Intelligent monitoring and maintenance method and system for large volume parenteral production equipment

Abstract

The invention discloses an intelligent monitoring maintenance method and system for large volume parenteral production equipment, and relates to the technical field of intelligent manufacturing and industrial automation, and the method comprises the steps: synchronously collecting equipment operation state data, process parameter data and environment monitoring data through a multi-source sensor array; performing space-time alignment and semantic fusion processing on the collected data, and extracting a multi-dimensional depth feature vector based on a deep convolutional neural network; inputting the feature vector into a hybrid model based on the combination of a long short-term memory network and an attention mechanism, and carrying out fault mode early recognition and residual service life prediction; generating an optimal maintenance strategy and a scheduling instruction according to the diagnosis result and the production batch plan; and continuously monitoring historical data of the equipment and dynamically adjusting model parameters. According to the method, full-dimension perception of the equipment state, high-precision fault prediction, optimal maintenance strategy generation and adaptive learning can be realized, and the availability and maintenance efficiency of the equipment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial automation, and particularly relates to an intelligent monitoring and maintenance method and system for large infusion production equipment. BACKGROUND

[0002] With the continuous improvement of the automation and intelligent level of the pharmaceutical industry, as the most basic and largest dosage type in clinical treatment, the production process of large infusion puts forward very high requirements on the stability of equipment, the continuity of process and the ability of sterile guarantee. Large infusion production equipment covers multiple key process units such as filling, sealing, sterilization, lamp inspection and packaging, and the running state of the equipment directly determines the quality consistency and production safety of the product. At present, the pharmaceutical industry generally adopts a regular maintenance strategy based on time or number of uses, and assists with manual inspection and a simple alarm mechanism to monitor the state of the equipment.

[0003] However, such traditional monitoring and maintenance mode exposes significant technical limitations in the face of large infusion production scenarios with high cleanliness, high continuity and high compliance: first, the state perception dimension of the equipment is single, only relying on a few fixed sensors to collect basic parameters such as temperature, pressure or speed, which cannot fully depict the hidden fault characteristics such as internal mechanical wear, sealing performance degradation or fluid passage blockage of the equipment; second, the monitoring system lacks the ability of spatio-temporal alignment and semantic fusion of multi-source heterogeneous data, resulting in that abnormal signals are covered by noise or misjudged, and it is difficult to achieve accurate identification of early faults; third, the maintenance decision is disconnected from production scheduling, and the maintenance instruction is generated by an independent system without linkage with process parameters such as batch plan, clean area pressure difference and sterilization cycle, which is easy to cause unplanned downtime or cross-contamination risk; finally, the existing system does not have self-adaptive learning ability and cannot dynamically optimize the monitoring threshold and maintenance strategy according to the historical running data of the equipment, environmental disturbance and process changes, resulting in waste of maintenance resources or missed detection of key component failures. SUMMARY

[0004] The purpose of the present application is to provide an intelligent monitoring and maintenance method and system for large infusion production equipment to solve the problems of inaccurate equipment fault identification, high maintenance cost and frequent unplanned downtime in the prior art.

[0005] To solve the above technical problems, the present application provides the following technical solutions: An intelligent monitoring and maintenance method and system for large infusion production equipment, comprising: The multi-source heterogeneous data acquisition module is configured to synchronously acquire equipment running state data, process parameter data and environment monitoring data from the filling unit, sealing unit, sterilization unit, lamp inspection unit and packaging unit of the large infusion production line; the multi-source heterogeneous data acquisition module is integrated with a vibration sensor array, an acoustic emission sensor, a pressure transmitter, a temperature sensor, a flow meter, a visual imaging unit and a clean area differential pressure sensor; The data fusion and feature extraction module is connected to the multi-source heterogeneous data acquisition module and is configured to perform time-space alignment and semantic fusion processing on the acquired multi-source heterogeneous data and extract a multi-dimensional deep feature vector representing the equipment health state based on a deep convolutional neural network; the data fusion and feature extraction module is internally provided with a time alignment unit, a wavelet packet transformation unit and a feature dimension reduction unit; The intelligent diagnosis and prediction module is connected to the data fusion and feature extraction module and is configured to receive the multi-dimensional deep feature vector and perform early identification of equipment fault modes and prediction of residual service life based on a hybrid model combining a long short-term memory network and an attention mechanism; The maintenance decision and scheduling optimization module is connected to the intelligent diagnosis and prediction module and an enterprise resource planning system and is configured to generate optimal maintenance strategies and scheduling instructions according to fault diagnosis results and residual service life prediction values, in combination with current production batch plans, clean area differential pressure dynamics and sterilization cycle constraints; the maintenance decision and scheduling optimization module is integrated with a constraint satisfaction problem solver and a multi-objective optimization algorithm; The adaptive learning and model updating module is connected to the intelligent diagnosis and prediction module and the maintenance decision and scheduling optimization module and is configured to continuously monitor equipment historical running data, environment disturbance data and process change records and dynamically adjust monitoring thresholds of the hybrid model and weight parameters of the multi-objective optimization algorithm based on an incremental learning mechanism.

