time compensation mapping table is a pre-established lookup table that stores the correspondence between phase spectrum topology offsets and ejection time fine-tuning values. This offset-to-time compensation mapping table was obtained through offline calibration experiments: under the baseline state of stable injection molding machine operation, a known degree of cooling state offset (such as adjusting mold temperature or cooling time) was artificially introduced. The corresponding phase spectrum topology offset and the optimal ejection time compensation value required to restore good demolding effect were recorded. Multiple sets of offset-compensation data pairs were repeatedly tested and constructed after discretization, grading, and interpolation. The table uses the offset direction (positive growth or negative decay) as the main index and the offset level range as the row index, storing the corresponding initial ejection time fine-tuning candidate values.

[0162] Receive the trend feature vector generated by the previous steps. This vector contains the monotonicity direction identifier of the topological offset and the cumulative increment value, which serves as the index key value for table lookup retrieval.

[0163] Logical analysis is performed on the monotonicity direction identifier in the trend feature vector to distinguish between two evolution paths: positive growth or negative decay of the offset, which correspond to physical states where cooling contraction is delayed or completed ahead of schedule, respectively.

[0164] Based on the direction identifier obtained from the parsing, the preset offset is locked to the corresponding positive or negative compensation sub-table in the time compensation mapping table to narrow the search range and improve the response speed.

[0165] Extract the cumulative increment value from the trend feature vector and quantify it into discrete offset level intervals, each interval corresponding to a specific phase spectrum topological deviation degree.

[0166] In the locked compensation sub-table, using the offset level range as the row index, perform an exact match retrieval operation to obtain candidate values ​​for the initial top-out time fine-tuning amount corresponding to the degree of deviation.

[0167] If the cumulative increment value is near the boundary of two preset level intervals, linear interpolation is used to calculate the weighted average of the fine-tuning amounts corresponding to adjacent levels to ensure the continuity and smoothness of the compensation amount.

[0168] By using table lookup and interpolation, the abstract phase spectrum topological deviation trend is transformed into specific physical time adjustment values, obtaining candidate values ​​for the initial ejection time fine-tuning, thus achieving a precise mapping from dynamic characteristics to execution parameters.

[0169] S5.4: Perform safety boundary constraint verification on the candidate value of the initial ejection time fine-tuning amount to ensure that the generated adjustment value is within the physical limit range allowed by the response of the injection molding machine hydraulic system, thereby outputting the final ejection time fine-tuning amount that conforms to the characteristics of the actuator.

[0170] The system receives the initial ejection time fine-tuning candidate value from the S5.3 output, reads the maximum response rate and minimum action interval parameters of the injection molding machine's hydraulic servo system, and constructs a physical execution boundary constraint range. Based on the minimum time resolution set by the injection molding machine controller, the initial fine-tuning amount is quantized and truncated to eliminate minute numerical fluctuations exceeding control precision. The system checks whether the absolute value of the initial fine-tuning amount exceeds the maximum allowable single adjustment threshold; if it does, it is clamped to the maximum allowable adjustment value to prevent mechanical shock caused by drastic parameter changes. Combining the remaining cooling time margin of the current production cycle, the system verifies whether the adjusted ejection trigger time is earlier than the safe mold opening time, ensuring that the ejection action does not interfere with the mold opening mechanism. Through the above multiple boundary constraint verifications, the theoretically calculated time compensation value is transformed into the final ejection time fine-tuning amount that conforms to the physical characteristics of the actuator, achieving both the safety and executability of the control command.

[0171] S5.5: Execute the control instruction selection logic based on the presence or absence of the final ejection time fine-tuning amount. If the final ejection time fine-tuning amount exists, generate a new ejection trigger instruction containing time offset parameters. Otherwise, maintain the original delayed ejection time instruction preset by the injection molding machine control system, thereby completing the dynamic calibration decision output.

[0172] Receive the final ejection time fine-tuning data output from step S5.4. This data is a numerical variable or null value identifier after safety boundary constraint verification.

