Online Detection Method for Injection Molding Fill Imbalance Based on Cavity Pressure Time-Sequence Analysis
By analyzing the timing of mold cavity pressure, the problems of insufficient acquisition of injection mold cavity pressure data and difficulty in comparing curves were solved, realizing the automation, stable detection and quality consistency recording of the injection filling process, and reducing the misjudgment rate.
Patent Information
- Application Number
- CN202611486794.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-23
- Publication Date
- 2026-11-17
AI Technical Summary
In the existing technology, the injection mold cavity pressure data is not sufficiently obtained, the sensor noise interference is serious, the mold cavity pressure curves cannot be directly compared, and there is a lack of common working condition benchmarks within the same injection cycle, resulting in a high rate of misjudgment and false alarm in filling imbalance detection.
By acquiring the cavity pressure timing signal, using the injection start trigger signal to extract the filling stage data, performing baseline correction and filtering to remove noise, mapping to the filling phase coordinates and aligning them, extracting multi-dimensional feature vectors, establishing a common working condition feature vector library, generating filling balance indexes, and adaptively adjusting the anomaly judgment threshold.
It improves the accuracy and comparability of data during the filling stage, reduces the false judgment rate, and realizes automated, stable detection and consistent quality recording of the injection molding process.
Smart Images

Figure CN123103997A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of injection mold cavity pressure signal analysis and intelligent detection of mold cavity filling status, and in particular to an online detection method for injection filling imbalance based on mold cavity pressure time sequence analysis. Background Technology
[0002] In the current field of injection mold cavity pressure signal analysis and intelligent detection of mold cavity filling status, mold cavity pressure sensors are widely used to collect pressure timing signals. By analyzing pressure changes during the filling stage, the filling status of the mold cavity can be monitored and its consistency evaluated. The mainstream technical solution focuses on using distributed sensors deployed in a multi-cavity injection mold, with each cavity equipped with an independent pressure detection module. Combined with the periodic trigger signals provided by the injection molding machine, the pressure curve of each cavity from the melt front to the completion of filling is captured and analyzed.
[0003] Currently, some systems also introduce characteristic parameter extraction of the cavity pressure curve, including indicators such as filling arrival time, pressure rise rate, pressure integral, and end pressure, and compare them with standard parameters or historical statistical ranges under known process conditions in the industry to determine abnormal conditions such as flow lag, increased resistance, or excessively fast filling between cavities.
[0004] However, existing technologies have the following problems: First, the acquisition and purification of pressure data during the filling stage are insufficient. The start time of the injection cycle is difficult to pinpoint precisely, and the collected pressure data often includes pressure data from both the zero-pressure segment before injection and the holding pressure segment, causing the true filling stage data to be contaminated by redundant data. At the same time, the pressure curve is distorted due to the influence of sensor zero drift, high-frequency impact noise, and non-physical spikes caused by electromagnetic interference or sensor transient failures, making it difficult to accurately determine the arrival time of the melt front and the end time of filling.
[0005] Secondly, the pressure curves of different cavities cannot be directly compared laterally. The physical flow channel lengths of each cavity in a multi-cavity mold are inconsistent, and there are inherent differences in the time when the melt arrives at each cavity and the filling duration of each cavity. This causes the pressure curves of each cavity to be misaligned on the original time axis and the time scale to be inconsistent. Directly comparing them on the original time axis will mix these inherent differences into the judgment results of filling imbalance, resulting in misjudgment.
[0006] Third, there is a lack of common operating condition benchmarks within the same injection molding cycle. Factors such as material batch fluctuations, changes in ambient temperature, and equipment response delays can cause an overall drift in the pressure curves of all mold cavities. Because there is a lack of a comparative benchmark that uses the common level of characteristics of each mold cavity within the same injection molding cycle as a reference, this overall operating condition drift can be misjudged as an abnormal filling of individual mold cavities. Summary of the Invention
[0007] This invention provides an online detection method for injection molding filling imbalance based on cavity pressure time-series analysis, aiming to solve the problems mentioned in the background art, such as insufficient acquisition and purification of pressure data during the filling stage, inability to directly compare pressure curves of different cavities laterally, and lack of common working condition benchmarks within the same injection cycle. The online detection method for injection molding filling imbalance based on cavity pressure time-series analysis provided by this invention specifically includes: Acquire the pressure timing signal and injection start trigger signal of each mold cavity, extract the pressure data of each mold cavity filling stage according to the injection start trigger signal, and generate a set of pressure curves for the filling stage. Baseline correction, filtering and noise reduction, and outlier identification are performed on the set of pressure curves during the filling stage to determine the arrival time of the melt front and the end time of filling, thereby generating a clean filling pressure curve sequence. Map the pressure time axis of each mold cavity from the arrival time of the melt front to the end time of filling to the filling phase coordinate and perform phase alignment to generate a standardized set of filling pressure curves. The arrival time difference, pressure rise slope, pressure integral value, end pressure value and curve shape deviation vector are extracted from the standardized filling pressure curve set. Common working condition feature vectors are generated using the median value of each cavity feature in the same period to form a multi-dimensional time series feature vector set. A benchmark library of allowable fluctuation ranges of multidimensional features is established based on historical normal cycles. The multidimensional time series feature vector set is compared with the common working condition feature vector and the benchmark library of allowable fluctuation ranges of multidimensional features to obtain the lateral deviation and the longitudinal deviation, and then fused to generate a filling balance index. The filling balance index is compared with the adaptive anomaly determination threshold. When the filling balance index exceeds the adaptive anomaly determination threshold, the abnormal cavity with filling lag or filling too fast and the degree of abnormality are determined, and a filling imbalance detection result containing the abnormal cavity number, the degree of abnormality and the type of abnormal feature is generated.
[0008] The online detection method for injection filling imbalance based on cavity pressure time-series analysis further includes: Based on the filling imbalance detection results, an alarm signal or process adjustment suggestion instruction is output. When the current cycle is confirmed to be in a normal state based on the filling balance index trend of multiple consecutive cycles, an exponentially weighted moving average is used to update the multidimensional feature allowable fluctuation range benchmark library with the multidimensional time series feature vector set of the current cycle, and the adaptive anomaly judgment threshold of subsequent cycles is updated accordingly.
[0009] The online detection method for injection filling imbalance based on cavity pressure time sequence analysis provided by this invention has the following beneficial effects: (1) By locking the injection cycle start point by triggering the injection start signal and extracting the pressure data of the filling stage, the interference of the zero pressure segment and the holding pressure segment before injection is eliminated. Then, the influence of sensor zero drift, high frequency impact noise and non-physical spikes is eliminated by baseline correction, filtering and noise reduction and abnormal point identification, so that the determination of the arrival time of the melt front and the end time of filling is more accurate. On this basis, the filling phase alignment method based on the arrival time of the melt front effectively overcomes the problem of timing signal misalignment caused by the difference in the start time of filling of each cavity in the multi-cavity injection molding process, and significantly improves the comparability between pressure curves of different cavities. Combined with the multi-dimensional feature extraction mechanism such as "arrival time difference, pressure rise slope, pressure integral value, end pressure value and curve shape deviation vector", it can comprehensively characterize the dynamic consistency of filling behavior between cavities. Compared with the single index evaluation method that only relies on peak pressure or filling time difference, this scheme has stronger sensitivity and accuracy in identifying small but trending filling deviations, and effectively solves the problem of early abnormality missed detection and misjudgment caused by insufficient feature characterization in the prior art.
[0010] (2) The median value of each mold cavity feature within the same period is introduced as a common working condition feature vector, and this vector is used as a reference benchmark to realize adaptive compensation for overall process drift (such as material batch fluctuation, environmental temperature change, equipment response delay, etc.), avoiding the technical defects of traditional methods that frequently trigger false alarms due to excessive reliance on historical standard curves; further, the lateral deviation obtained from the common working condition feature vector is combined with the longitudinal deviation obtained from the multidimensional feature allowable fluctuation range benchmark library to construct a filling balance index that has both spatial comparison and temporal stability, so that the system can continuously and stably evaluate filling consistency under complex working conditions; and through the consistency fusion constraint of arrival time difference, pressure rise slope, pressure integral value, end pressure value and curve shape deviation vector, transient interference (such as sensor jitter, electromagnetic noise) and real structural imbalance problems such as flow stagnation and gate blockage are effectively distinguished, significantly reducing the false alarm rate and improving the reliability and engineering usability of the diagnostic results.
[0011] (3) Since the judgment is based on an adaptive abnormal judgment threshold rather than a fixed threshold, and the filling balance index supports quantitative sorting and threshold judgment, the generated filling imbalance detection results can directly give the abnormal cavity number, abnormality degree and abnormal feature type (such as filling lag, filling too fast), providing intuitive and interpretable judgment basis for on-site process personnel; the entire detection process is automatically completed based on the pressure timing signal of each cavity in the same injection cycle, without the need for manual presetting of alignment windows or manual interpretation of pressure curves, and has good automated execution capability and production line adaptability; the output abnormal cavity number, abnormality degree and abnormal feature type can be directly used for quality consistency recording and traceability of multi-cavity molding process. Attached Figure Description
[0012] Figure 1 This is the main flowchart of an online detection method for injection filling imbalance based on cavity pressure timing analysis.
[0013] Figure 2 This is a sub-flowchart of an online detection method for injection filling imbalance based on cavity pressure timing analysis.
[0014] Figure 3 This is another sub-flowchart of the online detection method for injection filling imbalance based on cavity pressure time sequence analysis. Detailed Implementation
[0015] 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.
[0016] 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.
[0017] like Figures 1 to 3 As shown, this invention provides an online detection method for injection molding filling imbalance based on cavity pressure time-series analysis, specifically including: S1: Acquire the pressure timing signals and injection start trigger signals for each mold cavity. Based on the injection start trigger signals, extract the pressure data for each mold cavity during the filling stage to generate a set of pressure curves for the filling stage. The time reference in this step is precisely locked by the trigger level transition edge, eliminating trigger delay phase errors. At the same time, it removes data contamination from the zero-pressure and holding pressure stages before injection, and the analysis only targets the filling stages that truly carry melt flow information. Furthermore, it unifies the timestamp and mold cavity identifier, forming a standardized input for multi-cavity parallel analysis.
[0018] S2: The set of pressure curves for the filling stage is subjected to baseline correction, filtering and noise reduction, and outlier identification to determine the arrival time of the melt front and the end time of filling, generating a clean filling pressure curve sequence. This step eliminates sensor zero drift; at the same time, it filters out high-frequency impact noise while retaining the steep characteristics of the pressure rise edge, eliminates non-physical spikes caused by electromagnetic interference / transient faults and maintains curve continuity; and adaptively positions the melt front using a local dynamic baseline to avoid false triggering by zero drift and low-frequency fluctuations; and accurately defines the start and end boundaries of the filling interval.
