An injection molding process intelligent fault diagnosis method, device and equipment

Abstract

The application discloses an injection molding process intelligent fault diagnosis method, device and equipment, and relates to the technical field of intelligent fault diagnosis. The method comprises the following steps: acquiring multi-dimensional operation data in an injection molding process; determining initial boundary time between adjacent process links in each molding cycle of the injection molding process based on the multi-dimensional operation data; refining the initial boundary time to obtain refined boundary time according to a variable point in the multi-dimensional operation data; dividing the multi-dimensional operation data into multi-dimensional operation sub-data corresponding to each process link by using the refined boundary time; constructing a special data fingerprint of the process link according to the multi-dimensional operation sub-data corresponding to the process link; and judging whether the process link is abnormal according to the special data fingerprint of the process link and a benchmark data fingerprint of the process link. The application can accurately identify the boundary of each process link and unique abnormalities.

Description

Technical Field

[0001] This application relates to the field of intelligent fault diagnosis technology, and in particular to an intelligent fault diagnosis method, apparatus and equipment for injection molding process. Background Technology

[0002] Injection molding is a highly efficient process for manufacturing complex plastic products. It involves injection, pressure holding, and solvent extraction to achieve material shaping. The production process involves multi-stage precision control, and fluctuations in process parameters (such as pressure, speed, and temperature) at each stage directly affect product quality. Even minor anomalies caused by equipment, material, or environmental disturbances, if not detected in time, can lead to a large number of scraps or equipment failures. Therefore, real-time anomaly detection in the injection molding process, ensuring parameters remain stable within allowable ranges, is crucial for maintaining production continuity and reducing scrap rates.

[0003] Some existing methods have begun to use machine learning models for fault identification, but they still have some shortcomings: First, they usually use fixed time windows or overall time-series data for modeling, but because the duration of the same process step in different molding cycles and different molding cycles varies, this method cannot accurately distinguish between different molding cycles and different process steps; Second, they usually use general features such as the mean, variance, maximum and minimum values ​​of the running data, combined with threshold alarms for fault diagnosis, but injection molding faults are often not manifested as a single abnormal running data, but rather as a coupling of abnormalities in multiple running data. Therefore, this method cannot accurately identify the specific abnormalities of each process step. Summary of the Invention

[0004] This application provides a method, apparatus, and equipment for intelligent fault diagnosis of injection molding process, in order to solve or at least partially solve the defects or deficiencies in related technologies.

[0005] Firstly, this application provides an intelligent fault diagnosis method for injection molding processes, the intelligent fault diagnosis method for injection molding processes comprising: During the injection molding process, multi-dimensional operational data is acquired; the injection molding process includes multiple molding cycles, and each molding cycle includes multiple process steps. Based on the multidimensional operational data, the initial boundary time between adjacent process steps in each molding cycle of the injection molding process is determined. Based on the changing points in the multidimensional operational data, the initial boundary time is refined to obtain the refined boundary time. The multidimensional operational data is divided into multidimensional operational sub-data corresponding to each process step in each molding cycle using the fine-tuning boundary moment. Based on the multi-dimensional operational sub-data corresponding to the process step, construct the exclusive data fingerprint of the process step; Based on the unique data fingerprint of the process step and the baseline data fingerprint of the process step, it is determined whether there is an anomaly in the process step.

[0006] Secondly, this application provides an intelligent fault diagnosis device for injection molding processes, the intelligent fault diagnosis device for injection molding processes comprising: The data acquisition module is used to acquire multi-dimensional operational data during the injection molding process; the injection molding process includes multiple molding cycles, and the molding cycle includes multiple process steps. A boundary determination module is used to determine the initial boundary time between adjacent process steps in each molding cycle of the injection molding process based on the multidimensional operating data. The boundary refinement module is used to refine the initial boundary time based on the variable points in the multidimensional running data to obtain the refined boundary time. The data partitioning module is used to divide the multidimensional operational data into multidimensional operational sub-data corresponding to each process step in each molding cycle using the fine-tuning boundary moment. The fingerprint construction module is used to construct a unique data fingerprint for the process step based on the multi-dimensional operational sub-data corresponding to the process step. The anomaly detection module is used to determine whether there is an anomaly in the process step based on the unique data fingerprint of the process step and the baseline data fingerprint of the process step.

[0007] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent fault diagnosis method for injection molding process described in the first aspect.

[0008] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides an intelligent fault diagnosis method, apparatus, and equipment for injection molding processes. During the injection molding process, multi-dimensional operational data is acquired. First, based on the multi-dimensional operational data, the initial boundary time between adjacent process steps in each molding cycle of the injection molding process is determined. Then, based on the change points in the multi-dimensional operational data, the initial boundary time is refined to obtain the refined boundary time, thereby accurately identifying the molding cycle boundary and process step boundary. This allows for precise location of abnormal process steps when an anomaly is detected in the injection molding process, achieving step-level fault location in the injection molding process. Next, the refined boundary time is used to divide the multi-dimensional operational data into multi-dimensional operational sub-data corresponding to each process step in each molding cycle. Then, based on the multi-dimensional operational sub-data corresponding to each process step, a unique data fingerprint is constructed for each process step to characterize the unique operational data features of each process step. Then, based on the unique data fingerprint and the baseline data fingerprint of the same process step, it is determined whether the process step has an anomaly. This method, based on the unique multi-dimensional operational data coupling characteristics of each process step, accurately identifies the unique anomalies of each process step, compared to the traditional threshold alarm mechanism which only alarms when a certain operational data exceeds a certain limit at a certain time. Attached Figure Description

[0009] Figure 1 A flowchart illustrating an intelligent fault diagnosis method for injection molding process provided in an embodiment of this application; Figure 2 A functional module diagram of an intelligent fault diagnosis device for injection molding process provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0010] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0011] In one exemplary embodiment, such as Figure 1 As shown, an intelligent fault diagnosis method for injection molding processes is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. The terminal can be, but is not limited to, various desktop computers, laptops, etc. The server can be a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0012] In the embodiments of this application, such as Figure 1 As shown, the intelligent fault diagnosis method for injection molding process includes the following steps 101 to 106. Wherein: Step 101: During the injection molding process, acquire multi-dimensional operational data; the injection molding process includes multiple molding cycles, and each molding cycle includes multiple process steps.

[0013] Injection molding machines are key pieces of equipment in the production of plastic products, and their operating status directly affects product quality, production efficiency, and equipment stability. A molding cycle of the injection molding process includes multiple process steps such as mold closing, injection, pressure holding, melting, mold opening, and ejection.

[0014] During the injection molding process using an injection molding machine, multi-dimensional operational data can be collected, including one or more of the following: pressure-related operational data (including injection pressure, holding pressure, ejection pressure, and / or back pressure, etc.), speed-related operational data (including injection speed, etc.), position-related operational data (including screw position, mold platen position, and / or ejection position, etc.), current-related operational data (including motor current and / or ejection current, etc.), temperature-related operational data (including barrel temperature and / or oil temperature, etc.), speed-related operational data (including screw speed, etc.), clamping force, and signal-related operational data (including ejection signal and / or mold opening signal, etc.). The "multi-dimensional" in multi-dimensional operational data refers to the number of features in the data. For example, injection pressure data is one feature in a multi-dimensional operational data set, holding pressure is another, and screw speed is yet another.

[0015] For multidimensional operational data with different sampling frequencies, resample at a uniform time step, such as 100ms or 200ms.

[0016] Linear interpolation or forward imputation is performed on missing data in multidimensional operational data, and Hampel filtering is used to process abnormal peaks.

[0017] Step 102: Based on multidimensional operational data, determine the initial boundary moments between adjacent process steps in each molding cycle of the injection molding process.

[0018] In this embodiment, a finite state machine (FSM) is established based on the injection molding machine's process actions, defining a state set: S = {S0, S1, S2, S3, S4, S5, S6}, where S0 represents the waiting / start of the molding cycle, S1 represents mold closing, S2 represents injection, S3 represents holding pressure, S4 represents melting, S5 represents mold opening, and S6 represents ejection. This state set represents all process steps in the molding cycle.