[0173] The final ejection time fine-tuning value is subjected to a non-empty validity determination logic to check whether its value is within the valid range that the injection molding machine control system can resolve.

[0174] If the determination result indicates that there is a fine adjustment for the final ejection time, then the original delayed ejection time reference value corresponding to the current production cycle is read. This reference value is derived from the injection molding process parameter preset database.

[0175] Based on time offset superposition processing, the final ejection time fine-tuning amount is algebraically added to the origin fixed-delay ejection time reference value to generate a corrected ejection trigger time command containing dynamic compensation information.

[0176] The modified ejection trigger time instruction is format-encapsulated and converted into a standard communication protocol data packet that can be recognized by the injection molding machine PLC controller, which includes a timestamp field and an action trigger flag.

[0177] If the determination result is that there is no final ejection time fine-tuning amount, the origin fixed delay ejection time reference value is directly called to generate a regular ejection trigger instruction that maintains the original state, ensuring the continuity of control logic.

[0178] Through the above control instruction selection logic, the parameter optimization results of the previous step are transformed into specific execution layer trigger signals, thereby realizing the dynamic closed-loop calibration decision output of the delayed ejection time.

[0179] Step S6: Apply the ejection time fine-tuning amount to the trigger moment of the original delayed ejection time command to perform dynamic calibration, and obtain the product edge continuity judgment result after the ejection action is completed; wherein, the product edge continuity judgment result is obtained through an embedded vision sensor; the product edge continuity judgment result is used as demolding integrity detection data. Specifically, it includes: S6.1: Obtain the ejection time fine-tuning amount generated by the previous steps and the original delayed ejection time reference value, and perform dynamic correction processing on the original delayed ejection time reference value based on time offset superposition processing to generate the final ejection trigger time command containing compensation information.

[0180] The ejection time fine-tuning amount and the original delayed ejection time reference value output from the previous steps are read as input parameters for time offset superposition processing. Data type validation is performed on the original delayed ejection time reference value to ensure it is a positive real number and within the time setting range allowed by the injection molding machine controller. The system clock synchronization signal for the current production cycle is obtained, and the cooling end time is locked as the starting reference point for time calculation. Linear superposition logic is used to algebraically add the ejection time fine-tuning amount and the original delayed ejection time reference value to construct the calculation model for the final ejection trigger time. The final ejection trigger time is calculated using the following formula: Among them, T final T is the final ejection trigger time that includes compensation information. base The original delayed ejection time reference value is determined based on historical stable cycles, and Δt is the ejection time fine-tuning amount generated by phase spectrum topological offset mapping. A physical boundary constraint check is performed on the calculated final ejection trigger time to determine if it exceeds the minimum response time and maximum safe waiting time range of the injection molding machine's hydraulic ejection mechanism. If the final ejection trigger time exceeds the lower threshold, it is forcibly corrected to the minimum allowable delay time to prevent premature ejection leading to mold sticking or deformation. If the final ejection trigger time exceeds the upper threshold, it is forcibly corrected to the maximum allowable delay time to avoid excessive cooling causing increased energy consumption or production cycle delay. A standard control instruction data package containing the corrected final ejection trigger time is generated. This data package includes a timestamp identifier, instruction type code, and execution priority flag. Through time offset superposition and boundary constraint processing, the fine-tuning amount from the previous step is transformed into a final ejection trigger time instruction that conforms to the physical characteristics of the actuator, achieving precise dynamic calibration of the delayed ejection parameters.

[0181] S6.2: Receive the final ejection trigger time instruction and inject it into the injection molding machine control system. Based on the real-time clock synchronization mechanism, trigger the ejection actuator to perform the action when the time point specified by the final ejection trigger time instruction is reached, so as to complete the dynamic ejection operation of the product in the mold.

[0182] Receive the final ejection trigger time instruction containing time offset parameters, parse the absolute timestamp field and relative delay compensation value in the instruction, and map them to the high-precision real-time clock count value inside the injection molding machine controller.