[0019] S3: Map the pressure-time axis of each cavity from the arrival time of the melt front to the end time of filling to the filling phase coordinate and perform phase alignment to generate a standardized set of filling pressure curves. This step quantifies the difference in the filling start time of each cavity; at the same time, it eliminates the inherent time lag and duration difference caused by the inconsistency of physical flow channel length; and it makes the pressure curves of each cavity fall on the same dimensionless phase coordinate, which can be directly compared laterally.
[0020] S4: Extract the arrival time difference, pressure rise slope, pressure integral value, terminal pressure value, and curve shape deviation vector from the standardized filling pressure curve set. Generate a common operating condition feature vector using the median value of each cavity feature in the same period, forming a multi-dimensional time-series feature vector set. This step comprehensively characterizes the filling behavior using five complementary dimensions (time lag, flow rate, energy accumulation, terminal state, and topology), avoiding insufficient characterization and early missed detection using only peak pressure / filling time difference. At the same time, the median value in the same period is used as a reference to eliminate interference from extreme outliers and establish a horizontal comparison benchmark. Furthermore, it separates the overall operating condition drift from individual cavity anomalies.
[0021] S5: Establish a benchmark library of permissible fluctuation ranges for multidimensional features based on historical normal cycles. Compare the multidimensional time-series feature vector set with the common operating condition feature vector and the benchmark library of permissible fluctuation ranges for multidimensional features to obtain lateral deviation and longitudinal deviation, and fuse them to generate a fill balance index. This step uses a dual reference of lateral (spatial) and longitudinal (temporal) deviations, taking into account both inter-cavity comparison and cross-cycle stability, avoiding excessive reliance on historical standard curves and frequent false alarms. At the same time, the standardized deviation eliminates dimensional differences, realizing the quantitative classification of anomaly severity. Furthermore, multidimensional fusion highlights the dominant anomaly source, distinguishing between transient disturbances and structural imbalances.
[0022] S6: Compare the filling balance index with the adaptive anomaly judgment threshold. When the filling balance index exceeds the adaptive anomaly judgment threshold, determine the abnormal cavity with filling lag or filling too fast and the degree of abnormality, and generate a filling imbalance detection result containing the abnormal cavity number, degree of abnormality, and type of abnormality feature. This step replaces the fixed threshold with a threshold that automatically adjusts with the operating conditions, resulting in high robustness and low false alarms under varying operating conditions. Simultaneously, it moves from "whether it is abnormal" to "which cavity, how severe, and what type of feature," achieving rapid root cause identification and classification. Furthermore, it outputs a structured data frame that can be directly interfaced with the alarm / process adjustment module.
[0023] The beneficial effects of steps S1 to S6 are as follows: S1–S6 are interconnected, forming a complete detection chain from "data acquisition and purification during the filling stage to phase alignment and comparability, multi-dimensional feature extraction and common benchmark establishment, lateral / longitudinal deviation fusion, adaptive threshold judgment and structured output". This solves the three problems in the background technology: insufficient acquisition and purification of pressure data during the filling stage, inability to directly compare pressure curves of different mold cavities laterally, and lack of common working condition benchmarks within the same injection cycle. The entire detection process is automatically completed based on the pressure timing signals of each mold cavity within the same injection cycle, without the need for manual pre-setting of alignment windows or manual interpretation of pressure curves. It has good automated execution capabilities and production line adaptability. The output abnormal mold cavity number, abnormality degree, and abnormal feature type can be directly used for quality consistency recording and traceability in a multi-cavity molding process.
[0024] The online detection method for injection filling imbalance based on cavity pressure time-series analysis further includes: S7: Based on the filling imbalance detection results, output an alarm signal or process adjustment suggestion command. When the current cycle is confirmed to be in a normal state based on the filling balance index trend of multiple consecutive cycles, use exponential weighted moving average processing to update the multidimensional feature allowable fluctuation range benchmark library with the multidimensional time series feature vector set of the current cycle, and update the adaptive anomaly judgment threshold for subsequent cycles accordingly. This step connects the detection from "data judgment" to "control execution," providing immediate anomaly warnings and obtaining process intervention commands; at the same time, the benchmark library is only updated in normal cycles where non-real defects are confirmed, avoiding writing real defects into the benchmark; and the benchmark library and judgment threshold adaptively track and correct for long-term slow process drift in real time.
[0025] Step S7 upgrades the method from "one-time detection" to a closed loop of "detection-response-self-learning-threshold refitting". The judgment benchmark is continuously updated with the long-term drift of the production process, thereby improving the adaptability, long-term stability and judgment accuracy of the detection system to changes in operating conditions while maintaining a low false alarm rate.
[0026] Step S1: Acquire the original cavity pressure timing signals collected by the pressure sensors of each cavity in a multi-cavity injection mold and the injection start trigger signal provided by the injection molding machine. Based on the injection start trigger signal, extract the pressure data of each cavity during the filling stage to generate a set of filling stage pressure curves containing melt flow information. Specifically, this includes: S1.1: The analog voltage signals output by the cavity pressure sensors deployed in each cavity of a multi-cavity injection mold are subjected to high-frequency synchronous sampling processing to convert the continuous physical quantity into a discrete digital sequence, generating a set of original cavity pressure timing signals with a unified timestamp.
[0027] Piezoelectric or strain gauge pressure sensors deployed within each cavity of a multi-cavity injection mold output continuous analog voltage signals in real time, reflecting changes in melt pressure within the cavity. The data acquisition module performs analog-to-digital conversion on these analog voltage signals using a high-frequency synchronous sampling frequency of 1kHz to 20kHz, discretizing the continuous physical quantity into a digital sequence with fixed time intervals. The specific frequency is determined based on the sensor bandwidth, the shortest event duration during the filling stage, and the channel capacity of the data acquisition module, ensuring that the shortest event to be identified covers at least five sampling points. Hardware triggering or software synchronization mechanisms are employed to ensure that the sampling clock phase of all cavity channels is consistent, eliminating time errors between multiple channels. A unified timestamp based on the system's global clock is added to each sampling point, establishing a mapping relationship between data points and absolute time. A preliminary dimensional transformation is performed on the discretized digital sequence, converting the voltage values into standard pressure unit values based on the sensor sensitivity coefficient.
[0028] S1.2: Based on the level transition edge of the injection start trigger signal issued by the injection machine controller, perform a time base locking operation to determine the absolute start time of the current injection cycle and generate an injection cycle start time anchor point for subsequent data interception.
[0029] Receives the injection start trigger signal output by the injection molding machine controller. This signal is either a digital level pulse or an analog voltage transition signal.
[0030] High-frequency discrete sampling is performed on the trigger signal to obtain a discrete sequence of signal amplitude changes over time.
[0031] Calculate the absolute value of the voltage difference between adjacent sampling points to construct a voltage change rate sequence.
[0032] The voltage change rate sequence is compared with a preset jump sensitivity threshold to identify a set of candidate points for voltage change moments.
[0033] Perform de-jitter filtering on the candidate point set to remove spurious transition interference with a duration shorter than the minimum effective pulse width.
[0034] Select the first stable edge moment that meets the requirements of positive or negative polarity jump, and lock it as the physical instant when the injection action occurs.
[0035] Mark the timestamp corresponding to this physical instant as the absolute start time of the current injection molding cycle.
[0036] An injection cycle start time anchor is generated based on the absolute start time for subsequent data extraction.
[0037] S1.3: Using the injection cycle start time anchor point and the preset standard filling duration window, candidate filling intervals are extracted from the original mold cavity pressure timing signal set to remove the zero pressure segment and holding pressure stage data before injection, and generate an initial filling pressure data segment sequence that only contains the melt filling process.
[0038] Based on the injection cycle start time anchor point, a preset standard filling duration window parameter is read. This parameter is set during the mold trial stage based on statistical values of the filling end times of no less than thirty confirmed normal cycles: the 99th percentile of the filling duration is taken, plus a time margin of one to five sampling cycles as the window length; this window length must not exceed the time between the start of injection and the pressure holding switch. After changing the mold, material, or main injection parameters, the window parameter is recalculated and updated.
[0039] The left boundary of the candidate filling interval is determined by the injection cycle start time anchor point, and the right boundary is determined by the sum of the injection cycle start time anchor point and the standard filling duration window, thus constructing an initial time cutoff interval that covers the complete filling phase.
[0040] Perform time index mapping on the original cavity pressure time series signal set to filter out all discrete pressure sampling points whose timestamps are located within the initial time intercept interval.
[0041] Zero-pressure segment data with timestamps earlier than the injection cycle start time anchor point are removed to eliminate interference from pre-injection sensor static baseline and environmental noise on filling feature analysis.
[0042] Data points with timestamps later than the right boundary of the standard filling duration window are removed to eliminate the influence of pressure maintenance during the holding phase and pressure decay during the initial cooling phase on the extraction of filling dynamics features.
[0043] The selected discrete pressure sampling points are reassembled in chronological order to generate an initial filling pressure data segment sequence that contains only the melt filling process.
[0044] S1.4: Perform cavity number mapping and association processing on each data point in the initial filling pressure data segment sequence to establish a one-to-one correspondence between pressure values and specific physical cavity locations, and generate a structured filling pressure data stream carrying cavity identification.
[0045] The system receives the initial filling pressure data segment sequence generated in S1.3. This sequence contains pressure values sorted by time but lacks physical cavity attribution information. It parses the preset mapping table between data acquisition hardware channels and mold physical cavities, establishing a logical index dictionary from sensor channel IDs to cavity numbers. It iterates through each discrete sampling point in the initial filling pressure data segment sequence, reading its associated original sensor channel identifier. Based on the logical index dictionary, it converts the channel identifier into a unique cavity identification code, used to distinguish different independent cavities in a multi-cavity mold system. The converted cavity identification code is embedded as a metadata field into the corresponding pressure sampling point data structure, forming a triplet data unit containing a timestamp, pressure value, and cavity ID. After performing the above mapping operation on all sampling points, the data units are grouped and aggregated according to the cavity ID, ensuring that the data for the same cavity is logically continuous and independent.
[0046] S1.5: Based on the structured filling pressure data stream carrying the mold cavity identification, perform data encapsulation and formatting reorganization operations to form a standard data structure that meets the input requirements of the subsequent analysis module, and finally generate a set of filling stage pressure curves containing complete melt flow information.
[0047] Receive a structured filling pressure data stream carrying the mold cavity identification, parse the timestamp sequence and pressure value pairs in the data stream, and establish an independent data index table for each mold cavity.
[0048] Based on the preset sampling frequency parameters, the theoretical time interval between adjacent sampling points is calculated, the continuity and uniformity of the actual timestamp sequence are verified, and data segments with packet loss or time sequence disorder are marked.
[0049] For the marked abnormal data segments, a linear interpolation method is used to fill in the missing time point pressure values, ensuring that the pressure curve of each mold cavity is discretely distributed at equal intervals on the time axis.
