Steam sterilization process intelligent optimization and online verification method based on historical data

Abstract

The present application relates to the technical field of steam sterilization, in particular to a steam sterilization program intelligent optimization and online verification method based on historical data, which can generate a highly personalized and safe initial sterilization program for each new batch by fusing historical multi-source data to construct a process knowledge graph, overcoming the energy waste or sterilization deficiency risk caused by traditional fixed programs. In execution, the sterilization efficiency is evaluated in real time by using a data-driven prediction model, and dynamic parameter optimization is carried out in combination with model predictive control, thus actively saving energy and shortening the cycle under the premise of ensuring absolute sterilization safety. The whole scheme realizes a fundamental change from experience-driven to data-intelligent-driven, forms a complete closed loop of perception, decision-making, execution and self-learning, and can improve the reliability, economy and self-adaptive ability of the sterilization process.

Description

Technical Field

[0001] This invention relates to the field of big data analysis and mining technology, specifically a method for intelligent optimization and online verification of steam sterilization procedures based on historical data. Background Technology

[0002] Steam sterilization, as a classic and reliable terminal sterilization method, is widely used in critical industries related to life and health, such as pharmaceuticals, biological products, medical devices, and food. Its fundamental purpose is to completely kill all microorganisms, including bacterial spores, on products or items to achieve a level of sterility assurance. In actual industrial production, especially in large-scale biological product or high-end medical device manufacturing workshops, the objects to be sterilized are becoming increasingly complex and diverse, such as culture media containing proteins, surgical instrument packs with precise structures, and orthopedic implants with porous structures. The initial bioload of these items to be sterilized—that is, the types, quantities, and heat resistance of microorganisms—is significantly affected by various internal factors, such as the source of raw materials, the level of hygiene control in previous processing steps, personnel operation, and even the production batch, exhibiting dynamic fluctuations. Simultaneously, external environmental factors, such as… Seasonal changes in temperature and humidity in clean areas and fluctuations in boiler steam pressure at the plant level can interfere with the actual thermal state and steam saturation within the sterilization chamber. However, due to the absolute need for sterilization safety and conservative considerations for regulatory compliance, the industry has long adopted fixed sterilization programs designed based on worst-case scenarios, such as the classic 121°C for 30 minutes. This "one-size-fits-all" approach inevitably leads to excessive consumption of heat, water, and time when dealing with batches with light bioburden, increasing operating costs and shortening equipment lifespan. Furthermore, when handling exceptionally complex bioburden, special packaging, or difficult-to-penetrate cavity structures, fixed programs may pose a risk of sterilization failure due to uneven heat distribution or insufficient penetration, making it impossible to strike a balance between ensuring "zero defects" and achieving "lean production."

[0003] To optimize sterilization processes, existing technologies have attempted to move beyond simple fixed-program control. One mainstream approach relies on online accumulation of the physical parameter F0 for feedback control. This involves ending sterilization once the integral of the standard thermal inactivation time of microorganisms at the sterilization temperature reaches a preset threshold. While this method represents some progress, the F0 value calculation is based on an inactivation model of standard microorganisms under ideal conditions. This model's discrepancy with the complex microbial communities and non-ideal thermal conditions in actual loads constitutes a fundamental limitation. Another research approach focuses on establishing heat transfer and fluid dynamics simulation models of the sterilization chamber and load to simulate the temperature field and steam distribution, aiming for prediction and optimization during program development. However, the accuracy of such models, based on first-principles physicochemical principles, heavily depends on the precision of numerous boundary conditions. Precise settings, such as the precise geometry of the items to be sterilized, the thermophysical parameters of the materials, and the spatial distribution of the initial bioburden, are often difficult to obtain accurately or are highly simplified in complex industrial scenarios. More importantly, existing simulation models usually fail to deeply integrate and calibrate with massive amounts of historical operating data and final sterilization efficiency data that reflect the differences between different batches under real production conditions. Therefore, how to construct a method that does not rely on difficult-to-obtain precise physical parameters, but can make full use of multi-source historical data, adaptively learn the hidden mapping relationship between complex working conditions and final sterilization effect, and thus dynamically generate or adjust a safe, reliable, cost-effective, and personalized sterilization program for each specific batch has become a core problem that urgently needs to be solved in current technology. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent optimization and online verification of steam sterilization programs based on historical data. This method can construct a process knowledge graph by integrating historical multi-source data, generating highly personalized and safe initial sterilization programs for each new batch. This overcomes the energy waste or insufficient sterilization risks associated with traditional fixed programs. During execution, a data-driven predictive model is used to evaluate sterilization efficiency in real time, and dynamic parameter optimization is performed in conjunction with model predictive control. While ensuring absolute sterilization safety, this method proactively saves energy and shortens the cycle time. The entire solution represents a fundamental shift from experience-driven to data-driven intelligence, forming a complete closed loop of perception, decision-making, execution, and self-learning. This improves the reliability, economy, and adaptability of the sterilization process, thereby addressing the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent optimization and online validation of steam sterilization procedures based on historical data, specifically including the following steps: Step S1: In response to the optimization command, extract multimodal data corresponding to historical sterilization batches from the multi-source system; perform time-series alignment and feature extraction processing on the multimodal data to generate a structured feature dataset of historical sterilization batches, wherein the multimodal data includes process parameter time-series data, batch attribute data and environmental metadata; Step S2: Based on the structured feature dataset of historical sterilization batches, use clustering algorithm to mine efficiency patterns, and construct a sterilization efficiency knowledge graph based on the mining results. In the sterilization efficiency knowledge graph, the nodes represent different sterilization efficiency patterns, and the edges represent the association between patterns. Step S3: When the batch attribute data and environmental metadata of a new sterilization batch are obtained as initial features, the initial features are matched with the sterilization efficiency knowledge graph, and personalized initial sterilization program parameters for the new sterilization batch are generated based on the matching results. Step S4: Execute the sterilization process based on the personalized initial sterilization program parameters generated in step S3. During this process, a pre-trained sterilization efficiency prediction model is used to process the real-time collected process parameter time series data to obtain the predicted sterilization result. Based on the predicted sterilization result, the personalized initial sterilization program parameters are adjusted online adaptively using a model predictive control method to generate optimized sterilization program parameters. As a further aspect of the present invention: In step S1, multimodal data is extracted from a multi-source system, which includes a sterilization equipment control system, a manufacturing execution system, a laboratory information management system, and an environmental monitoring system; process parameter time-series data is extracted from the sterilization equipment control system, including temperature-time curves, pressure-time curves, and equipment state sequences measured by temperature sensors; batch attribute data is extracted from the manufacturing execution system, including the type of loaded items, loading density, and packaging materials; a final sterilization result label is obtained from the laboratory information management system, which is used to identify whether sterilization was successful or failed; and environmental metadata is extracted from the environmental monitoring system, including environmental temperature and humidity and supplied steam pressure.

[0006] As a further aspect of the present invention, the specific process of performing time-series alignment processing on multimodal data is as follows: From the temperature-time curves included in the process parameter time series data, select a baseline temperature-time curve for each historical sterilization batch; For this historical sterilization batch, a dynamic time bending algorithm is used to calculate the optimal bending path between the pressure-time curve, equipment state sequence, and other temperature-time curves in the process parameter time series data of this batch and the reference temperature-time curve. Based on the calculated optimal bending paths, the time axes of the pressure-time curve, equipment state sequence, and other temperature-time curves are nonlinearly scaled and adjusted to align the adjusted pressure-time curve, equipment state sequence, and other temperature-time curves with the baseline temperature-time curve in the time dimension. This results in an aligned time-series data set for each historical sterilization batch, which includes all adjusted temperature-time curves, adjusted pressure-time curves, and adjusted equipment state sequences.

