Machine Vision-Based Nozzle Clog Recognition Method and System

Abstract

This invention provides a machine vision-based nozzle clogging identification method and system, belonging to the field of image data analysis and processing technology. It employs multi-nozzle thermal vision data, deformation vision data, and acoustic data to collect and extract multi-physics dynamic features, monitoring the overall state of material flow, thermodynamic behavior, and structural changes within the nozzle. Thermal vision data analyzes heat propagation anomalies through temperature field sequences, deformation vision data captures structural responses through surface deformation sequences, and acoustic data identifies material flow obstacles through vibration spectra, improving detection accuracy. A multi-modal temporal fusion network is constructed for deep learning and fusion analysis of multi-physics dynamic features. Based on the root cause and mechanism type of clogging, clogging risk levels are classified and self-healing control strategies are implemented, achieving accurate identification and judgment of minor and initial clogging, improving clogging monitoring accuracy, reducing production interruptions caused by false alarms, and enhancing the efficiency and economy of the injection molding process.

Description

Technical Field

[0001] This invention relates to the field of image data analysis and processing technology, specifically to a nozzle clogging identification method and system based on machine vision. Background Technology

[0002] In injection molding hot runner systems, leakage at the connection between the manifold and the hot runner nozzle, as well as the issue of nozzle breakage during maintenance, have long been persistent technical challenges. While replaceable nozzle assembly designs have reduced maintenance costs due to breakage to some extent, these mechanical improvements have not solved another core problem: how to detect initial blockage defects inside the nozzle in a timely and accurate manner. Existing technologies generally employ methods such as manual experience judgment, sensor monitoring, and machine vision image recognition for nozzle blockage detection. Manual inspection indirectly judges nozzle blockage by observing defects in the injection molded product, which is not only inefficient but also wastes injection molded products. Sensor monitoring mainly uses pressure sensors to detect abnormal pressure in the runner and temperature sensors to detect temperature changes, but these are easily affected by factors such as changes in raw material viscosity and temperature, resulting in insufficient detection accuracy. Traditional machine vision image recognition involves comparing nozzle images after acquisition, generally through threshold segmentation or region detection. However, due to factors such as uneven lighting, plastic residue interference, and complex image backgrounds during nozzle image acquisition, the blockage recognition rate is low and the false alarm rate is high. In summary, existing technologies for nozzle blockage identification have the following technical problems:

[0003] Problem 1: Manual identification is highly subjective and cannot detect early blockages in real time, resulting in waste of finished products. Sensor and machine vision image recognition are easily affected by external environmental interference, leading to insufficient recognition accuracy and a high false alarm rate, which is not conducive to the accurate identification and monitoring of nozzle blockages.

[0004] The second problem is that existing nozzle clogging identification methods often rely on static image recognition and comparison, focusing solely on the geometric changes inside the nozzle. They fail to perceive the impact of differences in internal material flow and thermodynamic behavior on nozzle clogging. Furthermore, clogging issues are identified only after they have occurred, lacking the ability to dynamically monitor and warn of the early clogging process. This passive response cannot meet the usage requirements. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: a nozzle clogging identification method based on machine vision, the method comprising:

[0006] Acquire processing parameters and multimodal sensing data of the nozzle, wherein the multimodal sensing data includes thermal visual data, deformation visual data and acoustic data;

[0007] Multi-physics dynamic features are extracted from multimodal sensing data, including thermal gradient propagation features, deformation recovery features, and acoustic vibration features.

[0008] A causal network for blockage risk is constructed based on the dynamic characteristics of multiphysics fields, and the root causes and mechanism types of blockage are identified based on the causal network for blockage risk.

[0009] A multimodal time-series fusion network is constructed based on multimodal sensor data. Risk assessment and prediction are performed based on the multimodal time-series fusion network to generate congestion risk levels and predictive maintenance decisions.

[0010] Obtain the blockage risk level for graded response, select self-healing control strategies based on the root cause and mechanism type of blockage, and adjust the processing parameters based on the self-healing control strategies.

[0011] Furthermore, the acquisition of processing parameters and multimodal sensing data of the nozzle includes:

[0012] Acquire thermal visual data, deformation visual data and acoustic data of the nozzle during the processing, and establish spatial coordinate mapping relationship after time stamp synchronization to form multimodal sensing data;

[0013] The thermal visual data includes temperature field distribution data and heat flow propagation sequence data; the deformation visual data includes deformation sequence data and optical flow field data; and the acoustic data includes vibration spectrum data and sound wave propagation data.

[0014] Furthermore, the extraction of multi-physics dynamic features from multimodal sensing data includes:

[0015] Extract thermal gradient propagation features from thermal visual data, extract deformation recovery features from deformation visual data, and extract acoustic vibration features from acoustic data;

[0016] Multi-scale spatiotemporal characteristic analysis was performed on the thermal gradient propagation characteristics, deformation recovery characteristics, and acoustic vibration characteristics to form multi-physics field dynamic characteristics.

[0017] Furthermore, the construction of the causal network for congestion risk based on the dynamic characteristics of multiphysics includes:

[0018] Processing parameters and multiphysics dynamic characteristics are obtained as network nodes. The causal relationships between network nodes are analyzed to establish directed edges. Connection weights are set based on the directed edges. A causal network of blockage risk is constructed based on network nodes, connection weights, and directed edges.

[0019] Analyze the impact of each network node on congestion risk, and analyze the root causes of congestion based on connection weights and impact strength;

[0020] The mechanism type is determined based on the root cause of the blockage, and the mechanism types include material degradation type, foreign object blockage type, thermal runaway type, and mechanical wear type.

[0021] Furthermore, the construction of a multimodal temporal fusion network based on multimodal sensor data, and the risk assessment and prediction based on the multimodal temporal fusion network to generate congestion risk levels and predictive maintenance decisions, includes:

[0022] Thermodynamic features are extracted from thermal visual data and acoustic data, and a thermodynamic branch network is constructed based on the thermodynamic features.

[0023] Deformation dynamics features are extracted from deformation visual data, and a deformation dynamics branch network is constructed based on these features.