[0050] The corrected pressure data of each mold cavity are reorganized into a matrix according to the mold cavity number to construct a two-dimensional pressure data matrix with rows corresponding to sampling times and columns corresponding to mold cavity numbers.
[0051] Metadata information for the current injection cycle, including mold number, material batch code, and absolute time anchor point of the injection start trigger signal, is injected into the matrix header to form a standardized data structure object.
[0052] Step S2: The set of pressure curves for the filling stage is subjected to baseline correction, filtering and noise reduction, and outlier identification to eliminate sensor zero drift and high-frequency impact noise interference. The arrival time of the melt front and the end time of filling for each mold cavity are determined, generating a pre-processed clean filling pressure curve sequence. Specifically, this includes: S2.1: Perform sliding window mean baseline estimation processing on the original pressure data segments of each cavity in the pressure curve set of the filling stage, calculate the static pressure offset before injection as the zero drift compensation value, and subtract the zero drift compensation value from the original pressure data segments point by point to generate a baseline corrected pressure data sequence that eliminates the influence of sensor zero drift.
[0053] Receive an initial filling pressure data segment sequence carrying the mold cavity identification, which includes static background noise before injection and sensor zero drift components.
[0054] For each mold cavity pressure data segment, a time window is used for positioning. The baseline window length is set to 20 to 500 sampling points, and the standard deviation of pressure within the window must not exceed one percent of the corresponding sensor's full scale. If this condition is not met, the window is extended forward until a stability condition is met, with the upper limit of the extension being the length of the injection cycle idle period. The static period data that meets the conditions is used as the baseline estimation interval.
[0055] Within the baseline estimation interval, the local pressure average is calculated using the sliding window averaging method to eliminate transient electrical noise interference.
[0056] The static pressure offset of the mold cavity in the current cycle is obtained by arithmetically averaging the local mean values of each sliding window.
[0057] Calculate the zero drift compensation value using the following formula: in, This is the zero-drift compensation value. To estimate the total number of sampling points within the baseline interval, For the first The original pressure values at each sampling point.
[0058] The calculated zero-drift compensation value is deducted point by point from each sampling point of the initial filling pressure data segment.
[0059] Boundary checks are performed on the data after deduction, and negative pressure values less than zero are forcibly set to zero to conform to the non-negative characteristics of physical pressure.
[0060] S2.2: Based on the baseline-corrected pressure data sequence, wavelet threshold denoising filtering is performed. The signal frequency band is decomposed using a preset wavelet basis function and an adaptive threshold is set to separate the high-frequency impulse noise component. The retained low-frequency effective signal component is reconstructed into a smooth signal to generate a smooth filled pressure curve sequence that filters out high-frequency interference.
[0061] The baseline-corrected pressure data sequence is received as input. The db4 wavelet basis function from the Daubechies wavelet family is selected to perform a multi-scale discrete wavelet transform on the signal, decomposing the time-domain pressure signal into approximation coefficients and detail coefficients of different frequency sub-bands. Based on the physical characteristics of the injection molding filling process, the decomposition layer is set to 5 layers to separate high-frequency mechanical vibration noise from low-frequency melt flow trends. The median absolute deviation of the detail coefficients of each layer is calculated, and an adaptive hard threshold parameter is derived based on a general threshold rule. The threshold setting range of the wavelet threshold is 1-100. When the threshold calculated by the general threshold rule exceeds this range, the boundary value of that range is used. The following formula is used to calculate the... Threshold for layer detail factor: in, For the first Layer detail coefficient vector, The number of signal sampling points is denoted as . Detail coefficients below a threshold are set to zero to suppress impulse noise, while coefficients above the threshold are retained to maintain the steep characteristics of the pressure rise time. An inverse discrete wavelet transform is performed on the processed approximation coefficients and the corrected detail coefficients to reconstruct the time-domain pressure signal.
[0062] S2.3: Taking the smooth filling pressure curve sequence as input, first calculate the first-order difference slope of adjacent sampling points, then calculate the ratio of the difference between adjacent slopes to the absolute value of the previous slope as the slope mutation rate; compare the slope mutation rate with the physical limit threshold, mark abnormal sampling points after neighborhood continuity verification, and use the effective data before and after the abnormal points for linear interpolation repair to output a continuous clean filling pressure curve sequence.
[0063] Receive a smoothed filling pressure curve sequence after wavelet threshold denoising. This sequence has eliminated high-frequency impact noise but may retain non-physical spikes caused by electromagnetic interference or sensor transient failures.
[0064] Perform point-by-point first-order difference operations on discrete data points in the smoothed pressure curve sequence to calculate the rate of pressure change between adjacent sampling times, and construct a slope time series that reflects the intensity of dynamic pressure changes.
[0065] Based on the rheological properties of injection-molded melt and the response limits of sensors, a physically achievable maximum pressure rise slope threshold and a maximum pressure fall slope threshold are set, collectively referred to as the physical limit thresholds. These serve as the boundary for identifying abnormal abrupt change points. The physical limit thresholds are determined based on two constraints: the first is the 99.9th percentile of the slope change rate over at least thirty normal cycles, and the second is the upper limit calculated from the sensor's maximum response rate and sampling interval. The smaller of the two values is taken as the initial threshold, and it is saved after verification with known normal cycles and samples containing spikes.
[0066] The calculated rate of pressure change at each moment is compared point by point with the physical limit threshold, and the sampling points where the absolute value of the rate of change exceeds the threshold are marked as suspected abnormal peaks or drop points.
[0067] The marked outliers are checked for neighborhood continuity. Points with temporary high slopes caused by normal rapid filling are removed, and isolated extreme points that violate physical laws are retained as outliers that need to be repaired.
[0068] A linear interpolation method is used to reconstruct the pressure value at the outlier by using the pressure values and timestamps of the valid normal data points adjacent to the outlier, thus filling the data gaps and maintaining the continuity of the curve.
[0069] S2.4: Based on the clean filling pressure curve sequence, perform melt front arrival time detection processing based on dynamic pressure gradient. By scanning the rising edge of the pressure curve and identifying the moment when the pressure change rate first continuously exceeds the preset dynamic trigger threshold, mark the moment when the melt front arrives and generate a semi-structured pressure curve sequence with melt front arrival time marking.
[0070] Read the discrete pressure sampling point data of each mold cavity in the clean filling pressure curve sequence within a preset time window; the preset time window is set in such a way that its length is 1 / 10 to 1 / 5 of the total number of data points in the filling stage, and its starting point is located before the arrival time of the melt front, so as to completely cover the pressure rise edge.
[0071] A first-order difference operation is performed on adjacent sampling points to calculate the instantaneous pressure change rate, which characterizes the dynamic propulsion speed of the melt flow front.
[0072] Using a sliding window statistical method, the mean and standard deviation of the pressure change rate of the previous N sampling points at the current time are calculated to construct a local dynamic baseline.
[0073] Set the dynamic trigger threshold to the local baseline mean plus... The standard deviation is multiplied to adaptively filter background noise and capture significant pressure surge signals.
[0074] in, for Dynamic trigger threshold at any time is the mean rate of pressure change within the sliding window, and is the standard deviation. This is the sensitivity coefficient.
[0075] Scan the pressure change rate sequence along the time axis in the forward direction to identify the moment when the first M consecutive sampling points all exceed the dynamic trigger threshold.
[0076] Mark this moment as the arrival time of the melt front, and confirm it as the physical starting reference for the filling stage.
[0077] For example, the sliding window length is set to 10 sampling points, the sensitivity coefficient is 3.0, and the number of consecutive confirmation points is 5. For a certain mold cavity pressure data, the average pressure change rate of the 50th to 59th sampling points is calculated to be 0.5 MPa / ms, and the standard deviation is 0.1 MPa / ms, so the dynamic threshold is 0.8 MPa / ms. When the pressure change rates of the 60th to 64th sampling points are 0.9, 1.2, 1.5, 1.1, and 0.95 MPa / ms respectively, and all are greater than 0.8 MPa / ms, the time corresponding to the 60th sampling point is determined to be the arrival time of the melt front. This method effectively distinguishes the pressure fluctuations in the initial stage of injection from the actual melt filling signal, significantly improving the accuracy of the time reference for feature extraction.
[0078] S2.5: Perform filling end time determination processing based on pressure holding switching logic on the semi-structured pressure curve sequence with melt front arrival time marker. Combine the pressure holding switching signal provided by the injection molding machine or the inflection point feature of the pressure curve slope changing from positive to negative, determine the termination time point of melt filling completion, extract the effective data segment between the start and end times and encapsulate the timestamp information, and finally generate the pre-processed clean filling pressure curve sequence.
[0079] Read the semi-structured pressure curve sequence marked with the arrival time of the melt front, and extract the holding pressure switching signal level status output by the injection molding machine controller and the pressure sampling data of the current cycle. [Terminology needs to be standardized] A first-order difference operation is performed on the semi-structured pressure curve sequence marked with the arrival time of the melt front to calculate the pressure change rate between adjacent sampling points and generate a slope time series characterizing the dynamic trend of pressure.
[0080] A sliding window detection method is used to scan the slope time series and identify the inflection point region where the slope changes from a continuous positive value to a negative value or zero value, which is then marked as a potential candidate set of pressure extreme points.
[0081] By combining the rising edge timestamp of the pressure holding switching signal, it is determined whether the signal is located in the time neighborhood of the potential pressure extreme point candidate set. If it exists, the time of the signal is locked as the filling end time.
[0082] If the pressure holding switching signal is missing or abnormal, the moment when the slope crosses zero and the pressure value reaches a local maximum value is selected from the candidate set as the filling end time based on the curve shape judgment.
[0083] By using the determined filling end time and the aforementioned melt front arrival time, a closed time interval is constructed, and the clean pressure data segment within this interval is extracted.
[0084] The extracted data segments are labeled with start and end timestamps and encapsulated into structured data units containing complete information about the filling process.
[0085] Step S3: Map the pressure-time axis of each mold cavity from the arrival time of the melt front to the end time of filling to the filling phase coordinates and perform phase alignment to generate a standardized set of filling pressure curves. Specifically, this includes: S3.1: Read and analyze the arrival time of the melt front of each cavity in the clean filling pressure curve sequence to extract the absolute delay time value of each cavity relative to the anchor point of the injection cycle start time, and generate a set of cavity time offsets characterizing the flow channel transport lag.
[0086] The S3.1 sub-step aims to extract the absolute time information of the arrival of the melt front in each cavity from the pre-processed clean filling pressure curve sequence, providing a time reference for subsequent phase alignment.
[0087] Read the data structure corresponding to each mold cavity in the clean filling pressure curve sequence, and parse the melt front arrival time field encapsulated therein. This time represents the moment when the rising edge of the pressure signal first continuously exceeds the dynamic trigger threshold.
[0088] Obtain the injection cycle start time anchor point of the current injection cycle. This anchor point is generated by locking the level transition edge of the injection start trigger signal and serves as a unified time zero point reference for the entire mold.