[0183] Establish a microsecond-level clock synchronization link between the main control unit of the injection molding machine and the ejection servo drive module, and lock the global time base of the current production cycle through a hardware timer interrupt mechanism to eliminate timing jitter errors between multiple controllers.

[0184] The target trigger time after analysis is continuously compared with the real-time incrementing system clock counter. Digital phase-locked loop technology is used to dynamically adjust the sampling frequency of the waiting state to ensure that a high-precision trigger pulse is generated at the moment the clock count value matches.

[0185] When the system clock count reaches the preset final ejection trigger time, it immediately sends an opening signal to the ejection hydraulic proportional valve, driving the ejector plate to perform the ejection action according to the predetermined speed curve, thereby achieving precise physical execution of the demolding time.

[0186] Through the aforementioned real-time clock synchronization and hard triggering mechanism, the abstract time fine-tuning quantity is transformed into millisecond-level physical action execution, realizing rapid response and precise closed-loop control of the delayed ejection parameter to changes in the mold cooling state.

[0187] S6.3: Monitor the completion status of the ejection actuator and activate the embedded vision sensor within a preset time window after the action ends. Based on edge continuity pattern recognition processing, perform feature extraction processing on the collected product surface image data to obtain the edge continuity discrimination result reflecting the demolding integrity.

[0188] Monitor the completion status of the ejector mechanism's action. By reading the ejection position signal bit in the internal register of the injection molding machine's PLC, confirm that the ejector plate has reached the preset end position and stopped moving.

[0189] When the rising edge of the ejection position signal transitions from low to high is detected, a high-precision hardware timer is started, with a delay window set to 100 milliseconds to wait for the release of residual stress on the product surface and vibration attenuation.

[0190] After the delay window ends, the embedded industrial camera is triggered to execute a single exposure acquisition command to obtain grayscale image data containing the key edge areas of the product, ensuring uniform illumination and no motion blur.

[0191] The acquired raw grayscale image data is processed by Gaussian smoothing filtering, and a 3x3 convolution kernel is used to eliminate sensor noise interference while preserving high-frequency components at the edges.

[0192] The Canny edge detection operator is applied to calculate the gradient of the filtered image, and the edge contour is thinned by non-maximum suppression to generate a binarized edge mask.

[0193] Extract the edge pixel coordinate sequence along the demolding direction of the product, calculate the Euclidean distance change rate between adjacent pixels, and construct the edge continuity feature vector.

[0194] Based on a preset continuity threshold criterion, logistic regression classification is performed on the feature vector to determine whether there are breaks or step-like abrupt changes at the edges.

[0195] By processing edge continuity pattern recognition, the image data from the previous step is transformed into edge continuity discrimination results that reflect the integrity of demolding, thus realizing non-contact quantitative evaluation of ejection quality.

[0196] S6.4: Associate the edge continuity discrimination result with the final ejection trigger time instruction executed this time, and perform semantic parsing processing on the edge continuity discrimination result based on defect mapping logic to generate demolding integrity detection data that identifies whether there is edge tearing or top white defect.

[0197] Receive the edge continuity discrimination result data output by step S6.3 and the final ejection trigger time command parameters generated by step S6.1.

[0198] Morphological opening operations are performed on the pixel gradient distribution matrix in the edge continuity discrimination result to eliminate isolated noise points caused by uneven lighting or surface reflection, and a smoothed product contour binarization mask is obtained.

[0199] Based on the smoothed binary mask, the contour closure index is calculated. By tracking the connectivity of contour pixels, the presence of breakpoints or abnormal gaps is identified, and a preliminary geometric integrity status indicator is generated.

[0200] The geometric integrity status identifier is mapped to a preset defect semantic dictionary. If a contour breakpoint is detected and the breakpoint length exceeds 0.5mm, it is marked as an edge tear defect. If the area ratio of a local gray-scale abrupt change region exceeds a threshold, it is marked as a top white defect.

[0201] By combining the timestamp information of the final ejection trigger instruction, a structured demolding integrity detection data package containing defect type code, defect severity level and corresponding ejection time parameter is constructed.