[0024] Calculate the correlation weights between thermodynamic features and deformation kinetic features, and perform weighted fusion based on the correlation weights to form a comprehensive feature sequence;

[0025] Multi-scale time-dependent features are extracted based on comprehensive feature sequences. These features are then mapped to a congestion risk probability value R0. The congestion risk probability value R0 is compared with a preset risk threshold to classify the congestion risk level.

[0026] Furthermore, the congestion risk level is divided into four level ranges based on a preset risk threshold, which includes a first risk threshold R1, a second risk threshold R2, and a third risk threshold R3. The congestion risk probability value R0 is compared with the preset risk threshold.

[0027] When R0 < R1, it is judged as normal level; when R1 ≤ R0 < R2, it is judged as attention level; when R2 ≤ R0 < R3, it is judged as warning level; when R0 ≥ R3, it is judged as serious level.

[0028] Furthermore, the predictive maintenance decision is executed when the congestion risk level is determined to be severe, and the predictive maintenance decision includes:

[0029] Obtain the time series of the congestion risk probability value R0 and calculate the risk growth trend;

[0030] Predict the remaining life of the nozzle based on processing parameters;

[0031] Obtain the production plan for nozzle processing, calculate the optimal maintenance time window, and generate maintenance work orders;

[0032] By integrating risk growth trends, nozzle remaining life, and maintenance work orders, predictive maintenance decisions can be made.

[0033] Furthermore, the step of obtaining the blockage risk level for graded response, selecting a self-healing control strategy based on the root cause and mechanism type of the blockage, and adjusting the processing parameters based on the self-healing control strategy includes:

[0034] When the condition is determined to be at a normal level, the current processing parameters are maintained, and the probability of blockage risk is monitored and predicted based on a multimodal temporal fusion network.

[0035] When the level of concern is determined, the frequency of multimodal sensor data acquisition is increased and trend monitoring is performed;

[0036] When the warning level is determined, the processing parameters are adjusted based on the self-healing control strategy, and the adjustment process is recorded.

[0037] When the condition is determined to be severe, a predictive maintenance decision is executed to generate a maintenance work order.

[0038] Furthermore, the step of adjusting the processing parameters based on the self-healing control strategy and recording the adjustment process, and the step of selecting the self-healing control strategy according to the root cause and mechanism type of the blockage, includes:

[0039] When the mechanism type is material degradation, select the formulation optimization strategy; when the mechanism type is foreign matter blockage, select the flow channel self-cleaning strategy; when the mechanism type is thermal runaway, select the temperature field reconstruction strategy; when the mechanism type is mechanical wear, select the parameter compensation strategy.

[0040] The selected self-healing control strategy is optimized for multiple objectives to generate the optimal combination of process parameters and adjust the processing parameters. The multi-objective optimization includes optimization of blockage risk indicators, product quality indicators, and production energy consumption indicators.

[0041] A machine vision-based nozzle blockage identification system, the system comprising:

[0042] The data acquisition module is used to acquire processing parameters and multimodal sensing data of the nozzle, including thermal vision data, deformation vision data and acoustic data.

[0043] The feature extraction module is used to extract multi-physics dynamic features from multimodal sensing data, including thermal gradient propagation features, deformation recovery features, and acoustic vibration features.

[0044] The causal analysis module constructs a causal network for blockage risk based on the dynamic characteristics of multiphysics fields, and identifies the root causes and mechanism types of blockage based on the causal network for blockage risk.

[0045] The risk prediction module constructs a multimodal time-series fusion network based on multimodal sensor data, performs risk assessment and prediction based on the multimodal time-series fusion network, and generates congestion risk levels and predictive maintenance decisions.

[0046] The self-healing control module acquires the blockage risk level and performs a graded response. It selects a self-healing control strategy based on the root cause and mechanism type of the blockage, and adjusts the processing parameters based on the self-healing control strategy.

[0047] This invention provides a machine vision-based method and system for nozzle clogging identification. It offers the following advantages:

[0048] This invention employs multi-nozzle thermal visual data, deformation visual data, and acoustic data to collect and extract multi-physics dynamic features, monitoring the overall state of material flow, thermodynamic behavior, and structural changes within the nozzle. Thermal visual data analyzes heat propagation anomalies through temperature field sequence analysis, deformation visual data captures structural responses through surface deformation sequences, and acoustic data identifies material flow obstacles through vibration spectrum analysis. This effectively eliminates environmental interference such as uneven lighting and plastic residues, improving detection accuracy. By constructing a multi-modal temporal fusion network for deep learning and fusion analysis of multi-physics dynamic features, it classifies blockage risk levels based on the root cause and mechanism type of blockage and implements a self-healing control strategy. This achieves accurate identification and judgment of minor and initial blockages, improving blockage monitoring accuracy, reducing production interruptions caused by false alarms, and avoiding the subjectivity of manual inspection and waste of finished products, thus improving the efficiency and economy of the injection molding process.

[0049] This invention employs a causal network for clogging risk and a self-healing control strategy for analysis and execution. By identifying the root causes and mechanism types of clogging, anomalies can be detected in the early stages of clogging formation. By analyzing abnormal temperature oscillation frequencies in thermal gradient propagation characteristics, changes in deformation hysteresis curves in deformation recovery characteristics, and vibration dominance shifts in acoustic vibration characteristics, it can capture internal material flow states and thermodynamic behavior differences that are imperceptible to traditional static image recognition. A multimodal temporal fusion network is used to further analyze time-series data, predict the probability of clogging risk, generate clogging risk levels, and provide graded early warnings. Dynamic perception and early warning are performed throughout the entire clogging process, providing ample time for proactive intervention. Based on the mechanism type, a self-healing control strategy is selected to automatically adjust processing parameters to suppress clogging development and achieve self-healing. Predictive maintenance decisions are generated based on the clogging risk level, realizing proactive early warning and proactive defense. This changes the traditional passive response maintenance mode, achieving predictive maintenance, effectively reducing maintenance costs and scrap rates, enabling accurate identification and maintenance of nozzle clogging, and improving the continuity and intelligence of nozzle processing. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the steps of the nozzle clogging identification method based on machine vision of the present invention.

[0051] Figure 2This is a flowchart of the nozzle clogging identification method based on machine vision according to the present invention;

[0052] Figure 3 This is an architecture diagram of the machine vision-based nozzle blockage recognition system of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1, as Figures 1 to 2 As shown, a machine vision-based nozzle clogging identification method includes:

[0055] Step S100: Acquire processing parameters and multimodal sensing data of the nozzle. The multimodal sensing data includes thermal vision data, deformation vision data, and acoustic data.

