Multi-sensor adaptive control system and method for injection stretch blow molding machine

Abstract

The present invention relates to the technical field of industrial control systems, specifically to a multi-sensor adaptive control system and method for an injection stretch blow molding machine, comprising a sensor module that collects multimodal data; an intelligent diagnosis module that uses a layered architecture for rapid detection and precise working condition identification, and outputs a diagnostic identifier and confidence level; a decision module that integrates expert knowledge and online learning results, and determines an optimized response strategy in combination with confidence levels; a control parameter set that is dynamically loaded according to the strategy; an adaptive control module that calculates the original instruction based on the parameter set and data; a final signal generated by smoothing the instruction based on the confidence level; and an execution interface module that sends the signal to an actuator to implement adaptive control. The present invention also discloses a corresponding method. The system and method can achieve accurate diagnosis, differentiated decision-making, and precise strategy execution, and ensure smooth control transition, significantly improving control performance, product quality, and operational robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial control systems, and in particular to a multi-sensor adaptive control system and method for an injection stretch blow molding machine. Background Art

[0002] In the automated control of complex industrial processes, the use of program-based control systems to automate basic processes and the introduction of adaptive control technologies to address process parameter changes and external disturbances have become common approaches to improving production stability and efficiency. The application of multi-sensor technology also makes it possible to obtain richer process information.

[0003] However, despite progress, existing technologies still face a significant core technical bottleneck when dealing with highly complex, nonlinear, time-varying industrial processes that are susceptible to a variety of unexpected disturbances. This bottleneck is particularly prominent in precision molding processes such as injection stretch blow molding, which require extremely high product accuracy and consistency. Specifically, current adaptive control systems, even those leveraging multi-sensor information, typically lack the ability to deeply understand and distinguish between different operating conditions. This makes it difficult for systems to intelligently identify the root causes of performance deviations from complex data.

[0004] Precisely because of this lack of differentiated cognitive capabilities based on precise operating condition diagnosis, existing adaptive control adjustment strategies often appear overly stereotyped. They struggle to dynamically and specifically adjust their core control logic, optimization objectives, and adaptation mechanisms based on specific problem root causes diagnosed in real time. This blind and crude adaptability significantly limits the optimal performance and long-term robustness of control systems in real, volatile industrial environments. Therefore, developing a control method that can accurately diagnose operating conditions and implement intelligent, differentiated adaptive adjustments based on these conditions has become a critical and pressing issue in the field of advanced industrial control.

[0005] Therefore, a multi-sensor adaptive control system and method for injection stretch blow molding machine is proposed. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-sensor adaptive control system and method for an injection stretch blow molding machine. By achieving accurate diagnosis, differentiated decision-making, and precise execution of strategies, and ensuring smooth control transition, the control performance, product quality, and operational robustness are significantly improved. First, multimodal data is collected through the sensor module. Then, a hierarchical architecture is used by the intelligent diagnosis module to perform rapid detection and fine working condition identification, outputting diagnostic identification and confidence. Next, the decision module integrates expert knowledge and online learning results, and determines the optimized response strategy based on the confidence. The control parameter set is dynamically loaded according to the strategy. Subsequently, the adaptive control module calculates the original instruction based on the parameter set and data. The instruction is smoothed according to the confidence to generate the final signal. Finally, the signal is sent to the actuator through the execution interface module.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] Multi-sensor adaptive control system for injection stretch blow molding machine, including:

[0009] Sensor module, collects multimodal process data and performs preprocessing;

[0010] Data management module, which stores and retrieves process data, models, strategies, knowledge bases and learning results required for operation;

[0011] Interaction module, monitors system status and receives parameter input;

[0012] The intelligent diagnostic module analyzes data from the sensor module, uses the machine learning model stored in the data management module to identify operating conditions, and generates a diagnostic identifier and diagnostic confidence level;

[0013] The decision module receives the diagnostic identifier and combines the expert knowledge stored in the data management module with the online learning results to determine the optimized adaptive response strategy; based on the determined response strategy, it dynamically adjusts the relevant control parameters used for subsequent calculations;

[0014] The adaptive control module calculates the original control command based on the adjusted relevant control parameters and the data from the sensor module; based on the diagnostic confidence and the original control command, generates a final control signal that is smoothed to ensure a stable transition;

[0015] The execution interface module receives the final control signal and sends it to the corresponding actuator of the equipment.

[0016] Preferably, the intelligent diagnostic module adopts a hierarchical diagnostic structure including: a first diagnostic layer for rapid anomaly detection and a second diagnostic layer for fine working condition classification; the intelligent diagnostic module integrates an explainable artificial intelligence analysis function to provide explanatory information on input features and sensor sources that have a significant impact on the results after diagnosis.

[0017] Preferably, the adaptive response strategy for determining optimization includes:

[0018] Retrieving a benchmark response strategy that matches the current diagnostic condition from an expert knowledge base stored in a data management module based on the diagnostic identifier generated by the intelligent diagnostic module;

[0019] Inputting the diagnostic identifier as status information into an online learning module stored in the data management module to obtain a recommended response strategy for the current diagnostic operating condition learned by the online learning module based on historical operating data and long-term performance feedback;

[0020] Based on a preset fusion algorithm and taking into account the confidence score of the diagnostic identifier, the baseline coping strategy and the recommended coping strategy are integrated to generate an optimized adaptive coping strategy as a final output.

[0021] Preferably, the preset fusion algorithm adopts a dynamic weighted fusion method, and the logic for determining its weight coefficient includes: preliminarily determining a baseline weight value based on the current diagnostic confidence score, and using the real-time analysis results of historical performance data to generate an adjustment factor, which is used to adjust the baseline weight value.