[0202] By using defect mapping logic and semantic parsing, the original image features obtained in the previous step are transformed into standardized demolding integrity detection data, realizing the quantitative characterization and causal relationship of demolding quality status, and providing accurate quality feedback basis for subsequent adaptive model updates.

[0203] S6.5: Integrate the demolding integrity detection data with the corresponding final ejection trigger time command, and package the two into a standard quality feedback record for subsequent statistical analysis and model update based on the structured data encapsulation protocol.

[0204] Step S7: Associate the actual ejection time after each compensation with the demolding integrity detection data, and count the number of defects occurring within consecutive compensation cycles. If edge tearing or whitening defects are still detected after 5 consecutive compensations, trigger the adaptive model update flag. Specifically, this includes: S7.1: Obtain the actual ejection time data after the ejection action is completed and the original image frame data collected by the embedded vision sensor. Perform edge gradient operator convolution processing on the original image frame data to extract gray-scale change information of the product outline, and then generate the product edge continuity discrimination result containing edge continuity status identifier.

[0205] It receives the ejection action completion signal and the raw RGB image frame data acquired by the embedded vision sensor as the initial input for the surface quality analysis of the product.

[0206] The original image frame data is subjected to grayscale conversion processing, and the three-dimensional color space is mapped to a single-channel brightness matrix to eliminate the interference of ambient light color temperature fluctuations on edge detection.

[0207] A Gaussian smoothing filter is used to perform convolution operations on the grayscale image to suppress high-frequency noise and preserve the main contour structure, generating a denoised preprocessed image matrix.

[0208] The Sobel edge gradient operator is applied to perform convolution calculations on the preprocessed image matrix in the horizontal and vertical directions to extract the intensity of gray-level abrupt changes in pixels.

[0209] The overall gradient magnitude of each pixel is calculated using the following formula: Where G is the overall gradient magnitude, Gx is the horizontal gradient component, and Gy is the vertical gradient component.

[0210] A dynamic threshold is set to binarize the comprehensive gradient magnitude matrix, and pixels with values ​​higher than the threshold are selected to form a candidate edge set.

[0211] A morphological closing operation is performed on the candidate edge set to fill the gaps between edge breaks and connect discontinuous contours, forming a continuous outer contour line of the product.

[0212] Traverse the curvature change rate of continuous contour lines, identify breakpoints or distorted areas where curvature abruptly exceeds a preset threshold, and mark them as potential defect locations.

[0213] The distribution density and number of continuity interruptions of potential defect locations are statistically analyzed. If the number of interruptions exceeds the allowable tolerance range, it is judged as edge tearing. If there is abnormal local gray-scale aggregation, it is judged as top whitening.

[0214] Generate binary discrimination results that include edge continuity status indicators, and clearly mark the type of defects such as qualified, edge tear, or top white defect.

[0215] Through the above image processing and feature extraction process, unstructured visual image data is transformed into structured edge continuity discrimination results, realizing a quantitative evaluation of demodulation integrity and providing accurate quality feedback for subsequent adaptive model updates.

[0216] For example, a 1280×720 resolution image of the top of the product is acquired, converted to an 8-bit grayscale image, and filtered using a 3×3 Gaussian kernel (σ=1.5). The Sobel operator is applied to calculate the gradient, and a dynamic threshold is set to the mean plus 1.5 times the standard deviation. Analysis of the image after a certain ejection reveals a gradient interruption of 15 pixels in length on the right edge of the product, accompanied by a local grayscale value that is 20 units lower than the surrounding area, which is determined to be a slight edge tear. The system output status is marked as "edge tear," and this result is stored in the historical compensation archive to trigger subsequent model update logic.

[0217] S7.2: Based on the actual ejection time data and the product edge continuity discrimination result, perform a timestamp alignment operation, bind the two into an ejection quality association record with causal relationship, and write it into the historical compensation archive in a non-volatile storage medium to form a structured defect tracking dataset.

