Barreled water whole-process production monitoring management system and method thereof

Abstract

The invention relates to the technical field of monitoring management, solves the technical problem that the influence degree of each link on the water quality in the whole-process production of barreled water is difficult to quantitatively monitor in a targeted manner, and particularly discloses a monitoring management system and method for the whole-process production of barreled water, and the system comprises a whole-process data collection module and a data preprocessing module. A deep learning analysis module, a decision and resource allocation module and an equipment linkage control module; the method comprises the following steps: S100, system deployment and initialization; s200, full-process data acquisition is carried out; s300, carrying out data preprocessing and feature engineering; s400, training a deep learning model; s500, carrying out real-time monitoring and analysis; and S600, decision making and execution are carried out. The system is used for barreled water whole-process integrated monitoring management, and through real-time data acquisition, deep learning analysis and decision making, the production process is controllable, and the water quality is stably guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring and management, and particularly to a full-process production monitoring and management system and method for barreled water. Background Art

[0002] Barreled water refers to a product obtained by using tap water or extracting groundwater and filling it into PC barrels through a filling production line by modern industrial technology. The quality of barreled water mostly depends on the production process. Therefore, it is very necessary to monitor the production of barreled water in real time. At present, for barreled water manufacturers, there is a lack of a platform for unified management of each production stage and monitoring link on the production line, and the information of each detection link cannot be centrally analyzed and processed, resulting in incomplete supervision and management of product quality. Analyzing the online monitoring and management of the entire production process of barreled water, including well water extraction, purification, sterilization, filling, and packaging, can ensure stable production quality control. Summary of the Invention

[0003] To solve the technical problem in the prior art that it is difficult to quantitatively monitor the influence degree of each link on water quality in the full-process production of barreled water, the present invention provides a full-process production monitoring and management system and method for barreled water.

[0004] The technical solution adopted by the present invention is as follows:

[0005] A full-process production monitoring and management system for barreled water includes:

[0006] A full-process data acquisition module, which is used to collect water quality parameters, equipment operation parameters, and environmental parameters of the production process, and output the original collected data to the preprocessing module;

[0007] A data preprocessing module, which is used to receive the data collected by the full-process data acquisition module, and perform data cleaning, standardization, and feature extraction and screening, and output the preprocessed feature data;

[0008] A deep learning analysis module, which is used to receive the feature data output by the data preprocessing module, and establish a water quality influence degree evaluation sub-model and a water quality compliance prediction sub-model through a CNN-LSTM attention hybrid model, input the feature data for water quality influence degree evaluation, water quality compliance prediction, and risk point location, and output the influence weight value and the prediction result;

[0009] A decision-making and resource allocation module, which is used to receive the analysis results of the deep learning analysis module, and perform monitoring resource allocation and generate a process parameter adjustment strategy, and output a control instruction and a resource allocation plan;

[0010] An equipment linkage control module, which is used to receive the control instruction of the decision-making and resource allocation module, and output an equipment operation signal to execute process adjustment, abnormal shutdown, and alarm triggering.

[0011] Furthermore, the full-process data acquisition module includes: a well water extraction acquisition unit, a purification treatment acquisition unit, a sterilization and disinfection acquisition unit, and a filling and packaging acquisition unit, all of which are connected to the data preprocessing module via signal transmission.

[0012] The well water extraction and acquisition unit is used to collect water temperature, pH value, turbidity, total hardness, heavy metal content, total bacterial count, and coliform bacteria through sensors.

[0013] The purification and processing acquisition unit is used to acquire turbidity, TDS value, filtration pressure, number of regeneration cycles of softening resin, and filter membrane pressure difference through sensors.

[0014] The sterilization and disinfection data collection unit is used to collect data such as sterilization time, sterilization temperature, ozone concentration or ultraviolet intensity, and microbial residue through sensors.

[0015] The filling and packaging data acquisition unit is used to collect data on barrel pressure, filling volume, sealing performance, bottle mouth cleanliness, and secondary contamination indicators through sensors.

[0016] Furthermore, the water quality impact assessment sub-model of the deep learning analysis module takes into account the standardized time-series data of parameters from each production stage, extracts the spatial correlation features of parameters from each stage using CNN, captures the temporal variation features using LSTM, and strengthens the weights of key parameters based on an attention mechanism to establish the following model:

[0017]

[0018] in, : No. The weight of water quality impact at each production stage;

[0019] : No. Step, First The spatial feature contribution value of each link is calculated based on the correlation of various parameters of the corresponding production link, through the output of the CNN layer.

