Intelligent fault diagnosis and repair system of full-biodegradable composite material production line

Abstract

The invention relates to the technical field of high polymer material processing intelligent control, and discloses an intelligent fault diagnosis and repair system of a full biodegradable composite material production line, which comprises a physical production line layer, an edge calculation and intelligent control layer and a cloud platform data center layer. The edge layer constructs a time-synchronized unified data lake through a data acquisition module; a digital twin engine fuses the non-isothermal non-Newtonian fluid transport mechanism and the deep learning model to generate a dynamic health reference curve; the fault diagnosis module calculates a residual sequence of the actual monitoring data and the reference curve, and identifies a fault mode through multi-modal feature fusion; and the self-repairing decision module generates a control compensation instruction based on the fault mode, and issues and executes closed-loop control after safety verification. Through the mechanism and data driving fusion technology, the diagnosis problem caused by material characteristic fluctuation is solved, early fault early warning, process parameter self-adaptive compensation and predictive maintenance in the production process are achieved, and the product quality consistency and the production stability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for polymer material processing, specifically to an intelligent fault diagnosis and repair system for a fully biodegradable composite material production line. Background Technology

[0002] Currently, the preparation of fully biodegradable multiphase composite materials such as PBAT and PLA mainly adopts reactive extrusion technology. This production process integrates multi-component feeding, melt plasticizing, and reactive grafting units, involving multiple coupled thermal, mechanical, and rheological fields. The process window is usually narrow, and the physical properties of different batches of raw materials vary. The stability of the production process directly determines the degradation performance and mechanical strength of the final product.

[0003] For the aforementioned production process, existing control systems are mostly based on PLC architecture. The system collects data such as temperature, pressure, and main unit current in real time through sensors, and uses PID algorithms to adjust the heating zone temperature and screw speed based on feedback. Fault monitoring mainly relies on preset upper and lower thresholds; when the monitored values ​​exceed the set range, an alarm or shutdown interlock is triggered. Operators monitor the operating status based on the data curves displayed on the interface and manually adjust process parameters based on offline monitoring results.

[0004] However, existing technologies have shortcomings in dealing with complex operating conditions. Fixed threshold monitoring methods are insensitive to gradual faults and are unlikely to issue warnings in the early stages of filter clogging or minor screw wear, often delaying detection until a shutdown occurs. Due to the lack of perception and dynamic compensation capabilities for fluctuations in the intrinsic properties of raw materials, traditional PID control cannot eliminate the impact of changes in raw material viscosity or moisture content on the melt flow index of the product, resulting in poor batch-to-batch quality consistency. Furthermore, multivariate coupling makes it difficult to trace the root cause of faults, relying solely on manual experience for troubleshooting is inefficient, and the lack of a scientific lifespan prediction mechanism keeps maintenance costs high, preventing proactive predictive maintenance.

[0005] Therefore, the present invention provides an intelligent fault diagnosis and repair system for a fully biodegradable composite material production line to address the shortcomings of the prior art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent fault diagnosis and repair system for a fully biodegradable composite material production line, which solves the problems of delayed fault warning, poor product quality consistency due to raw material fluctuations and equipment performance degradation, and high passive maintenance costs in the existing production of fully biodegradable composite materials.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: an intelligent fault diagnosis and repair system for a fully biodegradable composite material production line, comprising a physical production line layer, an edge computing and intelligent control layer, and a cloud platform data center layer; The physical production line layer includes a reactive extrusion unit. The physical production line layer collects equipment status time-series data, process parameter time-series data, and online quality time-series data reflecting the reactive extrusion process of fully biodegradable materials through a sensor network. The edge computing and intelligent control layer includes: The data acquisition and preprocessing module is used to receive the equipment status time-series data, the process parameter time-series data, and the online quality time-series data to construct a unified data lake with time synchronization. A digital twin engine is used to generate a health status baseline curve based on the unified data lake by utilizing a hybrid model driven by a mechanism based on the principle of non-isothermal non-Newtonian fluid transport and a data-driven model. The fault diagnosis and health management module is used to output fault modes using the residual sequence of actual monitoring data in the unified data lake and the health status baseline curve; The process parameter self-repair decision module is used to generate control compensation instructions based on the fault mode and send them to the physical production line layer to execute closed-loop control.

[0008] By adopting the above technical solution, and utilizing a digital twin architecture that integrates physical mechanisms and data-driven approaches, a health benchmark based on the principle of non-isothermal, non-Newtonian fluid transport is established to address the high sensitivity of the rheological properties of fully biodegradable materials to temperature and shear rate. Residual analysis is used to identify faults, and closed-loop control enables parameter self-repair. Therefore, this solves the problem that traditional threshold monitoring cannot handle fluctuations in intrinsic material properties and complex rheological behaviors, achieving accurate real-time diagnosis and adaptive compensation in the production process, and improving the consistency of composite material product quality and the stability of production line operation.

[0009] Preferably, the process by which the data acquisition and preprocessing module receives the equipment status time-series data, the process parameter time-series data, and the online quality time-series data to construct a time-synchronized unified data lake includes: using a unified clock source to label the equipment status time-series data, the process parameter time-series data, and the online quality time-series data from the physical production line layer; using an interpolation algorithm to perform microsecond-level timestamp alignment operations on data sources with different sampling rates; and aggregating heterogeneous multi-source data to construct the unified data lake that is strictly synchronized in the time dimension.

[0010] By adopting the above technical solution, the phase deviation of multi-source heterogeneous sensor data in the time dimension was eliminated by using microsecond-level alignment technology, which ensured the temporal consistency of pressure, temperature and mass data collected at different frequencies in subsequent multivariate coupling analysis, laying a data foundation for high-precision modeling.

[0011] Preferably, in the process of generating a health status benchmark curve based on the unified data lake using a hybrid model of mechanism and data driven by the digital twin engine based on the principle of non-isothermal non-Newtonian fluid transport, the steps of the mechanism model path include: establishing a one-dimensional distributed parameter model of the screw extrusion process, discretizing the screw along the axial direction into multiple finite control volumes; calculating the theoretical viscosity distribution along the screw axial direction using the rheological state equation describing the relationship between the apparent viscosity of the material and the shear rate and temperature function based on the real-time collected screw speed and temperature feedback values; and outputting the theoretical melt pressure curve and theoretical torque curve under the current operating conditions as the physical benchmark of the health status benchmark curve by integrating the pressure gradient and viscous dissipation along the flow channel.