[0089] Regarding the first For each mold cavity, extract the arrival time of its melt front, perform time difference calculation, and calculate the absolute delay time of the mold cavity relative to the injection cycle start time anchor point.
[0090] The following formula is used to calculate the first... Absolute delay time of each mold cavity: in, For the first The absolute delay time of each mold cavity, For the first The arrival time of the melt front in each mold cavity This is the anchor point for the start time of the injection cycle.
[0091] Iterate through all N cavities in a multi-cavity mold, repeatedly performing the time difference calculation described above, to generate a set of N cavity time offsets. }={ , ,..., }
[0092] The validity of the set of time offset values of the mold cavity is verified, and null values or abnormally large values caused by sensor failure or signal loss are removed to ensure that the data in the set reflects the true flow channel transmission hysteresis characteristics.
[0093] S3.2: Based on the minimum value in the set of mold cavity time offsets as the global zero phase reference point, perform linear translation transformation on the original time axis of each mold cavity to eliminate the absolute time difference caused by the inconsistency of physical flow channel length and generate a relative time coordinate sequence with a unified flow starting point.
[0094] Read the arrival time data of the melt front edge of each cavity from the cavity time offset set, and use it as the time reference input for phase alignment.
[0095] Iterate through the time offset values of all mold cavities, perform the minimum value search method, and determine the mold cavity in the current injection cycle where the melt front arrival event occurs first and its corresponding timestamp.
[0096] The minimum time offset is set as the global zero-phase reference point, which represents the filling start time under ideal conditions, eliminating the absolute time lag caused by the difference in the physical length of the flow channel.
[0097] For each mold cavity, a linear translation transformation is performed at the original pressure sampling time point to subtract the value of the global zero-phase reference point from the original time coordinate.
[0098] The relative time coordinates of each mold cavity in a unified coordinate system are calculated using the following formula: in, This is the original sampling timestamp. For the global zero-phase reference point timestamp, The coordinates are the relative time coordinates after translation.
[0099] The monotonicity of the translated time series is checked to ensure that the filling start time of all cavities is aligned to the zero point or positive value range in the relative coordinate system.
[0100] The processed relative time coordinates are re-correlated with the pressure values of each mold cavity to construct time-pressure data pairs with a unified flow starting point.
[0101] S3.3: Perform normalization scaling operation on the relative time coordinate sequence with a unified flow starting point, and use the difference between the filling end time of each mold cavity and the arrival time of the melt front as the denominator to map the time axis to the dimensionless filling phase interval from zero to one, and generate a standardized filling phase coordinate axis.
[0102] Receives a relative time coordinate sequence with a unified flow starting point and data on the arrival time of the melt front and the end time of filling for each mold cavity.
[0103] Read the number of injection cycles in the current injection cycle The effective filling duration of a cavity is calculated by taking the arrival time of the melt front and the end time of filling for each cavity, and used as the denominator for normalization scaling.
[0104] Construct a linear mapping function from relative time to dimensionless filling phase, establish the correspondence between the time axis and the phase axis, and ensure that different cavities have a unified phase identifier under the same physical filling progress.
[0105] The normalized scaling operation is performed using the following formula: in, To standardize the fill phase, the value ranges from 0 to 1; The relative timestamp of the current sampling point; This refers to the arrival time of the melt front; This is to fill in the end time.
[0106] By iterating through each discrete sampling point in the clean filling pressure curve sequence, and substituting its original relative time value into the above mapping function, the corresponding dimensionless phase value is calculated.
[0107] The calculated phase value is bound to the original pressure value to form a new data pair with the filling phase as the independent variable and the pressure value as the dependent variable, thus eliminating the time scale inconsistency caused by the difference in flow channel length.
[0108] Generate a filling phase coordinate axis covering the interval from 0 to 1. This coordinate axis represents the relative filling process of the melt in the mold cavity, rather than the absolute elapsed time.
[0109] S3.4: Based on the filling phase coordinate axis, perform resampling interpolation processing on the discrete pressure sampling points in the clean filling pressure curve sequence, and use the cubic spline interpolation method to reconstruct the non-uniformly distributed pressure data into a pressure value sequence with equal phase intervals, generating a discrete pressure data point set after phase alignment.
[0110] The system receives normalized filling phase coordinate axes and discrete sampling point data from the clean filling pressure curve sequence of each cavity. Addressing the uneven distribution and phase misalignment of the original sampling points in the time domain, a resampling mapping model based on cubic spline interpolation is constructed. The standardized filling phase interval [0,1] is divided into N equally spaced phase nodes to determine the target phase vector for resampling. For any cavity, two adjacent known sampling points in its clean pressure curve are selected as interpolation basis nodes. A cubic polynomial function is used to fit the pressure change trajectory between the two basis nodes, ensuring continuous and smooth pressure values, first derivatives, and second derivatives at the nodes. The coefficient matrix of each piecewise cubic polynomial is determined by solving a system of linear equations, establishing an accurate mapping relationship between phase and pressure values. Based on the target phase vector, the interpolated pressure value corresponding to each equally phase node is calculated one by one. The calculated pressure values are physically verified to eliminate non-monotonic abnormal fluctuations caused by boundary effects. The interpolated pressure values of all cavities at the same phase node are indexed and arranged according to the cavity number. By using cubic spline interpolation resampling, the non-uniform discrete pressure data from the previous step is transformed into a set of standard discrete pressure data points with equal phase intervals, thereby achieving strict alignment of the multi-cavity pressure curves in the morphological dimension and eliminating morphological comparison errors caused by differences in sampling rates.
[0111] S3.5: Perform data structure reorganization and cavity identification association processing on the phase-aligned discrete pressure data point set, encapsulate the pressure values of each cavity under the same filling phase into a matrix form, and finally generate a standardized filling pressure curve group after phase alignment for subsequent feature extraction module to call.
[0112] The system receives the phase-aligned discrete pressure data point set and the identification identifiers of each mold cavity, and constructs a two-dimensional matrix structure to store the standardized filling pressure data. The row index of this matrix corresponds to the standardized filling phase sampling point, and the column index corresponds to the specific mold cavity number.
[0113] Traverse all mold cavities to be tested within the current injection cycle. For each cavity, extract the equal-phase interval pressure value sequence generated in step S3.4. Fill the extracted pressure value sequence into the corresponding column vector positions in the matrix according to the phase order, ensuring that the data in the same row represents the instantaneous pressure value of different mold cavities under the same filling phase.
[0114] Perform an integrity check on the filled two-dimensional pressure matrix to check for missing columns or invalid data rows due to sensor malfunction. If data is missing, mark the cavity as having a sensor malfunction and remove the corresponding column to prevent abnormal data from interfering with subsequent consistency evaluation.
[0115] The validated two-dimensional pressure matrix is associated with a physical location mapping table of the mold cavities to generate a standardized data structure of filling pressure curves containing spatial topology information. This data structure not only contains pressure values but also preserves the relative positional relationships between the mold cavities, providing a foundation for subsequent anomaly tracing based on spatial distribution.
[0116] Step S4: Extract the arrival time difference, pressure rise slope, pressure integral value, end pressure value, and curve shape deviation vector from the standardized filling pressure curve set. Generate a common working condition feature vector using the median value of each cavity feature in the same period, forming a multi-dimensional time-series feature vector set. Specifically, this includes: S4.1: Perform extreme value search processing based on dynamic time window on the discrete pressure data point set of each cavity in the standardized filling pressure curve group after phase alignment, so as to extract the arrival time difference value of each cavity relative to the global zero phase reference point and the end pressure value at the end of filling, and generate a set of key point features of the cavity containing time lag information and terminal status information.
[0117] Step S4.1 aims to extract key time points and terminal state features from the standardized filling pressure curve set after phase alignment. It reads the discrete pressure data point set after phase alignment, which contains the pressure value sequence of each cavity under a unified filling phase coordinate. A dynamic time window scan is performed on the pressure sequence of each cavity, setting the initial search window to cover the phase coordinate interval from 0 to 0.05 to locate the initial pressure response point after the melt front arrives. The absolute value of the first-order pressure difference within the sliding window is calculated, and the phase point where the pressure change rate first exceeds a preset noise threshold is identified and marked as the valid filling start phase. The offset of this start phase relative to the global zero-phase reference point is recorded as the arrival time difference value of the cavity, characterizing the channel transmission delay difference. Simultaneously, terminal state extraction is performed, setting the termination search window to cover the phase coordinate interval from 0.95 to 1.0, and retrieving the phase point corresponding to the maximum pressure value within this window. This maximum pressure value is extracted as the terminal pressure value at the end of filling, reflecting the resistance state of the cavity when filling is complete. The extracted time difference of arrival (TLOA) and terminal pressure values are structurally encapsulated according to the cavity number, generating a set of key feature points for each cavity that includes time lag information and terminal status information. Through the aforementioned dynamic window extreme value search process, the original pressure sequence is transformed into key indicators that quantify time lag and terminal pressure, enabling precise capture of the start and end states of filling in each cavity and providing boundary constraints for subsequent dynamic feature calculations.
[0118] S4.2: Based on the effective filling interval determined by the set of key features of the mold cavity, perform piecewise linear regression fitting on the standardized filling pressure curve set after phase alignment to calculate the pressure rise slope of each mold cavity from the melt front to the end of filling interval, and perform trapezoidal integral operation on the pressure values in the same interval to obtain the pressure integral value, thereby generating a set of mold cavity dynamic features characterizing the flow rate and energy accumulation.
[0119] Read the discrete pressure data point set of each cavity in the standardized filling pressure curve group after phase alignment, and extract the melt front arrival phase point and filling end phase point corresponding to each cavity according to the cavity key point feature set generated by S4.1, and define the effective data subsequence containing only the melt filling process.
[0120] For the effective data subsequence, piecewise linear regression fitting is performed using the least squares method to construct a linear model of pressure change with filling phase. The slope parameter of the best fitted line is determined by solving the problem of minimizing the sum of squared residuals. This slope directly characterizes the average pressure rise rate of the melt in the mold cavity.
[0121] The pressure rise slope k_i of the i-th cavity is calculated using the following formula: in, For the first The mold cavity in the first Fill phase of each sampling point For the corresponding pressure value, This represents the total number of valid data points for the mold cavity. and These are the average values of phase and pressure, respectively.
[0122] Based on the effective interval boundary determined by the fitting, a trapezoidal numerical integration operation is performed on the original discrete pressure data. The trapezoidal areas formed between adjacent sampling points are accumulated, and the total area under the pressure-phase curve is calculated. This area physically corresponds to the amount of work or energy accumulated by the melt overcoming flow resistance during the filling stage.
[0123] The following formula is used to calculate the first... Pressure integral value of each mold cavity: in, and The pressure values are those of adjacent sampling points. and This corresponds to the fill phase value.