[0020] : No. Step, First The attention weights of each step are output through the attention mechanism;

[0021] , : No. Step, First The temporal stability coefficient of the process, output through the LSTM layer, reflects the impact of parameter temporal fluctuations on water quality; the smaller the fluctuation, the better. The closer to 1;

[0022] Total number of production stages; Production process number;

[0023] The water quality impact assessment sub-model outputs the impact weights of each production stage on the final water quality.

[0024] Furthermore, the water quality compliance prediction sub-model of the deep learning analysis module takes preprocessed real-time parameter data and historical compliance data as input. Based on the deep learning model learning the mapping relationship between parameters and water quality compliance, and combined with quantitative calculations, it predicts in real time the probability of water quality compliance and potential exceedance risk points in the current production process, establishing the following model:

[0025]

[0026] in:

[0027] Probability of water quality meeting standards;

[0028] Sigmoid activation function;

[0029] : No. The weight of water quality impact in each stage;

[0030] : No. Real-time comprehensive characteristic values ​​of the process;

[0031] : No. Historical feature dataset of the process;

[0032] : No. The historical compliance mapping coefficient of each stage is obtained by fitting the training set data and is used to reflect the correlation strength between the characteristics of the corresponding production stage and the compliance status.

[0033] : No. The deviation correction term for the process is obtained through optimization during model training and is used to compensate for the effects of sensor errors and process fluctuations. Its value range is [-0.1, 0.1].

[0034] The water quality compliance prediction sub-model outputs the probability of water quality compliance, the risk level of exceeding the standard, and key risk parameters.

[0035] Furthermore, the decision-making and resource allocation module includes: a process adjustment strategy generation unit, used to generate strategies based on water quality impact weights. and the probability of water quality meeting standards Develop targeted solutions;

[0036] The equipment linkage control module includes:

[0037] The water source parameter adjustment unit is used to activate the backup water source and increase the number of pretreatment filtration stages;

[0038] The purification adjustment unit is used to switch to the backup filter membrane and adjust the filtration pressure.

[0039] The sterilization adjustment unit is used to adjust the sterilization time and ozone generator power.

[0040] The filling adjustment unit is used to perform shutdown checks and trigger secondary cleaning of the bottle neck;

[0041] The signal output terminal of the process adjustment strategy generation unit is connected to the signal input terminals of the water source parameter adjustment unit, the purification adjustment unit, the sterilization adjustment unit, and the filling adjustment unit, respectively.

[0042] A method for monitoring and managing the entire production process of bottled water, based on the aforementioned monitoring and management system for the entire production process of bottled water, includes the following steps:

[0043] S100, System Deployment and Initialization;

[0044] S200, full-process data acquisition;

[0045] S300, data preprocessing and feature engineering;

[0046] S400, deep learning model training;

[0047] S500, real-time monitoring and analysis;

[0048] S600, Decision-making and Execution.

[0049] Furthermore, the S300 specifically includes the following steps:

[0050] S301, Data Cleaning, including:

[0051] Outliers are removed by using the 3σ criterion to eliminate extreme data caused by sensor malfunctions;

[0052] Missing value imputation: linear interpolation is used to imput a small number of missing data. A sensor calibration warning is triggered when the missing value rate exceeds 10%.

[0053] Data deduplication involves removing redundant data that has been collected repeatedly.

[0054] S302. Data standardization, including Z-score standardization of parameters with different dimensions;

[0055] S303, Feature extraction, including:

[0056] Temporal feature extraction, calculating the sliding window mean, variance, maximum value, and minimum value of each parameter;

[0057] Feature extraction is performed, correlation coefficients between parameters in each stage are calculated, and spatial feature support for F_i calculation is strengthened.

[0058] Engineering feature extraction and derivative feature construction;

[0059] S304. Feature selection, including using mutual information to select features that are strongly correlated with water quality compliance and removing irrelevant features; outputting the selected feature set for deep learning model training and real-time analysis.