[0012] By adopting the above technical solution, starting from the rheological mechanism of fully biodegradable materials, and through discretization calculation and integral solution, a physical benchmark that can quantitatively reflect the viscosity evolution, pressure build-up and energy dissipation process during material processing is constructed, providing an interpretable theoretical basis for distinguishing between equipment failure and reasonable process fluctuations.

[0013] Preferably, in the process of generating a health status baseline curve based on the unified data lake using a hybrid model combining the mechanism of non-isothermal non-Newtonian fluid transport principle and a data-driven approach, the steps of running the data-driven pathway include: using a long short-term memory network enhanced by an attention mechanism to receive the equipment status time-series data and the process parameter time-series data from the unified data lake as input vectors; using long short-term memory network units to capture long-term dependencies and dynamic lag effects in the input vectors; using an attention mechanism module to identify the nonlinear influence weights of key process parameters on quality indicators; and outputting predicted values ​​of key quality parameters to correct the physical baseline.

[0014] By adopting the above technical solution, deep neural networks are used to capture complex nonlinear dynamic characteristics and environmental micro-perturbations that are difficult to quantify by simplified mechanistic models. Attention mechanisms are used to focus on key influencing factors, thereby improving the fitting accuracy and robustness of the health baseline curve to actual dynamic working conditions.

[0015] Preferably, in the process of the fault diagnosis and health management module outputting fault modes using the residual sequence of the actual monitoring data in the unified data lake and the health status benchmark curve, the step of calculating the residual sequence includes: receiving the actual monitoring data from the data acquisition and preprocessing module and the data from the health status benchmark curve from the digital twin engine; calculating the difference between the actual values ​​and the theoretical health benchmark values ​​for key variables such as host torque, melt pump inlet pressure and outlet pressure; and generating the residual sequence reflecting equipment performance degradation or abnormal disturbances by stripping away parameter fluctuations caused by normal process setting adjustments.

[0016] By adopting the above technical solutions, the parameter changes caused by normal operating condition adjustments and the fluctuations caused by abnormal faults are effectively separated. The residual analysis technology reduces the false alarm rate of faults in the process of changing operating conditions, enabling the diagnostic system to focus on the actual equipment performance degradation and process anomalies.

[0017] Preferably, in the process of the fault diagnosis and health management module outputting fault modes using the residual sequence of the actual monitoring data in the unified data lake and the health status benchmark curve, the step of extracting features from the residual sequence includes: processing in parallel through three channels: the first channel performs windowing processing on the residual sequence to extract the mean, variance, skewness, kurtosis, and trend slope as time-domain statistical features; the second channel performs fast Fourier transform on the periodic fluctuation signal to extract the energy proportion of a preset frequency band as frequency-domain energy features; and the third channel uses a dynamic time warping algorithm to calculate the similarity distance between the actually acquired spectral curve and the spectral curve of the standard qualified product as trajectory anomaly features.

[0018] By adopting the above technical solution, fault symptoms can be comprehensively extracted from three dimensions: time-domain statistical characteristics, frequency-domain energy distribution, and sequence trajectory morphology. This can effectively capture weak fault signals that are difficult to detect from a single dimension, thus improving the completeness of feature engineering.

[0019] Preferably, in the process of the fault diagnosis and health management module outputting fault modes using the residual sequence of the actual monitoring data in the unified data lake and the health status baseline curve, the step of outputting the fault mode based on the extracted features includes: fusing the time-domain statistical features, the frequency-domain energy features, and the trajectory anomaly features to form a multimodal feature vector; inputting the multimodal feature vector into a pre-trained gradient boosting tree classifier, which maps the multimodal feature vector to a predefined fault mode space and outputs the confidence probability of each fault mode; setting a probability threshold, and determining that the fault mode has occurred when the predicted probability of a certain type of fault mode exceeds the probability threshold.

[0020] By adopting the above technical solution, and using ensemble learning algorithms to perform nonlinear fusion and spatial mapping of multimodal features, high-precision classification of complex fault modes is achieved, and quantitative confidence assessment is provided, providing a reliable basis for subsequent decision-making.

[0021] Preferably, during the process of generating control compensation instructions based on the fault mode, the process parameter self-repair decision module performs the rule-based deterministic repair step, which includes: when the fault mode is filter clogging and the confidence probability exceeds a preset threshold, automatically generating the control compensation instruction to control the variable frequency motor of the melt gear pump to increase the speed and maintain the die head pressure within the set value range; and calculating the remaining service life of the filter based on the current differential pressure rise rate and the pump speed increase.

[0022] By adopting the above technical solutions, a deterministic closed-loop compensation strategy is used to maintain production stability for gradual failures with clear mechanisms, such as filter clogging, and predictive maintenance of filter replacement cycles is achieved by combining state simulation.

[0023] Preferably, in the process of the process parameter self-healing decision module generating control compensation instructions according to the fault mode, the step of executing the compensation decision based on reinforcement learning includes: for the quality deviation diagnosed by the fluctuation of the intrinsic properties of the raw materials, calling a deep deterministic policy gradient network as a reinforcement learning compensator; receiving the complete process state vector and the quality deviation value at the current moment as input states, and outputting an action vector containing multi-dimensional adjustment amounts, wherein the action vector corresponds to the adjustment amount of the main screw speed and the set temperature of the reaction zone, to compensate for the rheological property changes caused by the material properties.

[0024] By adopting the above technical solution and utilizing the optimization capability of reinforcement learning in the continuous action space, the nonlinear quality deviation problem caused by batch fluctuations of raw materials was solved, and adaptive dynamic optimization and compensation of multidimensional process parameters were realized.

[0025] Preferably, during the process of the process parameter self-healing decision module issuing the control compensation command to the physical production line layer to execute closed-loop control, the step of executing the safety interaction mechanism includes: before the control compensation command is issued, the edge computing and intelligent control layer verifies whether the control compensation command is within the preset process safety envelope; a highlighted prompt appears on the operation interface and a countdown waiting period is started; if the countdown ends and no rejection operation is received from the operator, the control compensation command is automatically issued and executed.