[0006] Preferably, in the multi-source heterogeneous data acquisition module, the vibration sensor array acquires equipment mechanical vibration signals at a sampling frequency of 5120 Hz, the acoustic emission sensor has an operating frequency range of 100 kHz to 1 MHz, the pressure transmitter has an accuracy level of 0.075%, the temperature sensor adopts a PT100 platinum resistance, the flow meter is a Coriolis mass flow meter, the visual imaging unit has a resolution of 2048x2048 pixels and a frame rate of 60 fps, and the clean area differential pressure sensor has a range of 0 to 50 Pa.

[0007] Furthermore, the time alignment unit of the data fusion and feature extraction module uses a dynamic time warping algorithm to perform time synchronization on sensor data from different sampling periods, and a cumulative distance cost function of a warping path thereof is defined as: wherein, Indicates the first The time point and the first Euclidean distance between data sequences at different time points.

[0008] Furthermore, the wavelet packet transform unit of the data fusion and feature extraction module performs a four-level decomposition of the vibration signal, extracting the wavelet packet energy entropy of each node in the fourth level as a time-frequency domain feature. The calculation formula is: in, Indicates the first Within each frequency band The energy percentage of each wavelet coefficient This represents the total number of wavelet coefficients within this frequency band.

[0009] Preferably, the hybrid model of the intelligent diagnosis and prediction module is composed of a bidirectional long short-term memory network and a multi-head self-attention mechanism connected in series. The bidirectional long short-term memory network has 128 hidden layer units, and the multi-head self-attention mechanism has 8 heads. Its attention weight calculation adopts scaled dot product attention.

[0010] Furthermore, the multi-objective optimization algorithm of the maintenance decision and scheduling optimization module simultaneously optimizes three objectives: equipment availability, maintenance cost, and product quality risk. Its objective function is expressed as minimizing the weighted sum of total unplanned downtime, total maintenance cost, and batch product defect rate. The constraints include the maximum allowable maintenance time window, the minimum cleanroom pressure differential maintenance value, and the sterilizer cooling time.

[0011] Furthermore, the incremental learning mechanism of the adaptive learning and model update module adopts an online sequential extreme learning machine, whose hidden layer node number is dynamically adjusted from 100 to 500, and whose initial learning rate is 0.01 and decays exponentially with each training round.

[0012] A method for intelligent monitoring and maintenance of large-volume parenteral solution production equipment, the specific steps of which are as follows: Step S110: Through a multi-source sensor array deployed in each process unit of the large-volume parenteral solution production line, equipment operating status data, process parameter data and environmental monitoring data are collected synchronously. Step S120: Spatiotemporal alignment and semantic fusion processing are performed on the collected multi-source heterogeneous data, and multi-dimensional deep feature vectors representing the health status of the device are extracted based on deep convolutional neural networks. Step S130: Input the multidimensional deep feature vector into a hybrid model based on a combination of long short-term memory network and attention mechanism to perform early identification of equipment failure modes and prediction of remaining service life. Step S140: Based on the fault diagnosis results and the predicted remaining service life, combined with the current production batch plan, the dynamic pressure difference in the clean area and the sterilization cycle constraints, the optimal maintenance strategy and scheduling instructions are generated through a multi-objective optimization algorithm. Step S150: Continuously monitor historical operating data of the equipment, environmental disturbance data, and process change records, and dynamically adjust the monitoring threshold of the hybrid model and the weight parameters of the multi-objective optimization algorithm based on the incremental learning mechanism.

[0013] Compared with the prior art, the beneficial technical effects of the present invention are as follows: This invention achieves full-dimensional perception of equipment status and accurate characterization of latent fault features through multi-source heterogeneous data acquisition and deep fusion, significantly improving the accuracy and timeliness of early fault identification.