[0019] State transitions (representing a jump from one process stage to the next) are triggered by events. The following are example triggering events: S0→S1: The template position begins to move in the mold closing direction, and the mold closing signal is 1; S1→S2: The screw position change rate exceeds the threshold, and the injection pressure begins to rise rapidly; S2→S3: The injection speed decreases significantly, and the injection pressure changes from a rapid increase to a stable maintenance. S3→S4: The pressure holding end signal appears, or the injection pressure enters the decreasing range, and the screw starts to rotate (i.e. the screw speed changes from 0 to a non-zero value). S4→S5: Mold opening signal triggered, mold position moves in the mold opening direction; S5→S6: The ejection signal is triggered, and the ejection current or ejection pressure increases; S6→S0: The ejection and retraction are completed, and the next forming cycle begins.

[0020] In step 102, the aforementioned finite state machine is first obtained. This finite state machine includes trigger events for transitioning from one process stage to the next, with the trigger events set based on changes in multidimensional operational data. Then, based on this finite state machine, a target trigger event for transitioning from the first process stage to the second process stage is monitored during the injection molding process. The first process stage can be any process stage, and the second process stage is the next process stage after the first. If the target trigger event is detected, its occurrence time is determined as the initial boundary time between the first and second process stages. Based on the initial boundary time between each process stage and the next, the time interval, start time, and end time of each process stage can be determined.

[0021] Step 103: Based on the changing points in the multidimensional running data, refine the initial boundary time to obtain the refined boundary time.

[0022] In this step, the initial boundary time is further refined by utilizing the changing points in the multidimensional running data, specifically including the following steps 1031 to 1035.

[0023] Step 1031: Detect the changing points in the multidimensional running data using a time series change point detection algorithm; Step 1032: Determine the search window with the initial boundary time as the midpoint; Step 1033: Search for change points in the search window to obtain the set of candidate change points corresponding to the initial boundary time. Step 1034: For each candidate variable in the candidate variable set corresponding to the initial boundary time, determine the comprehensive confidence level of the candidate variable based on multidimensional running data; Step 1035: Select the candidate variable with the highest comprehensive confidence from the candidate variable set corresponding to the initial boundary time as the refined boundary time corresponding to the initial boundary time.

[0024] In step 1031, change point detection can be performed on each type of operational data in the multidimensional operational data using time series change point detection algorithms such as CUSUM (Cumulative Sum Control Chart) and / or PELT (Pruned Exact Linear Time).

[0025] For example, let's take the detection of a change point in the injection pressure data curve P in multidimensional operational data using the CUSUM algorithm as an example: Calculate the change in injection pressure between adjacent time points: ΔP k =P k -P k-1 ; Determine the positive cumulative change at time k: C k + =max(0,C k-1 + +ΔP k -μ p -h p ); Determine the negative cumulative change at time k: C k - =max(0,C k-1- -ΔP k +μ p -h p ); Where: μ p This represents the average variation in normal injection pressure; h p The allowed fluctuation term represents the normal fluctuation range that the algorithm tolerates; C k + Used to detect rapid rises in the injection pressure data curve; C k - Used to detect rapid decreases in the injection pressure data curve.

[0026] When C is satisfied k + >T p At that time, it is assumed that the injection pressure data curve shows an upward turning point near time k, when C is satisfied. k - >T p At that time, it was assumed that the injection pressure data curve showed a decreasing inflection point near time k, where T pAs a threshold for cumulative change, all rising and falling change points of the injection pressure data curve are defined as the set of change points of the injection pressure data curve.

[0027] In step 1032, for any initial boundary time bfsm, the corresponding search window W=[bfsm-δ,bfsm+δ], where δ is the preset duration.

[0028] In step 1033, the search window is used to perform a change point search for each type of running data in the multidimensional running data to obtain the candidate change point set corresponding to the initial boundary time bfsm.

[0029] In step 1034, for any candidate variable b in the candidate variable set corresponding to the initial boundary time bfsm, based on the principle of multi-channel consistency, the comprehensive confidence of candidate variable b based on multi-dimensional operational data is determined as Conf(b) = Σw. i ·C i (b), where w i w represents the weight corresponding to the i-th type of operational data in the multidimensional operational data. i The value of is in the range of 0 to 1, and satisfies Σw i =1, C i (b) represents the change confidence of the i-th type of running data at candidate change point b.

[0030] Taking multi-dimensional operating data, including holding pressure data curve, injection speed data curve, template position data curve, motor current data curve and action signals (including mold opening signal, holding pressure end signal, ejection signal, etc.), as an example, Conf(b)=wpCp(b)+wvCv(b)+wlCl(b)+wiCi(b)+waCa(b).

[0031] Wherein, Cp(b) is the confidence level of the change point of the holding pressure data curve at point b, Cv(b) is the confidence level of the change point of the injection speed data curve at point b, Cl(b) is the confidence level of the change point of the template position data curve at point b, Ci(b) is the confidence level of the change point of the motor current data curve at point b, and Ca(b) is the confidence level of the change point of the action signal at point b.

[0032] For example, the confidence level of a signal-type data curve, such as Ca(b), can be determined by the time distance between point b and the trigger time of the action signal. For the confidence level of other data curves at point b, such as Cp(b), Cv(b), Cl(b), and Ci(b), it can be obtained by linearly weighting or multiplying the changes in the mean and slope of the data within the search window corresponding to point b, and the probability that point b is a change point on that data curve (this probability can be output by time series change point detection algorithms).

[0033] wp, wv, wl, wi, and wa represent different channel weights. All weights wp, wv, wl, wi, and wa are non-negative, ranging from 0 to 1, and satisfying wp + wv + wl + wi + wa = 1. In a preferred embodiment, wp is 0.25, wv is 0.15, wl is 0.25, wi is 0.10, and wa is 0.25. In other embodiments, each weight can be adjusted according to the importance of different process phases. The PLC action signal weight wa is preferably not less than 0.10, and the sum of the pressure curve weight wp and the position curve weight wl is preferably not less than 0.30. When data for a certain channel is missing, the weight of that channel can be set to 0, and the remaining weights are normalized.

[0034] In step 1035, the candidate variable with the highest comprehensive confidence is selected from the candidate variable set corresponding to the initial boundary time bfsm as the refined boundary time corresponding to the initial boundary time bfsm.

[0035] At this point, the refinement boundary moments between all adjacent process steps can be obtained, thus obtaining the precise boundaries between all adjacent process steps.

[0036] Step 104: Using the fine-tuning boundary moment, the multidimensional operation data is divided into multidimensional operation sub-data corresponding to each process step in each molding cycle.

[0037] In this step, based on the finishing boundary time between adjacent process steps determined in the previous step, the multidimensional operation data is divided into multidimensional operation sub-data corresponding to each process step in each molding cycle.

[0038] Step 105: Construct a unique data fingerprint for each process step based on the multi-dimensional operational sub-data corresponding to the process step.

[0039] In step 105, for any process step, the multidimensional operation sub-data corresponding to the process step can first be time-aligned with the baseline template sub-data corresponding to the process step; then, based on the time-aligned multidimensional operation sub-data corresponding to the process step, a unique data fingerprint for the process step is constructed.

[0040] The duration of the same process step may differ in different molding cycles. For example, the injection process may last 1.2 seconds in one molding cycle and 1.5 seconds in another. Direct point-by-point comparison would cause data curve misalignment. Therefore, the Dynamic Time Warping (DTW) algorithm can be used to align the multi-dimensional running sub-data corresponding to each process step with the baseline template sub-data corresponding to that process step in time, as detailed below.

[0041] First, obtain the baseline template data for the molding cycle, which includes the baseline template sub-data for each process step. The baseline template sub-data for each process step includes normal operation data with the same characteristics as the multidimensional operation data.