[0022] Preferably, the decision module implements dynamic adjustment including:

[0023] The data management module pre-stores a plurality of different control parameter configuration sets corresponding to different response strategy identifiers; the decision module retrieves and loads the corresponding complete parameter configuration set from the data management module according to the received current response strategy identifier.

[0024] Preferably, the adaptive control module adopts a model predictive control algorithm, which calculates the original control instructions by solving an online optimization problem in each control cycle. The objective function and constraints of the optimization problem are set in real time based on the relevant control parameter configuration set retrieved and loaded by the decision module.

[0025] Preferably, the process of the adaptive control module generating the final control signal includes:

[0026] receiving the original control command calculated by the adaptive control module and receiving the diagnostic confidence score generated by the intelligent diagnostic module;

[0027] Based on the received diagnostic confidence score, dynamically determining key adjustment parameters required for signal smoothing in a subsequent step in real time; the key adjustment parameter determination logic is configured such that a lower diagnostic confidence score corresponds to a parameter value calculated to result in a smoother control signal change;

[0028] The key adjustment parameters are dynamically determined in real time, and signal smoothing processing is performed on the original control instructions;

[0029] The signal obtained after processing is output to the execution interface module as the final control signal.

[0030] Preferably, the multi-sensor adaptive control method for an injection stretch blow molding machine comprises:

[0031] Collect multimodal process data and perform preprocessing; analyze data from sensor modules, use a layered architecture for rapid detection and fine-grained condition identification, and output diagnostic identification and confidence levels; integrate expert knowledge with online learning results, and determine an optimized response strategy based on confidence levels to dynamically load control parameter sets according to the strategy; calculate original instructions based on the parameter set and data, and generate a final signal based on the confidence level smoothing instructions; apply the final control signal to the actuator of the target device to achieve adaptive control.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention utilizes a layered architecture, multimodal fusion, and an intelligent diagnostic module with explainable artificial intelligence, significantly enhancing the ability to extract information from complex sensor data. It can accurately identify specific operating conditions that are difficult to distinguish with existing technologies, and provide reliable confidence assessments and interpretations. This improves the shortcomings of unclear diagnosis and lack of deep situational awareness, providing an accurate and reliable foundation for subsequent differentiated adaptive adjustments.

[0034] 2. The decision-making module constructed by this invention, which integrates expert knowledge and online learning, intelligently transforms the precise information output by the diagnostic module into the most optimized response strategy. This mechanism dynamically selects based on specific diagnostic results and continuously optimizes using historical feedback. This overcomes the problem of existing control systems' modeled adjustment strategies and lack of differentiated response capabilities, ensuring the targeted and effective control measures.

[0035] 3. Through dynamic target management, the application of model predictive control algorithms, and smooth transition execution with confidence adjustment, this invention accurately and stably implements the differentiated strategies of upper-level decision-making into specific equipment control operations. This complete, intelligent, and adaptive closed-loop system, from diagnosis and decision-making to execution, improves the system's overall performance in suppressing complex disturbances and ensuring product quality consistency, ultimately achieving the goal of intelligent, differentiated, and highly robust adaptive regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the structure of a multi-sensor adaptive control system for an injection stretch blow molding machine according to an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the structure of the intelligent diagnosis module according to an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of the flow of a multi-sensor adaptive control method for an injection stretch blow molding machine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] See also Figures 1 to 3 , the present invention provides the case name, the technical solution is as follows:

[0041] Example 1: As an embodiment of the present invention, refer to Figure 1 , the control system of this embodiment includes:

[0042] The sensor module collects multimodal process data and performs preprocessing. The sensor module integrates a variety of sensors based on different physical principles, including:

[0043] The temperature sensor, which uses an infrared thermal imager, is placed near the entrance of the heating station and the blow molding station to monitor the temperature distribution on the surface of the bottle blank in real time.

[0044] Pressure sensor, real-time measurement of the axial resistance of the material during the stretching process.

[0045] Laser displacement scanning sensor, online monitoring of the geometric profiles of preforms and bottles in key processes.

[0046] Position and speed sensors are used to obtain the real-time kinematic status and synchronization information of the equipment.

[0047] Environmental condition sensors monitor ambient temperature and humidity that affect equipment operation and process stability.

[0048] After collecting raw multimodal process data from various sensors, the sensor module performs a series of automated preprocessing operations to improve data quality, standardize the data format, and make it suitable for subsequent intelligent diagnosis and control calculations. First, it uses a median filter to smooth noisy sensor signals, remove random fluctuations and high-frequency interference, and extract clearer signal trends. Second, it detects and processes any obvious anomalies and erroneous outliers in the data stream, using a simple sliding window detection algorithm to identify and remove these outliers. Finally, based on the sensor calibration information, the raw electrical signal readings are converted to meaningful engineering units. Furthermore, because different sensors may have different sampling frequencies, data streams from different sources must be timestamp-aligned to ensure temporal synchronization. Finally, the sensor data is normalized to eliminate the impact of varying physical dimensions and numerical ranges on subsequent algorithms.

[0049] The data management module provides unified data storage and access services for the system. In this embodiment, it is implemented based on internal storage media. The data management module is primarily responsible for persistently storing information critical to system operation and efficiently retrieving it in response to requests from other modules. Its storage primarily includes: historical process data and real-time data cache collected and preprocessed by the sensor module; machine learning files and their configuration for the intelligent diagnosis module; the expert knowledge base and adaptive response strategy set for the decision-making module; learning status, intermediate results, and updated strategies and model parameters generated by the online learning mechanism; system operation logs, alarm records, and operator settings.