[0007] As a further aspect of the present invention, the specific process of feature extraction processing for multimodal data is as follows: Based on the aligned time-series data set, physical guidance features are constructed for each historical sterilization batch. To construct the physical guidance feature, firstly, based on the temperature-time curves in the aligned time-series data set, the cumulative heat lethality curve of the historical sterilization batch at the temperature measurement location corresponding to each temperature-time curve is calculated; then, the dynamic time bending distance between each pair of cumulative heat lethality curves corresponding to different temperature measurement locations is calculated, and the arithmetic mean of the dynamic time bending distance between each pair of cumulative heat lethality curves is obtained. This arithmetic mean is used as the heat distribution uniformity feature characterizing the spatial heat distribution uniformity of the sterilization process of the historical sterilization batch. Simultaneously, based on the pressure-time curve and the reference temperature-time curve in the aligned time-series data set, the theoretical saturated steam pressure corresponding to each temperature data point on the reference temperature-time curve is queried to obtain the theoretical saturated steam pressure curve; the absolute value of the difference between the pressure-time curve value at each moment and the theoretical saturated steam pressure curve value at the corresponding moment is calculated within the preset exposure time period of the sterilization process, and this absolute value is integrated over the entire exposure time period to obtain the first integral value; the first integral value is normalized by dividing it by the length of the exposure time period, and the normalized value is used as the steam quality dynamic deviation feature characterizing the steam quality status of the sterilization process of this historical sterilization batch; Finally, the heat distribution uniformity characteristics, steam quality dynamic deviation characteristics, conventional statistical characteristics extracted from the aligned time-series data set, batch attribute data, and environmental metadata are combined to form the final structured feature vector of the historical sterilization batch. The structured feature vectors of all historical sterilization batches and their corresponding final sterilization result labels constitute the structured feature dataset of historical sterilization batches.

[0008] As a further aspect of the present invention: in step S2, the specific operation of performance pattern mining using a clustering algorithm is as follows: Based on the structured feature dataset of historical sterilization batches, the composite distance between any two historical sterilization batches is calculated. The composite distance is obtained by weighted summation of the first sub-distance and the second sub-distance. The first sub-distance is calculated based on a subset of features in the structured feature vector of the historical sterilization batches, including batch attribute data, environmental metadata, and conventional statistical features, and is calculated using standardized Euclidean distance. The second sub-distance is calculated based on the heat distribution uniformity feature and steam quality dynamic deviation feature in the structured feature vector of the historical sterilization batches, and is calculated using Mahalanobis distance. After obtaining the composite distances between all pairs of historical sterilization batches, the batch-to-batch distance matrix formed by these composite distances is used to divide all historical sterilization batches into multiple clusters, each cluster corresponding to a preliminary performance model.

[0009] As a further aspect of the present invention: the specific operation of constructing a sterilization efficacy knowledge graph based on the mining results is as follows: For each obtained cluster, a performance evaluation is performed to form a final performance pattern; for each cluster, its average sterilization success rate is calculated, which is the proportion of the number of batches whose final sterilization result is marked as successful out of all historical sterilization batches in the cluster. Calculate the performance index vector for the cluster, where each component of the performance index vector represents the average heat lethality, average energy consumption, and average program duration of all historical sterilization batches within the cluster. Extract the central feature vector of the cluster, which is the arithmetic mean of the structured feature vectors of all historical sterilization batches within the cluster; An effectiveness pattern is defined by its central feature vector, average sterilization success rate, and effectiveness index vector. Then, each effectiveness pattern is used as a node in the knowledge graph, and the node is configured with attributes containing its central feature vector, average sterilization success rate, and effectiveness index vector. Then, the sequential relationship between historical sterilization batches is analyzed. The sequential relationship refers to the sequence formed by arranging historical sterilization batches according to their execution time, i.e., the historical sequence. When the historical sterilization batch data contains historical sequences arranged by time or process logic, the number of times a historical sterilization batch belonging to the previous efficiency mode immediately follows a historical sterilization batch belonging to the next efficiency mode is counted in the historical sequence. This number is recorded as the number of transfers from the previous efficiency mode node to the next efficiency mode node. The quotient obtained by dividing this number of transfers by the total number of transfers from the previous efficiency mode node to all possible subsequent efficiency mode nodes is used as the weight. A directed edge is established between the corresponding two efficiency mode nodes, thereby constructing a sterilization efficiency knowledge graph containing multiple nodes and weighted directed edges. When historical sterilization batch data does not contain explicit sequence information, the weights are calculated based on the distance between the central eigenvectors of the two performance modes in the feature space. This distance is obtained by calculating the composite distance between the central eigenvectors of the two performance modes.

[0010] As a further aspect of the present invention: the specific process of matching the initial features with the sterilization efficacy knowledge graph in step S3 is as follows: The input to the matching process consists of two parts: first, the sterilization efficiency knowledge graph output in step S2, where each efficiency pattern node has attributes including a central feature vector, average sterilization success rate, and efficiency index vector; second, the batch attribute data and environmental metadata of the new sterilization batch, which constitute the initial features of the new sterilization batch; first, the initial features of the new sterilization batch are zero-padded and aligned to generate a query feature vector with the same dimension as the central feature vector of the efficiency pattern node. The graph attention network is used for processing. First, for each efficiency pattern node in the sterilization efficiency knowledge graph, its central feature vector is concatenated with the average sterilization success rate and efficiency index vector. Then, through a learnable linear transformation, it is mapped to a unified feature space of a preset dimension to obtain the initial hidden state of each efficiency pattern node. Next, the query feature vector is treated as a virtual node and mapped to a unified feature space of a preset dimension through a learnable linear transformation; then, the attention coefficient between the virtual node and each performance mode node is calculated. Then, the unnormalized attention coefficients corresponding to all performance mode nodes are normalized to obtain the final attention weights between the query feature vector and each performance mode node. Finally, the final representation of the virtual node is updated by aggregating the information of all effective mode nodes: the initial hidden states of each mapped effective mode node are weighted and summed using the final attention weight as a coefficient, and the result of the weighted sum is processed by a non-linear activation function to obtain the updated virtual node representation.

[0011] As a further aspect of the present invention: the process of generating personalized initial sterilization procedure parameters for a new sterilization batch based on the matching results specifically comprises: For each efficiency mode node in the sterilization efficiency knowledge graph, a set of historical best program parameters is associated. This set of historical best program parameters comes from the clusters that constitute the efficiency mode node, and consists of the actual program parameters of multiple historical sterilization batches that were ultimately labeled as successful and had the best overall efficiency index. The actual program parameters include at least the target sterilization temperature and the holding time. Then, based on the query feature vector obtained from the matching calculation in step S3 and the final attention weight between each performance mode node, the historical best program parameter set associated with each performance mode node is weighted and fused to generate personalized initial sterilization program parameters for the new sterilization batch; the specific process of weighted fusion to generate the baseline value of personalized initial sterilization program parameters is as follows: First, for each performance mode node associated with the historical best program parameter set, calculate the arithmetic mean of each program parameter component to obtain an average program parameter vector; then, using the final attention weight between the query feature vector of the new sterilization batch calculated in step S3 and the performance mode node as a weighting coefficient, perform a weighted summation operation on the average program parameter vectors corresponding to all performance mode nodes, and use the result of the weighted summation as the benchmark value for the generated personalized initial sterilization program parameters; Meanwhile, the specific process of generating the allowable fluctuation range of personalized initial sterilization program parameters through weighted fusion is as follows: First, calculate the variance of each program parameter component in the historical best program parameter set associated with each performance mode node; then, calculate the square of the difference between the average program parameter vector of each performance mode node and the baseline value of the personalized initial sterilization program parameters for each corresponding component; next, add the variance of each performance mode node to the square of the corresponding difference to obtain an intermediate vector; then, using the final attention weight as the weighting coefficient, perform a weighted summation on the intermediate vectors corresponding to all performance mode nodes; finally, perform a square root operation on each component of the weighted summation vector, and use the result as the allowable fluctuation range of the generated personalized initial sterilization program parameters. The baseline values ​​and allowable fluctuation ranges of the personalized initial sterilization procedure parameters together constitute a complete description of the personalized initial sterilization procedure parameters for a new sterilization batch.