[0014] This invention is based on a hybrid model that combines long short-term memory networks and attention mechanisms to effectively capture long-term dependencies and key failure modes in device status data, thereby achieving high-precision prediction of remaining service life.

[0015] This invention deeply couples maintenance decisions with production scheduling and process parameters, and generates an optimal maintenance strategy that balances equipment availability, maintenance economy and product quality through multi-objective optimization, thereby minimizing the risks of unplanned downtime and cross-contamination.

[0016] This invention introduces an adaptive learning mechanism, enabling the system to dynamically adjust monitoring thresholds and decision parameters based on equipment operating history and environmental changes, thereby continuously improving the accuracy of monitoring and maintenance and the efficiency of resource utilization. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical architecture of an intelligent monitoring and maintenance method and system for large-volume infusion production equipment proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of multi-source heterogeneous data acquisition and intelligent diagnosis and prediction in this invention. Detailed Implementation

[0018] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0019] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0020] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).

[0021] Example 1 In the large-volume parenteral solution production workshop of a large pharmaceutical company, a multi-source heterogeneous data acquisition module comprehensively collects data from the filling, sealing, sterilization, light inspection, and packaging units. A vibration sensor array continuously acquires mechanical vibration signals from the filling pump at a sampling frequency of 5120 Hz, accurately capturing the high-frequency resonant components at the initial stage of bearing wear. Acoustic emission sensors monitor stress wave signals generated by fatigue of the sterilization cabinet's inner wall material within the operating frequency range of 100 kHz to 1 MHz, effectively identifying microcrack propagation characteristics. Pressure transmitters track pressure pulsations in the filling pipeline in real time with an accuracy class of 0.075%, temperature sensors using PT100 platinum resistance thermometers accurately measure the heat distribution during sterilization, and Coriolis mass flow meters simultaneously monitor filling accuracy deviations. A visual imaging unit continuously captures ampoule sealing integrity images at a resolution of 2048×2048 pixels and a frame rate of 60 frames per second, while a cleanroom differential pressure sensor dynamically monitors environmental pressure fluctuations within a range of 0 to 50 Pa. All sensor data is synchronized at the microsecond level via the Industrial Ethernet protocol, forming a real-time monitoring data stream containing 128 dimensions.

[0022] After receiving heterogeneous data from multiple sources, the data fusion and feature extraction module uses a dynamic time warping algorithm in its temporal alignment unit to synchronize the sampling data across different frequencies. This algorithm constructs the optimal warping path by calculating the Euclidean distance matrix between data sequences from different sensors, minimizing the cumulative distance cost. Specifically, during the warping process, the cumulative distance cost of the warping path at the current moment is determined by the state of the warping path at the previous moment, mathematically expressed as the sum of the Euclidean distance between the current data points and the historical minimum cumulative distance. The spatiotemporally aligned data stream enters the wavelet packet transform unit for four-level signal decomposition, extracting the wavelet packet energy entropy of each node in the fourth level as time-frequency domain features. The energy distribution of wavelet coefficients within each frequency band is quantified using information entropy, specifically calculated by calculating the negative logarithmic weighted sum after statistically analyzing the energy proportion of each wavelet coefficient. The feature dimensionality reduction unit uses principal component analysis to compress the 2048-dimensional original features into a 32-dimensional core feature vector, retaining over 95% of the original information.

[0023] After receiving multi-dimensional deep feature vectors, the intelligent diagnosis and prediction module's hybrid model first captures the long-term dependencies of device state data through a bidirectional long short-term memory network. This network has 128 hidden layer units, with the forward and backward layers learning historical and future state features, respectively. The network output, after layer normalization, is input into a multi-head self-attention mechanism, where eight attention heads calculate the association weights between different feature dimensions in parallel. The attention weights are calculated using a scaled dot product method, dividing the dot product of the query vector and key vector by the square root of the dimension and then normalizing it using a softmax function. Finally, the attention output and the value vector are weighted and summed to form a 128-dimensional state representation vector, which is then input into the fault classifier and regression predictor. The fault classifier uses a fully connected layer and a softmax activation function to output the probabilities of seven typical fault modes, while the regression predictor outputs the predicted remaining service life through a linear layer, with the prediction error controlled within ±8 hours.