[0026] By adopting the above technical solutions, a safety mechanism combining automatic control and manual supervision was constructed. By utilizing process envelope constraints and human-machine interaction confirmation, extreme instructions were prevented from being output by the decision model, thus ensuring the safety of the production line operation in unmanned or minimally manned conditions.

[0027] This invention provides an intelligent fault diagnosis and repair system for a fully biodegradable composite material production line. It has the following beneficial effects: 1. This invention utilizes a digital twin engine that deeply integrates mechanistic models and data-driven processes to generate a health baseline curve that dynamically follows changes in operating conditions. Furthermore, it employs residual sequence analysis technology to effectively isolate parameter fluctuations caused by normal process adjustments. This method can keenly detect early, subtle signs of equipment performance degradation, such as filter clogging, transforming fault diagnosis from traditional threshold-over-limit alarms to proactive warnings based on trend analysis. This effectively avoids unplanned downtime and improves the continuity of production line operations.

[0028] 2. This invention utilizes a deep reinforcement learning-based process parameter self-healing decision module to establish an adaptive compensation mechanism for fluctuations in the intrinsic properties of raw materials and degradation in equipment performance. By sensing changes in rheological properties in real time and dynamically optimizing key control variables such as screw speed and temperature zone, the system can effectively suppress the impact of external disturbances on the stability of the extrusion process, ensuring batch stability of key quality indicators such as melt index of fully biodegradable composite materials and improving the yield of superior products.

[0029] 3. This invention achieves high-precision identification and confidence assessment of fault modes through multimodal feature fusion technology. Combined with automatic closed-loop control and remaining life prediction functions under a safety interaction mechanism, it realizes the transformation of production line operation and maintenance mode from passive emergency repair to predictive maintenance. This not only reduces the reliance on the experience of highly qualified field engineers, but also helps to optimize spare parts inventory management and maintenance scheduling, thereby reducing the overall operation and maintenance cost of the production line while ensuring process safety. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the intelligent fault diagnosis and process parameter self-repair system for a fully biodegradable composite material production line according to an embodiment of the present invention; Figure 2 This is a flowchart of the intelligent fault diagnosis and process parameter self-repair method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the digital twin hybrid model according to an embodiment of the present invention; Figure 4 This is a flowchart of the multimodal feature extraction and classification process of the fault diagnosis module in this embodiment of the invention.

[0031] The system comprises: 10. Physical production line layer; 20. Edge computing and intelligent control layer; 21. Data acquisition and preprocessing module; 22. Digital twin engine; 23. Fault diagnosis and health management module; 24. Process parameter self-healing decision module; 25. Human-computer interaction and log module; and 30. Cloud platform data center layer. Detailed Implementation

[0032] The technical solutions in 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.

[0033] See attached document Figure 1 This invention provides an intelligent fault diagnosis and repair system for a fully biodegradable composite material production line. It adopts a layered distributed architecture design, logically and physically divided into a physical production line layer 10, an edge computing and intelligent control layer 20, and a cloud platform data center layer 30. These three layers are interconnected through an industrial communication network, forming a closed-loop control system.

[0034] The physical production line layer 10 constitutes the underlying hardware foundation of the system, encompassing all processing equipment and sensor-execution networks required for the production of fully biodegradable multiphase composite materials. Specifically, the physical production line layer 10 includes a raw material pretreatment unit, a multi-channel precision feeding unit, a reactive extruder unit, and an online monitoring unit. The raw material pretreatment unit is used to dry and mix raw materials such as PBAT, PLA, and plant fibers. The multi-channel precision feeding unit is equipped with several high-precision loss-in-weight balances and liquid mass flow meters, used for the metering and conveying of solid resins, powder fillers, and liquid additives, respectively. The reactive extruder unit, as the core reaction site, typically employs a twin-screw extruder with configurable length-to-diameter ratios and screw combinations. Its barrel is divided axially into multiple independent temperature control zones, each equipped with an electric heater and a cooling medium flow channel. A melt gear pump and a screen changer are located at the end of the extruder to establish a stable pressure-building extrusion process and filter melt impurities.

[0035] To achieve comprehensive monitoring of the production process, a dense network of various sensor types is deployed in the physical production line layer 10. For equipment status data acquisition, torque sensors, three-phase current transformers, and speed encoders are installed at the main drive motor and side feed motor to monitor motor load and operating status. Vacuum sensors and current sensors are installed at the vacuum exhaust port to monitor devolatilization efficiency. High-temperature melt pressure sensors are installed at the inlet and outlet sides of the melt gear pump to measure the pressure before and after the pump in real time; the difference between these two pressure values ​​is used to calculate the pressure difference before and after the filter. For process parameter acquisition, thermocouples are embedded in each temperature zone to provide feedback on the barrel temperature; each loss-in-weight scale and flow meter provides real-time feedback on instantaneous flow rate and cumulative weight data; and the extruder controller provides feedback on the actual rotational speeds of the main screw and side feed screw. For online quality data acquisition, the online detection unit includes a near-infrared spectrometer probe installed at the die head flow channel for non-destructive acquisition of melt spectral information; and an online melt rheometer, which measures the complex viscosity and elastic modulus of the melt in real time via bypass flow. In addition, some high-frequency pressure sensors are used to collect pressure fluctuation signals for subsequent extraction of viscosity characteristic frequency amplitude.

[0036] The edge computing and intelligent control layer 20 is the core processing unit of this system, physically deployed in an industrial server or high-performance edge computing gateway within the workshop. This layer connects to the PLC controllers and smart instruments of the physical production line layer 10 via industrial Ethernet, and performs unified data reading and writing through the OPC UA protocol. The edge computing and intelligent control layer 20 is logically divided into five collaborative modules: a data acquisition and preprocessing module 21, a digital twin engine 22, a fault diagnosis and health management module 23, a process parameter self-healing decision-making module 24, and a human-machine interaction and logging module 25.