[0060] Furthermore, the S400 specifically includes the following steps:

[0061] S401, Dataset Construction, including:

[0062] Collect historical production data, including full-process parameter data for both qualified and unqualified products;

[0063] The data should be labeled, indicating whether the water quality meets the standards, does not meet the standards, and the reasons for exceeding the standards;

[0064] The training set, validation set, and test set are divided in a 7:2:1 ratio.

[0065] S402. Model Architecture Construction: Construct a CNN-LSTM attention hybrid model, where the CNN layer includes three convolutional and pooling layers to extract spatial correlation features of parameters at each stage and output... ;

[0066] The LSTM layer consists of two LSTM layers, capturing temporal variation features and outputting... ;

[0067] The attention layer employs the Bahdanau attention mechanism to enhance the weights of key parameters and output... ;

[0068] The output layer outputs both the weight of the impact on water quality and the probability of water quality meeting standards.

[0069] Furthermore, the S400 also includes the following steps:

[0070] S403, Model Training and Validation, including:

[0071] The Adam optimizer is used, and the loss function is the sum of cross-entropy loss and mean squared error loss. Early stopping is used during training to prevent overfitting, and training is stopped if the accuracy on the validation set does not improve for 5 consecutive epochs.

[0072] Test set validation was conducted, with validation requirements including a water quality compliance prediction accuracy of ≥95% and an impact weight assessment error of ≤5%.

[0073] S404, Model Optimization and Deployment, including:

[0074] Pruning and quantization of the trained model improves the real-time calculation speed of water quality impact weights and water quality compliance probability.

[0075] The model is deployed to edge computing nodes and cloud servers to calculate the water quality impact weight and the probability of water quality compliance in real time.

[0076] Configure the model call interface to achieve linkage with the data preprocessing module and the decision and resource allocation module.

[0077] Furthermore, the S500 real-time monitoring and analysis specifically includes the following steps:

[0078] S501, Real-time data input: The preprocessed feature data is input into the deep learning model deployed on the edge node in real time to calculate the real-time comprehensive feature value of each stage; The cloud server synchronously receives data for the subsequent iteration of the model's real-time monitoring and historical target mapping coefficients and deviation correction terms.

[0079] S502, Water quality impact analysis: The model calculates the real-time impact weight of each production link and dynamically updates the weight distribution.

[0080] S503, Water Quality Compliance Prediction: The model calculates the probability of water quality compliance and the risk level of exceeding the standard in real time, locates key risk points, and outputs the parameters that lead to the risk.

[0081] S504, data association storage, synchronizes and stores real-time parameters, water quality impact weights, water quality compliance probability and risk point information to a distributed database, and associates timestamps with production batch numbers.

[0082] The beneficial effects of this invention are:

[0083] This invention addresses the core issues of difficulty in targeted quantitative monitoring of water quality impacts at various stages of bottled water production, information isolation, and lack of unified management by constructing an integrated monitoring and management system, with full-process collaborative work and deep learning quantitative analysis as its core. During system operation, the full-process data acquisition module first comprehensively collects water quality parameters, equipment operating parameters, and environmental parameters from all stages of production, obtaining more comprehensive raw data and breaking down information barriers between different detection stages. Subsequently, the data preprocessing module cleans, standardizes, and extracts and filters features from the collected data, providing high-quality data support for subsequent analysis. The core deep learning analysis module establishes a water quality impact assessment sub-model and a water quality compliance prediction sub-model using a CNN-LSTM attention hybrid model. It performs in-depth analysis of the preprocessed feature data, accurately quantifying the impact weight of each production stage on water quality, while simultaneously predicting water quality compliance and locating risk points, ensuring that monitoring is no longer blind. Based on the analysis results, the decision-making and resource allocation module rationally allocates monitoring resources and generates targeted process parameter adjustment strategies, ensuring that monitoring resources are tilted towards key stages affecting water quality. Finally, the equipment linkage control module receives control commands and executes process adjustments, abnormal shutdowns, or alarm triggers, forming a complete closed-loop management system. This working principle integrates data from the entire process through a unified platform, replacing traditional fuzzy monitoring with quantitative analysis. This enables precise control and targeted management of the impact on water quality at each stage, ultimately keeping the entire bottled water production process under real-time control. This effectively avoids oversights in product quality supervision and management, ensures the stability of production quality control, and improves monitoring efficiency through the rational allocation of resources. It also reduces unnecessary monitoring costs and achieves efficient and accurate production quality control. Attached Figure Description