[0024] After receiving fault diagnosis results and remaining service life predictions, the maintenance decision-making and scheduling optimization module first verifies the current production batch plan, cleanroom differential pressure dynamics, and sterilization cycle constraints using a constraint-satisfying problem solver. A multi-objective optimization algorithm simultaneously calculates the Pareto front for three objectives: equipment availability, maintenance cost, and product quality risk. The objective function is specifically constructed to minimize the weighted sum of total unplanned downtime, total maintenance cost, and batch product defect rate, with weight coefficients dynamically adjusted based on the urgency of the current production task. The optimization process must meet hard constraints such as a maximum allowable maintenance time window of no more than 4 hours, a cleanroom differential pressure maintenance value of no less than 15 Pa, and a sterilizer cooling time of more than 25 minutes. The final output includes a detailed maintenance procedure arrangement, a spare parts allocation list, and personnel scheduling instructions, and is distributed to the Manufacturing Execution System via the OPC UA protocol.

[0025] The adaptive learning and model update module continuously monitors historical equipment operation data, environmental disturbance data, and process change records. Its incremental learning mechanism uses an online sequential extreme learning machine to dynamically adjust model parameters. The number of hidden layer nodes is adaptively adjusted within the range of 100 to 500 based on feature complexity. The initial learning rate is set to 0.01 and adopts an exponential decay strategy, decreasing to 0.95 times the original rate every 1000 training iterations. During model updates, the module focuses on monitoring the false negative and false positive rates. When these rates exceed limits for three consecutive batches, the model retraining process is automatically triggered. Simultaneously, the decay coefficient of the remaining service life prediction model is dynamically adjusted based on the actual load rate of the sterilization unit to ensure that the prediction accuracy remains within the 90% confidence interval.

[0026] In practice, when the system detects an abnormal increase in energy in the 2000 Hz frequency band in the vibration characteristic vector of the filling unit, the intelligent diagnosis and prediction module identifies an early wear fault in the filling pump bearing within 15 seconds, predicting that the remaining service life of this component is 126 hours. The maintenance decision and scheduling optimization module, combined with the remaining 32 hours of production plan for the current batch, calculates the optimal maintenance time as immediately after the current batch ends, performing preventative maintenance. This decision also considers the cleanroom pressure differential maintenance requirements, scheduling the standby filling unit to start preheating 30 minutes in advance to ensure uninterrupted production. The adaptive learning and model update module records the evolution of this fault and accordingly adjusts the vibration characteristic threshold from 0.35 to 0.28, improving the early detection sensitivity of similar faults.

[0027] In the sterilization unit monitoring scenario, the temperature sensor detected that the temperature uniformity index inside the sterilization cabinet dropped to 0.89, below the threshold requirement of 0.92. The data fusion and feature extraction module, combining pressure transmitter data and steam jet status images captured by the visual imaging unit, identified localized nozzle blockage. The intelligent diagnosis and prediction module, based on long short-term memory network analysis of historical maintenance records, predicted that this fault would cause fluctuations in the aseptic assurance level of the next three batches of products. The maintenance decision and scheduling optimization module immediately generated online cleaning instructions, prioritizing minimizing product quality risks by adjusting the weights of the multi-objective optimization algorithm, embedding the maintenance operation into the equipment cooling phase after the current batch sterilization cycle ends, avoiding additional production time occupation. Based on the maintenance effect, the adaptive learning and model update module shortened the monitoring cycle of the temperature uniformity index from 60 minutes to 45 minutes, improving the system's response speed to progressive faults.

[0028] Example 2 On the flexible large-volume parenteral solution production line, the multi-source heterogeneous data acquisition module has been enhanced for multi-product co-production scenarios. The vibration sensor array has been increased to 16 installation points in the filling unit, enabling differentiated monitoring of the filling process for containers of different sizes. The acoustic emission sensors adopt a ring array layout to accurately capture stress concentration signals from the equipment's mechanical structure during production transitions. The visual imaging unit has been upgraded to a 4096×4096 resolution, simultaneously acquiring sealing quality images of various packaging materials such as ampoules and vials. The density of differential pressure sensors in the clean area has been increased to one monitoring point per 10 square meters, enabling real-time construction of a dynamic model of airflow organization in the clean area.

[0029] The data fusion and feature extraction module incorporates a product specification recognition unit, which analyzes container feature images acquired by the visual imaging unit through a convolutional neural network to automatically identify the type of product currently in production. The temporal alignment unit enhances the product switching event detection algorithm, automatically reconstructing the data alignment benchmark when a product specification change is detected. The wavelet packet transform unit adaptively adjusts the number of decomposition layers for different product process parameters; ampoules use a 4-layer decomposition while vials use a 5-layer decomposition, ensuring the applicability of feature extraction. The feature dimensionality reduction unit uses supervised learning to construct a product-specific feature selector, retaining the most discriminative feature dimensions for different products.