[0124] The calculated pressure rise slope and pressure integral value of each cavity are structured and encapsulated, and associated with the corresponding cavity identifier to form a set of cavity dynamic features that include flow rate characteristics and energy accumulation characteristics.
[0125] S4.3: Using the pressure integral value in the set of cavity dynamic features as a weighting coefficient, perform normalization morphology vector construction processing on the standardized filling pressure curve group after phase alignment, and map the discrete pressure values of each cavity on the standard filling phase coordinate axis into curve morphology deviation vectors in high-dimensional space to generate a set of cavity morphology features that characterize the pressure distribution topology.
[0126] Read the standardized filling pressure curve set after phase alignment, and extract the discrete pressure value sequence of each cavity on the standard filling phase coordinate axis.
[0127] Obtain the pressure integral value from the set of cavity dynamic characteristics generated in step S4.2, and use it as a weighting coefficient to characterize the degree of energy accumulation in melt filling.
[0128] Normalization is performed on the discrete pressure value sequence of each mold cavity to eliminate the difference in absolute pressure amplitude caused by material batch or temperature fluctuations, and a dimensionless pressure distribution pattern sequence is generated.
[0129] At each standard filling phase node, the difference between the current normalized pressure value of the mold cavity and the median pressure value of the corresponding mold cavities in the same period is calculated, and multiplied by a weight obtained by normalizing the integral value of the mold cavity pressure. All difference components are arranged in the order of the phase nodes to form a curve shape deviation vector. The Euclidean norm of this vector is defined as the curve shape deviation norm, which is used as a derived scalar for subsequent severity calculations, but does not replace the original vector.
[0130] The following formula is used to calculate the first... The mold cavity in the first Morphological deviation components at each phase point: in, For the first Pressure integral weighting coefficient for each mold cavity For the first The mold cavity in the first Normalized pressure values at each phase point For the same period, the median cavity is in the first... Normalized pressure values at each phase point.
[0131] The morphological deviation components of each phase point are arranged in phase order to form a curve morphological deviation vector, and the Euclidean norm of the vector is calculated as the curve morphological deviation norm. The curve morphological deviation vector and the curve morphological deviation norm are stored separately, and the curve morphological deviation vector is used for subsequent field fusion.
[0132] The curve shape deviation vector and its curve shape deviation norm of each mold cavity are associated and encapsulated into a mold cavity shape feature set according to the mold cavity number; among them, the curve shape deviation vector is used as a component field of the multidimensional single mold cavity feature vector, and the curve shape deviation norm is only used for weight calculation and result interpretation.
[0133] S4.4: Perform field fusion processing by mold cavity number on the set of key feature points of the mold cavity, the set of dynamic feature features of the mold cavity, and the set of morphological feature features of the mold cavity. Encapsulate the arrival time difference value, pressure rise slope, pressure integral value, end pressure value, and curve shape deviation vector into a single data structure to generate a multi-dimensional single mold cavity feature vector characterizing the complete filling behavior of a single mold cavity.
[0134] Receive the set of key feature points of the mold cavity, the set of dynamic feature features of the mold cavity, and the set of morphological feature features of the mold cavity, and parse the independent feature data block corresponding to each mold cavity number.
[0135] The arrival time difference and terminal pressure values are extracted from the feature set of key points in the mold cavity and mapped to the time dimension component and terminal state component of a multidimensional feature vector.
[0136] Read the pressure rise slope and pressure integral value from the set of dynamic characteristics of the mold cavity, calculate their log-normalized values to eliminate dimensional differences, and map them to velocity dimension components and energy dimension components.
[0137] Obtain the curve shape deviation vector from the set of mold cavity shape features, directly use the curve shape deviation vector as the shape topology dimension component, and retain the corresponding curve shape deviation norm as a derived scalar.
[0138] Construct a unified structured data container and arrange the above components in the order of time, rate, energy, terminal state, and morphological topology.
[0139] Each component is combined into a column vector of fixed length and assigned a unique cavity identification index to generate a multidimensional single cavity feature vector representing the complete filling behavior of a single cavity.
[0140] S4.5: Perform a dimension-wise median statistical operation on the multidimensional single-cavity feature vectors generated by all cavities in the current injection cycle to eliminate extreme outlier interference and extract common working condition feature vectors representing the overall mold state. At the same time, associate and encapsulate each multidimensional single-cavity feature vector with the common working condition feature vector to finally generate a multidimensional time-series feature vector set representing the single-cycle filling characteristics.
[0141] Receive the multi-dimensional single-cavity feature vectors of all cavities in the current injection cycle, and construct a feature matrix containing the arrival time difference of each cavity, the pressure rise slope, the pressure integral value, the end pressure value, and the curve shape deviation vector.
[0142] Independent sorting is performed on the data in each dimension of the feature matrix, and the values of each cavity in the same feature dimension are arranged in ascending order to form an ordered sequence.
[0143] Median extraction is performed based on an ordered sequence. If the number of cavities is odd, the value at the middle position is taken; if it is even, the arithmetic mean of the two middle values is taken. The common working condition feature value of this dimension is then calculated.
[0144] Traverse all feature dimensions, repeat the above sorting and value taking operations, and combine to generate the median feature vector representing the overall mold filling state of the current cycle. This vector is the common working condition feature vector, which is composed of the median value of each feature dimension, and serves as a benchmark reference system to eliminate the interference of extreme outliers.
[0145] Calculate the Euclidean distance or absolute difference between the multidimensional single-cavity feature vector and the median feature vector of each cavity in each dimension to quantify the degree of deviation of a single cavity from the common state.
[0146] The calculated deviation data is appended to the corresponding multidimensional single-mode cavity feature vector to form an enhanced feature data structure containing lateral deviation information.
[0147] Step S5: Establish a benchmark library of permissible fluctuation ranges for multidimensional features based on historical normal cycles. Compare the multidimensional time-series feature vector set with the common operating condition feature vector and the benchmark library of permissible fluctuation ranges for multidimensional features to obtain lateral deviation and longitudinal deviation, and fuse them to generate a filling balance index. Specifically, this includes: S5.1: Perform statistical distribution analysis on the multidimensional single-cavity feature vectors stored in the historical normal injection molding cycle to extract the mean vector and standard deviation vector of each feature dimension under stable process conditions, and construct a benchmark library of multidimensional feature allowable fluctuation range containing dynamic tolerance intervals.
[0148] Receive a set of multi-dimensional single-cavity feature vectors stored in historical normal injection molding cycles. This set includes arrival time difference, pressure rise slope, pressure integral value, end pressure value, and curve shape deviation vector data of the filtered cycles without abnormal markers.
[0149] The multidimensional single-mode cavity feature vector set is grouped according to feature dimensions. Statistical distribution characteristic analysis is performed for each feature dimension. The arithmetic mean and standard deviation of each feature under stable process conditions are calculated to construct a set of basic statistical parameters characterizing central tendency and dispersion.
[0150] Based on the calculated mean vector and standard deviation vector, a dynamic tolerance coefficient is introduced, and the upper and lower bounds of the allowable fluctuation for each feature dimension are calculated using the following formula: in, and These represent the upper and lower bounds of the allowable fluctuations in the feature dimension, respectively. This represents the mean of this feature dimension during a historical normal cycle. Standard deviation, This is the tolerance coefficient set according to the confidence level requirement.
[0151] The calculated upper and lower bounds of the allowable fluctuations for each feature dimension are encapsulated into a multi-dimensional feature allowable fluctuation range benchmark library. This benchmark library defines the normal distribution range of each mold cavity feature value under the current process state.
[0152] S5.2: Based on the multi-dimensional time-series feature vector set and common working condition feature vector generated in the current injection cycle, perform dimension-by-dimensional difference operation to calculate the deviation value of each cavity feature relative to the median state in the same cycle, and generate a lateral deviation vector set characterizing the differences between cavities.
[0153] Read the multi-dimensional single-cavity feature vectors of all mold cavities in the current injection cycle and the common working condition feature vectors generated by median statistics.
[0154] By indexing and traversing each dimension of the feature vector, a mapping relationship between the cavity number and the feature value is established.
[0155] Regarding the first The first mold cavity For each eigenvalue, perform a subtraction operation to calculate the difference between it and the corresponding eigenvalue of the common operating condition.
[0156] The lateral deviation is calculated using the following formula: in, For the first The mold cavity in the first Lateral bias on dimensional features This is the actual characteristic value of the mold cavity. The first characteristic vector of common working conditions The median value of the dimension.
[0157] The calculated set of differences can be processed by absolute value manipulation or by retaining the sign to distinguish between lagging and overly fast trends.
[0158] The deviation values of each dimension are reorganized according to the mold cavity number to construct a lateral deviation vector containing multi-dimensional information of time, pressure and morphology.
[0159] S5.3: Perform boundary comparison processing using the lateral deviation vector set and the dynamic tolerance interval in the multi-dimensional feature allowable fluctuation range benchmark library to quantify the magnitude of each cavity feature exceeding the normal fluctuation range. The comparison result obtained by the boundary comparison processing is the longitudinal deviation mentioned in step S5 (wherein, the lateral deviation is referenced to the median value of each cavity feature within the same injection cycle, used to locate the relative difference between cavities; the longitudinal deviation is referenced to the multi-dimensional feature allowable fluctuation range benchmark library established in the historical normal cycle, used to determine whether each cavity feature exceeds the normal fluctuation range). The weighted result of the mapping value of each feature dimension of each cavity is recorded as the local imbalance coefficient of the cavity, and the local imbalance coefficients of each cavity are combined into a local imbalance coefficient set.
[0160] The system receives the lateral deviation vector set and the dynamic tolerance interval parameters from the multi-dimensional feature allowable fluctuation range benchmark library. It extracts the mean and standard deviation vectors of the corresponding feature dimensions from the benchmark library as statistical benchmarks for normalization calculations. For each cavity feature component in the lateral deviation vector set, it performs standardized deviation calculation based on the Mahalanobis distance concept. The following formula is used to calculate the... The mold cavity in the first Standardized deviation index on each feature dimension: in, The standardized deviation index, This is the lateral deviation value. For the first in the benchmark library Standard deviation of dimensional features To prevent the use of tiny constants with zero denominators, the calculated standardized deviation indices for each dimension undergo nonlinear mapping. The Sigmoid function is used to compress the deviation indices from the infinite domain to a finite interval of 0 to 1. The mapped values for each dimension are then fused using a weighted summation method, with weight coefficients dynamically allocated based on the sensitivity of the features to filling balance. Local imbalance coefficients characterizing the severity of anomalies in individual mold cavities are generated, and these local imbalance coefficients for each mold cavity are combined into a set. Through the above standardization and nonlinear mapping processes, the original deviations from the previous step are transformed into dimensionless and comparable local imbalance coefficients, enabling quantitative grading of the anomaly severity of each mold cavity and eliminating the interference of dimensional differences on the judgment results.