[0042] For example, the baseline template data for the molding cycle is A={a1,a2,...,an}, the multidimensional running data for the current molding cycle is B={b1,b2,...,bm}, and the distance matrix is ​​D(i,j)=||ai-bj|| 2 Here, both ai and bj can be multi-channel vectors. For example, ai includes the injection pressure, holding pressure, injection speed, screw position, template position, motor current, clamping force, and ejection signal at time i, and bj includes the injection pressure, holding pressure, injection speed, screw position, template position, motor current, clamping force, and ejection signal at time j. Then, the minimum cumulative distance R(i,j) = D(i,j) + min{R(i-1,j),R(i,j-1),R(i-1,j-1)} is calculated. By backtracking the minimum path, the optimal alignment path from the reference template data B of the molding cycle to the multi-dimensional running data A of the current molding cycle is obtained. Optionally, to prevent overstretching or misalignment, a path constraint can be set: |ij| ≤ r, where r is the maximum allowable time offset range. Optionally, r can be 5%–20%, 8%–15%, 10%, etc., of the time length of the reference template data of the molding cycle. After time alignment is completed, the time-aligned multidimensional operation sub-data corresponding to each process step is uniformly converted into a fixed time length, such as 128 time points or 256 time points.

[0043] After time alignment, the next step is to construct a unique data fingerprint for each process step based on the time-aligned multidimensional operational sub-data corresponding to each process step.

[0044] In this embodiment, instead of uniformly extracting general features such as mean, variance, and maximum value from all operational data, a unique data fingerprint is constructed for each different process stage to reflect the operational data characteristics of each stage. Specifically, for any given process stage, feature data with pre-defined correlations (e.g., feature data related to the injection process stage, including injection pressure, injection speed, screw position, etc.) are selected from the time-aligned multi-dimensional operational sub-data corresponding to that stage, and a unique data fingerprint for each process stage is constructed based on this feature data. For example: 1. Specific data fingerprints for the injection process: The injection molding process uses a dedicated data fingerprint to characterize the pressure, velocity, and position response relationships of the melt as it enters the mold cavity. The injection pressure data curve is P(t), the injection velocity data curve is V(t), and the screw position data curve is L(t). Based on these, one or more of the following data are calculated: Peak injection pressure: F1 = max(P(t)), used to reflect injection resistance; The maximum injection pressure rise slope: F2=max(dP(t) / dt), is used to identify nozzle blockage, flow channel blockage, and increased melt flow resistance; Injection speed following error: F3 = Σ|Vset(t) - Vactual(t)| / N, used to reflect the response capability of the injection actuator, where Vset(t): the injection speed setpoint at time t, i.e., the target injection speed required by the injection molding machine in the process parameters; Vactual(t): the actual injection speed at time t, i.e., the actual injection speed fed back by the injection molding machine; |Vset(t) - Vactual(t)|: the absolute error between the set speed and the actual speed at time t; N: the number of sampling points involved in the calculation within the injection process interval; Injection endpoint position deviation: F4=|Lend-Lset|, used to determine whether the injection volume is stable; Injection pressure-injection speed hysteresis loop area: Construct a two-dimensional closed trajectory from curves P(t) and V(t), and calculate the area of ​​the closed loop: F5 = 1 / 2 × |Σ(P(k)V(k+1)-P(k+1)V(k))|. If the injection pressure increases but the injection speed decreases, this area will increase significantly, indicating abnormal resistance during the injection process; The pressure work per unit injection displacement is: F6=∫P(t)dL / |Lstart-Lend|, which is used to measure the pressure load required to complete a unit injection displacement.

[0045] For example, the dedicated data fingerprint for the injection process is: Finj=[F1,F2,F3,F4,F5,F6], which can be used to diagnose nozzle blockage, flow channel blockage, abnormal injection valve response, abnormal melt flow, and abnormal pressure sensing.

[0046] 2. Specific data fingerprint for the pressure holding process: The specific data fingerprint of the holding pressure process is used to characterize the pressure holding capability and melt feeding capability. The holding pressure data curve within the holding pressure process is Ph(t), from which one or more of the following data are calculated: Mean holding pressure: H1 = mean(Ph(t)); Pressure holding attenuation rate: H2=[Ph(t)] Ph_start )-Ph(t Ph_end)] / Δt, where Ph(t) Ph_start ) represents the start time t of the pressure holding process. Ph_start The holding pressure, Ph(t) Ph_end ) represents the end time t of the pressure holding process. Ph_end The holding pressure, Δt=t Ph_end -t Ph_start ; Pressure stability coefficient: The pressure holding process is divided into the first half and the second half. H3 = mean(Ph second half) / mean(Ph first half), where mean represents the average value (this meaning applies to all dedicated data fingerprints). The lower the H3, the lower the pressure holding capacity. Pressure fluctuation coefficient: H4=std(Ph(t)) / mean(Ph(t)), where std represents the standard deviation (this meaning applies to all dedicated data fingerprints); V / P switching point offset: H5 = |Tvp_current - Tvp_baseline|, where Tvp_current is the actual time point during the injection molding process when switching from the speed control stage (Velocity, V) to the pressure control / holding stage (Pressure, P), and Tvp_baseline is the reference time point during the injection molding process when switching from the speed control stage (Velocity, V) to the pressure control / holding stage (Pressure, P) (determined under normal and stable production conditions). Residual pressure ratio at the end of pressure holding: H6 = Ph(t) Ph_start ) / max(Ph(t)), where max represents the standard deviation (this meaning applies to all dedicated data fingerprints).

[0047] For example, the dedicated data fingerprint for the pressure holding process is: Fhold=[H1,H2,H3,H4,H5,H6], which can be used to diagnose issues such as check ring wear, pressure holding valve leakage, internal leakage in the hydraulic system, insufficient pressure replenishment in the mold cavity, and product shrinkage risk.

[0048] 3. Dedicated data fingerprints for the melt adhesive process: The dedicated data fingerprint of the melt adhesive process is used to characterize screw rotation, back pressure control, plasticizing resistance, and motor load. Within the melt adhesive process, the screw speed data curve is R(t), the motor current data curve is I(t), the back pressure data curve is B(t), the screw position data curve is L(t), and the barrel temperature data curve is T(t). Based on this, one or more of the following data are calculated: Melting time: M1=t me_end -t me_start , where t me_start t is the start time of the melt adhesive process. me_endThis marks the end of the melt adhesive process. Motor current-motor speed-load ratio: M2=mean(I(t)) / mean(R(t)). When the motor speed remains basically constant but the motor current increases, it indicates that the motor's mechanical resistance or plastic resistance has increased. Back pressure fluctuation index: M3 = std(B(t)) / mean(B(t)); Energy consumption per unit melt displacement: M4=∫I(t)dt / |L end -L start |, where L start L represents the starting position of the screw in the melt adhesive process. end This indicates the screw's end position in the melt adhesive process. Screw retraction speed stability coefficient: M5=std(dL / dt) / mean(dL / dt); Barrel temperature deviation index: M6=Σ|Tzone_i-Tset_i| / n, where, Tzone_i: the actual temperature of the i-th barrel temperature zone, such as the actual temperature of the 1st, 2nd and 3rd zones of the barrel; Tset_i: the set temperature of the i-th barrel temperature zone, i.e. the target barrel temperature set in the process parameters; n: the number of barrel temperature zones involved in the calculation.

[0049] For example, the dedicated data fingerprint for the melt adhesive process is: Fmelt=[M1,M2,M3,M4,M5,M6], which can be used to diagnose screw wear, poor plasticization, abnormal back pressure control, abnormal heating coil, abnormal motor load, and abnormal raw materials.