[0030] The hybrid model of the intelligent diagnosis and prediction module incorporates a product context attention mechanism, introducing product type embedding vectors when calculating feature weights. The bidirectional long short-term memory network's hidden layers are expanded to 256 units, enhancing the model's ability to remember production sequences of multiple products. The remaining useful life predictor trains product-specific decay models separately, employing differentiated prediction strategies based on the equipment load characteristics of different products. When the system detects a product switch, it automatically loads the optimal prediction model for the corresponding product, ensuring that prediction accuracy is unaffected by product changes.

[0031] The maintenance decision-making and scheduling optimization module incorporates a product sequence optimization algorithm, simultaneously considering product changeover costs when formulating maintenance plans. A multi-objective optimization algorithm adds a product quality consistency index to ensure that maintenance decisions do not affect the quality stability between different products. The constraint satisfaction problem solver enhances the handling of cleanroom pressure differential maintenance constraints during product changeovers, dynamically adjusting air conditioning system operating parameters to ensure environmental parameters consistently comply with GMP standards. When the remaining service life of critical equipment is predicted to be close to an integer multiple of the product's production cycle, the module automatically suggests scheduling maintenance operations at product changeover nodes to maximize equipment utilization efficiency.

[0032] The adaptive learning and model update module establishes a product knowledge graph, recording the performance degradation patterns of equipment during the production process of different products. The incremental learning mechanism automatically adjusts the model update frequency based on product production volume, updating the model weekly for high-volume products and monthly for low-volume products. The online sequential extreme learning machine's hidden layer node number adjustment strategy is correlated with product complexity; the maximum number of nodes is expanded to 500 during the production of complex products and compressed to 200 during the production of simple products, achieving optimal allocation of computing resources. The system continuously analyzes the interaction effects between different products' equipment. When it detects that the production of a certain product accelerates the wear of specific components, it automatically adjusts the monitoring parameters during the production of that product, enabling proactive preventative maintenance.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring and maintenance of large-volume parenteral solution production equipment, characterized in that, The method includes the following steps: Step S110: Through a multi-source sensor array deployed in each process unit of the large-volume parenteral solution production line, equipment operating status data, process parameter data and environmental monitoring data are collected synchronously. Step S120: Spatiotemporal alignment and semantic fusion processing are performed on the collected multi-source heterogeneous data, and multi-dimensional deep feature vectors representing the health status of the device are extracted based on deep convolutional neural networks. Step S130: Input the multidimensional deep feature vector into a hybrid model based on a combination of long short-term memory network and attention mechanism to perform early identification of equipment failure modes and prediction of remaining service life. Step S140: Based on the fault diagnosis results and the predicted remaining service life, combined with the current production batch plan, the dynamic pressure difference in the clean area and the sterilization cycle constraints, the optimal maintenance strategy and scheduling instructions are generated through a multi-objective optimization algorithm. Step S150: Continuously monitor historical operating data of the equipment, environmental disturbance data, and process change records, and dynamically adjust the monitoring threshold of the hybrid model and the weight parameters of the multi-objective optimization algorithm based on the incremental learning mechanism.

2. The intelligent monitoring and maintenance method for large-volume parenteral solution production equipment according to claim 1, characterized in that, In step S120, when performing spatiotemporal alignment processing on the collected multi-source heterogeneous data, a dynamic time warping algorithm is used to synchronize the time of the sampled data at different frequencies. This algorithm constructs the optimal warping path by calculating the Euclidean distance matrix between data sequences from different sensors, so as to minimize the cumulative distance cost.

3. The intelligent monitoring and maintenance method for large-volume parenteral solution production equipment according to claim 1, characterized in that, In step S130, the hybrid model first captures the long-term dependencies of device state data through a bidirectional long short-term memory network. After the network output is processed by layer normalization, it is input into a multi-head self-attention mechanism. The correlation weights between different feature dimensions are calculated in parallel by eight attention heads. Finally, the attention output and the value vector are weighted and summed to form a 128-dimensional state representation vector, which is input into the fault classifier and the regression predictor, respectively.