[0056] Step S101: Acquire thermal visual data, deformation visual data, and acoustic data of the nozzle during the processing. Thermal visual data is acquired by an infrared thermal imager, reflecting the temperature distribution and thermal behavior of the nozzle, and is used to monitor temperature anomalies caused by nozzle blockage, such as local overheating or heat flow obstruction. Deformation visual data is acquired by a high-speed camera or laser displacement sensor, reflecting changes in nozzle shape and identifying abnormal mechanical stress or deformation caused by nozzle blockage. Acoustic data is acquired by acoustic sensors, such as accelerometers or microphones, reflecting nozzle vibration and acoustic characteristics, and is used to detect abnormal vibration or changes in acoustic signals caused by nozzle blockage.

[0057] Step S102: Assign a unified timestamp to thermal visual data, deformation visual data, and acoustic data to ensure data acquisition time alignment; establish spatial coordinate mapping relationship; align the image coordinates of thermal visual data and deformation visual data to the same physical coordinate system through calibration; map acoustic data to corresponding spatial points according to the location of acoustic sensors; preprocess the thermal visual data, deformation visual data, and acoustic data, including noise reduction, standardization, and outlier removal; integrate the synchronized and mapped data into a structured dataset, with each time point containing the corresponding temperature field, deformation field, and acoustic signal, forming multimodal sensing data;

[0058] Among them, the processing parameters refer to the controllable and adjustable operation variables during the processing, which are used to monitor and adjust the nozzle status, including melt temperature, screw speed, back pressure, injection speed curve, holding time, cooling water flow rate, etc.

[0059] Thermal visual data includes temperature field distribution data and heat flow propagation sequence data. Temperature field distribution data is a two-dimensional temperature matrix on the nozzle surface, used to identify abnormal temperature areas in real time. Heat flow propagation sequence data reflects the changes in temperature field distribution data over a time series, used to analyze heat propagation dynamics. When acquiring thermal visual data, an infrared thermal imager is aimed at the nozzle, and infrared images are acquired at a fixed frame rate, such as 10 frames / second. Through calibration, the pixel values ​​of the infrared images are converted into temperature values ​​to generate temperature field distribution data. Temperature field distribution data at multiple time points are continuously acquired and combined in chronological order to form heat flow propagation sequence data.

[0060] Deformation visual data includes deformation sequence data and optical flow field data. Deformation sequence data is the deformation data of feature points within a time series, used to detect permanent deformation or elastic recovery. Optical flow field data is the motion vector field of pixels on the nozzle surface, used to analyze local motion patterns. When acquiring deformation visual data, a high-speed camera is used to capture visible light images of the nozzle at a high frame rate, such as 1000 frames per second. The displacement of feature points in the visible light images is calculated using digital image correlation techniques to generate deformation sequence data. At the same time, based on the visible light images acquired in a continuous time series, the motion vectors of the pixels are calculated to form optical flow field data.

[0061] Acoustic data includes vibration spectrum data and sound wave propagation data. Vibration spectrum data is a frequency component analysis of the vibration signal, used to identify resonant frequency shifts. Sound wave propagation data is a sound pressure time series, used to analyze sound wave attenuation characteristics. During acoustic data acquisition, an accelerometer is installed on the nozzle surface to collect vibration signals, and a microphone is used to collect sound wave signals. The vibration signals are subjected to a fast Fourier transform to obtain vibration spectrum data. The sound wave signals are then sampled and filtered in the time domain to obtain sound wave propagation data.

[0062] Step S200: Extract multi-physics dynamic features from multi-modal sensing data. The multi-physics dynamic features include thermal gradient propagation features, deformation recovery features, and acoustic vibration features.

[0063] Step S201: Extract thermal gradient propagation features from thermal visual data. These features include heat flow propagation velocity, unsteady thermal resistance, and temperature oscillation frequency. Thermal gradient propagation features assess the heat transfer anomalies caused by blockage from a thermal perspective. Extract deformation recovery features from deformation visual data. These features include deformation hysteresis curve, recovery time constant, and deformation localization degree. Deformation recovery features assess material fatigue or stress concentration caused by blockage from a mechanical perspective. Extract acoustic vibration features from acoustic data. These features include vibration dominant frequency shift, acoustic attenuation coefficient, and resonant mode changes. Acoustic vibration features assess structural vibration changes caused by blockage from an acoustic perspective. Finally, nine basic feature indicators are obtained.

[0064] Step S202: Perform multi-scale spatiotemporal feature analysis on each basic characteristic indicator, including time-scale analysis and spatial-scale analysis; perform multi-scale spatiotemporal feature analysis on thermal gradient propagation characteristics, deformation recovery characteristics, and acoustic vibration characteristics. First, perform time-scale analysis, extract the time series of each basic characteristic indicator, decompose the time series, and set short-term (e.g., 1 period), medium-term (e.g., 10 periods), and long-term (e.g., 100 periods) sliding time windows. Calculate the statistics within each sliding time window and set statistical thresholds for comparison. By comparison, distinguish between instantaneous fluctuations and continuous fluctuations. For example, a fluctuation variance greater than the variance threshold indicates rapid fluctuations, and a fluctuation variance less than the variance threshold indicates slow changes. The statistics include mean, standard deviation of fluctuation, trend slope, fluctuation variance, and cumulative deviation. By calculating the statistics, each basic characteristic indicator is converted into a multi-scale feature vector. For example, the multi-scale feature vector = [short-term mean, short-term standard deviation, medium-term trend slope, long-term cumulative deviation];

[0065] Then, spatial scale analysis is performed. Global statistics, such as average temperature and maximum deformation, are calculated for thermal gradient characteristics and deformation recovery characteristics. Based on the global statistics, the region is divided into N sub-regions. Local statistics are calculated for each sub-region. By comparing the differences between global and local statistics and the differences between local statistics, a difference threshold is set for comparison. If the difference between global and local statistics exceeds the upper and lower limits of the difference threshold, the sub-region is marked as an abnormal region. If the difference between local statistics exceeds the upper and lower limits of the difference threshold, the sub-region corresponding to the local statistics is marked as a candidate abnormal region. Continuous candidate abnormal regions are merged to form an abnormal region, thereby locating the spatial range of the anomaly and forming spatial correlation characteristics.