[0050] 4. Exclusive data fingerprints for the mold-making process: The dedicated data fingerprint of the mold opening process is used to characterize the mold platen movement state, mold opening resistance, mold opening positioning, mold opening stability, and hydraulic / servo load state of the injection molding machine during the mold opening stage. The mold platen position data curve is S(t), the mold opening speed data curve is V(t), the mold opening pressure data curve is P(t), the servo current or motor current data curve is I(t), and the mold opening control command curve is C(t). Based on this, one or more of the following data are calculated: Mold opening time: O1=t open_end -t open_start , where t open_start t is the start time of the mold opening process. open_end O1 represents the end time of the mold opening process and is used to characterize the overall time taken for the mold opening action. Mold opening position deviation: O2=|Send-Sset| / Sset, where Send is the actual position of the template at the end of mold opening, and Sset is the target mold opening position, used to characterize whether the mold opening is in place; Mold opening speed fluctuation index: O3=std(V(t)) / mean(V(t)), used to characterize the stability of mold opening speed; The mold opening pressure-speed load ratio O4 = mean(P(t)) / mean(V(t)) or the current-speed load ratio O4 = mean(I(t)) / mean(V(t)) is used to characterize the load level per unit mold opening speed. The peak load index during mold opening is O5 = max(P(t)) / mean(P(t)), or O5 = max(I(t)) / mean(I(t)), which is used to characterize the abnormal peak load during the mold opening process. The mold opening impact index O6=max(|dV(t) / dt|) is used to characterize the degree of sudden speed change or mechanical impact during the mold opening process. Mold opening hysteresis index O7 = Duration(V(t))<Vmin and P(t)> Pthreshold) / O1, or O7=Duration(V(t)) / O1.<Vmin and I(t)> Ithreshold) / O1 is used to characterize the continuous proportion of high load and low speed state, where Duration(·) represents the time when condition · is met, Vmin represents the minimum threshold of mold opening speed, Pthreshold represents the maximum threshold of mold opening pressure, and Ithreshold represents the maximum threshold of servo current or motor current. The mold opening command response deviation O8 = mean(|V(t)-Vcmd(t)|) / mean(Vcmd(t)) is used to characterize the response deviation between the actual mold opening speed V(t) and the target mold opening speed Vcmd(t).

[0051] For example, the dedicated data fingerprint for the mold opening process is represented as Fopen=[O1,O2,O3,O4,O5,O6,O7,O8]. Fopen can be used to identify problems such as incomplete mold opening, abnormal mold opening time, mold opening jamming, mold sticking, wear of guide pillars and bushings, insufficient lubrication, mold opening impact, abnormal hydraulic valve response, abnormal servo load, abnormal position sensor operation, and mismatched mold opening control parameters.

[0052] 5. Exclusive data fingerprint for the ejection process: The ejection process uses a dedicated data fingerprint to characterize the product's demolding resistance, ejector pin smoothness, and ejection mechanism load. The ejection current is Ie(t), the ejection pressure is Pe(t), and the ejection displacement is Le(t). Based on these, one or more of the following data are calculated: E1 = t ej_end -t ej_start , where t ej_startt is the start time of the ejection process. ej_end This marks the end of the ejection process. Peak output current: E2 = max(Ie(t)); Peak ejection pressure: E3 = max(Pe(t)); Ejection delay: E4 = Tarrive_current - Tarrive_baseline, where Tarrive_current: the actual time taken for the ejection mechanism to reach the target position / end point signal trigger during the current forming cycle; Tarrive_baseline: the baseline time taken for the ejection mechanism to reach the target position / end point signal trigger during normal and stable production. Ejection resistance index: E5=max(Ie(t))×E1 / |Le_end-Le_start|, where Le_start is the ejection displacement at the start of the ejection process and Le_end is the ejection displacement at the end of the ejection process. Ejection retraction anomaly index: E6 = Treturn_current / Treturn_baseline, where Treturn_current: the actual time taken for the ejection mechanism to retract from the start of retraction to the final position in the current forming cycle; Treturn_baseline: the baseline time required for the ejection mechanism to retract from the start of retraction to the final position under normal and stable production conditions. For example, the dedicated data fingerprint for the ejection process is: Feject=[E1,E2,E3,E4,E5,E6], which can be used to diagnose ejector pin jamming, product sticking to the mold, abnormal ejection cylinder, poor mold lubrication, and wear of the ejection mechanism.

[0053] Thus, a unique data fingerprint is constructed for each process step in each molding cycle, reflecting the unique data characteristics of each process step in each molding cycle.

[0054] Step 106: Determine whether there are any abnormalities in the process step based on the unique data fingerprint of the process step and the baseline data fingerprint of the process step.

[0055] Since there are few on-site fault samples for injection molding machines, the embodiments of this application do not rely on a large number of fault tags, but instead use data from normal molding cycles to establish a baseline data fingerprint for each process step.

[0056] In one implementation, this step can be achieved by steps 1061 to 1064.

[0057] Step 1061: Input the dedicated data fingerprint of the process step into the codec model to obtain the reconstructed data fingerprint of the process step. The reconstruction error between the dedicated data fingerprint of the process step and the reconstructed data fingerprint of the process step is determined as the reconstruction anomaly score of the process step. The codec model is trained based on the benchmark data fingerprint of the process step.

[0058] The codec model includes an encoder and a decoder. During training, the encoder is used to encode and map the unique data fingerprints (i.e., reference data fingerprints) of each process step in the normal molding cycle. The decoder is used to reconstruct the unique data fingerprints (i.e., reference data fingerprints) of each process step in the normal molding cycle based on the encoder output, so that the codec model learns the reconstruction pattern of the unique data fingerprints of each process step in the normal molding cycle.

[0059] Optionally, the encoder of the encoder-decoder model includes a TCN (Temporal Convolutional Network), and the decoder of the encoder-decoder model includes the decoder part of the Auto Encoder.

[0060] The training data for the codec model includes a benchmark data fingerprint sequence generated based on multidimensional operation data of K consecutive normal molding cycles: Fcycle_k=[Finj_k,Fhold_k,Fmelt_k,Feject_k,...]. For each process step in the K consecutive normal molding cycles, a training matrix Z={F1,F2,...,FK} is formed. For example, the benchmark data fingerprint subsequence corresponding to the injection process step in the K consecutive normal molding cycles is: Zinj=[Finj_1,Finj_2,...,Finj_K].

[0061] The encoder-decoder model extracts temporal dependencies from the training matrix through the encoder, thus mapping the training matrix to latent variables z = Encoder(Z). Taking TCN as the encoder as an example, TCN uses one-dimensional causal convolution and dilated convolution to extract temporal dependencies from the training data. The convolutional output of the Lth layer of TCN is: H L (t)=σ(ΣW L (i)×H L-1 (td L ×i)+b L ), where H L (t) represents the hidden features of the Lth layer at time t, σ is the activation function, and W L Let d be the weight of the convolutional kernel in layer L, i be the index of the position inside the convolutional kernel in layer L, and d be the weight of the convolutional kernel in layer L. L Let b be the void expansion coefficient of the Lth layer. LLet be the bias term of the Lth layer. By using different hole expansion coefficients d, the encoder-decoder model can simultaneously capture both short-period fluctuations and long-period trends of the training matrix.

[0062] The codec model reconstructs the training matrix based on the latent variable z using the decoder. The reconstructed matrix Z' = Decoder(z). The reconstruction error between the training matrix and the reconstructed matrix is ​​calculated as: Erec = ||Z - Z'|| 2 The encoder model is adjusted based on the reconstruction error.

[0063] During model inference, the specific data fingerprint Fcurrent of the current molding cycle's process step is input into the encoder-decoder model to obtain the reconstructed data fingerprint Fcurrent' of that process step. The reconstructing error Erec_current between the specific data fingerprint Fcurrent of that process step and the reconstructed data fingerprint Fcurrent' of that process step is calculated as ||Fcurrent - Fcurrent'||. 2 The reconfiguration anomaly score for this process step was determined.

[0064] The codec model is used to determine whether the unique data fingerprint of the process step in the current molding cycle is difficult to reconstruct from the normal mode. If the unique data fingerprint of the process step in the current molding cycle conforms to the normal mode, the codec model can reconstruct it well with a small reconstruction error. If the unique data fingerprint of the process step in the current molding cycle is abnormal and does not conform to the normal mode, the reconstruction error will increase significantly.

[0065] The output of this step is: the refactoring anomaly score Erec_current for each process step in the current molding cycle.

[0066] Step 1062: Determine the probability that the specific data fingerprint of the process step belongs to the baseline distribution, and determine the distribution deviation anomaly score of the process step based on the probability. The baseline distribution is obtained by fitting the baseline data fingerprint of the process step.