4. An intelligent monitoring and maintenance system for large-volume parenteral solution production equipment, applied to the intelligent monitoring and maintenance method for the large-volume parenteral solution production equipment as described in any one of claims 1-3, characterized in that, It includes the following components: The multi-source heterogeneous data acquisition module is configured to simultaneously acquire equipment operating status data, process parameter data, and environmental monitoring data from the filling unit, sealing unit, sterilization unit, light inspection unit, and packaging unit of the large-volume infusion production line. The multi-source heterogeneous data acquisition module integrates a vibration sensor array, an acoustic emission sensor, a pressure transmitter, a temperature sensor, a flow meter, a visual imaging unit, and a clean area differential pressure sensor. The data fusion and feature extraction module is connected to the multi-source heterogeneous data acquisition module and is configured to perform spatiotemporal alignment and semantic fusion processing on the acquired multi-source heterogeneous data, and extract multi-dimensional deep feature vectors representing the health status of the device based on a deep convolutional neural network; the data fusion and feature extraction module has built-in time sequence alignment unit, wavelet packet transform unit and feature dimensionality reduction unit; The intelligent diagnosis and prediction module is connected to the data fusion and feature extraction module and is configured to receive the multi-dimensional deep feature vector and perform early identification of equipment failure modes and prediction of remaining service life based on a hybrid model combining long short-term memory network and attention mechanism. The maintenance decision and scheduling optimization module is connected to the intelligent diagnosis and prediction module and the enterprise resource planning system. It is configured to generate the optimal maintenance strategy and scheduling instructions based on the fault diagnosis results and the predicted remaining service life, combined with the current production batch plan, the dynamic pressure difference of the clean area, and the sterilization cycle constraints. The maintenance decision and scheduling optimization module integrates a constraint satisfaction problem solver and a multi-objective optimization algorithm. The adaptive learning and model update module is connected to the intelligent diagnosis and prediction module and the maintenance decision and scheduling optimization module. It is configured to continuously monitor the historical operating data of the equipment, environmental disturbance data and process change records, and dynamically adjust the monitoring threshold of the hybrid model and the weight parameters of the multi-objective optimization algorithm based on the incremental learning mechanism.

5. The intelligent monitoring and maintenance system for large-volume parenteral solution production equipment according to claim 4, characterized in that, In the multi-source heterogeneous data acquisition module, the vibration sensor array acquires the mechanical vibration signal of the equipment at a sampling frequency of 5120 Hz, the acoustic emission sensor operates at a frequency range of 100 kHz to 1 MHz, the pressure transmitter has an accuracy class of 0.075%, the temperature sensor uses PT100 platinum resistance thermometer, the flow meter is a Coriolis mass flow meter, the visual imaging unit has a resolution of 2048×2048 pixels and a frame rate of 60 frames per second, and the clean area differential pressure sensor has a range of 0 to 50 Pa.

6. The intelligent monitoring and maintenance system for large-volume parenteral solution production equipment according to claim 4, characterized in that, The time alignment unit of the data fusion and feature extraction module uses a dynamic time warping algorithm to synchronize sensor data from different sampling periods.

7. The intelligent monitoring and maintenance system for large-volume parenteral solution production equipment according to claim 4, characterized in that, The wavelet packet transform unit of the data fusion and feature extraction module performs a four-level decomposition of the vibration signal and extracts the wavelet packet energy entropy of each node in the fourth level as a time-frequency domain feature.

8. The intelligent monitoring and maintenance system for large-volume parenteral solution production equipment according to claim 4, characterized in that, The hybrid model of the intelligent diagnosis and prediction module is composed of a bidirectional long short-term memory network and a multi-head self-attention mechanism connected in series. The bidirectional long short-term memory network has 128 hidden layer units, and the multi-head self-attention mechanism has 8 heads. Its attention weight calculation adopts scaled dot product attention.

9. The intelligent monitoring and maintenance system for large-volume parenteral solution production equipment according to claim 4, characterized in that, The multi-objective optimization algorithm of the maintenance decision and scheduling optimization module simultaneously optimizes three objectives: equipment availability, maintenance cost, and product quality risk. Its objective function is expressed as minimizing the weighted sum of total unplanned downtime, total maintenance cost, and batch product non-conformity rate. The constraints include a maximum allowable maintenance time window of no more than 4 hours, a cleanroom differential pressure maintenance value of no less than 15 Pa, and a sterilizer cooling time of more than 25 minutes.

10. The intelligent monitoring and maintenance system for large-volume parenteral solution production equipment according to claim 4, characterized in that, The incremental learning mechanism of the adaptive learning and model update module adopts an online sequential extreme learning machine, whose hidden layer node number is dynamically adjusted from 100 to 500, and whose initial learning rate is 0.01 and decays exponentially with each training round.

Images