[0218] The results of the product edge continuity judgment generated in S7.1 and the timestamp data of the actual ejection time recorded in S6.2 are used as input sources for constructing causal relationship records.

[0219] High-precision clock synchronization calibration is performed on the actual ejection time timestamp to eliminate clock drift error between the injection molding machine controller and the vision sensor system, ensuring the alignment accuracy of multi-source data under a unified time reference.

[0220] Based on the calibrated timestamp, a temporal mapping index is established between the ejection action trigger event and the visual inspection completion event, and discrete control command execution points are logically bound to quality inspection result points.

[0221] By adopting a structured data encapsulation protocol, the bound ejection time parameter, edge continuity status identifier and corresponding production cycle number are combined into a single data object to form an ejection quality association record with complete semantic information.

[0222] Through a database transaction mechanism, the ejected quality association record is atomically written to the historical compensation archive in a non-volatile storage medium, ensuring the integrity and immutability of the data write.

[0223] Perform an index update operation on the historical compensation archive, maintain the orderly arrangement of records according to the production cycle time series, and form a traceable structured defect tracking dataset.

[0224] By using the timestamp alignment and structured storage processing methods described above, the scattered control data and quality inspection data from the previous step are transformed into historical archive data with causal logical relationships, enabling accurate traceability of demolding quality and ejection parameters, and providing a reliable data foundation for subsequent continuous defect statistics.

[0225] S7.3: Read the ejection quality association records of the most recent five consecutive compensation cycles in the historical compensation archive, perform logical traversal statistics on the edge continuity status identifiers in the ejection quality association records, and calculate the cumulative defect count value of consecutive edge tearing or top whitening defects.

[0226] The system retrieves the ejection quality correlation records for the five most recent consecutive production cycles from the historical compensation archive and extracts the edge continuity status identifier field from each record. It then logically traverses these edge continuity status identifiers to identify abnormal status markers labeled "edge tear" or "top white defect." The cumulative defect count is initialized to zero, and the sliding window pointer is set to point to the latest production cycle record. The system determines whether the edge continuity status identifier of the record currently pointed to belongs to a preset defect type set. If the determination result is true, the cumulative defect count is incremented by one, and the sliding window pointer is moved forward one cycle position. If the determination result is false, the cumulative defect count is immediately reset to zero, and the sliding window pointer is updated to continue checking the next adjacent cycle. This judgment and counting logic is repeated until a full traversal and statistical analysis of the five most recent consecutive cycles is completed. Through this chained traversal and conditional accumulation mechanism, discrete quality inspection results are transformed into a cumulative defect count value characterizing process stability, achieving a quantitative assessment of the risk of continuous demolding failures and providing a precise decision-making basis for triggering adaptive model updates.

[0227] S7.4: Determine whether the cumulative defect count value has reached the preset threshold of five consecutive defects. If the determination result is true, generate a high-priority adaptive model update flag signal. This signal serves as the only enabling condition for triggering the subsequent local retraining process of the reference phase evolution fingerprint.

[0228] Step S8: In response to the adaptive model update flag, freeze the current reference phase evolution fingerprint, and perform local retraining on the reference phase evolution fingerprint using only the latest acquired tail interval distribution data of the phase evolution fingerprint to complete closed-loop feedback optimization. Specifically, this includes: S8.1: Obtain the adaptive model update flag triggered by the previous step and the currently stored reference phase evolution fingerprint, perform a model freeze operation based on the adaptive model update flag, and generate a frozen reference phase evolution fingerprint in a locked state to prevent historical data from being accidentally overwritten.

[0229] Receive the adaptive model update flag signal generated in step S7.4, which serves as the sole enabling condition for triggering the local retraining process of the baseline phase evolution fingerprint.

[0230] The baseline phase evolution fingerprint data currently in an active state is read from the historical compensation archive of non-volatile storage media. The model includes probability density distribution parameters for the head, middle and tail regions.