[0037] The data acquisition and preprocessing module 21 is used to acquire all time-series data uploaded by the physical production line layer 10 in real time at a sampling frequency of no less than 100Hz. This module integrates a data cleaning algorithm to remove outliers caused by signal interference and uses an interpolation algorithm to align the timestamps of data sources with different sampling rates, constructing a time-series data stream with a uniform format. The digital twin engine 22 is the simulation core of the system, embedding a hybrid model that runs synchronously with the physical production line. This hybrid model includes a mechanistic model layer based on the principle of non-isothermal non-Newtonian fluid transport and a data-driven layer based on an attention-enhanced long short-term memory network. The digital twin engine 22 is used to calculate the theoretical state of the production line under ideal or normal operating conditions in real time based on the current process setpoints and boundary conditions, generating a health state baseline curve containing theoretical torque, theoretical pressure, and theoretical viscosity.

[0038] The fault diagnosis and health management module 23 receives actual monitoring data from the data acquisition and preprocessing module 21 and baseline health status data from the digital twin engine 22. This module first calculates the residual sequence between the actual and baseline values, and then extracts time-domain statistical features, frequency-domain energy features, and trajectory anomaly features from the raw data and the residual sequence. The fault diagnosis and health management module 23 internally deploys a pre-trained machine learning classifier, such as a gradient boosting tree classifier, to output the fault modes and their confidence probabilities in the current system based on the extracted multimodal feature vectors.

[0039] The process parameter self-healing decision module 24 generates corresponding control compensation instructions based on the diagnostic results output by the fault diagnosis and health management module 23 or the quality prediction deviation output by the digital twin engine 22. This module integrates a repair rule base and a reinforcement learning compensator. For clear mechanical or physical faults, such as filter clogging, this module calls the preset rule base to generate deterministic control instructions. For complex raw material fluctuations or quality deviations, this module calls a deep deterministic strategy gradient network to calculate the optimal process parameter adjustment strategy, such as adjusting the temperature setpoint of a specific temperature zone or changing the screw speed.

[0040] The human-machine interaction and log module 25 provides a visual operating interface, displaying the real-time health status of the production line, fault warning information, self-repair decision suggestions, and historical operation logs to the operator. This module features a safety interaction mechanism; before executing critical self-repair actions, a highlighted prompt will appear on the interface, awaiting operator confirmation, while ensuring that all automatic adjustment commands are confined within a preset process safety envelope.

[0041] The cloud platform data center layer 30 is located at the top layer of the system architecture and communicates with the edge computing and intelligent control layer 20 via IoT protocols such as MQTT. The cloud platform data center layer 30 does not directly participate in millisecond-level real-time control, but rather undertakes the long-term storage of massive historical data, offline training of complex models, and version update tasks. Utilizing the accumulated large-scale historical production data, the cloud platform data center layer 30 periodically retrains the neural network model in the digital twin engine 22 and the classifier in the fault diagnosis and health management module 23, and then distributes the optimized model parameters to the edge computing and intelligent control layer 20 to continuously improve the system's diagnostic accuracy and decision-making capabilities.

[0042] See attached document Figure 2This invention provides an intelligent fault diagnosis and repair method for a fully biodegradable composite material production line. The method first executes step S1, namely, the full-dimensional data perception and high-throughput fusion step. In this step, the data acquisition and preprocessing module 21 establishes communication connections with each of the 10 units in the physical production line layer, constructing a high-speed data transmission channel through Industrial Ethernet and the OPC UA protocol. To capture the rapidly changing rheological properties and transient fault characteristics of the fully biodegradable material during reactive extrusion, the data acquisition and preprocessing module 21 is used to poll and collect time-series data from over 200 data points across the entire production line in real time at a frequency of at least 100 Hz. Addressing the differences in response speed and transmission delay among different sensors, this module uses a unified clock source to tag all collected data packets, performing microsecond-level timestamp alignment operations, thereby aggregating heterogeneous multi-source data into a unified data lake that is strictly synchronized in the time dimension.

[0043] The data collected in step S1 is mainly divided into three dimensions: equipment status time-series data, process parameter time-series data, and online quality time-series data. Equipment status time-series data primarily reflects the physical health and operating load of the production line's hardware components. Specifically, the system collects real-time operating parameters of the main drive motor and side feed motor, including output torque, three-phase current, and real-time speed. These parameters directly reflect the mechanical load and energy consumption status of the extruder. The system also collects duty cycle data of the heaters in each temperature zone of the extruder barrel to assess the working efficiency of the heating elements and the heat compensation requirements. In the melt conveying stage, the system collects the values ​​from the inlet and outlet pressure sensors of the melt gear pump and calculates the pressure difference data before and after the filter screen. This pressure difference data is a key indicator for judging the degree of filter screen clogging. Furthermore, the operating current and vacuum level of the vacuum exhaust system are also recorded in real-time to monitor the stability of the volatile matter removal process.

[0044] Process parameter timing data primarily reflects the set conditions and actual control response during production, forming the basis for the system's understanding of the current operating conditions. The data acquisition and preprocessing module 21 reads the instantaneous flow rate and cumulative weight data from each loss-in-weight scale in real time, as well as the readings from the liquid mass flow meter, thereby accurately monitoring the actual proportioning accuracy of solid resin, powder filler, and liquid additives. For temperature control, the system synchronously records the temperature setpoints and actual temperature values ​​fed back by thermocouples for each temperature zone of the extruder to analyze the hysteresis and fluctuations of temperature control. Simultaneously, the setpoints and feedback values ​​of the main screw speed and the side feed screw speed are also included in the data acquisition range for subsequent analysis of material shearing history and residence time distribution.

[0045] Online quality time-series data is a key input for achieving high-precision diagnosis and closed-loop quality control in this invention, primarily derived from advanced sensors in the online detection unit. The system acquires full-spectrum data or characteristic peak data of specific bands of the melt in real time using an online near-infrared spectrometer. This spectral information contains fingerprint characteristics of the material's chemical composition and molecular structure. Simultaneously, the system collects complex viscosity and elastic modulus data measured by an online melt rheometer, directly characterizing the material's rheological properties. Furthermore, the system performs in-depth analysis of the signals from the high-frequency melt pressure sensor, acquiring viscosity characteristic frequency amplitudes extracted from melt pressure oscillation signals through Fourier transform. These multi-dimensional online quality data, strictly aligned with the aforementioned equipment status data and process parameter data on the time axis, collectively constitute the complete input vector required for subsequent digital twin modeling and fault diagnosis algorithms.