[0066] Step S203: Calculate the cross-spectral density in the frequency domain of the temperature oscillation frequency of the thermal gradient propagation characteristics and the vibration main frequency shift of the acoustic vibration characteristics to obtain the acoustic-thermal coupling characteristics. The acoustic-thermal coupling characteristics reflect the energy transfer relationship between thermal fluctuations and mechanical vibrations. For example, thermal oscillations at a constant thermal frequency excite the vibration of the structure, which makes it easier to determine whether thermal problems affect vibration or vibration affects temperature.

[0067] The time delay correlation between the deformation hysteresis curve of deformation recovery characteristics and the unsteady thermal resistance of thermal gradient propagation characteristics is calculated to obtain the thermo-mechanical coupling characteristics.

[0068] By integrating acoustic-thermal coupling features, thermo-mechanical coupling features, multi-scale feature vectors, and spatial correlation features, multi-physics dynamic features are formed. These features contain refined dynamic information within each physical field and coupling information of interactions between physical fields, providing comprehensive data support for the subsequent construction of a causal network for congestion risk and the classification of congestion risk levels.

[0069] The basic feature indicators extracted include:

[0070] When extracting heat propagation features, the heat flow velocity reflects the smoothness of melt flow. When the nozzle is blocked, the heat flow velocity will decrease. By tracking the position change of a specified isotherm through continuous heat flow propagation data, the displacement velocity is calculated to obtain the heat flow velocity. Unsteady-state thermal resistance represents heat exchange efficiency. Degradation or deposition of nozzle material will lead to an increase in unsteady-state thermal resistance and aggravated fluctuations. By setting virtual temperature measurement points in the nozzle inlet and outlet regions, the unsteady-state thermal resistance is calculated based on the instantaneous temperature difference and the estimated heat flow rate. Temperature oscillation frequency identifies thermal fluctuations of a specific frequency caused by flow instability or periodic partial blockage. Based on the temperature field distribution data, the temperature time series of multiple points in the nozzle is analyzed by spectrum analysis to generate the temperature oscillation frequency.

[0071] When extracting deformation recovery features, the deformation hysteresis curve is used to quantify the energy dissipation of the nozzle material within one cycle. An increase in area indicates nozzle material fatigue or increased viscosity. In a complete processing cycle, the pressure deformation curve is plotted, and the area of ​​the hysteresis loop is calculated to form the deformation hysteresis curve. The recovery time constant is used to represent the elastic recovery capability of the nozzle material. By performing an exponential fit on the deformation recovery curve during the pressure relief stage, the recovery time constant is obtained. An increase in the recovery time constant indicates a decrease in the elasticity of the nozzle material. The degree of deformation localization is used to identify stress concentration phenomena. The degree of deformation localization is judged by calculating the statistical variance of the single-frame deformation field data. A sudden increase in variance indicates that the nozzle has foreign object jamming or local structural damage.

[0072] During acoustic vibration feature extraction, the dominant frequency parameter is identified from the vibration spectrum data, and the drift is monitored to obtain the dominant frequency shift. The dominant frequency shift reflects the change in the effective mass or stiffness of the nozzle structure. Mass adhesion (such as carbon deposits) reduces the dominant frequency parameter, while stiffness loss (such as cracks) increases the dominant frequency parameter. The acoustic attenuation coefficient represents the propagation loss of sound waves in the nozzle structure. Internal deposits will change the acoustic impedance, resulting in a change in the attenuation coefficient. By identifying the frequency and amplitude changes of multiple resonance peaks in a specified frequency band in the vibration spectrum data, the resonant mode changes are obtained. The resonant mode changes comprehensively assess the changes in the structural dynamic characteristics, reflecting mechanical wear and structural blockage.

[0073] Multiscale spatiotemporal feature analysis is a method for analyzing feature changes at different time and spatial scales. It is used to capture congestion precursors from instantaneous events to long-term trends, identify spatial anomalies from microscopic local to macroscopic features, and achieve dynamic capture of congestion. Multiscale spatiotemporal feature analysis improves the comprehensiveness and accuracy of risk identification. Multiscale spatiotemporal feature analysis includes time scale analysis (such as short-term fluctuations and long-term trends) and spatial scale analysis (such as local details and overall distribution).

[0074] Step S300: Construct a causal network for blockage risk based on the dynamic characteristics of multiphysics fields, and identify the root causes and mechanism types of blockage based on the causal network for blockage risk;

[0075] Step S301: Obtain processing parameters and multi-physics dynamic features as network nodes. The network nodes are divided into process nodes composed of processing parameters and feature nodes composed of multi-physics dynamic features. Analyze the causal relationships between network nodes to establish directed edges. When establishing directed edges, first determine the preliminary causal direction based on thermodynamics and fluid mechanics principles, such as the effect of fluid temperature on heat propagation speed. Then, based on historical processing data generated during nozzle processing, use a Bayesian network structure to verify and supplement the preliminary causal direction to form the final directed edges. Then, quantify the influence intensity based on historical processing data, calculate the change magnitude of the network nodes as results when the network nodes as causes change, set the connection weight of each edge based on the change magnitude, and finally integrate all network nodes, directed edges, and connection weights to construct a clogging risk causal network. Historical processing data includes multimodal sensor data and processing parameters under normal and abnormal states. Train the clogging risk causal network using machine learning methods.

[0076] Step S302: Analyze the impact strength of each network node on congestion risk, and analyze the root cause of congestion based on connection weight and impact strength; if abnormal features are observed in the dynamic characteristics of the multiphysics field, that is, the state of the feature node deviates from the normal range, then the feature node is defined as an abnormal node. The abnormal node is input into the congestion risk causal network, and the network node that caused the abnormal node is found by reverse reasoning through the causal relationship of the congestion risk causal network. The network node is defined as the root cause node. The root cause node is generally a network node that is upstream of the abnormal node and connected to multiple downstream abnormal nodes.