[0231] Based on the adaptive model update flag signal, the mutex lock control logic is executed to set the read and write permissions of the reference phase evolution fingerprint to read-only state, and generate a frozen reference phase evolution fingerprint in a locked state.

[0232] By using memory mapping technology, the frozen model data is loaded into an independent secure cache area, isolating the write operation of real-time production data and preventing the historical stable feature data from being accidentally overwritten or tampered with due to concurrent access during subsequent local retraining.

[0233] Perform an integrity check on the frozen baseline phase evolution fingerprint to confirm that the hash values ​​of the data distributed in the head and middle intervals are consistent with the archived records, ensuring that the basic data used for splicing and fusion has not been damaged.

[0234] By using the above-mentioned model freezing and isolation methods, the update trigger signal of the previous step is transformed into protected static reference model data, which realizes the solidification and protection of historical stable process characteristics. This provides a safe and consistent data benchmark for subsequent lightweight local retraining only for the tail section, effectively avoiding the waste of computing resources and model oscillation risks caused by full retraining.

[0235] For example, when the system detects a top-white defect for five consecutive cycles, a high-level adaptive model update flag is generated. The PLC controller responds immediately, reading the reference phase evolution fingerprint from the Flash memory. This model consists of a probability distribution array of 100 bins. The controller sets a memory protection bit, marking the array as read-only, and copies it to the security buffer 0x2000-0x2190 in RAM. At this point, even if new demolding force data continues to flow in, the data in the first 80 bins (corresponding to the head and middle sections) in the buffer cannot be modified. A CRC32 checksum confirms that the checksum of the frozen model matches the archived value 0xA1B2C3D4, verifying data integrity. This process takes less than 2ms, ensuring the real-time performance and security of the model update, and significantly improving the system's robustness in process drift scenarios.

[0236] S8.2: Read the original demolding force signal sequence synchronously collected within the last five consecutive production cycles, and perform segmentation processing on the original demolding force signal sequence based on the time window interception rule from the end of cooling to before ejection triggering, to generate the latest phase evolution fingerprint tail interval original data set to be updated.

[0237] The original demolding force signal sequence of the most recent five consecutive production cycles is read from the historical compensation archive and used as a dynamic data source for local model updates.

[0238] Based on the time window mark from the end of cooling to the ejection trigger, the start point alignment operation is performed on the demolding force signal of each cycle to eliminate timing jitter caused by the hydraulic response fluctuation of the injection molding machine.

[0239] Linear interpolation is used to resample the aligned signal, and the number of sampling points in different periods is standardized to a fixed length N to ensure the consistency of subsequent frequency domain analysis dimensions.

[0240] Extract the high signal-to-noise ratio range from the standardized signal corresponding to the end of the cooling contraction phase, and remove the mechanical impact noise segment generated at the moment of the ejection action.

[0241] The five extracted periodic signal segments are stacked in chronological order to construct a multidimensional time-series data matrix, forming the original data set of the latest phase evolution fingerprint tail interval to be updated.

[0242] By using the segmentation and standardization methods described above, the update requirement triggered in the previous step is transformed into a set of original data for the tail interval with spatiotemporal consistency, thus achieving the expected technical effect of data preparation for local retraining of the benchmark model.

[0243] S8.3: Perform zero-phase filtering and short-time Fourier transform processing sequentially on the original data set of the latest phase evolution fingerprint tail interval to generate local time-frequency blocks, and then perform Hilbert transform and sliding cumulative distribution calculation based on the local time-frequency blocks to generate the latest phase evolution fingerprint tail interval distribution data with temporal consistency.

[0244] Zero-phase filtering is performed on the original data set of the tail interval of the latest phase evolution fingerprint. A bidirectional Butterworth filter is used to eliminate signal group delay and ensure that the phase information is retained without distortion in the time domain.

[0245] The filtered demolding force signal is input into a short-time Fourier transform algorithm. The Hanning window length is set to 64 sampling points and the overlap rate is 50%, generating a high-resolution local time-frequency energy distribution map.