[0067] Reconstruction error is used to determine whether the process steps in the current molding cycle are difficult to reconstruct in the normal mode. However, reconstruction error alone is not enough. Therefore, this embodiment further introduces GMM (Gaussian Mixture Model) to calculate the distribution deviation anomaly score.

[0068] A GMM model is established based on the benchmark data fingerprint subsequences corresponding to the process steps in a normal molding cycle: p(F)=Σπ j ×N(F|μ j ,Σj), used to characterize the baseline distribution of the process steps, i.e., the normal data fingerprint distribution, where F is the data fingerprint, π jLet μ be the weight of the j-th Gaussian distribution. j Let Σj be the mean vector of the j-th Gaussian distribution, Σj be the covariance matrix of the j-th Gaussian distribution, and N be the Gaussian probability density function.

[0069] For the specific data fingerprint Fcurrent of the process step in the current molding cycle, calculate the probability p(Fcurrent) that it belongs to the baseline distribution of that process step. The deviation score of the distribution of that process step is Egmm_current = -logp(Fcurrent). When Fcurrent falls in the low probability region of the baseline distribution, Egmm_current increases, indicating that the process step deviates from the normal operating condition.

[0070] The output of this step is: the distribution deviation anomaly score Egmm_current for each process step in the current molding cycle.

[0071] Step 1063: Based on the fingerprint features of the dedicated data fingerprint of the process step, and the standardized deviation of the fingerprint feature mean and standard deviation of the fingerprint feature of the benchmark data fingerprint of the process step, determine the anomaly score of the key process feature of the process step.

[0072] Key process feature anomaly score Ekey=Σw c_i ×|f i -μ i | / σ i , where f i For the i-th feature data of the dedicated data fingerprint of the process step, μ i Let σ be the mean of the i-th feature data of the baseline data fingerprint of this process step over K consecutive normal molding cycles. i w represents the standard deviation of the i-th feature data of the baseline data fingerprint of this process step over K consecutive normal molding cycles. c_i The weight of the i-th feature data is the sum of the weights of all feature data in the same process step, Σw. c_i The value is 1.

[0073] The output of this step is: the distribution deviation anomaly score Ekey_current for each process step in the current molding cycle.

[0074] Step 1064: Determine whether there are any abnormalities in the process steps based on the reconfiguration anomaly score, distribution deviation anomaly score, and key process feature anomaly score of the process steps.

[0075] In this step, the reconstructed anomaly score, distribution deviation anomaly score, and key process feature anomaly score of the process link are weighted and summed to obtain the comprehensive anomaly score of the process link; based on the comprehensive anomaly score of the process link and the preset score threshold, it is determined whether there is an anomaly in the process link.

[0076] The overall anomaly score is: Score = α × norm(Erec) + β × norm(Egmm) + γ × norm(Ekey), where α, β, and γ are weighting coefficients, α + β + γ = 1, and optionally, α takes the value of 0.30 to 0.50, β takes the value of 0.20 to 0.40, and γ takes the value of 0.20 to 0.40; for example, α = 0.40, β = 0.30, γ = 0.30, and norm represents the normalization process.

[0077] The overall anomaly score for each process step in the current molding cycle is Score_current = α × norm(Erec_current) + β × norm(Egmm_current) + γ × norm(Ekey_current).

[0078] For example, for the injection process: Score_inj = αErec_inj + βEgmm_inj + γEkey_inj.

[0079] For example, for the pressure holding process: Score_hold = αErec_hold + βEgmm_hold + γEkey_hold.

[0080] The output of this step is: the comprehensive anomaly scores of different process stages in each molding cycle, such as Score_inj, Score_hold, Score_melt, and Score_eject.

[0081] In one implementation, determining whether a process step is abnormal based on a comprehensive abnormality score and a preset score threshold may specifically include: If the overall abnormal score of a process step exceeds a preset score threshold, the process step is determined to be abnormal. If the overall abnormal score of a process step does not exceed the preset score threshold, it is determined that no abnormality has occurred in that process step.

[0082] In one implementation, determining whether a process step is abnormal based on a comprehensive abnormality score and a preset score threshold may specifically include: Based on the comprehensive anomaly score of the process step, determine the index-weighted moving average comprehensive anomaly score of that process step; If the exponentially weighted moving average comprehensive abnormal score of a process step exceeds the preset score threshold corresponding to that process step in N consecutive molding cycles, the process step is determined to be abnormal, where N is the preset value. If the index-weighted moving average comprehensive abnormal score of a process step does not exceed the preset score threshold corresponding to that process step in multiple consecutive molding cycles, it is determined that no abnormality has occurred in that process step.

[0083] Injection molding processes are subject to occasional disturbances. If an alarm is triggered for an anomaly in a single molding cycle, false alarms are likely. Therefore, this implementation uses an EWMA (Exponentially Weighted Moving Averages) model to confirm the trend of the comprehensive anomaly score in consecutive molding cycles.

[0084] In this embodiment, the EWMA model assigns a comprehensive anomaly score to a certain process step in the t-th molding cycle. t Perform the following calculations: EWMA t =λ×Score t +(1-λ)×EWMA t-1 Among them, EWMA t The exponentially weighted moving average comprehensive anomaly score of this process step in the t-th molding cycle is given, where λ is a smoothing coefficient, 0 < λ ≤ 1, optionally λ is 0.1 to 0.5, optionally λ is 0.2 to 0.4, and for example λ is 0.3. (EWMA) t-1 This is the exponentially weighted moving average comprehensive anomaly score of this process step in the (t-1)th molding cycle.

[0085] For a certain process step, if EWMA t If the threshold value is 'Threshold_phase' (which is the preset threshold value corresponding to the phase of this process step), and this condition is met for N consecutive molding cycles, then the process step is determined to be abnormal.

[0086] For example, if EWMA occurs for five consecutive molding cycles in the injection molding process... t If Threshold_inj exceeds the threshold, an abnormality is confirmed in the injection molding process. If only a single molding cycle meets EWMA... t If the threshold is set to Threshold_inj, but the subsequent molding cycle does not meet the threshold, then it is determined that there is no abnormality in the injection process, thereby filtering out occasional abnormalities and reducing the false alarm rate.

[0087] The output of this step is: the abnormal process step and the number of abnormal cycles.

[0088] This allows for the precise identification of the specific process step where an abnormality occurs during the injection molding process.

[0089] In an optional implementation, after identifying the abnormal process step, the intelligent fault diagnosis method for injection molding can further perform a detailed diagnosis of the injection molding fault through the following steps 107 to 110, that is, determine the suspected location and / or suspected cause of the abnormality.

[0090] Step 107: For the injection molding machine performing the injection molding process, obtain the fault template corresponding to each fine fault of the injection molding machine. The fault template includes the abnormal process link of the corresponding fine fault, the abnormal feature data in the exclusive data fingerprint of the abnormal process link, the change trend of the abnormal feature data in the continuous molding cycle, and the importance weight of each feature data in the exclusive data fingerprint of the abnormal process link for determining the fine fault.

[0091] For each fine fault of the injection molding machine, a corresponding fault template is established: Fault_j={phase,feature_set,trend_set,weight_set}, where phase is the abnormal process step of fine fault j, feature_set is the abnormal feature data in the dedicated data fingerprint of the abnormal process step, trend_set is the changing trend of the abnormal feature data in the continuous molding cycle, and weight_set is the importance weight of each feature data in the dedicated data fingerprint of the abnormal process step for determining the fine fault.

[0092] Step 108: Determine the matching degree between the multi-dimensional operation sub-data corresponding to each process step in the continuous molding cycle and the fault template corresponding to each fine fault.

[0093] For the multi-dimensional operational sub-data corresponding to process step i within a continuous molding cycle, the matching degree between it and the fault template corresponding to each fine fault is calculated using the following formula: Match j =Σw d_i ×g i (f d_i ), where Match j f represents the matching degree between the multi-dimensional operational sub-data corresponding to process step i within a continuous molding cycle and the fault template corresponding to the fine fault j. d_i For the abnormal feature data in the dedicated data fingerprint of process step i, g i (·) is a scoring function for determining whether anomalous feature data is anomalous, w d_i represents the weight of anomalous feature data in fine-grained fault j.