[0012] As a further aspect of the present invention: the specific process of processing the real-time acquired process parameter time-series data using a pre-trained sterilization efficiency prediction model in step S4 is as follows: The pre-trained sterilization efficiency prediction model receives real-time process parameter time-series data as input, which is a segment of real-time process parameters with a fixed time length taken forward from the current time as the end point. The segment of real-time process parameter time-series data includes temperature-time curve segments and pressure-time curve segments from multiple temperature measurement points. The pre-trained sterilization efficiency prediction model first processes the input through a bidirectional long short-term memory network layer. This bidirectional long short-term memory network layer consists of two independent long short-term memory networks, forward propagation and backward propagation, which process the input from the start time to the end time and from the end time to the start time, respectively. The network state at each time point is formed by splicing the network states at the corresponding time points of forward propagation and backward propagation, thereby obtaining the hidden state corresponding to each time point in the input, which integrates the preceding and following time-series information, forming a hidden state sequence. Next, the hidden state sequence is processed through an attention mechanism layer, which calculates the correlation score between the last hidden state in the hidden state sequence and each hidden state in the sequence. Then, all the calculated correlation scores are processed by a normalized exponential function to convert each correlation score into a value between zero and one. Each value obtained after conversion is used as the attention weight of the hidden state at the corresponding historical time. Then, using the attention weights as coefficients, the hidden states at each time step in the hidden state sequence are summed in a weighted manner to generate a comprehensive context vector. Finally, the last hidden state of the hidden state sequence is concatenated with the comprehensive context vector and passed through a fully connected neural network layer to output two predicted values: one is the predicted sterilization failure probability at the end of the program; the other is the predicted final thermal lethal dose at the end of the program.

[0013] As a further aspect of the present invention: the specific process of using a model predictive control method to adaptively adjust the personalized initial sterilization program parameters online based on the predicted sterilization results is as follows: In each control cycle of the sterilization process, a rolling time-domain optimization problem is constructed and solved. The optimization variable of the rolling time-domain optimization problem is the sequence of control increments for several consecutive future control cycles. The objective function of the rolling time-domain optimization problem is to minimize an objective function value, which is the sum of three weighted terms. The three weighted terms are calculated as follows: for each future moment in the prediction time domain of the rolling time-domain optimization problem, the first weight coefficient is multiplied by the sterilization failure probability output by the pre-trained sterilization efficiency prediction model at that moment, plus the second weight coefficient multiplied by the energy consumption predicted based on the control increment at that moment, plus the third weight coefficient multiplied by the square of the difference between the predicted key program parameters at that moment and the reference parameter values ​​at that moment, to obtain the contribution value of that future moment; then the contribution values ​​of all future moments in the prediction time domain are summed as the objective function value. The rolling time-domain optimization problem includes several constraints: The first constraint is the terminal sterilization guarantee constraint, requiring that the final thermal lethality predicted by the sterilization efficiency prediction model in step S4 at the very end of the prediction time domain is not less than a target thermal lethality set by the sterilization process requirements; the second constraint is the parameter safety constraint, requiring that the predicted values ​​of key program parameters at each future time in the prediction time domain must be within a dynamic safety range. The lower limit of this dynamic safety range is obtained by subtracting the component of the corresponding parameter in the allowable fluctuation range generated in step S3 from the baseline value of the corresponding time in the personalized initial sterilization program parameters, and the upper limit is obtained by adding the component of the corresponding parameter in the allowable fluctuation range to the baseline value of the corresponding time; the key program parameters include the target sterilization temperature; the third constraint is the control increment constraint, requiring that the value of the optimized variable, i.e., the control increment of each future control cycle, cannot exceed the maximum and minimum adjustment range allowed by the device actuator in a single operation; the fourth constraint is an optional process model constraint, used to describe the state change of the key program parameters from one time to the next time, which is determined by the parameter state of the previous time, the currently applied control increment, and possible disturbances; After solving the rolling time-domain optimization problem in each control cycle, the first control increment in the obtained future control increment sequence is applied to the actuator of the sterilization equipment to realize online adaptive adjustment of personalized initial sterilization program parameters and generate the actual execution trajectory of the optimized sterilization program parameters.

[0014] Compared with the prior art, the beneficial effects of the present invention are: Step S1 involves temporal alignment and physical-guided feature extraction of historical multimodal data, resulting in a high-quality structured feature dataset. Based on this, step S2 utilizes a composite distance metric and spectral clustering algorithm to mine different sterilization efficiency patterns from historical data and constructs a structured sterilization efficiency knowledge graph. The nodes of this graph encapsulate the central features and efficiency indicators of historical successful patterns. Step S3 utilizes a graph attention network to match the initial features of the new batch with the sterilization efficiency knowledge graph. By calculating attention weights and fusing multiple relevant historical optimal program parameter sets, initial sterilization program parameters and their allowable fluctuation range that are both personalized and robust are generated. This process enables each new batch to obtain a customized optimization starting point, fundamentally overcoming the risks of energy waste or insufficient sterilization caused by traditional fixed programs or control based solely on F0 values. In step S4, during program execution, a sterilization efficiency prediction model trained based on a bidirectional long short-term memory network and attention mechanism is used for real-time prospective evaluation, and a model predictive control method is employed for online rolling optimization. This optimization process targets the predicted sterilization failure probability and energy consumption, and is constrained by the terminal thermal lethality and the dynamic safety range defined by the allowable fluctuation range of the initial program parameters. Therefore, this method can proactively and in real-time adjust program parameters to save energy while rigidly ensuring absolute sterilization safety, and is expected to shorten unnecessary sterilization time. In summary, this invention achieves intelligent initialization by constructing a knowledge graph through historical data mining, and combines real-time prediction and model predictive control to achieve safe closed-loop optimization, forming a complete intelligent system that learns from historical experience, optimizes the current batch, and can continuously update itself. This method significantly improves the reliability of the steam sterilization process, enhances energy efficiency and economy through personalized and dynamic optimization, and strengthens the system's adaptability to different loads and operating conditions.

[0015] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In this embodiment of the invention, a method for intelligent optimization and online validation of steam sterilization procedures based on historical data is described below. Figure 1 As shown, the specific steps include: Step S1: In response to the optimization command, extract multimodal data corresponding to historical sterilization batches from the multi-source system; perform time-series alignment and feature extraction processing on the multimodal data to generate a structured feature dataset of historical sterilization batches, wherein the multimodal data includes process parameter time-series data, batch attribute data and environmental metadata; Step S2: Based on the structured feature dataset of historical sterilization batches, use clustering algorithm to mine efficiency patterns, and construct a sterilization efficiency knowledge graph based on the mining results. In the sterilization efficiency knowledge graph, the nodes represent different sterilization efficiency patterns, and the edges represent the association between patterns. Step S3: When the batch attribute data and environmental metadata of a new sterilization batch are obtained as initial features, the initial features are matched with the sterilization efficiency knowledge graph, and personalized initial sterilization program parameters for the new sterilization batch are generated based on the matching results. Step S4: Execute the sterilization process based on the personalized initial sterilization program parameters generated in step S3. During this process, a pre-trained sterilization efficiency prediction model is used to process the real-time collected process parameter time-series data to obtain the predicted sterilization result. Based on the predicted sterilization result, the personalized initial sterilization program parameters are adjusted online adaptively using a model predictive control method to generate optimized sterilization program parameters.