[0161] S5.4: Perform weighted aggregation operation on the set of local imbalance coefficients, and assign weight coefficients according to the sensitivity of pressure integral value and curve shape deviation vector, so as to integrate deviation information of time, rate and shape in multiple dimensions and generate a comprehensive filling balance index that characterizes the overall mold filling consistency.
[0162] It receives the set of local imbalance coefficients and the sensitivity weight parameters of each feature dimension, and constructs a multi-dimensional weighted aggregation model to generate a comprehensive filling balance index.
[0163] Based on the sensitivity of pressure integral values to changes in flow energy and the ability of curve shape deviation vectors to capture topological anomalies, a dynamic weight allocation strategy is set, in which the pressure integral weight reflects the difference in macroscopic resistance and the shape deviation weight characterizes the microscopic flow distortion.
[0164] The following formula is used to calculate the first... Overall imbalance score for each mold cavity: in, For the first The overall imbalance score of each mold cavity. For the first Normalized weight coefficients for each feature dimension. For the first The mold cavity in the first Local imbalance coefficients in each feature dimension This represents the total number of feature dimensions.
[0165] The maximum value extraction operation is performed on the comprehensive imbalance score of all cavities in the same period to identify the cavity that deviates most significantly from the common working condition in the current period as the dominant anomaly source.
[0166] The extracted maximum overall imbalance score is mapped to a dimensionless index in the range of 0 to 1. The closer the value is to 1, the worse the filling consistency is, and the closer the value is to 0, the more balanced the filling is.
[0167] S5.5: Based on the comprehensive filling balance index and the preset adaptive anomaly judgment threshold, perform logical mapping processing to determine the filling state level of the current injection molding cycle and associate it with the corresponding abnormal mold cavity number, and output the filling imbalance detection result containing the abnormal mold cavity number, the degree of abnormality and the type of abnormality feature.
[0168] The system receives the fill balance index generated by S5.4 and the preset adaptive anomaly judgment threshold. It performs a numerical comparison operation on the fill balance index and the adaptive anomaly judgment threshold, constructing a logical discriminant function to define the fill consistency status level of the current cycle. A segmented mapping mechanism is used to discretize the continuous fill balance index into three status indicators: normal, slightly unbalanced, and severely unbalanced. Based on the status indicator, the cavity tracing logic is triggered. When an unbalanced state is determined, the system retrieves the lateral deviation vector set generated by S5.2. The component with the largest absolute value in the lateral deviation vector set and its corresponding cavity index are extracted to locate the dominant abnormal cavity number. The ratio of the local imbalance coefficient of this dominant cavity to the baseline tolerance upper limit is calculated to generate a normalized anomaly degree quantification value. Combined with the arrival time difference sign bit extracted by S4, the anomaly type is determined to be either fill lag or fill too fast. The abnormal cavity number, anomaly degree quantification value, and anomaly feature type are encapsulated into a fill imbalance detection result and serialized using JSON format. By using logical mapping and multi-source data fusion processing, the quantitative indicators from the previous step are transformed into diagnostic results data with clear physical meaning, achieving the expected technical effect of closed-loop output from data statistics to engineering decision-making.
[0169] Step S6: Compare the filling balance index with the adaptive anomaly detection threshold. When the filling balance index exceeds the adaptive anomaly detection threshold, determine the abnormal cavity with filling lag or filling too fast and its degree of abnormality, and generate a filling imbalance detection result containing the abnormal cavity number, degree of abnormality, and type of abnormality feature. Specifically, this includes: S6.1: Based on the dynamic tolerance range generated by the statistics of historical normal injection molding cycles and the fluctuation rate of current process parameters, perform adaptive threshold dynamic reconstruction processing to calculate the filling imbalance judgment critical value that is automatically adjusted with the drift of working conditions, and generate an adaptive anomaly judgment threshold that is adapted to the current production environment.
[0170] The mean vector and standard deviation vector of the multidimensional feature allowable fluctuation range stored in the historical normal injection molding cycle are obtained as static basic parameters for adaptive threshold calculation.
[0171] Read the process parameter settings for the current injection molding cycle, including injection speed, melt temperature, and holding pressure, and calculate their deviation from the standard process parameters to generate the operating condition fluctuation coefficient.
[0172] Based on the operating condition fluctuation coefficient, the standard deviation vector in the benchmark library is dynamically scaled, and the adjustment factor for the dynamic tolerance interval is calculated using the following formula: in, To adjust the factor, This is the sensitivity adjustment coefficient. This is the operating condition fluctuation coefficient.
[0173] The adjustment factor is applied to the baseline standard deviation vector, and a dynamic standard deviation vector that changes with the operating condition drift is generated through multiplication to expand or shrink the tolerance interval boundary.
[0174] By combining the preset confidence level coefficient, the dynamic standard deviation vector and the mean vector are combined to construct a set of dynamic tolerance intervals containing upper and lower limits. The confidence level coefficient is set according to the allowable false alarm rate. When the tolerance interval is required to cover 99.7% of normal samples, the value is 3, and when it is required to cover 95% of normal samples, the value is 1.96. The lower the false alarm rate requirement, the larger the value and the wider the tolerance interval.
[0175] Smoothing filtering is applied to the dynamic tolerance interval set to eliminate threshold jitter caused by single-cycle data mutations and generate a stable adaptive anomaly detection threshold.
[0176] For example, setting the sensitivity adjustment coefficient to 0.5 and the current cycle operating condition fluctuation coefficient to 0.2, the calculated adjustment factor is 1.1. If the baseline standard deviation is 0.5 MPa, the dynamic standard deviation is adjusted to 0.55 MPa. At a 95% confidence level, the dynamic tolerance range is expanded from ±0.98 MPa to ±1.078 MPa. This adjustment effectively avoids false alarms caused by slight fluctuations in melt temperature and significantly improves the system's accuracy in determining varying operating conditions.
[0177] S6.2: Perform multi-dimensional logic comparison processing on the comprehensive filling balance index and the adaptive anomaly judgment threshold. By judging whether the local imbalance coefficient continues to exceed the boundary of the dynamic tolerance interval, the filling status level of the current injection cycle is determined, and a Boolean status flag bit representing whether the filling consistency is qualified or abnormal is generated.
[0178] The system receives the adaptive anomaly detection threshold generated by S6.1 and the comprehensive filling balance index and local imbalance coefficient set output by S5. The comprehensive filling balance index is mapped to a multi-dimensional feature space and compared with the adaptive anomaly detection threshold dimension by dimension. For each cavity's local imbalance coefficient, its deviation from the upper and lower limits of the dynamic tolerance interval is extracted. The normalized deviation of each feature dimension is calculated to construct a cavity state discrimination vector. A time persistence constraint mechanism is introduced, performing a sliding window integration operation on the deviation of N consecutive sampling points. N is set as follows: 3-10 sampling points are selected based on the injection cycle duration and sampling frequency, and the judgment condition is that the normalized deviation of all sampling points within the window is greater than 1.0. When the integral value exceeds a preset confidence threshold, the cavity is marked as a potential anomaly source. The confidence threshold is set as follows: based on the statistical distribution of the deviation integral value under normal cycles, it is taken as 3-5 times the mean of the distribution, used to exclude mislabeling caused by single-point instantaneous noise. The marked states of all cavities are combined, and a logical AND operation is performed to eliminate instantaneous noise interference. If at least one cavity is marked and the overall filling balance index exceeds the global threshold, the current cycle filling status is determined to be abnormal. A Boolean status flag representing whether the filling consistency is acceptable or abnormal is generated and associated with the cavity index list that triggered the abnormality. Through multi-dimensional logical comparison and continuous verification, the quantitative index is transformed into a deterministic status flag, achieving highly robust initial screening of abnormalities.
[0179] S6.3: Based on the result that the Boolean state flag is abnormal, extreme value search and directionality discrimination processing are performed on the transverse deviation vector set. By analyzing the positive and negative signs of the arrival time difference value and the deviation of the pressure rise slope, the dominant mold cavity that causes filling imbalance and its hysteresis or excessive speed attributes are identified, and a list of candidate abnormal mold cavities with abnormal flow direction indicators is generated.
[0180] Receive the set of lateral deviation vectors determined by S6.2 to be in an abnormal state. This set includes the arrival time difference, pressure rise slope and end pressure deviation value of each mold cavity relative to the common working condition.
[0181] Sign discrimination processing is performed on the arrival time difference components in the lateral deviation vector set. Positive deviations are marked as candidates for filling lag, and negative deviations are marked as candidates for filling too fast. That is, filling lag and filling too fast are determined by the positive and negative signs of the arrival time difference, and a time deviation label sequence with directional attributes is generated.
[0182] Based on the time deviation label sequence, the pressure rise slope deviation value of the corresponding mold cavity is extracted, and the ratio of the absolute value of the slope deviation to the baseline standard deviation is calculated to quantify the severity of the change in flow resistance.
[0183] An extreme value search method is used to traverse the weighted deviation scores of all cavities, identify the dominant cavity with the largest deviation, and eliminate the interference of small fluctuations caused by sensor noise.
[0184] The cavity anomaly direction index is calculated using the following formula: in, This is an abnormal direction index. The time difference symbol is for arrival. This represents the slope deviation value. is the standard deviation of the slope.
[0185] Based on the sign and magnitude of the abnormal direction index, determine the number and physical properties of the dominant abnormal cavity, namely, lag caused by runner blockage or excessive speed caused by excessive gate size.
[0186] The identified dominant cavity number, abnormal attribute label, and corresponding key feature deviation value are encapsulated to generate a candidate abnormal cavity list with clear flow direction identification.
[0187] S6.4: Perform multi-feature weighted fusion evaluation on each cavity in the candidate abnormal cavity list. Use the weighted product of the curve shape deviation vector and the local imbalance coefficient as a weight factor to quantify the severity of deviation of each cavity from the common working condition and generate a quantitative value of the abnormality degree characterizing the severity of the fault of a single cavity.
[0188] The system receives a list of candidate abnormal cavities and their corresponding sets of lateral deviation vectors, local imbalance coefficients, and curve morphology deviation vectors. For each abnormal cavity in the list, the curve morphology deviation norm calculated and stored in S4.3 is read to characterize the degree to which the overall topology of the pressure curve deviates from the common operating condition. The product of the morphology deviation norm and a preset morphology sensitivity factor is calculated to generate a morphology weight coefficient, which amplifies the contribution of subtle morphological distortions in the assessment. The preset morphology sensitivity factor is set as follows: based on the indicative ability of curve morphology deviation to structural defects, it is set to 1.0-3.0, preferably 2.0. The larger the value, the higher the weight of morphological distortion in the anomaly assessment. The local imbalance coefficient of the cavity is obtained, which quantifies the extent to which scalar characteristics such as time, rate, and end pressure exceed the dynamic tolerance range. The morphology weight coefficient and the local imbalance coefficient are nonlinearly coupled to construct an anomaly quantification model. The anomaly quantification value is calculated using the following formula: in, To quantify the degree of abnormality, As a local imbalance regulator, This is the local imbalance coefficient. This is the shape sensitivity gain coefficient. This represents the curve shape deviation norm. The calculated... The values are normalized and mapped to a dimensionless range of 0 to 100. This means the quantified anomaly severity value ranges from 1 to 100; a higher value indicates a more severe filling imbalance in the mold cavity, eliminating the magnitude differences caused by different dimensional features. Based on the normalized E-value, three anomaly levels—slight, moderate, and severe—are defined, each corresponding to different process intervention strategies. Through multi-feature weighted fusion evaluation, the candidate anomaly information from the previous step is transformed into a quantified anomaly severity value, achieving precise classification and localization of the severity of the filling imbalance fault.