[0094] Among them, g i (f d_i This is used to calculate the degree of matching between the current feature data and the corresponding abnormal feature data in the target fault template. First, it is based on the mean μ of the normal samples of this feature data. i and standard deviation σi Calculate the standardized deviation z i =(f d_i -μ i ) / (σ i +ε), where ε is a non-negative constant (e.g., 10). -6 10 -8 To prevent the denominator from being zero, the abnormal direction coefficient s is preset in the target fault template. i Calculate the deviation intensity d i When d i Less than the abnormal start threshold θ i At that time, g i (f d_i )=0; when d i Greater than the abnormal saturation threshold ρ i At that time, g i (f d_i )=1; when d i Located at θ i With ρ i Between, g i (f d_i The value increases linearly. This scoring function characterizes the degree to which the current feature data supports the fine-grained faults of the target.

[0095] Step 109: Normalize the matching degree between the multi-dimensional operation sub-data corresponding to the process links in the continuous molding cycle and the fault template corresponding to each fine fault to obtain the fault confidence of each fine fault currently occurring.

[0096] The current fault confidence for a fine fault j is: Conf j =Match j / ΣMatch j .

[0097] Step 110: Identify the fine fault with the highest fault confidence as the current fault.

[0098] The fine fault with the highest fault confidence is selected as the current fault, that is, the current fault is suspected to be this fine fault.

[0099] For example: Rising anomaly function: Zup(x) = max(0, (x-μ) x ) / (σ x +ε)) indicates that the current feature data x is higher than the mean μ of x under normal conditions. x The more abnormal points, the higher the anomaly score, σ x Let x be the standard deviation of x under normal conditions, and ε be a constant to ensure that the denominator is not zero.

[0100] Downward anomaly function: Zdown(x) = max(0, (μ x -x) / (σ x +ε)) indicates that the current feature data x is lower than the mean μ of x under normal conditions. x The more abnormal points, the higher the anomaly score, σ x Let x be the standard deviation of x under normal conditions, and ε be a constant to ensure that the denominator is not zero.

[0101] Matching degree for "nozzle blockage / flow channel blockage" fault: Match_nozzle=w n1 ×Zup(F1)+w n2 ×Zup(F2)+w n3 ×Zup(F5)+w n4 ×Zup(F6)+w n5 ×Zdown(F3).

[0102] Matching degree for "Check ring wear / pressure holding leakage" fault: Match_checkring=w c1 ×Zup(H2)+w c2 ×Zdown(H3)+w c3 ×Zdown(H6)+w c4 ×Zup(H5)+w c5 ×Zup(H4).

[0103] Matching degree for the "screw wear / abnormal plasticization" fault: Match_screw=w s1 ×Zup(M1)+w s2 ×Zup(M2)+w s3 ×Zup(M4)+w s4 ×Zup(M3)+w s5 ×Zup(M6).

[0104] Matching degree for "ejector jamming / product sticking to mold" fault: Match_eject=w e1 ×Zup(E1)+w e2 ×Zup(E2)+w e3 ×Zup(E3)+w e4 ×Zup(E4)+w e5 ×Zup(E5)+w e6 ×Zup(E6).

[0105] Detailed fault diagnosis methods: When Match_nozzle is at its highest, the output is: Suspected nozzle blockage or flow channel blockage; When Match_checkring is at its highest, the output is: suspected check ring wear, pressure holding valve leakage, or internal hydraulic leakage. When Match_screw is at its highest, the output will be: suspected screw wear, poor plasticization, abnormal back pressure control, or abnormal raw materials. When Match_eject is at its highest, the output is: suspected ejector pin jamming, product sticking to the mold, or abnormal ejection mechanism.

[0106] The following provides some alternative solutions for the exemplary embodiments of this application.

[0107] 1. Alternative solutions for process boundary identification: The exemplary embodiments of this application employ a finite state machine + CUSUM / PELT change point detection algorithm to identify process boundaries in the injection molding cycle. In other embodiments, Bayesian change point detection, sliding window slope detection, wavelet transform, hidden Markov model, or LSTM (Long Short-Term Memory) / Transformer model can also be used instead.

[0108] 2. Alternatives to running data curve alignment: The exemplary embodiments of this application employ the DTW (Dynamic Time Warping) algorithm to align the process data curves of process steps with inconsistent lengths in different molding cycles. In other embodiments, piecewise linear interpolation, key action point anchoring alignment, ShapeDTW, Soft-DTW, or attention-based temporal alignment models can also be used instead.

[0109] 3. Alternatives to normal data fingerprinting: The exemplary embodiments of this application employ TCN-AutoEncoder to learn data fingerprints under normal production conditions and identify anomalies through reconstruction errors. In other embodiments, models such as LSTM-AutoEncoder, GRU (Gated Recurrent Unit)-AutoEncoder, Transformer-AutoEncoder, VAE (Variational Autoencoder), DeepSVDD (Deep Support Vector Data Description), One-ClassSVM (One-Class Support Vector Machine), or IsolationForest can be used instead.

[0110] 4. Alternative methods for calculating distribution deviation outlier scores: The exemplary embodiments of this application employ a Gaussian Mixture Model (GMM) to calculate the deviation of the current proprietary data fingerprint from the normal data fingerprint distribution. In other embodiments, alternative methods such as Kernel Density Estimation (KDE), Mahalanobis distance, Local Outlier Factor (LOF), DBSCAN (Density-Based Spatial Clustering of Applications with Noise) outlier detection, Gaussian process models, or One-Class SVM boundary detection can be used.

[0111] 5. Alternative solutions for confirming continuous periodic anomalies: The exemplary embodiments of this application employ the EWMA (Exponentially Weighted Moving Average) algorithm to confirm trends in abnormal scores across multiple consecutive forming cycles. In other embodiments, alternative methods may be used, such as sliding window mean, N consecutive threshold exceedances, CUSUM trend accumulation, moving median filtering, Kalman filtering, or Bayesian online anomaly probability updates.

[0112] 6. Alternative solutions for diagnosing faulty components: The exemplary embodiments of this application perform detailed diagnosis of suspected faults based on abnormal processes and abnormal fingerprint feature data. In other embodiments, decision trees, random forests, XGBoost (eXtreme Gradient Boosting), Bayesian networks, rule engines, knowledge graph reasoning, multi-class neural networks, or fuzzy comprehensive evaluation models can also be used as alternatives.

[0113] 7. Alternatives to fingerprint feature data: The exemplary embodiments of this application use fingerprint feature data such as pressure-velocity hysteresis loop area, pressure holding stability coefficient, current-speed load ratio, and ejection resistance index. In other embodiments, fingerprint feature data such as pressure integral area, current integral energy consumption, speed response delay, clamping force establishment slope, back pressure stability coefficient, and template position response error can be added or replaced according to different injection molding machine types, sensor configurations, and process scenarios. As long as these feature data can be used to characterize the operating status of the corresponding process steps and participate in anomaly detection and fine fault diagnosis, they can all be used as alternatives or extensions to fingerprint feature data.

[0114] The exemplary embodiments of this application can achieve the following technical effects.

[0115] 1. Improve the accuracy of process boundary identification: In some exemplary embodiments of this application, a finite state machine combined with a time-series change-point detection algorithm is used to achieve initial identification and fine-tuning of the boundaries of process stages such as mold closing, injection, pressure holding, melting, mold opening, and ejection within the injection molding cycle, thereby accurately locating the boundaries of process stages. Compared to fixed time window segmentation, some exemplary embodiments of this application can accurately locate the boundaries of process stages based on the actual changes in multi-channel data such as pressure, speed, position, current, and action signals, achieving stage-level positioning, avoiding data mixing between different process stages, and improving the accuracy of fault diagnosis.

[0116] 2. Improve the comparability of data between different molding cycles: The duration of each process step in an injection molding machine varies depending on the mold, material, product, and cycle time. In some exemplary embodiments of this application, the DTW (Dynamic Time Warping) algorithm is used to align the data curves of the same process steps in different molding cycles. This allows for a unified comparison of data curves for injection, holding pressure, melting, and ejection, which have different lengths and cycle times, reducing misjudgments caused by variations in molding cycle length.