[0246] Based on the frequency constraint condition of 8 to 25 Hz, the complex spectral components of the corresponding frequency band are extracted from the time-frequency energy distribution map to construct a local time-frequency block matrix that reflects the micro-vibration characteristics of the final cooling stage.

[0247] Perform a Hilbert transform on the local time-frequency block matrix to construct an analytic signal to extract instantaneous phase information. The instantaneous phase sequence is calculated using the following formula: in, This is the filtered signal. For Hilbert operators.

[0248] Based on the instantaneous phase sequence, the phase difference between adjacent sampling points is calculated by sliding with a fixed step size, thereby generating a difference sequence that characterizes the phase evolution rate.

[0249] Histogram statistics are performed on the phase difference sequence to map the phase difference values ​​to a discretized interval and normalize them to form a probability density distribution curve.

[0250] By employing zero-phase filtering, time-frequency truncation, and Hilbert phase calculation, the original demolding force signal is transformed into the latest phase evolution fingerprint tail interval distribution data with temporal consistency, thereby achieving high-precision feature reconstruction that resists noise interference.

[0251] S8.4: Extract the head interval distribution data and middle interval distribution data from the frozen reference phase evolution fingerprint, and splice and fuse the head interval distribution data, middle interval distribution data and tail interval distribution data of the latest phase evolution fingerprint to generate an updated reference phase evolution fingerprint containing historical stable features and recent dynamic features.

[0252] S8.5: Replace the original reference phase evolution fingerprint in the system with the updated reference phase evolution fingerprint, and reset the adaptive model update flag to complete the closed-loop feedback optimization process and generate a delayed ejection time calibration reference with the latest process adaptability.

[0253] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0254] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0255] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction, characterized in that, include: S1: Obtain the original demolding force signal sequence within a preset time window from the end of the cooling phase of the mold shaping stage to the ejection trigger; S2: Perform zero-phase filtering on the original demolding force signal sequence, and extract local time-frequency blocks in a preset frequency range based on the filtered original demolding force signal sequence; S3: Perform Hilbert transform based on the local time-frequency block to obtain the instantaneous phase sequence, and calculate the cumulative distribution histogram of adjacent phase differences to construct a reference phase evolution fingerprint; S4: During the real-time operation phase, continuously collect the current demolding force signal within the same preset time window and synchronously generate the current phase evolution fingerprint, and calculate the topological offset between the current phase evolution fingerprint and the reference phase evolution fingerprint; S5: If the topology offset continues to exceed the preset threshold, the ejection time fine-tuning amount is generated by looking up the table based on the monotonicity trend direction and cumulative increment of the topology offset; otherwise, the original delayed ejection time command preset by the injection molding machine control system is maintained. S6: Apply the ejection time fine-tuning amount to the triggering time of the original delayed ejection time command to perform dynamic calibration operation, and after the ejection action is completed, use the product edge continuity discrimination result as demolding integrity detection data.

2. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, After step S6, the following is included: S7: Associate the actual ejection time after each compensation with the demolding integrity detection data, count the number of defects occurring within the continuous compensation cycle, and if the number of defects occurs reaches the preset trigger threshold, trigger the adaptive model update flag. S8: In response to the adaptive model update flag, freeze the current reference phase evolution fingerprint and perform local retraining on the reference phase evolution fingerprint using only the latest acquired tail interval distribution data of the phase evolution fingerprint.

3. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, Step S3 specifically includes: The complex analytic signal within the local time-frequency block is subjected to Hilbert transform processing to obtain a complex analytic signal sequence containing instantaneous amplitude information and instantaneous phase information; Based on the ratio of the real part to the imaginary part of the complex analytic signal sequence, an arctangent operation is performed to obtain an instantaneous phase sequence that changes continuously with time. The instantaneous phase sequence is subjected to sliding window truncation processing according to a preset fixed step size to generate multiple overlapping short-time phase segments. Then, the phase difference between adjacent sampling points in each short-time phase segment is calculated to obtain the instantaneous phase difference sequence. Statistical frequency accumulation processing is performed on the instantaneous phase difference sequence to obtain the original phase difference cumulative distribution histogram data. The original phase difference cumulative distribution histogram data is subjected to normalized integral processing to construct the reference phase evolution fingerprint.

4. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 3, characterized in that, The original phase difference cumulative distribution histogram data includes discretized bins intervals and their corresponding frequency count values. The original phase difference cumulative distribution histogram data is normalized and integrated to construct the reference phase evolution fingerprint, including: Perform a summation operation on the frequency count values ​​corresponding to each bins interval to obtain the total number of valid phase difference samples in the current sampling window; Based on the total number of effective phase difference samples, a division operation is performed on the frequency count value corresponding to each bins interval to convert the absolute frequency into a relative frequency value, thereby obtaining a relative frequency distribution sequence. The relative frequency distribution sequence is accumulated using the trapezoidal numerical integration method to construct a monotonically increasing cumulative distribution function curve; Perform a smoothing filter on the cumulative distribution function curve to form a probability density feature curve; The probability density feature curve is mapped to a standardized vector space representation and given a unique topological structure identifier, which serves as the reference phase evolution fingerprint.

5. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, Step S4 specifically includes: Acquire the original signal sequence of the current demolding force within the same preset time window from the end of the cooling phase to ejection during the real-time operation phase; The original demolding force signal sequence is subjected to zero-phase filtering, and the filtered current demolding force signal sequence is input into a short-time Fourier transform algorithm to generate the current time-frequency energy distribution map. Based on the current time-frequency energy distribution map, a current local time-frequency block with a frequency range limited to the preset frequency interval is extracted, and a Hilbert transform is performed on the current local time-frequency block to obtain the current instantaneous phase sequence; Based on the current instantaneous phase sequence, the cumulative distribution histogram of adjacent phase differences is calculated by sliding with a fixed step size, and the current phase evolution fingerprint is constructed. The topological offset between the current phase evolution fingerprint and the reference phase evolution fingerprint is calculated using the Wasserstein distance as a metric.

6. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, Step S5 specifically includes: The topological offset is subjected to a sliding window continuity determination process to filter out valid offset events that continuously exceed a preset threshold for three sampling periods, thereby obtaining an abnormal state marker sequence. Based on the topological offset corresponding to the effective offset event marker sequence, perform monotonic trend analysis processing to identify the cumulative increment direction and rate of change of the topological offset over time, and generate a trend feature vector. The trend feature vector is retrieved by using a preset offset-to-time compensation mapping table to obtain candidate values ​​for the initial top-out time fine-tuning. Perform safety boundary constraint verification on the initial candidate values ​​of the ejection time fine-tuning, and output the final ejection time fine-tuning. The control instruction selection logic is executed based on the presence or absence of the final ejection time fine-tuning amount. If the final ejection time fine-tuning amount exists, a new ejection trigger instruction containing time offset parameters is generated; otherwise, the original delayed ejection time instruction preset by the injection molding machine control system remains unchanged.

7. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, Based on the filtered original demolding force signal sequence, local time-frequency blocks within a preset frequency range are extracted, including: Based on the filtered demolding force original signal sequence, a time-frequency energy distribution map is generated according to the short-time Fourier transform algorithm, and a local time-frequency block with a frequency range limited to the preset frequency interval is extracted.

8. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 7, characterized in that, The preset frequency range is 8-25Hz.

9. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, Determining whether to trigger adaptive model updates based on the frequency of defect occurrences includes: The actual ejection time after each compensation is associated with the demolding integrity detection data, and the number of defects occurring within a continuous compensation cycle is counted. If edge tearing or whitening defects are still detected after 5 consecutive compensations, the adaptive model update flag is triggered.

10. The method for optimizing injection molding delayed ejection parameters based on demolding force feature extraction according to claim 1, characterized in that, The preset time window is a fixed duration window that starts from the end of the cooling phase and ends before the ejection trigger, and the original demolding force signal sequence is derived from the collected data of multiple consecutive stable production cycles.

Images