[0046] See attached document Figure 3 and attached Figure 2 The digital twin hybrid model of this invention is in Figure 2 The digital twin engine 22 shown is used to execute step S2, namely the hybrid modeling and health baseline generation step based on the digital twin. This invention achieves deep simulation and state prediction of the production process by constructing and running a digital twin synchronized with the physical production line in real time. This digital twin adopts a hybrid architecture combining mechanism-guided and data-driven approaches, with parallel computation of the mechanism model path and the data-driven path, ultimately fusing and outputting a health status baseline under the current operating conditions.

[0047] In the mechanistic model pathway, the digital twin engine 22 establishes a simplified one-dimensional distributed parameter model of the screw extrusion process based on the transport principle of non-isothermal non-Newtonian fluids. This model discretizes the screw along the axial direction into multiple finite control volumes and, based on the laws of conservation of mass, momentum, and energy, calculates in real-time the theoretical melt state distribution along the screw axis under the current process settings. To accurately describe the rheological behavior of fully biodegradable multiphase composite materials in complex flow fields, a rheological equation of state is introduced into the mechanistic model layer. In this embodiment, a power-law model is used to describe the functional relationship between the material's apparent viscosity, shear rate, and temperature. The core formula is as follows: ; in, γ represents the apparent viscosity of the material, and γ represents the shear rate. Represents absolute temperature. The activation energy of viscous flow represents the material. Represents the ideal gas constant. Represents non-Newtonian exponents. This represents a material constant. Among them, the activation energy for viscous flow is... Non-Newtonian exponents and material constants Fixed parameters obtained through prior offline rheological experiments are input into the model as the basis for calculation. The mechanistic model calculates the local shear rate of each micro-element based on real-time collected screw rotation speed, and combines this with temperature feedback values ​​from each temperature zone to calculate the theoretical viscosity distribution along the screw axis using the aforementioned formula. Furthermore, the model outputs the theoretical melt pressure curve and theoretical torque curve under the current operating conditions by integrating the pressure gradient and viscous dissipation along the flow channel. These theoretical curves represent the system response under ideal equipment conditions and standard material properties, constituting a physical benchmark for judging the health status of the equipment.

[0048] In the data-driven pathway, the digital twin engine 22 employs an attention-enhanced Long Short-Term Memory (LSTM) network to train and extrapolate massive amounts of historical data under normal operating conditions. This network receives time-series data on equipment status and process parameters collected in step S1 as input vectors. These input vectors are processed by the LSM network units to capture long-term dependencies and dynamic lag effects in the time-series data, such as the delayed impact of temperature changes on melt quality. Subsequently, the attention mechanism module weights the hidden layer states output by the LSM network, automatically assigning higher weights to the input features most relevant to the current output, thereby identifying the nonlinear impact weights of key process parameters on quality indicators. The data-driven pathway aims to learn complex mapping relationships that are difficult for mechanistic models to capture, such as the specific mixing effect of screw combinations and the impact of trace impurities on the reaction process, ultimately outputting predicted values ​​for key quality parameters, including component content predicted by near-infrared spectroscopy and viscosity values ​​measured by a rheometer.

[0049] During the health baseline generation phase, the digital twin engine 22 integrates the calculation results from the mechanistic model with the prediction results from the data-driven model. Using the theoretical curves calculated by the mechanistic model as a framework, and combining them with the data-driven model's corrective predictions for specific operating conditions, the system generates a set of health baseline curves that include theoretical torque curves, theoretical melt pressure curves, and theoretical viscosity values. This health baseline curve dynamically reflects the standard state that a fully healthy production line should exhibit under the current specific raw material batch, process settings, and environmental conditions, providing a precise benchmark for subsequent fault diagnosis.

[0050] See attached document Figure 4 and attached Figure 2 The multimodal feature extraction and classification process of the fault diagnosis module in this embodiment of the invention is as follows: Figure 2 The fault diagnosis and health management module 23 shown is executed to implement step S3, namely the multimodal feature extraction and dynamic fault diagnosis step. This invention performs deep processing on the health benchmark generated from the original sensing data and digital twin to construct a multi-dimensional feature space, and utilizes machine learning algorithms to identify potential equipment faults or process anomalies.

[0051] During the residual sequence generation phase, the fault diagnosis and health management module 23 receives real-time actual parameters from the data acquisition module and corresponding health baseline values ​​from the digital twin engine 22. The system calculates the residual sequence for key variables, including but not limited to main engine torque, melt pump inlet pressure, and outlet pressure. The calculation of the residual sequence follows the formula below: ; in, express The residual value at time step, express The actual physical quantity values ​​collected by the time sensor. express The theoretical health baseline value is calculated from the real-time digital twin. By calculating the residuals, the system can isolate parameter fluctuations caused by normal process setting adjustments, such as changes in screw speed or feed rate, thereby highlighting the deviation components caused by equipment performance degradation or abnormal disturbances.

[0052] In the multimodal feature extraction stage, the system extracts features from the raw data and residual sequence in parallel through three channels to construct a high-dimensional feature vector for fault diagnosis. The first channel is the time-domain statistical feature extraction channel. The system performs windowing processing on the residual sequence and calculates statistical indicators within the window, including mean, variance, skewness, and kurtosis. The mean reflects the central trend of the deviation, the variance reflects the severity of fluctuations, and the skewness and kurtosis are used to capture sudden shocks in non-Gaussian distributions. In addition, the system performs linear regression analysis on the residual data within the past 5-minute sliding window and extracts the slope of the regression line as a trend feature. This slope feature is indicative of gradual faults such as progressive filter blockage or slow wear of screw components, and can quantify the rate at which parameters drift over time.