[0117] 3. Improve the ability to express fault characteristics: In some exemplary embodiments of this application, instead of simply extracting general features such as mean, maximum value, and variance, specific data fingerprints are constructed for different process stages. These specific data fingerprints can more accurately reflect the unique operating status of the corresponding process stage and improve the identification of fault characteristics.

[0118] 4. Reduce reliance on a large number of fault samples: Injection molding machine sites typically have abundant normal data but few fault samples, making it difficult for traditional supervised learning models to obtain sufficient labeled samples. In some exemplary embodiments of this application, an encoder-decoder model learns the fingerprint features of different process steps under normal production conditions, and identifies abnormal states by reconstructing errors. This approach does not rely on a large amount of fault label data and can achieve fault diagnosis even in scenarios with insufficient fault samples.

[0119] 5. Improve the accuracy and robustness of anomaly detection: In some exemplary embodiments of this application, the reconstruction anomaly score, distribution deviation anomaly score, and key process feature anomaly score are fused to form a comprehensive anomaly score. Specifically, the codec model is used to identify timing pattern anomalies, the baseline distribution model is used to identify deviations from normal distributions, and the fingerprint feature standardization deviation is used to identify core feature anomalies. The combined judgment of multiple anomaly information reduces the problem of single model failure or misjudgment, improving the stability of fault diagnosis.

[0120] 6. Reduce false alarms caused by single-cycle fluctuations: During injection molding, there may be fluctuations in raw materials, temperature disturbances, instantaneous pressure fluctuations, or occasional delays in action. In some exemplary embodiments of this application, the EWMA exponentially weighted moving average algorithm is used to confirm the trend of the comprehensive anomaly score of multiple consecutive molding cycles. Fault diagnosis is only triggered when the comprehensive anomaly score of a certain process step continuously exceeds a threshold, thereby filtering out occasional anomalies and reducing the false alarm rate.

[0121] 7. Achieve detailed fault type diagnosis: Traditional injection molding machine alarms typically only indicate an abnormality in a single parameter, making it difficult to directly pinpoint the source or location of the fault. In some exemplary embodiments of this application, based on abnormal process steps and abnormal fingerprint feature data, fault diagnosis is upgraded from a machine-wide alarm level to a three-tiered alarm system: abnormal process step identification → abnormal fingerprint analysis → suspected faulty component / cause location.

[0122] 8. Improve the intelligent operation and maintenance capabilities of injection molding machines: In some exemplary embodiments of this application, online analysis of continuous operation data of injection molding machines can be performed to promptly detect early abnormal trends, assist maintenance personnel in assessing fault risks in advance, reduce downtime for troubleshooting, and improve equipment maintenance efficiency. Therefore, some exemplary embodiments of this application can upgrade injection molding machine fault diagnosis from traditional manual experience-based judgment and threshold alarms to automated, process-level, and component-level intelligent diagnosis based on artificial intelligence-based process fingerprinting and multi-channel time-series fusion.

[0123] Based on the same inventive concept, this application also provides an intelligent fault diagnosis device for injection molding processes to implement the intelligent fault diagnosis method for injection molding processes described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent fault diagnosis device for injection molding processes provided below can be found in the limitations of the intelligent fault diagnosis method for injection molding processes described above, and will not be repeated here.

[0124] In one exemplary embodiment, such as Figure 2 As shown, an intelligent fault diagnosis device for injection molding process is provided. The intelligent fault diagnosis device for injection molding process includes: The data acquisition module is used to acquire multi-dimensional operational data during the injection molding process; the injection molding process includes multiple molding cycles, and each molding cycle includes multiple process steps. The boundary determination module is used to determine the initial boundary moments between adjacent process steps in each molding cycle of the injection molding process based on multi-dimensional operational data. The boundary refinement module is used to refine the initial boundary time based on the variable points in the multidimensional running data, so as to obtain the refined boundary time. The data partitioning module is used to divide the multidimensional operational data into multidimensional operational sub-data corresponding to each process step in each molding cycle by using the fine-tuning boundary moment. The fingerprint construction module is used to construct a unique data fingerprint for each process step based on the multi-dimensional operational sub-data corresponding to each process step. The anomaly detection module is used to determine whether there are any anomalies in a process step based on the process step's unique data fingerprint and the process step's baseline data fingerprint.

[0125] Optionally, based on the unique data fingerprint of each process step and the baseline data fingerprint of each process step, it can be determined whether there are any anomalies in the process step, including: The specific data fingerprint of the process step is input into the codec model to obtain the reconstructed data fingerprint of the process step. The reconstruction error between the specific data fingerprint of the process step and the reconstructed data fingerprint of the process step is determined as the reconstruction anomaly score of the process step. The codec model is trained based on the benchmark data fingerprint of the process step. Determine the probability that the specific data fingerprint of a process step belongs to the baseline distribution, and determine the distribution deviation anomaly score of the process step based on the probability. The baseline distribution is obtained by fitting the baseline data fingerprint of the process step. Based on the characteristic data of the dedicated data fingerprint of the process step, and the standardized deviation of the fingerprint feature mean and standard deviation of the fingerprint feature based on the benchmark data fingerprint of the process step, the anomaly score of the key process feature of the process step is determined. Based on the reconfiguration anomaly score, distribution deviation anomaly score, and key process characteristic anomaly score of the process steps, it is determined whether there are any anomalies in the process steps.

[0126] Optionally, based on the reconfiguration anomaly score, distribution deviation anomaly score, and key process characteristic anomaly score of the process step, it can be determined whether there is anomaly in the process step, including: The reconfiguration anomaly score, distribution deviation anomaly score, and key process feature anomaly score of the process are weighted and summed to obtain the comprehensive anomaly score of the process. Based on the overall abnormality score of the process step and the preset score threshold, it is determined whether there is an abnormality in the process step.

[0127] Optionally, based on the overall anomaly score of the process step and a preset score threshold, it can be determined whether an anomaly exists in the process step, including: Based on the comprehensive anomaly score of each process step, determine the index-weighted moving average comprehensive anomaly score of that process step. If the index-weighted moving average comprehensive abnormal score of a process step exceeds the preset score threshold corresponding to the process step in N consecutive molding cycles, the process step is determined to be abnormal, where N is the preset value. If the index-weighted moving average comprehensive abnormal score of the process step does not exceed the preset score threshold corresponding to the process step in multiple consecutive molding cycles, it is determined that no abnormality has occurred in the process step.

[0128] Optionally, the intelligent fault diagnosis device for injection molding processes is also used for: For an injection molding machine performing an injection molding process, obtain a fault template corresponding to each fine fault of the injection molding machine. The fault template includes the abnormal process step of the corresponding fine fault, the abnormal feature data in the exclusive data fingerprint of the abnormal process step, the change trend of the abnormal feature data in the continuous molding cycle, and the importance weight of each feature data in the exclusive data fingerprint of the abnormal process step for determining the fine fault. Determine the matching degree between the multi-dimensional operation sub-data corresponding to each process step in the continuous molding cycle and the fault template corresponding to each fine fault; The matching degree between the multi-dimensional operation sub-data corresponding to the process steps in the continuous molding cycle and the fault template corresponding to each fine fault is normalized to obtain the fault confidence of each fine fault currently occurring. The fine-grained fault with the highest fault confidence is identified as the current fault.

[0129] Optionally, based on the changing points in the multidimensional runtime data, the initial boundary time is refined to obtain the refined boundary time, including: The time series change point detection algorithm is used to detect change points in multidimensional running data; Determine the search window with the initial boundary time as the midpoint; Search for change points in the search window to obtain the set of candidate change points corresponding to the initial boundary time. For each candidate variable in the candidate variable set corresponding to the initial boundary time, determine the comprehensive confidence level of the candidate variable based on multidimensional running data; The candidate change point with the highest overall confidence is selected from the set of candidate change points corresponding to the initial boundary time as the refined boundary time corresponding to the initial boundary time.