[0189] S6.5: Perform structured data encapsulation processing based on the abnormality degree quantification value and abnormal flow direction identifier, encapsulate the abnormal cavity number, abnormality degree quantification value and abnormal feature type of filling lag or filling too fast into filling imbalance detection result, and output the filling imbalance detection result according to the standard communication protocol format.
[0190] Receive the set of abnormal cavity numbers generated in step S6.4, the corresponding abnormality quantification value, and the abnormal flow direction identifier that represents the filling lag or too fast.
[0191] The abnormal mold cavity number is converted into a binary code and mapped to a fixed-length device address code to ensure compatibility with the addressing protocol of the injection molding machine control system.
[0192] The abnormality level is quantified and formatted using fixed-point numbers, preserving the preset decimal places of precision, eliminating floating-point transmission errors, and generating a standardized severity data field.
[0193] Based on the logical state of the abnormal flow identifier, a table is looked up to match the predefined fault type code, mapping the filling delay to a specific error code and the filling too fast to another specific error code, generating an abnormal feature type identifier.
[0194] Construct a data frame structure containing timestamp, mold ID, abnormal mold cavity address code, severity data field, and abnormal feature type identifier, and serialize and encapsulate it according to the JSON or Modbus TCP standard communication protocol format.
[0195] A checksum field is added to the encapsulated data frame, and a cyclic redundancy check is performed to ensure the integrity and anti-interference capability of data transmission, and to generate a fill imbalance detection result.
[0196] Step S7: Based on the filling imbalance detection result, output an alarm signal or process adjustment suggestion instruction. When the current cycle is confirmed to be in a normal state based on the filling balance index trend of multiple consecutive cycles, use exponential weighted moving average processing to update the multidimensional feature allowable fluctuation range benchmark library with the multidimensional time series feature vector set of the current cycle, and update the adaptive anomaly judgment threshold for subsequent cycles accordingly. Specifically, this includes: S7.1: Perform protocol parsing and status bit extraction on the filling imbalance detection results to identify Boolean status flags and associated abnormality quantification values, and generate a structured control instruction set containing abnormal cavity numbers and abnormal feature types.
[0197] Receive the filling imbalance detection result generated from step S6; the filling imbalance detection result includes at least the abnormal cavity number, the abnormality degree quantification value and the abnormal feature type, and saves the Boolean state flag, the comprehensive filling balance index and the local imbalance coefficient.
[0198] The data stream is subjected to protocol header verification and payload unpacking. Key fields are extracted according to the predefined communication protocol format, and Boolean flags representing the overall state of the cycle and quantitative numerical sequences representing the deviation of each cavity are separated.
[0199] When the Boolean status flag indicates an anomaly, the set of local imbalance coefficients is traversed to identify the cavity index whose value exceeds the boundary of the dynamic tolerance interval and mark it as an anomaly object to be processed.
[0200] Read the abnormal flow direction identifier of the corresponding index in the candidate abnormal cavity list, and determine the filling lag or filling too fast attribute by combining the positive or negative sign of the arrival time difference, and determine the abnormal feature type code.
[0201] The abnormal cavity number, abnormality degree quantification value, and abnormal feature type are encoded and encapsulated into a structured control instruction set, and the instruction priority and execution timestamp are defined.
[0202] S7.2: Based on the structured control instruction set, perform multi-channel signal mapping processing, convert abnormal mold cavity numbers into specific audible and visual alarm codes or map abnormal feature types into injection molding machine process parameter adjustment suggestion instructions, and generate real-time response signals that can be directly executed by the human-machine interface or the underlying controller.
[0203] Receive a structured control instruction set containing the abnormal cavity number, the quantified value of the abnormality degree, and the abnormality feature type.
[0204] The abnormal characteristic type field in the parsing instruction set is used to distinguish specific fault modes such as filling lag, filling too fast, or abnormal flow resistance.
[0205] For the audible and visual alarm channel, a mapping dictionary between the physical location of the cavity and the alarm code is established, converting the abnormal cavity number into a specific LED flashing frequency code. The LED flashing frequency code, together with the following buzzer drive signal parameters, constitutes the audible and visual alarm code, used to distinguish the location and severity of the abnormal cavity through both light flashing and sound.
[0206] The abnormality level is quantified into three grades: mild, moderate, and severe, each matched with a buzzer drive signal parameter of different volume. The criteria for determining the grade are as follows: a quantitative abnormality value between 1 and 30 is considered mild, between 31 and 70 is considered moderate, and between 71 and 100 is considered severe. The volume of the buzzer drive signal parameter increases sequentially with the grade, allowing users to judge the severity of the abnormality by the sound intensity without needing to consult a host computer interface.
[0207] For the process adjustment suggestion channel, the system searches the preset process correction strategy library based on the anomaly characteristic type to obtain the corresponding injection molding machine parameter adjustment instruction template.
[0208] If the filling is determined to be delayed, an incremental adjustment command is generated to increase the injection speed, increase the melt temperature, or extend the holding time; if the filling is determined to be too fast, a reduction adjustment command is generated to reduce the upper limit of the injection pressure or reduce the mold temperature.
[0209] The alarm codes, buzzer drive parameters, and process adjustment instructions are encapsulated into standard Modbus TCP communication messages or PLC register write instructions.
[0210] Real-time response signals are sent to the injection molding machine's human-machine interface and the underlying controller execution unit via Ethernet or fieldbus interface.
[0211] S7.3: Perform normal state secondary verification processing on the multi-dimensional time-series feature vector set of the current injection cycle after the real-time response signal is triggered. Combined with the trend analysis of the filling balance index of multiple consecutive cycles, confirm whether the current cycle belongs to the normal fluctuation range without real defects, and generate candidate updated feature vectors with normal state confirmation marks.
[0212] The system receives the multi-dimensional temporal feature vector set of the current injection molding cycle and the fill balance index sequence of N consecutive historical cycles after receiving the real-time response signal trigger. A trend analyzer based on a sliding time window is constructed to extract the comprehensive fill balance index of the most recent M cycles to form a short-term trend vector. The ratio of the standard deviation to the mean of this short-term trend vector is calculated as the fluctuation coefficient to quantify the dispersion of the recent process state. A dynamic stability threshold is set; when the fluctuation coefficient is lower than this threshold, the current cycle is determined to be in a steady-state operating range. The dynamic stability threshold is set by taking 2-3 times the mean of the statistical distribution of the fluctuation coefficient of historical steady-state cycles. Isolated forest anomaly detection is performed on the lateral deviation vector of the current cycle, and anomaly scores for each cavity feature are calculated. If the anomaly scores of all cavities are lower than the preset isolation threshold, and the fluctuation coefficient meets the steady-state condition, the current cycle is marked as a normal fluctuation without real defects. A normal state confirmation flag is added to the multi-dimensional temporal feature vector set of the current cycle to generate candidate update feature vectors. By employing a dual verification mechanism, the results of the previous step are transformed into high-confidence benchmark update data, thereby achieving the expected technical effect of eliminating transient interference and optimizing adaptive thresholds.
[0213] S7.4: Based on the candidate updated feature vector with normal state confirmation mark, perform sliding window statistical reconstruction processing, and use the exponential weighted moving average method to fuse the mean vector and standard deviation vector of the new cycle data with the historical multidimensional feature allowable fluctuation range benchmark library to generate dynamic tolerance interval update parameters that reflect the latest process drift characteristics.
[0214] Receive candidate updated feature vectors with a normal state confirmation flag. These vectors contain multi-dimensional temporal feature data for each mold cavity within the current injection molding cycle. Extract the historical mean vector and historical standard deviation vector stored in the multi-dimensional feature allowable fluctuation range benchmark library as the initial state for statistical updates. Set a smoothing factor for the exponentially weighted moving average method (i.e., the weighting coefficient of the exponentially weighted moving average processing). The smoothing factor ranges from 0.1 to 0.8 and is dynamically adjusted according to the process stability level to balance the sensitivity of new data and the inertia of historical data. Construct a mean update model by linearly weighting the candidate feature vector of the current cycle with the historical mean vector. Calculate the updated mean vector using the following formula: in, for The mean vector updated at each time step. As a smoothing factor, This is the candidate feature vector for the current period. for The historical mean vector at each time step. A variance update model is constructed, calculating the squared deviation between the current feature vector and the updated mean vector, and then weighted and fused with the historical variance. The updated variance vector is calculated using the following formula: in, for The variance vector updated at each time step. for The historical variance vector at each moment is used. The updated variance vector is square-rooted to generate a dynamic standard deviation vector reflecting the dispersion of the latest process. Based on the updated mean vector and dynamic standard deviation vector, the dynamic tolerance interval boundaries for each feature dimension are recalculated. Through exponentially weighted moving average processing, the candidate feature vectors from the previous step are transformed into dynamic tolerance interval update parameters reflecting the drift characteristics of the latest process, achieving the expected technical effect of adaptive tracking and real-time correction of long-term slow process drift by the benchmark library.
[0215] S7.5: Perform benchmark library version iteration writing processing on the dynamic tolerance interval update parameters, replace the old statistical distribution model in the multi-dimensional feature allowable fluctuation range benchmark library, complete the detection loop closure, and generate an adaptive anomaly judgment threshold adapted to the next injection molding cycle to optimize the judgment accuracy of subsequent cycles.
[0216] Receive the dynamic tolerance interval update parameters generated by S7.4, including the updated mean vector and standard deviation vector.
[0217] Version locking and read / write permission verification are performed on the multi-dimensional feature allowable fluctuation range benchmark library to ensure that no other processes modify the data concurrently during the data writing process, thus guaranteeing data consistency.
[0218] An atomic replacement strategy is used to mark the old version of the statistical distribution model as pending archiving and to instantiate the newly calculated mean vector and standard deviation vector into a new benchmark model object.
[0219] Based on the standard deviation vector in the new benchmark model, and combined with the preset confidence coefficient, the dynamic tolerance upper and lower bounds of each feature dimension are recalculated. The confidence coefficient is set as follows: 1.96-3 according to the false alarm rate requirement, corresponding to a confidence interval of 95%-99.7% respectively.