[0019] In this embodiment, it is specifically necessary to explain step S1, in which multimodal data is extracted from a multi-source system. The multi-source system includes a sterilization equipment control system, a manufacturing execution system, a laboratory information management system, and an environmental monitoring system. These systems are connected to the data acquisition module through an industrial communication protocol, and the data is associated by batch identifier. Process parameter time-series data is extracted from the sterilization equipment control system. This process parameter time-series data includes temperature-time curves, pressure-time curves, and equipment state sequences measured by temperature sensors (the equipment state sequence is a discrete-time signal characterizing the sterilization cabinet door opening / closing state, the main steam valve opening / closing state, and the vacuum pump start / stop state). The system extracts batch attribute data from the Manufacturing Execution System (MES), including the type of loaded items, loading density, and packaging materials. Loading density is defined as the ratio of the total volume of the items to the effective volume of the sterilization chamber. It also retrieves the final sterilization result label from the Laboratory Information Management System (LIMS), which indicates whether sterilization was successful or failed (a label value of 1 corresponds to success, and a label value of 0 corresponds to failure). Furthermore, it extracts environmental metadata from the Environmental Monitoring System, including ambient temperature and humidity, and supplied steam pressure. Ambient temperature and humidity are measured at preset time points before the sterilization process begins, and supplied steam pressure is the average pressure value during the process. The specific process of performing time-series alignment on multimodal data is as follows: From the temperature-time curves included in the process parameter time series data, a baseline temperature-time curve is selected for each historical sterilization batch. This baseline temperature-time curve is derived from the chamber master thermometer. The master thermometer is usually located near the geometric center of the sterilization chamber or a recognized cold point, and its curve best represents the master control logic set in the program. For this historical sterilization batch, a dynamic time bending algorithm is used to calculate the optimal bending path between the pressure-time curve, equipment status sequence, and other temperature-time curves in the process parameter time series data of this batch and the reference temperature-time curve. The dynamic time bending algorithm effectively solves the asynchronous problem caused by the deviation of the sampling start time or the difference in local rate by finding the path with the minimum cumulative distance between two time series. Based on the calculated optimal bending paths, the time axes of the pressure-time curve, equipment state sequence, and other temperature-time curves are nonlinearly scaled and adjusted to align the adjusted pressure-time curve, equipment state sequence, and other temperature-time curves with the baseline temperature-time curve in the time dimension. This yields an aligned time-series data set for each historical sterilization batch. This aligned time-series data set includes all adjusted temperature-time curves, adjusted pressure-time curves, and adjusted equipment state sequences, where data points under the same time index correspond to the same physical moment. The time index is a sequence of equally spaced time points generated based on the sampling interval of the baseline temperature-time curve. The specific process of feature extraction from multimodal data is as follows: Based on the aligned time-series data set, physical guidance features are constructed for each historical sterilization batch. The physical guidance feature is constructed by first calculating the cumulative heat lethality curve at each temperature measurement location corresponding to the temperature-time curve in the aligned time-series data set for the historical sterilization batch. The cumulative heat lethality curve is obtained by integrating the equivalent lethality of microorganisms with time at a specified temperature, and its value reflects the cumulative sterilization efficiency. Next, the dynamic time bending distance between each pair of cumulative heat lethality curves corresponding to different temperature measurement locations is calculated, and the arithmetic mean of the dynamic time bending distance between each pair of cumulative heat lethality curves is obtained. This arithmetic mean is used as the heat distribution uniformity feature characterizing the spatial heat distribution uniformity of the sterilization process of the historical sterilization batch. The smaller the value of this heat distribution uniformity feature, the more synchronous the sterilization efficiency accumulation process at each point in the cavity is during the entire sterilization process, and the more uniform the spatial temperature field distribution is. It is an important quantitative indicator for assessing whether there are local "cold spots" or uneven heat penetration. Simultaneously, based on the pressure-time curve and reference temperature-time curve in the aligned time-series data set, the theoretical saturated vapor pressure curve is obtained by querying the corresponding theoretical saturated vapor pressure for each temperature data point on the reference temperature-time curve; the query basis is the thermodynamic property table of water or the saturated vapor pressure calculation formula; the absolute value of the difference between the pressure-time curve value and the corresponding theoretical saturated vapor pressure curve value at each moment within the preset exposure time period of the sterilization process is calculated, and this absolute value is integrated over the entire exposure time period to obtain the first integral value; the exposure time period usually refers to the time set in the sterilization program parameters, during which the temperature is maintained at the target sterilization temperature. The time interval with bacterial counts (e.g., 121°C) above the target value is defined by the process, for example, from the time the temperature first reaches the target value until the cooling process begins. The first integral value is divided by the length of the exposure time and normalized. This normalized value is used as the steam quality dynamic deviation characteristic characterizing the steam quality status of the sterilization process of this historical sterilization batch. The smaller the value of this steam quality dynamic deviation characteristic, the closer the actual working pressure is to the theoretical saturated steam pressure at the corresponding temperature, the lower the content of non-condensable gases in the steam, the better the steam quality, and the higher the heat transfer efficiency. Conversely, it may indicate that the steam is too humid, contains air, or is superheated, which will significantly affect the sterilization effect. Finally, the heat distribution uniformity characteristics, steam quality dynamic deviation characteristics, conventional statistical features extracted from the aligned time-series data set, batch attribute data, and environmental metadata are combined to form the final structured feature vector of the historical sterilization batch. The conventional statistical features include at least: the average heating rate during the heating phase, the maximum heating rate during the heating phase, the average temperature during the holding phase, the standard deviation of the temperature during the holding phase, and the duration of the entire sterilization process. This final structured feature vector is a one-dimensional array whose dimension is equal to the sum of the number of all features. The structured feature vectors of all historical sterilization batches and their corresponding final sterilization result labels constitute the structured feature dataset of historical sterilization batches. This structured feature dataset is stored in tabular or matrix form, with each row corresponding to a historical sterilization batch and each column corresponding to a feature or label, for subsequent machine learning model training.

[0020] In this embodiment, the specific operation of using clustering algorithm to mine performance patterns in step S2 is as follows: Based on the structured feature dataset of historical sterilization batches, the composite distance between any two historical sterilization batches is calculated. The composite distance is obtained by weighted summation of the first sub-distance and the second sub-distance. The first sub-distance is calculated based on a subset of features in the structured feature vector of historical sterilization batches, including batch attribute data, environmental metadata, and conventional statistical features. The first sub-distance is calculated using standardized Euclidean distance, which eliminates the influence of features with different dimensions on the distance calculation, making features with large numerical differences, such as loading density and ambient temperature, comparable. The second sub-distance is calculated based on the heat distribution uniformity feature and steam quality dynamic deviation feature in the structured feature vector of historical sterilization batches. The second sub-distance is calculated using Mahalanobis distance, which, when calculated, incorporates the covariance matrix estimated from the heat distribution uniformity feature and steam quality dynamic deviation feature of all historical sterilization batches to simultaneously consider the scale of the heat distribution uniformity feature and the steam quality dynamic deviation feature, as well as the correlation between these two features. In the weighted summation operation, the weight coefficient used to weight the first sub-distance is a preset harmonic parameter between 0 and 1. This harmonic parameter can be determined through cross-validation. For example, setting it to 0.5 indicates that static attribute similarity is given the same importance as dynamic process morphological similarity. The weight coefficient used to weight the second sub-distance is 1 minus this harmonic parameter. After obtaining the composite distances between all pairs of historical sterilization batches, a spectral clustering algorithm is used to divide all historical sterilization batches based on the batch distance matrix formed by these composite distances, resulting in multiple clusters. Each cluster corresponds to a preliminary performance pattern. The spectral clustering algorithm first converts the distance matrix into a similarity matrix, usually achieved through a Gaussian kernel function. Then, it calculates the Laplacian matrix of the similarity matrix and performs eigenvalue decomposition. Finally, it performs K-means clustering on the eigenvectors, thereby achieving the division of the original data in a low-dimensional feature space. This method can identify non-spherically distributed clusters and is more suitable for pattern discovery in complex process data. The specific steps for constructing a sterilization efficacy knowledge graph based on the mining results are as follows: For each cluster obtained, a performance evaluation is performed to form a final performance pattern; for each cluster, its average sterilization success rate is calculated, which is the proportion of the number of batches whose final sterilization result is marked as successful out of all historical sterilization batches in the cluster; this proportion directly quantifies the historical reliability of this pattern. Calculate the efficiency index vector for this cluster. Each component of the efficiency index vector represents the average heat lethality, average energy consumption, and average program duration of all historical sterilization batches within the cluster. The average heat lethality is obtained by averaging the values ​​of each historical sterilization batch within the cluster based on its cumulative heat lethality curve at the end of the sterilization program. The value of the average heat lethality comprehensively reflects the sterilization intensity under this mode. The average energy consumption is the average of the energy consumption measurements recorded or calculated for each historical sterilization batch within the cluster throughout the sterilization process. The energy consumption measurements are obtained from the sterilization equipment control system or manufacturing execution system and are usually measured in terms of steam consumption or electrical energy consumption. The average program duration is the average of the total duration of the sterilization program for each historical sterilization batch within the cluster from start to finish. The value of the average program duration reflects the time efficiency of this mode. Extract the central feature vector of the cluster. The central feature vector is the arithmetic mean of the structured feature vectors of all historical sterilization batches in the cluster. The central feature vector serves as the "standard profile" or prototype of this pattern. An effectiveness pattern is defined by its central feature vector, average sterilization success rate, and effectiveness index vector. Then, each effectiveness pattern is used as a node in the knowledge graph, and the node is configured with attributes containing its central feature vector, average sterilization success rate, and effectiveness index vector. Thus, each node in the knowledge graph encapsulates a type of reusable process knowledge and effectiveness expectation. Then, the sequential relationship between historical sterilization batches is analyzed. The sequential relationship refers to the sequence formed by arranging historical sterilization batches according to their execution time, i.e., the historical sequence. When the historical sterilization batch data contains historical sequences arranged by time or process logic, the number of times a historical sterilization batch belonging to the previous efficiency mode immediately follows a historical sterilization batch belonging to the next efficiency mode is counted in the historical sequence. This number is recorded as the number of transfers from the previous efficiency mode node to the next efficiency mode node. The quotient obtained by dividing this number of transfers by the total number of transfers from the previous efficiency mode node to all possible subsequent efficiency mode nodes is used as the weight. A directed edge is established between the corresponding two efficiency mode nodes. This weight encodes the empirical probability of switching between different process scenarios in production practice, providing sequence context information for subsequent recommendations, thereby constructing a sterilization efficiency knowledge graph containing multiple nodes and weighted directed edges. When historical sterilization batch data does not contain explicit sequence information, the weights are calculated based on the distance between the central eigenvectors of two efficacy patterns in the feature space. This distance is obtained by calculating the composite distance between the central eigenvectors of the two efficacy patterns. The weight value is proportional to the negative exponential function of the distance between the central eigenvectors of the two efficacy patterns calculated here, specifically exp(-d / σ), where d is the composite distance between the two central eigenvectors and σ is a scaling parameter used to control the rate at which the weight decays with distance, for example, the median of the distances between all patterns.