[0077] The categories of root causes of blockage include uncontrolled key process parameters, deterioration of component performance, external interference factors, and changes in material properties. Uncontrolled key process parameters include excessively high or fluctuating melt temperature, persistently low back pressure, and cooling system malfunctions. Deterioration of component performance includes heater power attenuation and wear of the screw or check ring. External interference factors include excessive impurity content in raw material batches and drastic changes in ambient temperature and humidity. Changes in the properties of the nozzle material itself include decreased polymer melt stability. When determining the category of root cause of blockage, the determination of each category is based on preset thresholds, including preset thresholds for each parameter category of each processing parameter and preset thresholds for multi-physics dynamic characteristics. The specific types and parameters of preset thresholds are set according to design specifications and production requirements, and will not be elaborated here.

[0078] Step S303: Determine the mechanism type based on the root cause of the blockage. The mechanism types include material degradation type, foreign object blockage type, thermal runaway type, and mechanical wear type. Match the identified root cause nodes with abnormal features to determine the category of the root cause of the blockage. Compare the category of the root cause of the blockage with the mechanism type to determine the specific mechanism type.

[0079] Mechanism type is a classification of the inherent physical, chemical, or mechanical processes involved in blockage. It is used to classify the root cause of blockage and facilitate the determination of self-healing control strategies. For example, when the root cause of blockage points to "excessively high melt temperature" or "poor thermal stability of raw materials," and is accompanied by abnormal features such as "increased recovery time constant," "slowly widening deformation hysteresis curve area," and "increased unsteady thermal resistance," the mechanism type is determined to be material degradation type. When there is no clear root cause of blockage with specific processing parameters, but abnormal features such as "sharp increase in the degree of deformation localization," "abrupt change in acoustic attenuation coefficient," and "sudden drop in heat flow propagation velocity at the throat" suddenly appear, the mechanism type is determined to be foreign matter blockage type.

[0080] Among them, material degradation type refers to the decomposition, cross-linking, or molecular weight reduction of processed materials under high temperature and high shear, resulting in deterioration of rheological properties, and is used to explain the poor flow caused by changes in the material's own properties; foreign matter blockage type refers to solid particles such as impurities, coke, and metal debris generated externally or internally, causing physical blockage in narrow flow channels, and is used to explain sudden, localized complete or partial blockage; thermal runaway type refers to the dynamic imbalance of the heating or cooling system, causing the temperature of the local area of ​​the nozzle to continuously deviate from the set value, resulting in premature solidification or excessive degradation of the material, and is used to explain periodic or gradual blockage caused by thermal management failure; mechanical wear type refers to the wear of moving parts or pressure-bearing parts such as screws and nozzle bodies due to long-term use, resulting in changes in gaps, pressure leakage, or shape changes, which in turn affect the stability of melt delivery, and is used to explain the slowly developing blockage mode related to the mechanical life of the equipment.

[0081] Step S400: Construct a multimodal time-series fusion network based on multimodal sensor data, perform risk assessment and prediction based on the multimodal time-series fusion network, and generate congestion risk level and predictive maintenance decision;

[0082] Step S401: Acquire thermal visual data and acoustic data. Extract thermal visual features based on the temperature field distribution data and heat flow sensing sequence data of the thermal visual data. Thermal visual features include temperature field change features, heat flow dynamic features, and thermal oscillation analysis. Extract temperature-related frequency components from the vibration spectrum data of the acoustic data, such as the correlation between low-frequency vibration and thermal expansion. Calculate the correlation coefficient between vibration energy and temperature change to obtain vibration-temperature coupling features. Analyze the spatiotemporal relationship between sound wave propagation data and temperature field distribution data, perform acoustic-thermal correlation, and obtain sound wave propagation features. Integrate the thermal visual features, vibration-temperature coupling features, and sound wave propagation features into a multivariate time series to form thermodynamic features. Each time point corresponds to a feature vector. Construct a thermodynamic branch network based on the thermodynamic features and acoustic vibration features. The thermodynamic branch network is constructed based on a convolutional neural network (CNN) or a recurrent neural network (RNN) to learn the temporal patterns of thermal and acoustic data. For example, thermodynamic features = [average temperature, temperature gradient, heat flow velocity, temperature oscillation frequency, vibration-temperature correlation coefficient];

[0083] Deformation sequence features and optical flow field features are extracted from visual deformation data. Deformation sequence features include statistical features, deformation rate features, and deformation recovery features. Optical flow field features include motion vector features. These features are integrated into a multivariate time series to form deformation dynamics features. A deformation dynamics branch network is constructed based on these features, using either a convolutional neural network (CNN) or a recurrent neural network (RNN) to learn the temporal patterns of the deformation data. Statistical features are obtained by calculating the deformation field statistics at each time point, such as the average... Deformation, maximum deformation, deformation variance, etc.; deformation rate characteristics are obtained by calculating the rate of change of deformation over time (first derivative) to identify the acceleration or deceleration phase of deformation; deformation recovery characteristics are obtained by extracting the recovery time constant and the area of ​​the deformation hysteresis curve from the deformation sequence data; motion vector characteristics are directly extracted from optical flow field data; flow field consistency characteristics are obtained by calculating the spatial consistency (such as variance) and temporal consistency (such as autocorrelation) of motion vectors to identify abnormal motion regions; for example, deformation dynamics characteristics = [mean deformation, deformation rate, recovery time constant, motion vector amplitude, flow field consistency];

[0084] The correlation weights between thermodynamic features and deformation kinetic features are calculated. The correlation weights represent the contribution of thermodynamic features and deformation kinetic features to the current blockage risk. The correlation weights are dynamically set through a pre-established weight lookup table. A set of basic weights is first set to introduce a dynamic fine-tuning mechanism. The correlation weights are allocated by analyzing the fluctuation range of thermodynamic features and deformation kinetic features in real time. The correlation weights with larger fluctuation ranges are increased. For example, when the melt temperature is abnormally high, the correlation weight of thermodynamic features is increased. The correlation weights are weighted and fused to form a comprehensive feature sequence.

[0085] Step S402: Extract multi-scale time-dependent features based on the comprehensive feature sequence. Set short-term, medium-term and long-term time windows, and extract the change features of the comprehensive feature sequence in different time windows as multi-scale time-dependent features. Input the multi-scale time-dependent features into a regression model trained based on historical processed data, such as a support vector regression model or a neural network regression model. Map the multi-scale time-dependent features to a congestion risk probability value R0 between [0,1]. The regression model uses historical processed data for supervised learning and compares the congestion risk probability value R0 with a preset risk threshold to classify the congestion risk level.