[0220] Calculate the first using the following formula Adaptive anomaly detection thresholds for each feature dimension: in, The threshold boundary for the i-th dimension feature is... This is the updated mean. The updated standard deviation, The confidence level coefficient is set according to the false alarm rate requirement.
[0221] The threshold boundaries of each feature dimension are encapsulated into a standard data structure for the adaptive anomaly determination threshold, and associated with the current mold number and material batch identifier.
[0222] The encapsulated threshold set is written into the configuration cache of the real-time detection engine to activate the anomaly detection logic for the next injection molding cycle.
[0223] Release memory resources occupied by the old version model to complete the version iteration and closed-loop update of the benchmark library.
[0224] By using atomic substitution and threshold recalculation, the dynamic tolerance interval update parameters are transformed into adaptive anomaly judgment thresholds that adapt to the latest operating conditions, enabling the detection system to continuously optimize the real-time tracking and judgment accuracy of process drift.
[0225] 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.
[0226] Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention 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” involved in the embodiments of this invention refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0227] The above description is merely an exemplary embodiment of the present invention, but the scope of protection of the present invention 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 the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. An online detection method for injection molding filling imbalance based on cavity pressure time-series analysis, characterized in that, Includes the following steps: Acquire the pressure timing signal and injection start trigger signal of each mold cavity, extract the pressure data of each mold cavity filling stage according to the injection start trigger signal, and generate a set of pressure curves for the filling stage. Baseline correction, filtering and noise reduction, and outlier identification are performed on the set of pressure curves during the filling stage to determine the arrival time of the melt front and the end time of filling, thereby generating a clean filling pressure curve sequence. Map the pressure time axis of each mold cavity from the arrival time of the melt front to the end time of filling to the filling phase coordinate and perform phase alignment to generate a standardized set of filling pressure curves. The arrival time difference, pressure rise slope, pressure integral value, end pressure value and curve shape deviation vector are extracted from the standardized filling pressure curve set. Common working condition feature vectors are generated using the median value of each cavity feature in the same period to form a multi-dimensional time series feature vector set. A benchmark library of allowable fluctuation ranges of multidimensional features is established based on historical normal cycles. The multidimensional time series feature vector set is compared with the common working condition feature vector and the benchmark library of allowable fluctuation ranges of multidimensional features to obtain the lateral deviation and the longitudinal deviation, and then fused to generate a filling balance index. The filling balance index is compared with the adaptive anomaly determination threshold. When the filling balance index exceeds the adaptive anomaly determination threshold, the abnormal cavity with filling lag or filling too fast and the degree of abnormality are determined, and a filling imbalance detection result containing the abnormal cavity number, the degree of abnormality and the type of abnormal feature is generated.
2. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, Also includes: When the current period is confirmed to be in a normal state based on the trend of the filling balance index over multiple consecutive periods, an exponentially weighted moving average is used to update the multidimensional feature allowable fluctuation range benchmark library with the multidimensional time series feature vector set of the current period, and the adaptive anomaly judgment threshold for subsequent periods is updated accordingly.
3. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, The process of mapping the pressure-time axis of each mold cavity from the arrival time of the melt front to the end time of filling to the filling phase coordinates and performing phase alignment to generate a standardized filling pressure curve set specifically includes: The arrival time of the melt front of each mold cavity in the clean filling pressure curve sequence is read and analyzed. The absolute delay time value of each mold cavity relative to the injection cycle start time anchor point is extracted to generate a set of mold cavity time offsets. Based on the minimum value in the set of cavity time offsets as the global zero phase reference point, a linear translation transformation is performed on the original time axis of each cavity to eliminate the absolute time difference caused by the inconsistency of physical flow channel length and generate a relative time coordinate sequence. Normalization scaling is performed on the relative time coordinate sequence. The time axis is mapped to the dimensionless filling phase interval from zero to one by using the difference between the filling end time and the arrival time of the melt front of each cavity as the denominator, thus generating the filling phase coordinate axis. Based on the filling phase coordinate axis, resampling interpolation is performed on the discrete pressure sampling points in the clean filling pressure curve sequence. The cubic spline interpolation method is used to reconstruct the non-uniformly distributed pressure data into a pressure value sequence with equal phase intervals, generating a discrete pressure data point set. The discrete pressure data point set is reorganized into a data structure and associated with the cavity identifier. The pressure values of each cavity under the same filling phase are encapsulated into a matrix form, and finally a standardized filling pressure curve set is generated.
4. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, The process involves extracting arrival time difference, pressure rise slope, pressure integral value, end pressure value, and curve shape deviation vector from the standardized filling pressure curve set. Common operating condition feature vectors are then generated using the median values of the features of each mold cavity in the same period, forming a multi-dimensional time-series feature vector set. Specifically, this includes: An extreme value search process based on a dynamic time window is performed on the discrete pressure data point set of each cavity in the standardized filling pressure curve group to extract the arrival time difference value of each cavity relative to the global zero phase reference point and the end pressure value at the end of filling, and generate a set of key point features of the cavity. Based on the effective filling interval determined by the set of key features of the mold cavity, piecewise linear regression fitting is performed on the standardized filling pressure curve set to calculate the pressure rise slope of each mold cavity from the arrival time of the melt front to the end time of filling, and trapezoidal integral operation is performed on the pressure values within the same time to obtain the pressure integral value, thereby generating a set of dynamic features of the mold cavity. Using the pressure integral value in the set of mold cavity dynamic features as a weighting coefficient, the standardized filling pressure curve group is processed by normalizing the morphological vector construction process. The discrete pressure values of each mold cavity on the standard filling phase coordinate axis are mapped to the curve morphological deviation vector in high-dimensional space to generate a set of mold cavity morphological features. The key feature set, dynamic feature set, and morphological feature set of the mold cavity are fused by the mold cavity number. The arrival time difference, pressure rise slope, pressure integral value, end pressure value, and curve shape deviation vector are encapsulated into a single data structure to generate a multi-dimensional single mold cavity feature vector. Based on the multidimensional single-cavity feature vectors generated by all cavities in the current injection molding cycle, perform a dimension-wise median statistical operation to remove extreme outlier interference and extract common working condition feature vectors representing the overall mold state. At the same time, associate and encapsulate each multidimensional single-cavity feature vector with the common working condition feature vector to generate a multidimensional time-series feature vector set.
5. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, The process involves establishing a multi-dimensional feature allowable fluctuation range benchmark library based on historical normal cycles, comparing the multi-dimensional time-series feature vector set with the common operating condition feature vector and the multi-dimensional feature allowable fluctuation range benchmark library respectively, obtaining lateral deviation and longitudinal deviation, and fusing them to generate a fill balance index. Specifically, this includes: Statistical distribution analysis is performed on the multidimensional single-cavity feature vectors stored in the historical normal injection molding cycle to extract the mean vector and standard deviation vector of each feature dimension under stable process conditions, and to construct a benchmark library of allowable fluctuation ranges of multidimensional features. Based on the multi-dimensional time-series feature vector set and common working condition feature vector generated in the current injection cycle, perform dimension-by-dimensional difference operation to calculate the deviation value of each cavity feature relative to the median state in the same cycle, and generate a lateral deviation vector set. The boundary comparison process is performed by comparing the lateral deviation vector set with the dynamic tolerance interval in the multidimensional feature allowable fluctuation range benchmark library to quantify the amplitude ratio of each cavity feature exceeding the normal fluctuation range and generate a set of local imbalance coefficients. A weighted aggregation operation is performed on the set of local imbalance coefficients. Weight coefficients are assigned based on the sensitivity of the pressure integral value and the curve shape deviation vector to integrate deviation information in multiple dimensions such as time, rate and shape, and generate a comprehensive filling balance index. Based on the comprehensive filling balance index and the preset adaptive anomaly judgment threshold, a logical mapping process is performed to determine the filling state level of the current injection molding cycle and associate it with the corresponding abnormal mold cavity number, outputting the final filling imbalance detection result data stream containing the quantification value of the anomaly degree.
6. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, The process involves comparing the fill balance index with an adaptive anomaly detection threshold. When the fill balance index exceeds the adaptive anomaly detection threshold, an abnormal cavity with either delayed or excessive filling, along with its degree of abnormality, is identified. A fill imbalance detection result is then generated, containing the abnormal cavity number, degree of abnormality, and type of abnormality characteristics. Specifically, this includes: Based on the dynamic tolerance range generated by the statistics of historical normal injection molding cycles and the fluctuation rate of current process parameters, an adaptive threshold dynamic reconstruction process is performed to calculate the critical value for filling imbalance judgment that is automatically adjusted with the drift of working conditions, and to generate an adaptive anomaly judgment threshold that is adapted to the current production environment. A multi-dimensional logical comparison process is performed on the comprehensive filling balance index and the adaptive anomaly judgment threshold. By judging whether the local imbalance coefficient continues to exceed the boundary of the dynamic tolerance interval, the filling status level of the current injection cycle is determined, and a Boolean status flag bit representing whether the filling consistency is qualified or abnormal is generated. Based on the result that the Boolean state flag is abnormal, extreme value search and directionality discrimination processing are performed on the lateral deviation vector set. By analyzing the positive and negative signs of the arrival time difference value and the deviation of the pressure rise slope, the dominant cavity that causes filling imbalance and its hysteresis or excessive speed attributes are identified, and a list of candidate abnormal cavities with abnormal flow direction indicators is generated. For each cavity in the candidate abnormal cavity list, a multi-feature weighted fusion evaluation process is performed. The weighted product of the curve shape deviation vector and the local imbalance coefficient is used as a weighting factor to quantify the severity of deviation of each cavity from the common working condition and generate a quantitative value of the abnormality degree characterizing the severity of the fault of a single cavity. Structured data encapsulation processing is performed based on the abnormality degree quantification value and abnormal flow direction identifier. The abnormal cavity number, abnormality degree quantification value and abnormal feature type of filling lag or filling too fast are encapsulated into filling imbalance detection result, and the filling imbalance detection result is output according to the standard communication protocol format.
7. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, The pressure timing signal is obtained by performing high-frequency synchronous sampling processing on the analog voltage signals of each cavity, and the high-frequency synchronous sampling frequency is 1-20kHz.
8. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 1, characterized in that, The baseline correction uses sliding window mean estimation, the filtering and denoising uses wavelet thresholding, and the outlier identification compares the slope mutation rate with the physical limit threshold and uses linear interpolation to repair the gaps; the threshold setting range for the wavelet threshold is 1-100.
9. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 2, characterized in that, The weighting coefficient for the exponentially weighted moving average processing is between 0.1 and 0.8; the alarm signal includes an audible and visual alarm code; and the process adjustment suggestion instruction includes suggestions for adjusting injection pressure, injection speed, holding time, or mold temperature.
10. The online detection method for injection molding filling imbalance based on cavity pressure time sequence analysis according to claim 6, characterized in that, The range of the abnormality quantification value is 1-100.