[0050] Each intelligent functional module of the system needs to interact with the data management module to read the required data, models and knowledge, and write new results and records into storage.

[0051] The interactive module provides a human-machine interface for operators to monitor system status and make necessary settings. In this embodiment, it uses an industrial touch screen. This module primarily displays key operating data processed by the sensor module, system status information, and summary information generated by the intelligent diagnostic module. It also receives basic operating instructions and process parameter settings entered by the operator.

[0052] Intelligent diagnostic module: Analyzes data from the sensor module, uses the machine learning model stored in the data management module to identify operating conditions, and generates diagnostic identification and diagnostic confidence levels;

[0053] Specifically, the intelligent diagnosis module structure diagram is as follows Figure 2As shown, the intelligent diagnosis module adopts a hierarchical diagnostic architecture, which is divided into a first diagnostic layer and a second diagnostic layer. The first diagnostic layer deploys an algorithm with low computational complexity and fast response speed. In this embodiment, a statistical process control method based on key performance indicators is used. The first diagnostic layer determines whether the system deviates from the normal operating mode through the statistical process control method. Specifically, it first calculates the mean and standard deviation of these indicators based on the historical data of key performance indicators collected when the system is operating normally, and sets the center line and control limit of the statistical control chart accordingly. During the real-time monitoring process, once it is detected that the current value of any KPI exceeds its preset control limit, or the KPI data sequence shows a violation of the preset statistical operating rules, it is determined that the system state has significantly deviated from the normal operating mode, thereby triggering the second diagnostic layer and simultaneously issuing a preliminary alarm to the interactive module.

[0054] Furthermore, upon receiving an abnormality trigger signal from the first diagnostic layer, the second diagnostic layer activates and calls upon a trained machine learning model stored in the data management module. In this embodiment, the system calls upon a pre-built, computationally efficient, linearized state-space model from the data management module, based on specific physics knowledge. This model is constructed based on a deep understanding of the heat transfer physics during the preform heating phase in an injection stretch blow molding machine. Specifically, based on the thermodynamic law of conservation of energy, and taking into account key physical effects such as radiative heat transfer from the heating lamps, absorption and internal heat conduction of the preform material, and convective heat dissipation between the preform surface and the surrounding environment, a nonlinear differential equation is established to describe how the temperature of key preform locations varies with heating power and environmental conditions. Subsequently, a typical stable operating point is selected and this nonlinear equation is linearized, ultimately deriving a linearized state-space representation. During operation, this model receives the current power command value of the heating lamps and the reading of the ambient temperature sensor, and predicts in real time the expected temperature dynamic trajectories of several key monitoring points on the preform under ideal heating conditions.

[0055] Furthermore, the actual temperature measurements at key monitoring points of the preform, collected by the sensor module, are subtracted point by point from the expected temperature dynamic trajectory predicted in real time by the model. This generates a multidimensional residual time series vector containing multiple temperature deviation signals. Each element in this vector represents the real-time deviation between the actual temperature at a specific monitoring point and its physically expected temperature, quantifying the degree to which the heating process deviates from the ideal physical model. The system processes a data segment of the multidimensional residual time series within a preset time window. For each residual signal within this data segment, the system calculates a set of predefined statistical eigenvalues, including the mean, standard deviation, peak-to-peak value, skewness, kurtosis, zero-crossing rate, and signal energy of the residual signal within the time window. These statistical eigenvalues ​​for the residual signals at all monitoring points are combined to form a fixed-length residual feature vector. This feature vector is then input into a pre-trained, offline gradient boosting decision tree model stored in the data management module. The gradient boosting decision tree model has been trained to map specific residual feature vector patterns to specific equipment operating conditions or fault categories. The model efficiently classifies the input residual feature vector by integrating the judgments of multiple decision trees. Table 1 shows some key features and examples of the residual patterns they usually reflect.

[0056] Table 1 Predefined statistical characteristic values ​​and their descriptions

[0057] Feature Name Calculation Description Mainly reflected residual pattern average value The mean of the residual signal within the time window The degree of continuous bias of the signal Standard deviation Standard deviation of the residual signal within the time window Signal fluctuation amplitude Peak-to-peak value The difference between the maximum and minimum values ​​of the signal within the time window The amplitude of the signal oscillation or sudden jump Skewness Asymmetry of signal distribution Imbalance in signal deviation direction Signal energy The sum of squares of the residual signal within the time window Overall signal deviation strength Kurtosis The sharpness of the signal distribution Frequency characteristics of the signal Zero-crossing rate Number of times the signal crosses the zero line within the time window The frequency of extreme values ​​in the signal

[0058] Furthermore, the gradient boosting decision tree model outputs a specific operating condition event identifier based on the input residual feature vector. This identifier is one of the predefined category labels, such as "heating lamp aging and reduced efficiency," "abnormal preform feeding temperature," "impact of sudden ambient temperature change," or "temperature sensor drift in a specific area." Simultaneously, the gradient boosting decision tree model's internal ensemble learning mechanism calculates and outputs a diagnostic confidence score based on the vote ratio of all base decision trees for the optimal category, quantifying the reliability of the diagnosis result.