[0086] Step S403: The blockage risk level is divided into four level intervals based on preset risk thresholds. The preset risk thresholds include a first risk threshold R1, a second risk threshold R2, and a third risk threshold R3. The preset risk thresholds are set based on historical processing data and self-healing control strategies, analyzing effective early intervention points in numerous experiments. For example, R1=0.3, R2=0.6, and R3=0.8. The blockage risk probability value R0 is compared with the preset risk thresholds:

[0087] When R0 < R1, it is judged as normal level; when R1 ≤ R0 < R2, it is judged as attention level; when R2 ≤ R0 < R3, it is judged as warning level; when R0 ≥ R3, it is judged as serious level.

[0088] Step S404, Predictive maintenance decision-making, is executed when the congestion risk level is determined to be severe. Predictive maintenance decisions include:

[0089] First, obtain the time series of the blockage risk probability value and calculate the risk growth trend; obtain the blockage risk probability value R0 within a set period to form a time series, and perform curve fitting on the time series. First, perform linear fitting and calculate the slope. If the linear fitting degree is low, perform exponential fitting or nonlinear model fitting. The slope obtained from the fitting and the curve itself jointly represent the risk growth trend. Set a standard value for the slope for comparison. If the slope is positive and greater than the set standard value, it indicates that the blockage risk is accumulating at an accelerated pace.

[0090] Secondly, the remaining nozzle life is predicted based on processing parameters. A relationship model between health indicators and remaining life is derived in advance using physical mechanisms, based on historical processing data and nozzle life tests. This model can be a physics-based degradation equation, a data-based statistical model, etc. By determining the current mechanism type, the current clogging risk probability value R0, the slope of the risk growth trend, and the dynamic characteristics of the multiphysics field are input into the relationship model to obtain the predicted remaining nozzle life. For example, for a mechanism type of mechanical wear, a life equation based on the cumulative area of ​​the deformation hysteresis curve is established. When the cumulative amount reaches a set threshold, the life is determined to end, and the relationship model outputs an estimated remaining life value T expressed as "remaining production cycles" or "remaining operating days". L ;

[0091] Then, obtain the production plan and remaining life estimate T for nozzle processing. L Calculate the optimal maintenance time window, and obtain it from the production plan starting from the current time T. N The time span to the future is greater than T L Find a time window T among all planned shutdown periods. W The start time T of this window W1 and end time T W2 Satisfy: T W1 < (T) N +T L ) < T W2 Ensure that maintenance is completed before the remaining lifespan expires; if multiple time windows meet the condition, select the one closest to T. N +T L The sum of time windows is used to utilize the equipment for as long as possible before the risk of congestion occurs, while ensuring that maintenance can be completed. This time window [T] is then output. W1 ,T W2 As the optimal maintenance time window, maintenance work orders are generated based on this window. Each work order includes basic work order information, equipment information (such as nozzle number and equipment location number), maintenance type (such as disassembly and cleaning, nozzle replacement, etc.), optimal maintenance time window, required resources, and safety precautions. The basic work order information includes a unique work order number, work order generation time, and work order priority. Work order priority determines the execution order among maintenance work orders, generally set according to the work order generation time. Based on the work order generation time, urgent maintenance work orders, such as those with nozzle damage or where processing cannot proceed, are given priority.

[0092] Finally, the risk growth trend, nozzle remaining life, and maintenance work orders are integrated to form predictive maintenance decisions. Predictive maintenance decisions also include parameters such as the probability value of clogging risk, mechanism type, current processing parameters, self-healing control strategies already adopted, and current multimodal sensor data.

[0093] The production plan is a task arrangement information set before processing for a specified future time, including product orders to be produced, order quantities, delivery dates, production sequence, and production time periods. The production plan is obtained directly from the processing system and is used to ensure that maintenance activities are scheduled within the planned production gaps, mold change times, or predetermined downtime when calculating the optimal maintenance time window, thereby minimizing the impact on the delivery schedule and the resulting production losses.

[0094] Step S500: Obtain the blockage risk level and implement a graded response; select a self-healing control strategy based on the root cause and mechanism type of the blockage; and adjust the processing parameters based on the self-healing control strategy.

[0095] Step S501: Real-time acquisition of the congestion risk probability value R0 and congestion risk level generated in S400 for graded response, including:

[0096] When the condition is determined to be normal, the current processing parameters are maintained, and the process returns to step S400 to continue monitoring and predicting the probability value of blockage risk based on the multimodal temporal fusion network.

[0097] When the level of concern is determined, the frequency of multimodal sensor data acquisition is increased and trend monitoring is performed, and the process returns to step S400 for continued monitoring and analysis.

[0098] When the warning level is determined, according to the root cause and mechanism type of the blockage in step S300, the corresponding self-healing control strategy is selected from the self-healing strategy library, the processing parameters are adjusted based on the self-healing control strategy, and the process of adjusting the processing parameters is recorded.

[0099] When the condition is determined to be severe, step S404 is executed to generate a maintenance work order based on the predictive maintenance decision.

[0100] Step S502: When the level is determined to be a warning, a self-healing control strategy is selected based on the root cause and mechanism type of the blockage, including:

[0101] When the mechanism type is material degradation, select the formulation optimization strategy, adjust the melt temperature and screw speed, such as reducing the melt temperature by 5-10℃ and increasing the screw speed by 5-10%. Through the combination of temperature and mechanical, a mild processing environment is provided for the unstable melt, thus delaying material degradation.

[0102] When the mechanism type is foreign object blockage, select the flow channel self-cleaning strategy, adjust the back pressure and optimize the injection speed curve, such as increasing the back pressure by 10-15%, and use the backflow to flush away the tiny foreign objects attached to the screw end or nozzle wall.

[0103] When the mechanism type is thermal runaway, a temperature field reconstruction strategy is selected to redistribute the heater power and adjust the cooling water flow rate.