[0084] Figure 1 This is a system logic flowchart of the present invention. Detailed Implementation

[0085] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0086] Example 1

[0087] This embodiment is based on a full-process data acquisition module and a data preprocessing module. The full-process data acquisition module includes a well water extraction acquisition unit, a purification treatment acquisition unit, a sterilization and disinfection acquisition unit, and a filling and packaging acquisition unit. Each unit collects key parameters of its corresponding stage through sensors. The well water extraction acquisition unit collects water temperature, pH value, turbidity, total hardness, heavy metal content, total bacterial count, and coliform bacteria. The purification treatment acquisition unit collects turbidity, total dissolved solids, filtration pressure, number of regeneration cycles of softening resin, and filter membrane pressure difference. The sterilization and disinfection acquisition unit collects sterilization time, sterilization temperature, ozone concentration or ultraviolet intensity, and microbial residue. The filling and packaging acquisition unit collects tank pressure, filling volume, sealing performance, bottle mouth cleanliness, and secondary pollution indicators. All collected data are transmitted to the data preprocessing module. The data preprocessing module processes data according to a predetermined process. First, data cleaning is performed, using the three-standard-deviation criterion to remove extreme data caused by sensor malfunctions. Linear interpolation is used to fill in a small number of missing data. When the missing rate exceeds 10%, a sensor calibration warning is triggered, and redundant data collected repeatedly is deleted. Next, data standardization is performed, using Z-score standardization for parameters of different dimensions to eliminate the influence of differences in dimensions. Then, feature extraction is carried out, calculating the sliding window mean, variance, maximum, and minimum values ​​of each parameter to obtain time-series features, calculating the correlation coefficients between parameters in each stage to strengthen spatial feature support, and constructing derived features. Finally, the mutual information method is used to screen features strongly correlated with water quality compliance, remove irrelevant features, and output the filtered feature set.

[0088] This embodiment collects parameters from the entire production process comprehensively and in stages, breaking the traditional isolation of information from each detection stage. Through multi-step data preprocessing to eliminate invalid data, standardize data, and extract effective features, it provides high-quality, highly relevant data support for subsequent deep learning analysis. Its beneficial technical effects include achieving comprehensive collection of key parameters throughout the bottled water production process, avoiding analytical biases caused by missing or poor-quality data. Simultaneously, feature filtering simplifies data dimensions, improving the efficiency and accuracy of subsequent model analysis, and laying a solid data foundation for the precise operation of the entire monitoring and management system.

[0089] Example 2

[0090] This embodiment is based on the water quality impact assessment sub-model in the Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM) attention hybrid model. The system first receives standardized time-series parameter data for each production stage from the data preprocessing module. Then, the model extracts the spatial correlation features of parameters in each stage through the convolutional neural network layer, i.e., the mutual influence relationships between different parameters within the same stage, and outputs the spatial feature contribution value. The LSTM network layer captures the temporal characteristics of parameter changes over time, reflecting the fluctuations of parameters during the production process, and outputs the temporal stability coefficient. The Badanao attention mechanism strengthens the weights of parameters critical to water quality impact, highlighting the role of core parameters, and outputs the attention weights. Finally, these three values ​​are substituted into the water quality impact weight calculation formula to calculate the impact weight of each production stage on the final water quality.

[0091] The principle of this model is to comprehensively evaluate the "importance" of each production link to water quality. First, it examines the degree of correlation between the parameters within each link. The closer the correlation, the more stable the parameter system of that link, and the higher its spatial characteristic contribution value. Next, it examines the fluctuation of parameters during the production process. The smaller the fluctuation, the more stable the operation of the link, and the closer the temporal stability coefficient is to 1. Finally, it emphasizes the parameters that play a key role in water quality, highlighting the influence of important parameters. The product of these three factors and the result is divided by the sum of the products of all three factors across all links to obtain the influence percentage of each link. Furthermore, the sum of the weights of all links is 1, ensuring that the results are objectively quantified.