[0059] Furthermore, to enhance the comprehensibility of diagnostic results and assist in troubleshooting, the intelligent diagnostic module performs an explainable artificial intelligence analysis process. First, after the gradient boosted decision tree model completes the operating condition event classification and outputs a diagnostic identifier and confidence level, the system initiates the feature impact analysis step. In this step, the system analyzes the trained gradient boosted decision tree model, quantifying the contribution of each input feature to the final classification result by, for example, calculating the total information gain of each input feature used for node splitting during the model's decision tree construction process. Based on this, the system identifies and ranks several key residual features that have the most significant impact on the current diagnostic result. Next, the feature tracing step is performed. The system associates the key residual features identified in the previous step with their physical sources. Each feature in the residual feature vector is derived from specific sensor data and is calculated by combining its deviation from the physical model's prediction. The system can trace back to determine which specific statistical feature indicators and their corresponding original sensor data sources, such as the standard deviation of the mean residual calculated from temperature sensor data at a specific location or the peak-to-peak residual calculated from pressure sensor data, contribute most to the current diagnostic result. Finally, the intelligent diagnosis module generates and outputs explanatory information, integrating the results of feature impact analysis and feature traceability to form an explanatory information report. This report clearly lists the key input features that play a decisive role in the diagnosis conclusion and their corresponding sensor source information. This explanatory information is output along with the diagnostic identifier and confidence score, allowing operators to intuitively understand the cause of the fault and provide data support for subsequent equipment maintenance and control strategy optimization.

[0060] The intelligent diagnosis module effectively improves the problems mentioned in the background technology, such as the difficulty of existing systems in deeply understanding and distinguishing working conditions, the inability to intelligently identify the root causes of faults, and the patterned and blind adaptive adjustment strategies. By accurately quantifying the deviation of key physical quantities from ideal behavior and performing efficient pattern recognition on these information-rich residual features, this solution can not only achieve fine classification and cause tracing of specific working conditions, improving the interpretability of diagnosis, but also the clear diagnostic identification output by it can drive subsequent decisions and decision modules to perform differentiated and targeted adaptive control, thereby significantly improving the optimal performance and long-term operational robustness of the entire system in complex and changing environments.

[0061] Decision-making module: Receives diagnostic identification, combines expert knowledge stored in the data management module with online learning results, and determines the optimized adaptive response strategy; based on the determined response strategy, dynamically adjusts the relevant control parameters used for subsequent calculations.

[0062] Specifically, the decision module receives the diagnostic identifier and corresponding diagnostic confidence score output by the intelligent diagnostic module in real time. Using the received diagnostic identifier, it searches the expert knowledge base stored in the data management module for a predefined baseline response strategy that matches the identifier. This baseline strategy represents a standardized treatment recommendation for a specific operating condition, based on engineering experience and theoretical analysis. The decision module first uses the received diagnostic identifier as a search keyword to perform a search and match query within the expert knowledge base stored in the data management module. The expert knowledge base, in the form of a rule library and lookup table, contains standardized treatment plans defined by domain experts for various fault diagnostic identifiers. Upon successful matching, the decision module extracts the baseline response strategy corresponding to the current diagnostic identifier. This baseline strategy is a standardized treatment recommendation, based on engineering experience and theoretical analysis, designed to quickly stabilize the situation and provide initial compensation. For example, if the diagnostic identifier is "significant drop in ambient temperature," the corresponding baseline strategy is: "Temporarily increase the power settings of all heating zones by 1% until the ambient temperature stabilizes."

[0063] Furthermore, the decision module uses the received diagnostic identifier as a key representation of the current system state and inputs it into the online learning module that is also stored in the data management module. In this embodiment, the online learning module is specifically implemented using a reinforcement learning method based on value iteration. Its operation process is as follows: the online learning module maintains a value function Q-value function internally, which is used to evaluate the long-term expected benefits of taking different response strategy adjustments under a specific system state. The value function is updated online based on the experience accumulated during the continuous operation of the system. Specifically, after the system executes control according to a certain response strategy, the online learning module will receive performance feedback information related to the execution effect, such as a reward signal composed of quantified product quality, production efficiency, and energy consumption indicators. Using the reward signal and based on the selected reinforcement learning algorithm, the module will update its internal value function, thereby continuously refining the value assessment of taking different strategy adjustment actions under different states. When the decision module requests a recommended strategy, the online learning module, based on its most recently learned value function, determines the policy adjustment action that will yield the highest long-term expected benefit for the state represented by the input diagnostic identifier. The online learning module then converts this internally determined optimal action into a clearly defined output format, for example, mapping it to a set of specific control parameter adjustment recommendations. This formatted output is the recommended response strategy provided to the decision module for subsequent integration. This recommended strategy reflects the adjustment plan that the system has autonomously learned based on its operating history and performance feedback, and that tends to optimize long-term goals.

[0064] The decision-making module acquires two complementary response strategies: a baseline solution from an expert knowledge base representing reliable engineering experience, ensuring a proven, understandable, and reliable backup option. A recommended solution from the online learning module reflects the actual operating performance of the equipment and potential optimization directions, providing data-driven, continuously adaptable adjustment suggestions. By simultaneously preparing these two sources of strategies, the decision-making module obtains more comprehensive and balanced decision-making information, encompassing both static, rule-based stability considerations and dynamic, performance-based optimization considerations. This provides the necessary and high-quality input for subsequent intelligent, risk-controlled fusion judgments, avoiding the potential bias or lack of adaptability associated with reliance on a single information source.

[0065] After obtaining the baseline and recommended response strategies, the decision module performs a strategy fusion step to generate the final optimized response strategy. This fusion process uses a dynamic weighted fusion algorithm based on diagnostic confidence. The core of this algorithm is to dynamically determine the contribution weights of the baseline and recommended response strategies in the final strategy formation based on diagnostic confidence scores and historical performance feedback.