[0104] When the mechanism type is mechanical wear, select the parameter compensation strategy, adjust the injection pressure curve, optimize the holding time and pressure, bypass the performance degradation of mechanical components, and maintain the stability of the final output function by adjusting the control parameters;

[0105] Multi-objective optimization is performed on the processing parameters involved in the selected self-healing control strategy to generate the optimal combination of process parameters for adjustment. This multi-objective optimization includes optimization of blockage risk indicators, product quality indicators, and production energy consumption indicators. Blockage risk indicator optimization aims to minimize the risk probability value R0; product quality indicator optimization aims to minimize deviations of key quality characteristics such as size, weight, and internal stress from nominal values; and production energy consumption indicator optimization aims to minimize the total energy consumption throughout the entire cycle. This multi-objective optimization generates the optimal combination of process parameters, avoiding the introduction of new quality defects or energy waste in order to solve the blockage problem. During multi-objective optimization, the processing parameters corresponding to the self-healing control strategy are defined as a parameter search space. Sampling and iteration are performed within this search space to predict the optimal values ​​of the three indicators for each parameter combination. A multi-objective optimization algorithm, such as a genetic algorithm or gradient descent algorithm, is used to find an optimal solution—the parameter combination that achieves the best balance among the three indicator optimization values. This optimal combination of process parameters is then executed to adjust the processing parameters.

[0106] The process parameters are adjusted in real time and monitored in real time through step S500. After the adjustment is completed, the adjusted multimodal sensing data is obtained. The blockage risk probability value R0 is recalculated through the multimodal temporal fusion network to verify the effectiveness of the self-healing control strategy. After verification, a new blockage risk level is determined and a new cycle is implemented.

[0107] The self-healing control strategy is a predefined, automated process parameter adjustment scheme. When early signs of clogging risk are detected, the processing conditions are actively and appropriately adjusted to counteract or delay the development of the clogging mechanism, so that the nozzle can be restored or maintained in a healthy state. The specific self-healing control strategy is determined based on the mechanism type and root cause of clogging determined in step S300. The self-healing control strategy is stored in the self-healing strategy library. The self-healing strategy library is based on domain knowledge, such as polymer materials science, injection molding process science and historical maintenance data. It associates specific failure mechanisms with the most likely effective process intervention methods, and after verifying their effectiveness through simulation and experiments, it is solidified into If-Then rules in the self-healing strategy library.

[0108] In this embodiment, the method for identifying nozzle blockage is mainly described in the injection molding process. In actual use, the method and system for identifying nozzle blockage in this application are not limited to injection molding, but can also be applied to other processes in which nozzles are used for processing.

[0109] In this embodiment, multi-physics dynamic features are collected and extracted using thermal visual data, deformation visual data, and acoustic data from multiple nozzles. This monitors the overall state of material flow, thermodynamic behavior, and structural changes within the nozzle. Thermal visual data analyzes heat flow anomalies through temperature field sequence analysis, deformation visual data captures structural responses through surface deformation sequences, and acoustic data identifies material flow obstacles through vibration spectrum analysis. This effectively eliminates environmental interference such as uneven lighting and plastic residues, improving detection accuracy. By constructing a multi-modal temporal fusion network for deep learning and fusion analysis of multi-physics dynamic features, the system classifies blockage risk levels based on the root cause and mechanism type of blockage and implements a self-healing control strategy. This achieves accurate identification and judgment of minor and initial blockages, improving blockage monitoring accuracy, reducing production interruptions caused by false alarms, and avoiding the subjectivity of manual inspection and waste of finished products, thereby improving the efficiency and economy of the injection molding process.

[0110] This study employs a causal network for clogging risk and a self-healing control strategy for analysis and execution. By identifying the root causes and mechanism types of clogging, anomalies can be detected in the early stages of clogging formation. By analyzing abnormal temperature oscillation frequencies in thermal gradient propagation characteristics, changes in deformation hysteresis curves in deformation recovery characteristics, and vibration dominance frequency shifts in acoustic vibration characteristics, it can capture internal material flow states and thermodynamic behavior differences that cannot be perceived by traditional static image recognition. A multimodal temporal fusion network is used to further analyze time-series data, predict the probability of clogging risk, generate clogging risk levels, and provide graded early warnings. Dynamic perception and early warning are implemented throughout the entire clogging process, providing ample time for proactive intervention. Simultaneously, the self-healing control module automatically adjusts processing parameters according to the mechanism type to suppress clogging development and perform self-healing, achieving proactive early warning and proactive defense. This changes the traditional passive response maintenance mode, enabling predictive maintenance, effectively reducing maintenance costs and scrap rates, and improving the continuity and intelligence of nozzle processing.

[0111] Example 2, as Figure 3 As shown, a machine vision-based nozzle clogging recognition system includes:

[0112] The data acquisition module is used to acquire processing parameters and multimodal sensing data of the nozzle. The multimodal sensing data includes thermal vision data, deformation vision data and acoustic data.

[0113] The feature extraction module is used to extract multi-physics dynamic features from multimodal sensing data. These multi-physics dynamic features include thermal gradient propagation features, deformation recovery features, and acoustic vibration features.

[0114] The causal analysis module constructs a causal network for blockage risk based on the dynamic characteristics of multiphysics fields, and identifies the root causes and mechanism types of blockage based on the causal network for blockage risk.

[0115] The risk prediction module constructs a multimodal time-series fusion network based on multimodal sensor data, performs risk assessment and prediction based on the multimodal time-series fusion network, and generates congestion risk levels and predictive maintenance decisions.

[0116] The self-healing control module acquires the blockage risk level and performs a graded response. It selects a self-healing control strategy based on the root cause and mechanism type of the blockage, and adjusts the processing parameters based on the self-healing control strategy.

[0117] Example 3: This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code, which, when executed by the one or more processors, can perform the machine vision-based nozzle blockage recognition method and system described above.

[0118] The methods and systems according to the embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the machine vision-based nozzle blockage recognition method and system provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs.

[0119] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision-based method for identifying nozzle blockages, characterized in that, The method includes: Acquire processing parameters and multimodal sensing data of the nozzle, wherein the multimodal sensing data includes thermal visual data, deformation visual data and acoustic data; Multi-physics dynamic features are extracted from multimodal sensing data, including thermal gradient propagation features, deformation recovery features, and acoustic vibration features. A causal network for blockage risk is constructed based on the dynamic characteristics of multiphysics fields, and the root causes and mechanism types of blockage are identified based on the causal network for blockage risk. A multimodal time-series fusion network is constructed based on multimodal sensor data. Risk assessment and prediction are performed based on the multimodal time-series fusion network to generate congestion risk levels and predictive maintenance decisions. Obtain the blockage risk level for graded response, select self-healing control strategies based on the root cause and mechanism type of blockage, and adjust the processing parameters based on the self-healing control strategies.