[0053] The second channel is the frequency domain energy feature extraction channel, primarily targeting signals with periodic fluctuation characteristics, especially the main engine torque signal and melt pressure pulsation signal. The system performs a Fast Fourier Transform on these time-series signals, converting the time-domain signals into frequency-domain spectra. The system predefines specific frequency bands related to the mechanical structure, such as those related to the screw meshing frequency and its harmonics. The system calculates the proportion of energy within these specific frequency bands in the total energy as frequency domain features. When wear or scraping occurs in the screw components, the energy amplitude at specific frequencies changes significantly; therefore, this feature is mainly used to identify physical damage to mechanical parts and abnormal fit clearances.

[0054] The third channel is the trajectory anomaly feature extraction channel, primarily processing high-dimensional online quality inspection data, especially full-spectrum data acquired by a near-infrared spectrometer. Due to the time-varying nature of the production process, spectral curves may undergo nonlinear scaling or translation along the time axis. The system employs a dynamic time warping algorithm to calculate the similarity distance between the currently acquired near-infrared spectral curve and the near-infrared spectral curve of a standard qualified product. By finding the optimal matching path between the two time series, the dynamic time warping algorithm can avoid the influence of time axis distortion and accurately quantify the degree of difference between the current chemical composition or molecular structure of the material and the standard state. This similarity distance, as a trajectory anomaly feature, can sensitively detect quality shifts caused by raw material batch fluctuations or insufficient reaction.

[0055] During the dynamic fault diagnosis phase, the system fuses the aforementioned time-domain statistical features, frequency-domain energy features, and trajectory anomaly features to form a multimodal feature vector. This feature vector is input into a pre-trained lightweight gradient boosting tree classifier. The gradient boosting tree classifier is an ensemble learning model that constructs a strong classifier by combining multiple weak classifiers, i.e., decision trees, and has the ability to handle nonlinear relationships and high-dimensional data. The classifier maps the input feature vector to a predefined fault mode space and outputs the probability of occurrence for each fault mode. Fault modes cover equipment faults such as filter clogging and kneading block wear, as well as process anomalies such as high resin intrinsic viscosity and excessive moisture content. The system sets a probability threshold; when the predicted probability of a certain type of fault exceeds the threshold, the fault is determined to have occurred. This classifier is trained offline using historically accumulated fault case data and has online learning capabilities, continuously updating model parameters as new fault samples appear to adapt to new operating conditions.

[0056] See attached document Figure 2 This invention provides a detailed implementation of the process parameter self-repair decision module 24 and the human-machine interaction and log module 25 of an intelligent fault diagnosis and process parameter self-repair system for a fully biodegradable composite material production line. This part corresponds to... Figure 2 The system architecture shown illustrates the decision-making and execution levels, specifically step S4 in the execution method flow, which is the process parameter self-repair decision-making and execution based on reinforcement learning. The process parameter self-repair decision module 24 integrates a preset repair rule base and a deep learning-based reinforcement learning compensator, aiming to implement hierarchical control strategies for different types of production anomalies.

[0057] For deterministic equipment faults or process deviations caused by single variables identified by the fault diagnosis module, the process parameter self-repair decision module 24 adopts a rule-based deterministic repair strategy. This repair rule base pre-stores standard handling logic for specific fault modes. When the probability of a specific fault output by the fault diagnosis module exceeds a preset threshold, the system triggers a corresponding repair action. For example, when the diagnosis result shows that the probability of filter clogging exceeds 80%, the system determines that the melt flow resistance has increased, leading to a drop or fluctuation in die pressure. At this time, the repair decision module automatically generates control commands to control the variable frequency motor of the melt gear pump, gradually increasing its speed at a rate of 0.5 revolutions per minute. The goal of this adjustment is to maintain the die pressure within a set value of ±0.2 MPa, thereby ensuring the stability of the extrusion volume. Simultaneously, the system calculates the remaining service life of the filter based on the current rate of pressure increase and pump speed increase, and sends this data to the maintenance management system to arrange subsequent filter replacement plans.

[0058] For complex quality deviations caused by fluctuations in the intrinsic properties of raw materials, such as when the predicted melt flow index is consistently below the lower limit of the specification and the diagnostic cause is high intrinsic viscosity of the resin, simple rule-based control is insufficient to achieve nonlinear coordination of multiple parameters. In this case, the process parameter self-healing decision module 24 invokes a deep deterministic policy gradient network as a reinforcement learning compensator. The deep deterministic policy gradient network is an algorithm based on an actor-critic architecture, suitable for control problems in continuous action spaces. This network receives the complete process state vector and quality deviation value at the current moment as input states. The process state vector includes data such as the actual temperature of each temperature zone, screw speed, feed rate, and melt pressure; the quality deviation value is the difference between the online predicted quality index and the target setpoint.

[0059] The deep deterministic policy gradient network, after extensive interactive training in a digital twin simulation environment, learns the complex mapping relationship between various process parameters and product quality. Based on the current input state, the network outputs an action vector containing multi-dimensional adjustments. This action vector directly corresponds to the setpoint adjustments of key actuators on the production line, such as increasing the main screw speed by 3 revolutions per minute while simultaneously raising the setpoint temperature of the reaction zone (zones seven to nine) by 2 degrees Celsius. By increasing the speed to increase shear heat and raising the temperature of the reaction zone, the system can effectively reduce the flow resistance of high-viscosity materials, thereby compensating for the decrease in melt index caused by the high intrinsic viscosity of the raw materials, and ensuring that the rheological properties of the final product return to the center of the target range.

[0060] To ensure the safety and reliability of the self-healing process, all adjustment commands generated by the process parameter self-healing decision module 24 must be verified and executed through a secure interaction mechanism between the human-machine interface and the log module 25. The system pre-sets a strict process safety envelope, limiting the absolute upper and lower limits of allowable adjustments for each process parameter to prevent parameter adjustments from exceeding the equipment's capacity or causing thermal degradation of materials. Before executing any self-healing action, the system will display a highlighted message notification on the operator's human-machine interface. This notification details the current diagnostic conclusion, the system's suggested parameter adjustment values, and the expected quality improvement effect. Simultaneously, the system initiates a 10-second countdown waiting period for confirmation. During this period, the operator has the highest priority veto right and can cancel or modify the system's suggested action at any time. If the countdown ends and the operator does not intervene, the system will automatically send the adjustment command to the underlying programmable logic controller for execution, thereby achieving closed-loop control. This mechanism maximizes the system's automated response speed while ensuring production safety.