[0021] In this embodiment, the specific process of matching the initial features with the sterilization efficacy knowledge graph in step S3 is as follows: The input to the matching process consists of two parts: first, the sterilization efficiency knowledge graph output in step S2, where each efficiency pattern node has attributes including a central feature vector, average sterilization success rate, and efficiency index vector; second, the batch attribute data and environmental metadata of the new sterilization batch, which constitute the initial features of the new sterilization batch. Since the initial features of the new sterilization batch do not contain process-related features, their dimension is smaller than the dimension of the central feature vector of the efficiency pattern node. Therefore, the initial features of the new sterilization batch are first zero-padded and aligned to generate a query feature vector with the same dimension as the central feature vector of the efficiency pattern node. The positions corresponding to the heat distribution uniformity feature, steam quality dynamic deviation feature, and conventional statistical features extracted from the time series data of process parameters are filled with zero or a preset identifier. The preset identifier can be a special value (such as -1) used to distinguish it from the real value in subsequent calculations. The graph attention network is used for processing. First, for each efficiency pattern node in the sterilization efficiency knowledge graph, its central feature vector is concatenated with the average sterilization success rate and efficiency index vector. Then, through a learnable linear transformation, it is mapped to a unified feature space of a preset dimension to obtain the initial hidden state of each efficiency pattern node. The feature space of the preset dimension (e.g., 128 dimensions) is a hyperparameter that is set before model training. Next, the query feature vector is treated as a virtual node and mapped to a unified feature space of a preset dimension through a learnable linear transformation. Then, the attention coefficient between this virtual node and each performance mode node is calculated. The calculation process of the attention coefficient is as follows: the mapped query feature vector and the initial hidden state of the mapped performance mode node are concatenated, a learnable attention vector is used to perform a linear transformation on the concatenated vector, and then it is processed by a non-linear activation function to obtain a basic score reflecting the correlation between the two features. The non-linear activation function is preferably the LeakyReLU function, and its slope in the negative interval can be set to 0.2. On this basis, a term consisting of the average sterilization success rate of the performance mode node multiplied by a preset balance factor is added to obtain the final unnormalized attention coefficient. The preset balance factor is used to adjust the contribution strength of the historical success rate in the attention calculation. Its value can be determined through cross-validation, for example, set to between 0.1 and 0.3, to ensure that feature correlation dominates while introducing reliability prior. Next, the unnormalized attention coefficients corresponding to all performance mode nodes are normalized to obtain the final attention weights between the query feature vector and each performance mode node; the normalization process uses the softmax function to make the sum of all final attention weights equal to 1. Finally, the final representation of the virtual node is updated by aggregating the information of all performance mode nodes: the initial hidden states of each mapped performance mode node are weighted and summed using the final attention weight as a coefficient, and the result of the weighted sum is processed by a nonlinear activation function to obtain the updated virtual node representation; the nonlinear activation function here can also be ReLU or LeakyReLU function to introduce nonlinear expressive power; the final attention weight reflects the correlation between the initial characteristics of the new sterilization batch and each historical performance mode at the feature level, and integrates the reliability prior knowledge of each historical performance mode itself; The process of generating personalized initial sterilization procedure parameters for the new sterilization batch based on the matching results is as follows: For each efficiency pattern node in the sterilization efficiency knowledge graph, a historical optimal program parameter set is associated. This historical optimal program parameter set comes from the actual program parameters of multiple historical sterilization batches in the clusters that constitute that efficiency pattern node, where the final sterilization result is labeled as successful and the overall efficiency index is optimal. The optimal overall efficiency index can be evaluated by defining an objective function, for example, objective function value = average sterilization success rate * w_1 - average energy consumption * w_2 - average program time * w_3, where w_1, w_2, and w_3 are preset positive weights. The program parameters of the top N batches (e.g., N=5) that maximize the objective function value from each cluster constitute the historical optimal program parameter set for that node. The actual program parameters include at least the target sterilization temperature and holding time. Then, based on the query feature vector obtained from the matching calculation in step S3 and the final attention weight between each performance mode node, the historical best program parameter set associated with each performance mode node is weighted and fused to generate personalized initial sterilization program parameters for the new sterilization batch; the specific process of weighted fusion to generate the baseline value of personalized initial sterilization program parameters is as follows: First, for each performance mode node, the arithmetic mean of the components of the historical best program parameters is calculated, resulting in an average program parameter vector. The arithmetic mean is calculated separately for each parameter (such as temperature and time) in the historical best program parameter set. Then, using the final attention weight between the query feature vector of the new sterilization batch and the performance mode node obtained in step S3 as a weighting coefficient, the average program parameter vectors corresponding to all performance mode nodes are weighted and summed. The result of the weighted sum is used as the baseline value for the generated personalized initial sterilization program parameters. This operation generates a "center of gravity" that is biased towards the typical program parameters corresponding to those performance modes that are more similar to the current new batch and have more reliable historical performance. Meanwhile, the specific process for generating the allowable fluctuation range of personalized initial sterilization program parameters through weighted fusion is as follows: First, calculate the variance of each program parameter component in the historical best program parameter set associated with each performance mode node; the variance is the sample variance, reflecting the dispersion of the optimal program parameter represented by that node in its historical batches; then, calculate the square of the difference between the average program parameter vector of each performance mode node and the baseline value of the personalized initial sterilization program parameters for each corresponding component; this reflects the deviation between the typical parameters of that node and the final recommended baseline value; finally, add the variance of each performance mode node to the corresponding square of the difference to obtain a... The intermediate vector is added to incorporate the uncertainties within the node and the systematic deviation between the node and the final recommended baseline. Then, the intermediate vectors corresponding to all effective mode nodes are summed using the final attention weight as the weighting coefficient. Finally, the square root of each component of the weighted summation vector is performed, and the result is used as the allowable fluctuation range of the generated personalized initial sterilization program parameters. This allowable fluctuation range can be statistically interpreted as a standard deviation estimate based on attention weighting. It defines a safe parameter adjustment boundary for subsequent online optimization steps. For example, the baseline temperature ± fluctuation range is the allowable operating range of temperature during online adjustment. The baseline values ​​and allowable fluctuation ranges of the personalized initial sterilization procedure parameters together constitute a complete description of the personalized initial sterilization procedure parameters for a new sterilization batch.