[0066] Specifically, the decision module first calculates the weight of the recommended strategy and the weight of the baseline strategy based on the input diagnostic confidence score, ensuring that the sum of these two weight values ​​is always equal to 1. The weight of the recommended strategy increases monotonically with the diagnostic confidence score through a predefined S-shaped function. A high confidence score will increase the weight of the recommended strategy, making the decision tend to adopt the optimization result of online learning, while a low confidence score will increase the weight of the baseline strategy, making the decision tend to be more conservative. The specific formula of the S-shaped function is as follows:

[0067]

[0068] in, represents the initial weight of the recommended strategy, Indicates the current diagnostic confidence. is a parameter that controls the steepness of the curve. is the center point of the curve. The value makes the weight The system is more sensitive to small changes in confidence. A value of 0 makes the transition smoother. These parameters are typically set in the middle of the confidence score range, such as 0.5 or 0.6, to define the turning point for weight distribution. The specific values ​​of these parameters can be adjusted through simulation experiments and expert experience to achieve a reasonable balance between online learning results and expert knowledge at different confidence levels.

[0069] Furthermore, the system continuously tracks and analyzes the sliding window average performance scores of the online learning module for specific diagnostic indicators. For example, it calculates performance metrics such as product qualification rate and process efficiency, compares these scores with the expected baseline, and generates a historical performance adjustment factor. The decision module uses this adjustment factor to adjust the weights. The initial recommended strategy weight is multiplied by the historical performance adjustment factor to obtain a temporary adjusted recommended weight, while the initial baseline strategy weight is maintained as the temporary baseline weight. To ensure that the final weights sum to 1, these two temporary weights are then normalized. The temporary adjusted recommended weight and the temporary baseline weight are divided by their sum, respectively, to obtain the final recommended strategy weight and the final baseline strategy weight. This final weight reflects both the confidence of the current diagnosis and the historical performance of the online learning module. Finally, based on these final weights, the strategy content is integrated. The optimized adaptive strategy is obtained by multiplying the adjustments to the corresponding parameters in the baseline and recommended strategies by their respective final weights.

[0070] The optimized adaptive response strategy calculated through the dynamic weighted fusion process above integrates expert knowledge, online learning experience, and balances risk based on the current diagnostic confidence. This optimized adaptive response strategy is passed to the decision module to guide the implementation of subsequent control actions.

[0071] The present invention intelligently integrates expert knowledge and online learning results through the decision-making module, and uses diagnostic confidence and historical performance for dynamic weighting and risk management, overcoming the defects of traditional adaptive control strategies such as patterning and blindness, and realizing differentiated, optimized and robust response strategy decisions for specific diagnostic conditions.

[0072] Furthermore, the optimized adaptive response strategy output by the decision module generates a set of relevant control parameters for subsequent adaptive control calculations. Specifically, after outputting a determined response strategy represented by a specific strategy identifier, the decision module uses this identifier to access information stored in the data management module. The data management module pre-stores multiple sets of control parameter configurations uniquely corresponding to each strategy identifier. These control parameter configurations are pre-determined to generate robust control parameter combinations for each foreseeable typical operating condition, such as raw material batch changes, large fluctuations in ambient temperature, and initial wear of specific components, through systematic offline analysis and optimization. These typical operating conditions are defined through historical data analysis, expert experience, and simulation experiments. Based on the strategy identifier, the decision module retrieves and locates the specific parameter configuration set to be applied. Each configuration set contains key parameters required for subsequent adaptive control calculations, such as the model predictive control algorithm's predictions, objective function weights, constraint limits, and underlying controller gains. The parameter configuration sets corresponding to different strategy identifiers are targeted in content. The decision module loads the retrieved parameter configuration set as the current effective configuration, replacing the original settings, and outputs this updated, currently effective parameter configuration set in its entirety to the adaptive control module for calculation based on the new parameters in the next control cycle, thereby achieving dynamic, precise adjustment and effective execution of the control strategy.

[0073] The decision-making module retrieves and loads pre-stored and optimized parameter configuration sets from the data management module based on the strategy identifier, achieving efficient and reliable decoupling between the upper-level response strategy decision and the specific implementation of the underlying control parameters, ensuring that different response strategies can be quickly, accurately and consistently converted into the specific parameter configurations required by the downstream computing module, thereby providing key support and implementation guarantees for the precise execution of differentiated adaptive control and ensuring the effective implementation of advanced decisions.

[0074] The adaptive control module calculates the original control command based on the adjusted relevant control parameters and the data from the sensor module; and generates a final control signal that is smoothed to ensure a stable transition based on the diagnostic confidence and the original control command.

[0075] Specifically, the module receives the currently effective control parameter configuration set from the decision module and the real-time sensor data from the sensor module in each control cycle, and accesses the internal process prediction model stored in the data management module. As the model used for MPC internal prediction, the linear time-invariant state space model is selected in this embodiment. The state space model is achieved by physically modeling and linearizing the key physical processes of the injection stretch blow molding machine at different working points. The specific parameters of the model are specified by the currently effective control parameter configuration set. The LTI model is adopted to significantly improve the computing efficiency while ensuring sufficient prediction accuracy to meet real-time requirements. In order to obtain the accurate current state information required by MPC, the module adopts a Kalman filter. The Kalman filter fuses the process model defined in step 2 with the real-time sensor measurement value, estimates the key state variables inside the system online, and provides the optimal state initial value for subsequent prediction and optimization.