2. The machine vision-based nozzle clogging identification method according to claim 1, characterized in that, The acquisition of processing parameters and multimodal sensing data of the nozzle includes: Acquire thermal visual data, deformation visual data and acoustic data of the nozzle during the processing, and establish spatial coordinate mapping relationship after time stamp synchronization to form multimodal sensing data; The thermal visual data includes temperature field distribution data and heat flow propagation sequence data; the deformation visual data includes deformation sequence data and optical flow field data; and the acoustic data includes vibration spectrum data and sound wave propagation data.

3. The machine vision-based nozzle clogging identification method according to claim 2, characterized in that, The extraction of multi-physics dynamic features from multimodal sensing data includes: Extract thermal gradient propagation features from thermal visual data, extract deformation recovery features from deformation visual data, and extract acoustic vibration features from acoustic data; Multi-scale spatiotemporal characteristic analysis was performed on the thermal gradient propagation characteristics, deformation recovery characteristics, and acoustic vibration characteristics to form multi-physics field dynamic characteristics.

4. The machine vision-based nozzle clogging identification method according to claim 1, characterized in that, The construction of the causal network for congestion risk based on the dynamic characteristics of multiphysics includes: Processing parameters and multiphysics dynamic characteristics are obtained as network nodes. The causal relationships between network nodes are analyzed to establish directed edges. Connection weights are set based on the directed edges. A causal network of blockage risk is constructed based on network nodes, connection weights, and directed edges. Analyze the impact of each network node on congestion risk, and analyze the root causes of congestion based on connection weights and impact strength; The mechanism type is determined based on the root cause of the blockage, and the mechanism types include material degradation type, foreign object blockage type, thermal runaway type, and mechanical wear type.

5. The machine vision-based nozzle clogging identification method according to claim 1, characterized in that, The process involves constructing a multimodal time-series fusion network based on multimodal sensor data, performing risk assessment and prediction based on this network, and generating congestion risk levels and predictive maintenance decisions, including: Thermodynamic features are extracted from thermal visual data and acoustic data, and a thermodynamic branch network is constructed based on the thermodynamic features. Deformation dynamics features are extracted from deformation visual data, and a deformation dynamics branch network is constructed based on these features. Calculate the correlation weights between thermodynamic features and deformation kinetic features, and perform weighted fusion based on the correlation weights to form a comprehensive feature sequence; Multi-scale time-dependent features are extracted based on comprehensive feature sequences. These features are then mapped to a congestion risk probability value R0. The congestion risk probability value R0 is compared with a preset risk threshold to classify the congestion risk level.

6. The machine vision-based nozzle clogging identification method according to claim 5, characterized in that, The congestion risk level is divided into four level ranges based on a preset risk threshold. The preset risk threshold includes a first risk threshold R1, a second risk threshold R2, and a third risk threshold R3. The congestion risk probability value R0 is compared with the preset risk threshold. When R0 < R1, it is judged as normal level; when R1 ≤ R0 < R2, it is judged as attention level; when R2 ≤ R0 < R3, it is judged as warning level; when R0 ≥ R3, it is judged as serious level.

7. The machine vision-based nozzle clogging identification method according to claim 6, characterized in that, The predictive maintenance decision is executed when the congestion risk level is determined to be severe. The predictive maintenance decision includes: Obtain the time series of the congestion risk probability value R0 and calculate the risk growth trend; Predict the remaining life of the nozzle based on processing parameters; Obtain the production plan for nozzle processing, calculate the optimal maintenance time window, and generate maintenance work orders; By integrating risk growth trends, nozzle remaining life, and maintenance work orders, predictive maintenance decisions can be made.

8. The machine vision-based nozzle clogging identification method according to claim 7, characterized in that, The process of obtaining the blockage risk level and implementing a graded response, selecting a self-healing control strategy based on the root cause and mechanism type of the blockage, and adjusting the processing parameters based on the self-healing control strategy includes: When the condition is determined to be at a normal level, the current processing parameters are maintained, and the probability of blockage risk is monitored and predicted based on a multimodal temporal fusion network. When the level of concern is determined, the frequency of multimodal sensor data acquisition is increased and trend monitoring is performed; When the warning level is determined, the processing parameters are adjusted based on the self-healing control strategy, and the adjustment process is recorded. When the condition is determined to be severe, a predictive maintenance decision is executed to generate a maintenance work order.

9. The machine vision-based nozzle clogging identification method according to claim 8, characterized in that, The selection of self-healing control strategies based on the root cause and mechanism type of blockage includes: When the mechanism type is material degradation, select the formulation optimization strategy; when the mechanism type is foreign matter blockage, select the flow channel self-cleaning strategy; when the mechanism type is thermal runaway, select the temperature field reconstruction strategy; when the mechanism type is mechanical wear, select the parameter compensation strategy. The selected self-healing control strategy is optimized for multiple objectives to generate the optimal combination of process parameters and adjust the processing parameters. The multi-objective optimization includes optimization of blockage risk indicators, product quality indicators, and production energy consumption indicators.

10. A machine vision-based nozzle clogging recognition system, characterized in that, The system includes: The data acquisition module is used to acquire processing parameters and multimodal sensing data of the nozzle, including thermal vision data, deformation vision data and acoustic data. The feature extraction module is used to extract multi-physics dynamic features from multimodal sensing data, including thermal gradient propagation features, deformation recovery features, and acoustic vibration features. The causal analysis module constructs a causal network for blockage risk based on the dynamic characteristics of multiphysics fields, and identifies the root causes and mechanism types of blockage based on the causal network for blockage risk. The risk prediction module constructs a multimodal time-series fusion network based on multimodal sensor data, performs risk assessment and prediction based on the multimodal time-series fusion network, and generates congestion risk levels and predictive maintenance decisions. The self-healing control module acquires the blockage risk level and performs a graded response. It selects a self-healing control strategy based on the root cause and mechanism type of the blockage, and adjusts the processing parameters based on the self-healing control strategy.

Images