[0022] In this embodiment, the specific process of processing the real-time acquired process parameter time-series data using a pre-trained sterilization efficiency prediction model in step S4 is as follows: The pre-trained sterilization efficiency prediction model receives real-time process parameter time-series data as input, with the current moment as the endpoint and a fixed time length preceding it. The fixed time length, i.e., the length of the time window for model input, can be set according to the typical duration of the sterilization procedure, for example, covering data 5 to 10 minutes backward from the current moment to ensure sufficient process dynamic information is included. This segment of real-time process parameter time-series data must contain at least temperature-time curve segments and pressure-time curve segments from multiple temperature measurement points; typically, it must include temperature data at least at the geometric center of the cavity, cold spots (such as drain outlets), and multiple representative locations. The pre-trained sterilization efficiency prediction model first... The input is processed through a bidirectional long short-term memory (LSTM) network layer. This bidirectional LSM network layer consists of two independent LSM networks, one for forward propagation and one for backward propagation. The forward propagation network processes the input from the start time to the end time and from the end time to the start time, respectively. The network state at each time step is formed by concatenating the network states at the corresponding time steps of forward and backward propagation. This results in the hidden state sequence corresponding to each time step in the input, which integrates the temporal information from the past and future. The bidirectional structure allows the model to utilize both "past" and "future" contextual information to better represent the features of the current time step, making it particularly suitable for analyzing processes with clear stages, such as heating, heat preservation, and cooling. The training process of the pre-trained sterilization efficiency prediction model is as follows: The model is trained based on the structured feature dataset of historical sterilization batches and their corresponding process parameter time series data generated in step S1. Specifically, from the complete process parameter time series data of each historical sterilization batch, multiple fixed-length data segments are slid out in chronological order. The end time of each data segment corresponds to a historical moment in the sterilization process of that batch. Each data segment is used as a training input sample, and the two true values ​​of the historical batch corresponding to the end time of the segment are used as training labels: one is the final sterilization result label of the historical batch (success or failure), and the other is the final heat lethality calculated when the program ends. In this way, a large training sample set is constructed. Its essence is to let the model learn the mapping relationship of predicting the final outcome of the batch from any historical moment's "process segment". During training, a combined loss function is used to simultaneously optimize the model's predictions for two objectives: the first is the prediction of the sterilization failure probability, which uses a binary cross-entropy loss function to make the probability value output by the model approximate the true success / failure label; the second is the prediction of the final thermal lethal dose, which uses a mean squared error loss function to reduce the numerical difference between the predicted thermal lethal dose and the actual thermal lethal dose. The total training loss is the weighted sum of the two loss functions mentioned above, and the weight coefficients can be adjusted according to actual needs, for example, all can be set to 1.0 to achieve balanced optimization. During training, the backpropagation algorithm and adaptive moment estimation optimizer are used to iteratively update all learnable parameters in the model, including the weight matrix and bias vector in the bidirectional long short-term memory network layer, attention mechanism layer and fully connected neural network layer. Training will continue for multiple rounds until the model’s prediction performance on independent validation datasets no longer improves significantly, thereby obtaining a stable pre-trained model with strong generalization ability. This trained model is then used for online real-time prediction in step S4. Next, an attention mechanism layer processes the hidden state sequence. This layer calculates the correlation score between the last hidden state and each hidden state in the sequence. The method for calculating the correlation score is as follows: First, the last hidden state and the hidden state at each historical time step are linearly transformed using their respective learnable weight matrices. Then, the two transformation results are added together and processed by a non-linear activation function to obtain an intermediate value. Finally, a dot product operation is performed between a learnable parameter vector and this intermediate value to obtain the corresponding correlation score. The non-linear activation function here is typically the tanh function. Afterwards... All calculated correlation scores are processed using a normalized exponential function, converting each correlation score into a value between zero and one, such that the sum of all converted values ​​equals one. Each converted value serves as the attention weight for the hidden state at the corresponding historical moment. The normalized exponential function is the softmax function, and the magnitude of this set of attention weights characterizes the importance of the state at different historical moments in the input to predicting the current future sterilization efficiency. For example, an abnormal pressure drop event that occurs early in the sterilization process may receive a higher attention weight, indicating that the model considers the event to have a significant impact on the final result. Then, using the attention weights as coefficients, the hidden states at each time step in the hidden state sequence are summed in a weighted manner to generate a comprehensive context vector. Finally, the last hidden state of the hidden state sequence is concatenated with the comprehensive context vector and passed through a fully connected neural network layer to output two predicted values: one is the predicted sterilization failure probability at the end of the program, which is between zero and one; this value can be interpreted as a risk estimate of failing to achieve the sterility assurance level at the end of the program under the current process state; the other is the predicted final heat lethal dose at the end of the program; this predicted value can be compared with the real-time calculated F0 value as an auxiliary verification of the reliability of the model prediction. Based on the predicted sterilization results, the specific process of using model predictive control methods to adaptively adjust the personalized initial sterilization program parameters online is as follows: For each control cycle executed during the sterilization process, a rolling time-domain optimization problem is constructed and solved. The control cycle is consistent with the sampling cycle of the sterilization equipment control system, for example, 1 second or 5 seconds. The optimization variable of the rolling time-domain optimization problem is the sequence of control increments for several consecutive control cycles in the future. The control increment refers to the adjustment of the control commands of the sterilization equipment actuator, such as the adjustment of the opening of the heating valve. The number of future control cycles M is usually less than the prediction time-domain length P. For example, M=5, P=20, which means optimizing the control actions for the next 5 cycles, but evaluating their impact on the next 20 cycles. The objective function of the rolling time-domain optimization problem is to minimize an objective function value, which is the sum of three weighted terms. The three weighted terms are calculated as follows: for each future moment in the prediction time-domain of the rolling time-domain optimization problem, the first weight coefficient is multiplied by the sterilization failure probability output by the pre-trained sterilization efficiency prediction model at that moment, and the second weight coefficient is multiplied by the control increment at that moment. The predicted energy consumption, plus the third weighting coefficient multiplied by the square of the difference between the predicted key program parameters at that moment and the reference parameter values ​​at that moment, yields the contribution value for that future moment. Then, the contribution values ​​for all future moments within the prediction time domain are summed to obtain the objective function value. The first, second, and third weighting coefficients need to be tuned during the system debugging phase to achieve a balance between safety, economy, and stability. For example, the first weighting coefficient could be 10.0, the second weighting coefficient 0.01, and the third weighting coefficient 1.0. The first factor drives the controller to reduce sterilization risk, the second factor drives the controller to reduce energy consumption, and the third factor drives the control process to smooth out the process and track the initial optimization results as much as possible. The reference parameter value is taken from the baseline value of the personalized initial sterilization program parameters generated in step S3 during the heat preservation phase of the sterilization process; during the heating or cooling phase, the corresponding value on the preset ideal process curve is taken. The ideal process curve defines the standard heating rate and the target cooling rate. The rolling time-domain optimization problem includes several constraints: The first constraint is the terminal sterilization guarantee constraint, requiring that the final heat lethality predicted by the sterilization efficiency prediction model in step S4 at the very end of the prediction time domain is not less than a target heat lethality set by the sterilization process requirements; for example, for an overkill procedure, the target heat lethality F0_target can be set to 12 minutes. The second constraint is the parameter safety constraint, requiring that the predicted values ​​of key program parameters at each future time in the prediction time domain must be within a dynamic safety range. The lower limit of this dynamic safety range is obtained by subtracting the component of the corresponding parameter in the allowable fluctuation range generated in step S3 from the baseline value of the corresponding time in the personalized initial sterilization program parameters, and the upper limit is obtained by adding the component of the corresponding parameter in the allowable fluctuation range to the baseline value of the corresponding time; for example, if the baseline temperature is 121... The allowable fluctuation range is ±1.0°C, so the dynamic safety range is [120.5°C, 122.5°C]. Key program parameters include the target sterilization temperature. The third constraint is a control increment constraint, requiring that the value of the optimized variable, i.e., the control increment for each future control cycle, cannot exceed the maximum and minimum single adjustment range allowed by the device actuator. For example, the heating valve opening adjustment is limited to ±5% each time. The fourth constraint is an optional process model constraint, used to describe the state change of key program parameters from one moment to the next. This change is determined by the parameter state of the previous moment, the currently applied control increment, and possible disturbances. This process model can be a simplified, energy-balance-based first-order inertial plus pure time-delay model used to quickly predict the temperature response to the control quantity. After solving the rolling time-domain optimization problem in each control cycle, the first control increment in the obtained future control increment sequence is applied to the actuator of the sterilization equipment to achieve online adaptive adjustment of personalized initial sterilization program parameters and generate the actual execution trajectory of the optimized sterilization program parameters. To improve control efficiency, an event triggering mechanism is introduced: when the absolute difference between the sterilization failure probability predicted in the current control cycle and the sterilization failure probability predicted in the previous control cycle exceeds a first preset threshold (e.g., the first preset threshold is set to 0.05), or the absolute value of the predicted value of the key program parameter deviates from the reference parameter value by more than a second preset threshold (e.g., for temperature, the second preset threshold is set to 2.0°C), or the sterilization process enters a preset critical stage, the critical stage includes the transition point from the heating stage to the holding stage, or the transition point from the holding stage to the cooling stage. The determination of the critical stage is usually based on whether the main control temperature has reached the set value and stabilized before restarting. The new execution complete rolling time-domain optimization problem is constructed and solved; otherwise, the control strategy obtained from the previous control cycle is used. In non-triggering cycles, only one step of the control sequence of the previous cycle can be shifted and the first control variable can be applied, which greatly reduces the computational load. After the sterilization process is completed, the actual process parameter time series data of the current batch, the final sterilization result label, and the actual execution trajectory of the optimized sterilization program parameters are used as new training samples to update the pre-trained sterilization efficiency prediction model in an incremental learning manner. The update can be carried out by online gradient descent or mini-batch update. Similarly, the historical optimal program parameter set associated with the efficiency pattern node that is closest to the characteristics of the new batch in the sterilization efficiency knowledge graph is updated. The update method can be to add the optimal program parameters of the batch (if successful and efficient) to the historical optimal program parameter set of the node, or to recalculate the average program parameter vector of the node to achieve continuous optimization of system efficiency.