[0076] Furthermore, the system's state trajectory is recursively calculated based on the LTI model over the future prediction horizon, where the prediction length is set by a parameter set. In each control cycle, an optimization problem is constructed to calculate the optimal control input sequence over the future control horizon. Because the LTI prediction model is used and linear constraints are typically set, the optimization problem is transformed into a quadratic programming problem. The optimization objective function (whose weights are set by the parameter set and aim to minimize tracking error and control energy consumption) and the constraints (whose limits are also set by the parameter set, such as actuator saturation and process boundaries) jointly define this QP problem. The system uses an efficient QP solver optimized for real-time embedded applications to solve this QP problem online. From the solved future optimal control input sequence, the first control action vector is extracted as the original control instruction to be issued in the current control cycle.

[0077] The adaptive control module adopts a model predictive control algorithm and dynamically sets its optimization objective function and constraints based on the parameter configuration set provided in real time by the upstream decision module. It can proactively calculate the optimal original control instructions that accurately execute the current specific response strategy intentions and explicitly meet the system operation constraints, thereby ensuring that advanced adaptive decisions can be effectively converted into high-performance and robust underlying control actions. It is the key computing core for achieving the intelligent and refined control goals of the entire system.

[0078] Furthermore, the module uses a predefined nonlinear function to dynamically calculate the core parameters required for the subsequent first-order low-pass filter, the filter time constant, based on the diagnostic confidence score received in real time. The function is designed so that the calculated filter time constant increases monotonically as the diagnostic confidence score decreases. When the diagnostic confidence score is 1, the function outputs a preset minimum time constant; when the diagnostic confidence score is 0, the function outputs a preset maximum time constant; between 0 and 1, the time constant smoothly transitions from the maximum value to the minimum value. The specific form of the function uses an exponential to achieve the desired transition curve and its maximum / minimum time constant parameters, which are determined through simulation and experiments during the system design and debugging phase, aiming to optimally balance response speed and stability. The specific calculation formula of the nonlinear function is as follows:

[0079]

[0080] in represents the filter time constant, Indicates the preset maximum time constant, Indicates the preset minimum time constant, Represents the diagnostic confidence score.

[0081] Furthermore, after determining the current filter time constant, the module uses a first-order digital low-pass filter to perform real-time filtering on the received original control instruction sequence. The output value of the filter is obtained by weighted average of the current input and the filter output value at the previous moment through a filter coefficient. The filter coefficient is determined in real time by calculating the filter time constant and the fixed control sampling period of the system. A larger time constant will calculate a smaller filter coefficient, making the current filter output more dependent on the output at the previous moment, thereby achieving a stronger smoothing effect; conversely, a smaller time constant corresponds to a larger filter coefficient, allowing the output to follow the input changes more quickly. The signal sequence processed by this first-order digital low-pass filter is the final control signal. The signal ensures the gentleness of the control action when the diagnostic uncertainty is high, and the final control signal is then output to the execution interface module.

[0082] The smoothness of control commands is dynamically adjusted based on the diagnostic confidence score. When diagnostic uncertainty is high, the system can automatically enhance the buffer to mitigate control actions and ensure system stability. When the confidence level is high, it allows for faster control response. This intelligent transition mechanism significantly improves the robustness and operational safety of the entire adaptive control system under various operating conditions.

[0083] The execution interface module receives the final control signal and sends it to the corresponding actuator of the equipment.

[0084] Specifically, the execution interface module forms the bridge between the adaptive control system and the physical actuators of the injection stretch blow molding machine. It is responsible for converting the final control intent generated by the control system's internal calculations into physical signals capable of driving actual equipment actions. This module receives the smoothed final control signal output from the adaptive control module. This signal is digitized and represents the precise control action status required for the current cycle. The core function of the execution interface module is to perform the necessary signal conversion, adaptation, and transmission.

[0085] Table 2 Example of target actuators and signals processed by the execution interface module

[0086] Target Implementation Agency Required control signal type Execute interface module operations Heating lamp power unit Analog voltage DAC conversion output Servo motor drive Industrial fieldbus messages Encode instructions and send bus data stepper motor driver Pulse / direction signal Generates a pulse train of specified frequency and number High / low pressure air blowing solenoid valve Digital switch DO (digital output) Proportional flow valve Simulated current DAC conversion output Frequency converter (fan / pump) Modbus communication Organize and send communication messages

[0087] Table 2 provides specific examples of how the execution interface module in this embodiment processes several common actuators. For a heating lamp power adjustment unit requiring analog control, the digital power command is converted into a standard 0-10V analog voltage signal via a digital-to-analog converter (DAC). For a servo motor driver, digital position, speed, and torque commands are packaged into data messages that comply with specific industrial fieldbus protocols. For controlling switching devices such as high-pressure and low-pressure air solenoid valves, the logic state is converted into corresponding digital output (DO) signals.

[0088] Furthermore, while the execution interface module sends final control signals to each device's actuators, it also monitors execution and communication status. For example, the execution interface module can capture error codes, execution timeouts, or communication failure signals returned by actuators (such as servo motor drivers or fieldbus devices). This feedback from the execution layer can directly alert operators through the interaction module to abnormalities on the device's execution side. Furthermore, this information is transmitted to the data management module for recording and can serve as a supplementary diagnostic information source for the intelligent diagnosis module, helping the system to more comprehensively assess device status and distinguish between process issues and actuator problems, thereby indirectly improving subsequent diagnostic accuracy and decision-making rationality. This feedback from the execution interface further enhances the system's closed-loop nature and its ability to perceive actual operating conditions.