[0061] To verify the technical effectiveness of the intelligent fault diagnosis and process parameter self-repair system and method for the fully biodegradable composite material production line of this invention in actual industrial production, the following describes in detail the specific operation process and parameter response of the system under two typical operating conditions. This section is based on the foregoing Figures 1 to 4 The system architecture and methodology shown demonstrate the system's closed-loop control capabilities in handling raw material batch fluctuations and gradual equipment failures.

[0062] In the first implementation scenario, the production line is continuously producing a composite material of poly(butylene adipate / terephthalate), polylactic acid, and bamboo powder. During this process, the raw material supply system switches to a new batch of polylactic acid resin, which has an intrinsic viscosity 0.1 deciliters per gram higher than the previous batch. At the initial stage of the raw material switchover, the data acquisition and preprocessing module 21 acquires spectral data from an online near-infrared spectrometer and rheological data from an online melt rheometer in real time. The attention-enhanced long short-term memory network in the digital twin engine 22 uses this data as input, combined with the current process settings, to perform forward-looking predictions. Fifteen minutes before the actual output of the finished product, the system predicts that the melt flow index of the final product will trend below the lower limit of the specification.

[0063] Subsequently, the fault diagnosis and health management module 23 performs a source analysis on the predicted quality anomalies. The module calculates the residual sequence between the actual value of the host torque and the health benchmark value, and finds that the torque residual shows a slight positive deviation, indicating that the melt viscosity is higher than theoretically expected. Considering the multimodal characteristics of the positive torque residual and the lower predicted melt index, the gradient boosting tree classifier outputs a diagnostic result, determining that the cause of the anomaly is the high intrinsic viscosity of polylactic acid resin, with a confidence probability of 88%.

[0064] Based on the diagnostic conclusion, the process parameter self-healing decision module 24 activates a deep deterministic strategy gradient network for compensation decision-making. This network calculates the optimal process parameter adjustment vector based on the current state space input. The system then issues control commands to increase the main screw speed from 450 rpm to 454 rpm, and simultaneously raise the set temperature of the reaction zone (eighth temperature zone) from 162 degrees Celsius to 164 degrees Celsius. The increase in main screw speed increases shear heat generation, and the increase in temperature accelerates the reaction rate and directly reduces melt viscosity; both work synergistically to compensate for the decreased fluidity caused by high-viscosity resin. After 20 minutes of these adjustments, the online predicted melt index value smoothly returned to the center of the target range. Throughout the self-healing process, the production line maintained continuous and stable operation without producing any defective products. In contrast, under the traditional fixed-parameter production mode without this system, the same fluctuation in raw material volume would cause the melt index of the entire batch of products to remain below the standard, ultimately resulting in product downgrading.

[0065] In the second implementation scenario, the system performs early warning and adaptive compensation for progressive clogging of the melt filter screen. The system continuously monitors pressure sensor data at the inlet and outlet of the melt gear pump and calculates the pressure difference before and after the filter. In the early stages of production, as impurities accumulate on the filter, the actual pressure difference begins to rise slowly. At a certain point, although the absolute pressure difference has not yet reached the alarm threshold of a traditional control system, the system detects that the residual between the actual pressure difference trajectory and the health baseline curve generated by the digital twin exceeds a preset sensitivity threshold. The system then issues an early warning signal, indicating a potential risk of increased flow resistance.

[0066] As the operating time progresses, the fault diagnosis module extracts the slope and frequency domain characteristics of the differential pressure residual sequence, confirming that the fault mode is filter blockage, with the probability judgment value rising to over 90%. At this point, the process parameter self-repair decision module 24 intervenes in the control, initiating an adaptive compensation strategy. To maintain stable die pressure and prevent extrusion volume fluctuations caused by increased filter resistance, the system automatically adjusts the speed of the melt gear pump. The melt pump speed curve exhibits a stepped or sloping upward trend corresponding to the increasing differential pressure. The system gradually increases the pump speed through closed-loop control to offset the pressure loss caused by increased flow resistance, ensuring that the die pressure is always maintained within ±0.2 MPa of the process setpoint.

[0067] While performing adaptive compensation, the system calculates the remaining service life of the filter in real time based on the current rate of pressure rise and the adjustment margin of the pump speed. When the predicted remaining service life approaches the critical value or the pump speed reaches the equipment's permissible safety limit, the system generates a planned shutdown suggestion, guiding maintenance personnel to perform filter replacement operations within an appropriate time window. This state-based maintenance strategy avoids over-pressure alarms and emergency shutdowns caused by complete filter blockage. In contrast, without this system, the pressure differential curve typically continues to rise without intervention until the safety interlock is triggered, leading to unplanned shutdowns and prolonged material degradation within the barrel.

[0068] Based on the above statistical data from actual operation, the system of this invention has achieved significant technical benefits. Regarding fault early warning, for progressive faults such as filter clogging and screw wear, the system can issue warnings 2 to 8 hours in advance, reducing unplanned downtime by more than 70%. In terms of product quality control, through proactive feedforward compensation for raw material fluctuations and equipment performance degradation, the batch-to-batch standard deviation of key product indicators such as melt flow index has been reduced by more than 50%, and the rate of superior products has been consistently maintained above 99.5%. Regarding maintenance costs, the predictive maintenance strategy reduces the frequency of emergency repairs and spare parts inventory pressure, while also reducing reliance on on-site troubleshooting by senior process engineers, resulting in an overall maintenance cost reduction of 25%.

Claims

1. An intelligent fault diagnosis and repair system for a fully biodegradable composite material production line, characterized in that, It includes a physical production line layer (10), an edge computing and intelligent control layer (20), and a cloud platform data center layer (30). The physical production line layer (10) includes a reactive extrusion unit. The physical production line layer (10) collects equipment status time-series data, process parameter time-series data and online quality time-series data reflecting the reactive extrusion process of fully biodegradable materials through a sensor network. The edge computing and intelligent control layer (20) includes: The data acquisition and preprocessing module (21) is used to receive the equipment status time series data, the process parameter time series data and the online quality time series data to construct a unified data lake with time synchronization; A digital twin engine (22) is used to generate a health status baseline curve based on the unified data lake by utilizing a hybrid model of mechanism and data driven by the principle of non-isothermal non-Newtonian fluid transport. The fault diagnosis and health management module (23) is used to output fault modes using the residual sequence of the actual monitoring data in the unified data lake and the health status baseline curve; The process parameter self-repair decision module (24) is used to generate control compensation instructions based on the fault mode and send them to the physical production line layer (10) to perform closed-loop control.

2. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 1, characterized in that, The process by which the data acquisition and preprocessing module (21) receives the equipment status time-series data, the process parameter time-series data, and the online quality time-series data to construct a time-synchronized unified data lake includes: Using a unified clock source, the equipment status timing data, process parameter timing data and online quality timing data from the physical production line layer (10) are tagged. Using an interpolation algorithm, microsecond-level timestamp alignment operations are performed on data sources with different sampling rates, and heterogeneous multi-source data are aggregated to construct a unified data lake that is strictly synchronized in the time dimension.

3. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 1, characterized in that, The digital twin engine (22) generates a health status baseline curve based on the unified data lake using a hybrid model of mechanism and data driven by the principle of non-isothermal non-Newtonian fluid transport. The steps of the mechanism model pathway include: A one-dimensional distributed parameter model of the screw extrusion process is established, and the screw is discretized into multiple finite control volumes along the axial direction; Based on the real-time collected screw speed and temperature feedback values, the theoretical viscosity distribution along the screw axis is calculated using the rheological state equation that describes the relationship between the material's apparent viscosity, shear rate, and temperature function. By integrating the pressure gradient and viscous dissipation along the flow channel, the theoretical melt pressure curve and theoretical torque curve under the current operating conditions are output as the physical reference for the health state reference curve.

4. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 3, characterized in that, The digital twin engine (22) generates a health status baseline curve based on the unified data lake using a hybrid model that combines the mechanism of non-isothermal non-Newtonian fluid transport principle with a data-driven approach. The steps of running the data-driven pathway include: An attention-enhanced Long Short-Term Memory network is used to receive the device status timing data and the process parameter timing data from the unified data lake as input vectors. The system utilizes a long short-term memory network unit to capture long-term dependencies and dynamic lag effects in the input vector, and uses an attention mechanism module to identify the nonlinear influence weights of key process parameters on quality indicators, outputting predicted values ​​of key quality parameters to correct the physical benchmark.

5. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 1, characterized in that, In the process of the fault diagnosis and health management module (23) outputting the fault mode using the residual sequence of the actual monitoring data in the unified data lake and the health status baseline curve, the steps for calculating the residual sequence include: Receive the actual monitoring data from the data acquisition and preprocessing module (21) and the health status baseline data from the digital twin engine (22); For key variables such as main unit torque, melt pump inlet pressure and outlet pressure, the difference between the actual value and the theoretical health benchmark value is calculated. By removing parameter fluctuations caused by normal process setting adjustments, the residual sequence reflecting equipment performance degradation or abnormal disturbances is generated.

6. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 5, characterized in that, In the process of the fault diagnosis and health management module (23) outputting fault modes using the residual sequence of the actual monitoring data in the unified data lake and the health status baseline curve, the step of extracting features from the residual sequence includes: Processed in parallel through three channels: The first channel performs windowing processing on the residual sequence to extract the mean, variance, skewness, kurtosis, and trend slope as time-domain statistical features; The second channel performs a fast Fourier transform on the periodic fluctuation signal to extract the energy ratio of the preset frequency band as the frequency domain energy feature; The third channel uses a dynamic time warping algorithm to calculate the similarity distance between the actual acquired spectral curve and the spectral curve of the standard qualified product as a trajectory anomaly feature.

7. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 6, characterized in that, In the process of the fault diagnosis and health management module (23) outputting the fault mode based on the extracted features during the process of using the residual sequence of the actual monitoring data in the unified data lake and the health status baseline curve to output the fault mode, the steps include: The time-domain statistical features, the frequency-domain energy features, and the trajectory anomaly features are fused to form a multimodal feature vector; The multimodal feature vector is input into a pre-trained gradient boosting tree classifier, which maps the multimodal feature vector to a predefined fault mode space and outputs the confidence probability of each fault mode. A probability threshold is set, and when the predicted probability of a certain type of fault mode exceeds the probability threshold, the fault mode is determined to have occurred.

8. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 1, characterized in that, During the process of generating control compensation instructions based on the fault mode, the process parameter self-repair decision module (24) performs the following steps for deterministic repair based on the rule base: When the fault mode is filter blockage and the confidence probability exceeds the preset threshold, the control compensation command is automatically generated to control the variable frequency motor of the melt gear pump to increase the speed and maintain the die pressure within the set value range. The remaining service life of the filter is calculated based on the current rate of increase in differential pressure and the magnitude of increase in pump speed.

9. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 1, characterized in that, In the process of generating control compensation instructions based on the fault mode, the self-healing decision module (24) of the process parameters performs the following steps for executing compensation decisions based on reinforcement learning: For the quality deviations diagnosed as being caused by fluctuations in the intrinsic properties of raw materials, a deep deterministic policy gradient network is invoked as a reinforcement learning compensator. The system receives the complete process state vector and quality deviation value at the current moment as input states, and outputs an action vector containing multi-dimensional adjustment amounts. The action vector corresponds to the adjustment amount of the main screw speed and the set temperature of the reaction zone, compensating for changes in rheological properties caused by material characteristics.

10. The intelligent fault diagnosis and repair system for a fully biodegradable composite material production line according to claim 1, characterized in that, During the process of the process parameter self-healing decision module (24) issuing the control compensation command to the physical production line layer (10) to execute closed-loop control, the steps of executing the safety interaction mechanism include: Before the control compensation command is issued, the edge computing and intelligent control layer (20) verifies whether the control compensation command is within the preset process safety envelope range; A highlighted prompt will pop up on the operation interface and a countdown will start to wait for confirmation. If the countdown ends and no rejection operation is received from the operator, the control compensation command will be automatically issued and executed.

Images