[0023] This invention provides a method for intelligent optimization and online verification of steam sterilization programs based on historical data. By integrating historical multi-source data to construct a process knowledge graph, it can generate highly personalized and safe initial sterilization programs for each new batch, overcoming the energy waste or insufficient sterilization risks caused by traditional fixed programs. During execution, a data-driven predictive model is used to evaluate sterilization efficiency in real time, and dynamic parameter optimization is performed in conjunction with model predictive control. Under the premise of ensuring absolute sterilization safety, it actively saves energy and shortens the cycle. The entire solution realizes a fundamental transformation from experience-driven to data-intelligent driven, forming a complete closed loop of perception, decision-making, execution, and self-learning, which can improve the reliability, economy, and adaptability of the sterilization process.

[0024] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0025] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent optimization and online validation of steam sterilization programs based on historical data, characterized in that, Specifically, the steps include the following: Step S1: In response to the optimization command, extract multimodal data corresponding to historical sterilization batches from the multi-source system; perform time-series alignment and feature extraction processing on the multimodal data to generate a structured feature dataset of historical sterilization batches, wherein the multimodal data includes process parameter time-series data, batch attribute data and environmental metadata; Step S2: Based on the structured feature dataset of historical sterilization batches, use clustering algorithm to mine efficiency patterns, and construct a sterilization efficiency knowledge graph based on the mining results. In the sterilization efficiency knowledge graph, the nodes represent different sterilization efficiency patterns, and the edges represent the association between patterns. Step S3: When the batch attribute data and environmental metadata of a new sterilization batch are obtained as initial features, the initial features are matched with the sterilization efficiency knowledge graph, and personalized initial sterilization program parameters for the new sterilization batch are generated based on the matching results. Step S4: Execute the sterilization process based on the personalized initial sterilization program parameters generated in step S3. During this process, a pre-trained sterilization efficiency prediction model is used to process the real-time collected process parameter time-series data to obtain the predicted sterilization result. Based on the predicted sterilization result, the personalized initial sterilization program parameters are adjusted online adaptively using a model predictive control method to generate optimized sterilization program parameters.

2. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 1, characterized in that: In step S1, multimodal data is extracted from a multi-source system, which includes a sterilization equipment control system, a manufacturing execution system, a laboratory information management system, and an environmental monitoring system. Process parameter time-series data is extracted from the sterilization equipment control system, including temperature-time curves, pressure-time curves, and equipment status sequences measured by temperature sensors. Batch attribute data is extracted from the manufacturing execution system, including the type of loaded items, loading density, and packaging materials. A final sterilization result label is obtained from the laboratory information management system, used to identify whether sterilization was successful or failed. Environmental metadata is extracted from the environmental monitoring system, including ambient temperature and humidity, and supplied steam pressure.

3. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 2, characterized in that: The specific process for performing time-series alignment on multimodal data is as follows: From the temperature-time curves included in the process parameter time series data, select a baseline temperature-time curve for each historical sterilization batch; For this historical sterilization batch, a dynamic time bending algorithm is used to calculate the optimal bending path between the pressure-time curve, equipment state sequence, and other temperature-time curves in the process parameter time series data of this batch and the reference temperature-time curve. Based on the calculated optimal bending paths, the time axes of the pressure-time curve, equipment state sequence, and other temperature-time curves are nonlinearly scaled and adjusted to align the adjusted pressure-time curve, equipment state sequence, and other temperature-time curves with the baseline temperature-time curve in the time dimension. This results in an aligned time-series data set for each historical sterilization batch, which includes all adjusted temperature-time curves, adjusted pressure-time curves, and adjusted equipment state sequences.

4. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 3, characterized in that: The specific process for feature extraction from multimodal data is as follows: Based on the aligned time-series data set, physical guidance features are constructed for each historical sterilization batch. To construct the physical guidance feature, firstly, based on the temperature-time curves in the aligned time-series data set, the cumulative heat lethality curve of the historical sterilization batch at the temperature measurement location corresponding to each temperature-time curve is calculated; then, the dynamic time bending distance between each pair of cumulative heat lethality curves corresponding to different temperature measurement locations is calculated, and the arithmetic mean of the dynamic time bending distance between each pair of cumulative heat lethality curves is obtained. This arithmetic mean is used as the heat distribution uniformity feature characterizing the spatial heat distribution uniformity of the sterilization process of the historical sterilization batch. Simultaneously, based on the pressure-time curve and the reference temperature-time curve in the aligned time-series data set, the theoretical saturated steam pressure corresponding to each temperature data point on the reference temperature-time curve is queried to obtain the theoretical saturated steam pressure curve; the absolute value of the difference between the pressure-time curve value at each moment and the theoretical saturated steam pressure curve value at the corresponding moment is calculated within the preset exposure time period of the sterilization process, and this absolute value is integrated over the entire exposure time period to obtain the first integral value; the first integral value is normalized by dividing it by the length of the exposure time period, and the normalized value is used as the steam quality dynamic deviation feature characterizing the steam quality status of the sterilization process of this historical sterilization batch; Finally, the heat distribution uniformity characteristics, steam quality dynamic deviation characteristics, conventional statistical characteristics extracted from the aligned time-series data set, batch attribute data, and environmental metadata are combined to form the final structured feature vector of the historical sterilization batch. The structured feature vectors of all historical sterilization batches and their corresponding final sterilization result labels constitute the structured feature dataset of historical sterilization batches.

5. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 4, characterized in that: In step S2, the specific operation of performance pattern mining using clustering algorithm is as follows: Based on the structured feature dataset of historical sterilization batches, the composite distance between any two historical sterilization batches is calculated. The composite distance is obtained by weighted summation of the first sub-distance and the second sub-distance. The first sub-distance is calculated based on a subset of features in the structured feature vector of the historical sterilization batches, including batch attribute data, environmental metadata, and conventional statistical features, and is calculated using standardized Euclidean distance. The second sub-distance is calculated based on the heat distribution uniformity feature and steam quality dynamic deviation feature in the structured feature vector of the historical sterilization batches, and is calculated using Mahalanobis distance. After obtaining the composite distances between all pairs of historical sterilization batches, the batch-to-batch distance matrix formed by these composite distances is used to divide all historical sterilization batches into multiple clusters, each cluster corresponding to a preliminary performance model.

6. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 5, characterized in that: The specific steps for constructing a sterilization efficacy knowledge graph based on the mining results are as follows: For each obtained cluster, a performance evaluation is performed to form a final performance pattern; for each cluster, its average sterilization success rate is calculated, which is the proportion of the number of batches whose final sterilization result is marked as successful out of all historical sterilization batches in the cluster. Calculate the performance index vector for the cluster, where each component of the performance index vector represents the average heat lethality, average energy consumption, and average program duration of all historical sterilization batches within the cluster. Extract the central feature vector of the cluster, which is the arithmetic mean of the structured feature vectors of all historical sterilization batches within the cluster; An effectiveness pattern is defined by its central feature vector, average sterilization success rate, and effectiveness index vector. Then, each effectiveness pattern is used as a node in the knowledge graph, and the node is configured with attributes containing its central feature vector, average sterilization success rate, and effectiveness index vector. Then, the sequential relationship between historical sterilization batches is analyzed. The sequential relationship refers to the sequence formed by arranging historical sterilization batches according to their execution time, i.e., the historical sequence. When the historical sterilization batch data contains historical sequences arranged by time or process logic, the number of times a historical sterilization batch belonging to the previous efficiency mode immediately follows a historical sterilization batch belonging to the next efficiency mode is counted in the historical sequence. This number is recorded as the number of transfers from the previous efficiency mode node to the next efficiency mode node. The quotient obtained by dividing this number of transfers by the total number of transfers from the previous efficiency mode node to all possible subsequent efficiency mode nodes is used as the weight. A directed edge is established between the corresponding two efficiency mode nodes, thereby constructing a sterilization efficiency knowledge graph containing multiple nodes and weighted directed edges. When historical sterilization batch data does not contain explicit sequence information, the weights are calculated based on the distance between the central eigenvectors of the two performance modes in the feature space. This distance is obtained by calculating the composite distance between the central eigenvectors of the two performance modes.

7. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 6, characterized in that: In step S3, the specific process of matching the initial features with the sterilization efficacy knowledge graph is as follows: The input to the matching process consists of two parts: first, the sterilization efficiency knowledge graph output in step S2, where each efficiency pattern node has attributes including a central feature vector, average sterilization success rate, and efficiency index vector; second, the batch attribute data and environmental metadata of the new sterilization batch, which constitute the initial features of the new sterilization batch; first, the initial features of the new sterilization batch are zero-padded and aligned to generate a query feature vector with the same dimension as the central feature vector of the efficiency pattern node. The graph attention network is used for processing. First, for each efficiency pattern node in the sterilization efficiency knowledge graph, its central feature vector is concatenated with the average sterilization success rate and efficiency index vector. Then, through a learnable linear transformation, it is mapped to a unified feature space of a preset dimension to obtain the initial hidden state of each efficiency pattern node. Next, the query feature vector is treated as a virtual node and mapped to a unified feature space of a preset dimension through a learnable linear transformation; then, the attention coefficient between the virtual node and each performance mode node is calculated. Then, the unnormalized attention coefficients corresponding to all performance mode nodes are normalized to obtain the final attention weights between the query feature vector and each performance mode node. Finally, the final representation of the virtual node is updated by aggregating the information of all effective mode nodes: the initial hidden states of each mapped effective mode node are weighted and summed using the final attention weight as a coefficient, and the result of the weighted sum is processed by a non-linear activation function to obtain the updated virtual node representation.

8. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 7, characterized in that: The process of generating personalized initial sterilization procedure parameters for the new sterilization batch based on the matching results is as follows: For each efficiency mode node in the sterilization efficiency knowledge graph, a set of historical best program parameters is associated. This set of historical best program parameters comes from the clusters that constitute the efficiency mode node, and consists of the actual program parameters of multiple historical sterilization batches that were ultimately labeled as successful and had the best overall efficiency index. The actual program parameters include at least the target sterilization temperature and the holding time. Then, based on the query feature vector obtained from the matching calculation in step S3 and the final attention weight between each performance mode node, the historical best program parameter set associated with each performance mode node is weighted and fused to generate personalized initial sterilization program parameters for the new sterilization batch. The specific process of generating baseline values ​​for personalized initial sterilization program parameters through weighted fusion is as follows: First, for each performance mode node associated with the historical best program parameter set, calculate the arithmetic mean of each program parameter component to obtain an average program parameter vector; then, using the final attention weight between the query feature vector of the new sterilization batch calculated in step S3 and the performance mode node as a weighting coefficient, perform a weighted summation operation on the average program parameter vectors corresponding to all performance mode nodes, and use the result of the weighted summation as the benchmark value for the generated personalized initial sterilization program parameters; Meanwhile, the specific process of generating the allowable fluctuation range of personalized initial sterilization program parameters through weighted fusion is as follows: First, calculate the variance of each program parameter component in the historical best program parameter set associated with each performance mode node; then, calculate the square of the difference between the average program parameter vector of each performance mode node and the baseline value of the personalized initial sterilization program parameters for each corresponding component; next, add the variance of each performance mode node to the square of the corresponding difference to obtain an intermediate vector; then, using the final attention weight as the weighting coefficient, perform a weighted summation on the intermediate vectors corresponding to all performance mode nodes; finally, perform a square root operation on each component of the weighted summation vector, and use the result as the allowable fluctuation range of the generated personalized initial sterilization program parameters. The baseline values ​​and allowable fluctuation ranges of the personalized initial sterilization procedure parameters together constitute a complete description of the personalized initial sterilization procedure parameters for a new sterilization batch.

9. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 8, characterized in that: In step S4, the specific process of processing the real-time acquired process parameter time-series data using a pre-trained sterilization efficiency prediction model is as follows: The pre-trained sterilization efficiency prediction model receives real-time process parameter time-series data that is truncated to a fixed time length from the current time as input; the truncated real-time process parameter time-series data includes at least temperature-time curve segments and pressure-time curve segments from multiple temperature measurement points. The pre-trained sterilization efficiency prediction model first processes the input through a bidirectional long short-term memory network layer. This bidirectional long short-term memory network layer consists of two independent long short-term memory networks, forward propagation and backward propagation, which process the input from the start time to the end time and from the end time to the start time, respectively. The network state at each time step is formed by splicing the network states at the corresponding time steps of forward propagation and backward propagation, thereby obtaining the hidden state corresponding to each time step in the input, which integrates the temporal information before and after the input, forming a hidden state sequence. Next, the hidden state sequence is processed through an attention mechanism layer, which calculates the correlation score between the last hidden state in the hidden state sequence and each hidden state in the sequence. Then, all the calculated correlation scores are processed by a normalized exponential function to convert each correlation score into a value between zero and one. Each value obtained after conversion is used as the attention weight of the hidden state at the corresponding historical time. Then, using the attention weights as coefficients, the hidden states at each time step in the hidden state sequence are summed in a weighted manner to generate a comprehensive context vector. Finally, the last hidden state of the hidden state sequence is concatenated with the comprehensive context vector and passed through a fully connected neural network layer to output two predicted values: one is the predicted sterilization failure probability at the end of the program; the other is the predicted final thermal lethal dose at the end of the program.

10. The method for intelligent optimization and online validation of steam sterilization programs based on historical data according to claim 9, characterized in that: The specific process of using model predictive control to adaptively adjust the personalized initial sterilization program parameters online based on the predicted sterilization results is as follows: For each control cycle executed during the sterilization process, a rolling time-domain optimization problem is constructed and solved; The optimization variable for the rolling time-domain optimization problem is the sequence of control increments for several consecutive future control cycles. The objective function of the rolling time-domain optimization problem is to minimize a single objective function value, which is the sum of three weighted terms. The weighted sum is calculated as follows: For each future moment in the prediction time domain, the first weight coefficient is multiplied by the sterilization failure probability output by the pre-trained sterilization efficiency prediction model at that moment; the second weight coefficient is multiplied by the energy consumption predicted based on the control increment at that moment; and the third weight coefficient is multiplied by the square of the difference between the predicted key program parameters and the reference parameters at that moment. This yields the contribution value for that future moment. Then, the contribution values ​​for all future moments in the prediction time domain are summed to obtain the objective function value. The rolling time-domain optimization problem includes several constraints: the first constraint is the terminal sterilization guarantee constraint, which requires that the final heat lethality predicted by the sterilization efficiency prediction model in step S4 at the very end of the prediction time domain is not less than a target heat lethality set by the sterilization process requirements; the second constraint is the parameter safety constraint, which requires that the predicted values ​​of the key program parameters at each future time in the prediction time domain must be within a dynamic safety range. The lower limit of this dynamic safety range is obtained by subtracting the component of the corresponding parameter in the allowable fluctuation range generated in step S3 from the baseline value of the corresponding time in the personalized initial sterilization program parameters, and the upper limit is obtained by adding the component of the corresponding parameter in the allowable fluctuation range to the baseline value of the corresponding time. Key procedure parameters include the target sterilization temperature; The third constraint is the control increment constraint, which requires that the value of the optimized variable, namely the control increment of each future control cycle, cannot exceed the maximum and minimum adjustment range allowed by the device actuator in a single operation; the fourth constraint is an optional process model constraint, which describes the state change of key program parameters from one time to the next time. This change is determined by the parameter state of the previous time, the currently applied control increment, and possible disturbances. After solving the rolling time-domain optimization problem in each control cycle, the first control increment in the obtained future control increment sequence is applied to the actuator of the sterilization equipment to realize online adaptive adjustment of personalized initial sterilization program parameters and generate the actual execution trajectory of the optimized sterilization program parameters.

Images