[0089] The multi-sensor adaptive control system for the injection stretch blow molding machine proposed in the present invention obtains accurate working condition knowledge and confidence through intelligent hierarchical diagnosis combined with physical knowledge, effectively improving the problems of unclear diagnosis and difficulty in identifying the root cause. The decision-making module intelligently integrates expert knowledge and online learning and performs dynamic weighting based on confidence, realizing the formulation of differentiated optimal strategies under controllable risks, overcoming the drawbacks of adaptive adjustment modeling and blindness. The decision-making module loads a specific parameter set based on the strategy, driving the adaptive control module using model predictive control to accurately execute the optimization and constraint-satisfying control instructions under the intention of the strategy, ensuring the effective implementation of the decision. The adaptive control module smoothes the instructions based on confidence to ensure stable and safe operation. The system has completely constructed a closed-loop control process from accurate diagnosis to intelligent decision-making, and then to optimized execution and smooth transition, significantly improving the control accuracy, product quality consistency, operational robustness and overall intelligence level of complex industrial processes such as injection stretch blow molding.

[0090] Example 2: As an embodiment of the present invention, refer to Figure 3 , the control method of this embodiment includes:

[0091] Collect multimodal process data and perform preprocessing.

[0092] The injection stretch blow molding machine continuously collects multimodal process data during operation, covering information such as the temperature distribution in the preform heating zone, stretch rod force, blow pressure, key component positions and speeds, and ambient temperature and humidity. The sensor module performs necessary preprocessing on the collected raw data, including signal filtering, denoising, calibration, normalization, and timestamp alignment, to form a structured data stream for subsequent analysis.

[0093] Analyze data from sensor modules, use a layered architecture for rapid detection and precise condition identification, and output diagnostic identification and confidence levels.

[0094] Hierarchical diagnostics are performed on the preprocessed data stream. The first diagnostic layer employs a statistical process control (SPC) approach based on key performance indicators (KPIs), continuously monitoring key KPIs for significant deviations from statistical control limits established based on historical normal data. Once the SPC detects an anomaly, declaring the system has deviated from normal operation, it immediately generates an anomaly trigger signal and simultaneously issues a preliminary alert through the interaction module. Upon receiving the trigger signal from the first diagnostic layer, the second diagnostic layer of the intelligent diagnostic module is activated. This layer first utilizes a linearized state-space model based on the physical process of preform heating and heat transfer, stored in the data management module. Combined with current equipment inputs, it predicts the expected temperature trajectory of key monitoring points in real time. The actual temperature measured by the sensor module is then subtracted from this expected trajectory to calculate a multidimensional temperature residual time series. The system then processes this residual series within a pre-set time window, extracting its statistical features to form a residual feature vector. Finally, this feature vector is input into a pre-trained gradient boosting decision tree (GBDT) model, also stored in the data management module. The GBDT model classifies the residual feature patterns and outputs clear and specific working condition event identifiers and corresponding diagnostic confidence scores.

[0095] Furthermore, the diagnostic identifier and diagnostic confidence score output from the intelligent diagnostic module are received. The decision module executes the strategy determination process and uses the diagnostic identifier to retrieve the matching baseline response strategy in the expert knowledge base stored in the data management module; at the same time, the diagnostic identifier is input into the online learning module in the data management module to obtain the recommended response strategy learned based on historical operating data and long-term performance feedback. Then, the decision module uses a dynamic weighted fusion algorithm based on diagnostic confidence and historical performance adjustment factors to integrate the baseline strategy and the recommended strategy. Specifically, the final weight is calculated based on the confidence and historical performance, and the adjustment suggestions for the control parameters in the two strategies are fused through weighted averaging and other methods to generate an optimized adaptive response strategy.

[0096] The expert knowledge is integrated with the online learning results, and the confidence level is combined to determine the optimized response strategy and dynamically load the control parameter set according to the strategy.

[0097] The decision module receives the optimized adaptive response strategy output from the decision module. Based on this optimized adaptive response strategy, the decision module accesses the data management module to retrieve and load a pre-stored set of control parameter configurations unique to the optimized adaptive response strategy. This configuration set contains all the parameters required by the downstream adaptive control module. The decision module sets this parameter configuration set as the current effective configuration.

[0098] Furthermore, the module receives the currently effective control parameter configuration set from the decision module and real-time data from the sensor module. The module first uses a Kalman filter to perform state estimation to accurately determine the current system state. It then predicts the system's future behavior based on the linear time-invariant (LTI) process prediction model stored in the data management module and the current state. Next, based on the optimization objective function and constraints set by the current parameter configuration set, it constructs and solves a quadratic programming (QP) optimization problem, using an efficient real-time QP solver to calculate the optimal control input sequence in the future control time domain. The module extracts the first control action vector from this sequence and outputs it as the original control command.

[0099] The original instructions are calculated based on the parameter set and data, and the final signal is generated by smoothing the instructions according to the confidence.

[0100] The module receives raw control commands from the adaptive control module and a diagnostic confidence score from the intelligent diagnostic module. Using a predefined nonlinear function, it dynamically calculates the filter time constant of a first-order digital low-pass filter based on the confidence score. It then applies a filter coefficient determined by this dynamic time constant to perform real-time low-pass filtering on the raw control command sequence, generating a final control signal with smoother transitions and a stable transition.

[0101] The final control signal is applied to the actuator of the target device to achieve adaptive control.

[0102] The final control signal is received from the adaptive control module. This module is responsible for performing the necessary signal conversion and adaptation. The converted signal is sent to the corresponding physical actuators on the injection stretch blow molding machine, driving them to perform the corresponding physical actions, completing the entire adaptive control process.