[0130] Optionally, based on multidimensional operational data, the initial boundary moments between adjacent process steps in each molding cycle of the injection molding process are determined, including: Obtain a finite state machine, wherein the finite state machine includes trigger events for transitioning from one process stage to the next process stage, and the trigger events are set based on changes in multidimensional operational data; Based on a finite state machine, during the injection molding process, the target trigger event of jumping from the first process stage to the second process stage is monitored. Here, the first process stage is any process stage, and the second process stage is the next process stage after the first process stage. If a target triggering event is detected, the time of occurrence of the target triggering event is determined as the initial boundary time between the first process stage and the second process stage.

[0131] Optionally, based on the multi-dimensional operational sub-data corresponding to each process step, a unique data fingerprint for that process step can be constructed, including: Time alignment is performed between the multi-dimensional operational sub-data corresponding to the process step and the benchmark template sub-data corresponding to the process step. Based on the time-aligned multi-dimensional operational sub-data corresponding to the process steps, construct a unique data fingerprint for each process step.

[0132] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0133] The aforementioned computer device may be, for example, a server or a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements an intelligent fault diagnosis method for injection molding processes.

[0134] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0135] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0136] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0137] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0138] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0139] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0140] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for intelligent fault diagnosis in injection molding process, characterized in that, The intelligent fault diagnosis method for injection molding process includes: During the injection molding process, multi-dimensional operational data is acquired; the injection molding process includes multiple molding cycles, and each molding cycle includes multiple process steps. Based on the multidimensional operational data, the initial boundary time between adjacent process steps in each molding cycle of the injection molding process is determined. Based on the changing points in the multidimensional operational data, the initial boundary time is refined to obtain the refined boundary time. The multidimensional operational data is divided into multidimensional operational sub-data corresponding to each process step in each molding cycle using the fine-tuning boundary moment. Based on the multi-dimensional operational sub-data corresponding to the process step, construct the exclusive data fingerprint of the process step; Based on the unique data fingerprint of the process step and the baseline data fingerprint of the process step, it is determined whether there is an anomaly in the process step.

2. The intelligent fault diagnosis method for injection molding process according to claim 1, characterized in that, The step of determining whether there is an anomaly in the process step based on the unique data fingerprint of the process step and the baseline data fingerprint of the process step includes: The dedicated data fingerprint of the process step is input into the codec model to obtain the reconstructed data fingerprint of the process step. The reconstruction error between the dedicated data fingerprint of the process step and the reconstructed data fingerprint of the process step is determined as the reconstruction anomaly score of the process step. The codec model is trained based on the baseline data fingerprint of the process step. The probability that the specific data fingerprint of the process step belongs to the baseline distribution is determined, and the distribution deviation anomaly score of the process step is determined based on the probability, wherein the baseline distribution is obtained by fitting the baseline data fingerprint of the process step; Based on the characteristic data of the dedicated data fingerprint of the process step, and the standardized deviation of the fingerprint feature mean and fingerprint feature standard deviation of the baseline data fingerprint of the process step, the critical process feature anomaly score of the process step is determined. Based on the reconstructed anomaly score, the distribution deviation anomaly score, and the key process feature anomaly score of the process step, it is determined whether there is an anomaly in the process step.

3. The intelligent fault diagnosis method for injection molding process according to claim 2, characterized in that, The step of determining whether there is an anomaly in the process step based on the reconstruction anomaly score, the distribution deviation anomaly score, and the key process feature anomaly score includes: The reconstruction anomaly score, the distribution deviation anomaly score, and the key process feature anomaly score of the process step are weighted and summed to obtain the comprehensive anomaly score of the process step. Based on the comprehensive abnormality score of the process step and the preset score threshold, it is determined whether there is an abnormality in the process step.

4. The intelligent fault diagnosis method for injection molding process according to claim 3, characterized in that, The step of determining whether a process step is abnormal based on the comprehensive abnormality score of the process step and a preset score threshold includes: Based on the comprehensive anomaly score of the process step, determine the exponentially weighted moving average comprehensive anomaly score of the process step; If the exponentially weighted moving average comprehensive abnormal score of the process step exceeds the preset score threshold corresponding to the process step in N consecutive molding cycles, the process step is determined to be abnormal, where N is a preset value. If the exponentially weighted moving average comprehensive abnormal score of the process step does not exceed the preset score threshold corresponding to the process step in multiple consecutive molding cycles, it is determined that no abnormality has occurred in the process step.

5. The intelligent fault diagnosis method for injection molding process according to claim 1, characterized in that, The intelligent fault diagnosis method for injection molding process also includes: For an injection molding machine performing the injection molding process, obtain a fault template corresponding to each fine fault of the injection molding machine. The fault template includes the abnormal process step corresponding to the fine fault, the abnormal feature data in the exclusive data fingerprint of the abnormal process step, the change trend of the abnormal feature data in the continuous molding cycle, and the importance weight of each feature data in the exclusive data fingerprint of the abnormal process step for determining the fine fault. Determine the matching degree between the multi-dimensional operation sub-data corresponding to the process steps within the continuous molding cycle and the fault template corresponding to each fine fault; The matching degree between the multi-dimensional operation sub-data corresponding to the process steps in the continuous molding cycle and the fault template corresponding to each fine fault is normalized to obtain the fault confidence of each fine fault currently occurring. The fine-grained fault with the highest fault confidence is identified as the current fault.

6. The intelligent fault diagnosis method for injection molding process according to claim 1, characterized in that, The step of refining the initial boundary time based on the changing points in the multidimensional operational data to obtain the refined boundary time includes: The changing points in the multidimensional operational data are detected by a time series change point detection algorithm; Determine a search window with the initial boundary time as its midpoint; Search the variable point in the search window to obtain the set of candidate variable points corresponding to the initial boundary time. For each candidate variable point in the candidate variable point set corresponding to the initial boundary time, determine the comprehensive confidence level of the candidate variable point based on the multidimensional running data; The candidate change point with the highest overall confidence is selected from the set of candidate change points corresponding to the initial boundary time as the refined boundary time corresponding to the initial boundary time.

7. The intelligent fault diagnosis method for injection molding process according to claim 1, characterized in that, The determination of the initial boundary time between adjacent process steps in each molding cycle of the injection molding process based on the multidimensional operational data includes: Obtain a finite state machine, wherein the finite state machine includes a trigger event for transitioning from one process stage to the next process stage, the trigger event being set based on changes in the multidimensional operating data; Based on the finite state machine, during the injection molding process, a target trigger event is monitored to jump from the first process stage to the second process stage, wherein the first process stage is any one of the process stages, and the second process stage is the next process stage after the first process stage. If the target triggering event is detected, the time of occurrence of the target triggering event is determined as the initial boundary time between the first process step and the second process step.

8. The intelligent fault diagnosis method for injection molding process according to claim 1, characterized in that, The step of constructing a unique data fingerprint for each process step based on the multi-dimensional operational sub-data corresponding to that process step includes: Time-align the multi-dimensional operational sub-data corresponding to the process step with the baseline template sub-data corresponding to the process step. Based on the time-aligned multidimensional operational sub-data corresponding to the process step, a unique data fingerprint for the process step is constructed.

9. An intelligent fault diagnosis device for injection molding process, characterized in that, The intelligent fault diagnosis device for the injection molding process includes: The data acquisition module is used to acquire multi-dimensional operational data during the injection molding process; the injection molding process includes multiple molding cycles, and the molding cycle includes multiple process steps. A boundary determination module is used to determine the initial boundary time between adjacent process steps in each molding cycle of the injection molding process based on the multidimensional operating data. The boundary refinement module is used to refine the initial boundary time based on the variable points in the multidimensional running data to obtain the refined boundary time. The data partitioning module is used to divide the multidimensional operational data into multidimensional operational sub-data corresponding to each process step in each molding cycle using the fine-tuning boundary moment. The fingerprint construction module is used to construct a unique data fingerprint for the process step based on the multi-dimensional operational sub-data corresponding to the process step. The anomaly detection module is used to determine whether there is an anomaly in the process step based on the unique data fingerprint of the process step and the baseline data fingerprint of the process step.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the intelligent fault diagnosis method for injection molding process according to any one of claims 1-8.

Images