[0103] Through the cyclic execution of the above steps, this method realizes in-depth analysis, intelligent diagnosis, adaptive decision-making and optimized control of the multi-sensor data of the injection stretch blow molding machine, which can effectively cope with complex working conditions and improve production quality and efficiency.

[0104] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Multi-sensor adaptive control system for injection stretch blow molding machine, characterized in that: include: Sensor module, collects multimodal process data and performs preprocessing; Data management module, which stores and retrieves process data, models, strategies, knowledge bases and learning results required for operation; Interaction module, monitors system status and receives parameter input; The intelligent diagnostic module analyzes data from the sensor module, uses the machine learning model stored in the data management module to identify operating conditions, and generates a diagnostic identifier and diagnostic confidence level; The decision module receives the diagnostic identifier and combines the expert knowledge stored in the data management module with the online learning results to determine the optimized adaptive response strategy; based on the determined response strategy, it dynamically adjusts the relevant control parameters used for subsequent calculations; The decision module performs a strategy fusion step to generate the final optimized response strategy. The fusion step uses a dynamic weighted fusion algorithm based on diagnosis confidence. The core of the algorithm is to dynamically determine the contribution weights of the baseline response strategy and the recommended response strategy in the formation of the final strategy based on the diagnosis confidence score and historical performance feedback. The adaptive control module calculates the original control command based on the adjusted relevant control parameters and data from the sensor module. Based on the diagnostic confidence and the original control command, it generates a final control signal that is smoothed to ensure a stable transition. The module uses a predefined nonlinear function to dynamically calculate the core parameters and filter time constant required for the subsequent first-order low-pass filter based on the diagnostic confidence score received in real time. The specific calculation formula of the nonlinear function is as follows: ； in represents the filter time constant, Indicates the preset maximum time constant, Indicates the preset minimum time constant, represents the diagnostic confidence score; The execution interface module receives the final control signal and sends it to the corresponding actuator of the equipment.

2. The multi-sensor adaptive control system for the injection stretch blow molding machine according to claim 1, characterized in that: The intelligent diagnostic module adopts a hierarchical diagnostic structure, including a first diagnostic layer for rapid anomaly detection and a second diagnostic layer for fine-grained operating condition classification. The intelligent diagnostic module integrates an explainable artificial intelligence analysis function to provide explanatory information on input features and sensor sources that have a significant impact on the results after diagnosis.

3. The multi-sensor adaptive control system for the injection stretch blow molding machine according to claim 1, characterized in that: The adaptive response strategy determined and optimized includes: Retrieving a benchmark response strategy that matches the current diagnostic condition from an expert knowledge base stored in a data management module based on the diagnostic identifier generated by the intelligent diagnostic module; Inputting the diagnostic identifier as status information into an online learning module stored in the data management module to obtain a recommended response strategy for the current diagnostic operating condition learned by the online learning module based on historical operating data and long-term performance feedback; Based on a preset fusion algorithm and taking into account the confidence score of the diagnostic identifier, the baseline coping strategy and the recommended coping strategy are integrated to generate an optimized adaptive coping strategy as a final output.

4. The multi-sensor adaptive control system for the injection stretch blow molding machine according to claim 3, characterized in that: The preset fusion algorithm adopts a dynamic weighted fusion method, and the logic for determining its weight coefficient includes: preliminarily determining a baseline weight value based on the current diagnostic confidence score, and using the real-time analysis results of historical performance data to generate an adjustment factor, which is used to adjust the baseline weight value.

5. The multi-sensor adaptive control system for the injection stretch blow molding machine according to claim 1, characterized in that: The decision module implements dynamic adjustment including: The data management module pre-stores a plurality of different control parameter configuration sets corresponding to different response strategy identifiers; the decision module retrieves and loads the corresponding complete parameter configuration set from the data management module according to the received current response strategy identifier.

6. The multi-sensor adaptive control system for the injection stretch blow molding machine according to claim 5, characterized in that: The adaptive control module adopts a model predictive control algorithm, which calculates the original control instructions by solving an online optimization problem in each control cycle. The objective function and constraints of the optimization problem are set in real time based on the relevant control parameter configuration set retrieved and loaded by the decision module.

7. The multi-sensor adaptive control system for an injection stretch blow molding machine according to claim 1, characterized in that: The process of the adaptive control module generating the final control signal includes: receiving the original control command calculated by the adaptive control module and receiving the diagnostic confidence score generated by the intelligent diagnostic module; Based on the received diagnostic confidence score, dynamically determining key adjustment parameters required for signal smoothing in a subsequent step in real time; the key adjustment parameter determination logic is configured such that a lower diagnostic confidence score corresponds to a parameter value calculated to result in a smoother control signal change; The key adjustment parameters are dynamically determined in real time, and signal smoothing processing is performed on the original control instructions; The signal obtained after processing is output to the execution interface module as the final control signal.

8. A multi-sensor adaptive control method for an injection stretch blow molding machine, using the multi-sensor adaptive control system for an injection stretch blow molding machine according to claim 1, characterized in that: include: Collect multimodal process data and preprocess it; Analyze data from sensor modules, using a layered architecture for rapid detection and precise condition identification, outputting diagnostic identifiers and confidence levels. The system integrates expert knowledge with online learning results, and uses confidence levels to determine an optimized response strategy. The system dynamically loads control parameter sets based on the strategy. The system calculates raw instructions based on the parameter set and data, and smoothes the instructions based on confidence levels to generate the final signal. The final control signal is applied to the actuator of the target device to achieve adaptive control.

Images