[0092] As a computational example, the model operation process is as follows: Assume there are five production stages: well water extraction, purification, sterilization, filling and packaging. The spatial feature contributions of each stage are 0.8, 0.9, 1.0, 0.7, and 0.6, respectively; the attention weights are 0.7, 0.8, 0.9, 0.6, and 0.5, respectively; and the temporal stability coefficients are 0.9, 0.85, 0.95, 0.8, and 0.75, respectively. First, calculate the product of the three for each stage:

[0093] The well water extraction step costs 0.8 × 0.7 × 0.9 = 0.504.

[0094] The purification process takes 0.9 × 0.8 × 0.85 = 0.612 seconds.

[0095] The sterilization and disinfection process takes 1.0 × 0.9 × 0.95 = 0.855 seconds.

[0096] The filling and packaging process takes 0.7 × 0.6 × 0.8 = 0.336.

[0097] The packaging cost is 0.6 × 0.5 × 0.75 = 0.225;

[0098] The sum of the products of all components is 0.504 + 0.612 + 0.855 + 0.336 + 0.225 = 2.532;

[0099] Then, calculate the impact weight of each stage separately: well water extraction stage is 0.504÷2.532≈0.2, purification treatment stage is 0.612÷2.532≈0.242, sterilization and disinfection stage is 0.855÷2.532≈0.338, filling and sealing stage is 0.336÷2.532≈0.133, and packaging stage is 0.225÷2.532≈0.089.

[0100] This embodiment achieves precise quantification of the impact of each production stage on water quality, changing the current situation where the impact degree is vaguely judged in existing related monitoring and management technologies. It provides more accurate data basis for the rational allocation of subsequent monitoring resources, allowing monitoring resources to be concentrated on key stages with high impact weight, avoiding resource waste, and improving the pertinence and effectiveness of monitoring.

[0101] Example 3

[0102] This embodiment is based on the water quality compliance prediction sub-model in the Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM) attention hybrid model. Its operation is as follows: the system receives preprocessed real-time parameter data and historical compliance data. The model first learns the parameter records of qualified and unqualified products from historical data to understand the mapping relationship between parameters and water quality compliance, and then fits the historical compliance mapping coefficients for each stage. Simultaneously, during training, it optimizes and obtains deviation correction terms to compensate for sensor errors and process fluctuations. Subsequently, it combines the water quality impact assessment sub-model output for each stage... The influence weights are determined by substituting the real-time comprehensive feature values ​​of each stage into the water quality compliance probability calculation formula. First, the real-time comprehensive feature values ​​are standardized by subtracting the minimum value of the historical feature dataset for that stage and dividing by the difference between the maximum and minimum values ​​of the historical feature dataset. Then, the result is multiplied by the historical compliance mapping coefficient and a deviation correction term is added to obtain the contribution value of each stage. The contribution value of each stage is multiplied by the corresponding influence weight and summed. The result is then mapped to the 0-1 interval through the Sigmund activation function and converted into a percentage form to finally obtain the water quality compliance probability. At the same time, the exceedance risk level and key risk parameters are output.

[0103] The model works by standardizing and converting the real-time parameters of current processes based on parameters from past production of qualified and unqualified products, combined with the importance of each stage to water quality. After correcting for potential errors, the model comprehensively calculates the probability of current water quality meeting standards and identifies key parameters that could lead to a low probability of compliance. As a calculation example, the model's operation process is as follows:

[0104] Assume the water quality impact weights for the five production stages are 0.2, 0.242, 0.338, 0.133, and 0.089, respectively; the historical compliance mapping coefficients are 0.3, 0.25, 0.35, 0.2, and 0.15, respectively; and the deviation correction term is 0.05 for each stage.

[0105] The real-time comprehensive feature value of the well water extraction stage is 80. The maximum value of the historical feature dataset of this stage is 100 and the minimum value is 50. Its standardized value is (80-50)÷(100-50)=0.6, and its contribution value is 0.3×0.6+0.05=0.23.

[0106] The real-time comprehensive characteristic value of the purification process is 70, the historical maximum value is 90, the minimum value is 60, the standardized value is (70-60)÷(90-60)≈0.333, and the contribution value is 0.25×0.333+0.05≈0.133;

[0107] The real-time comprehensive characteristic value of the sterilization and disinfection process is 90, the historical maximum value is 100, the minimum value is 70, the standardized value is (90-70)÷(100-70)≈0.667, and the contribution value is 0.35×0.667+0.05≈0.283;

[0108] The real-time comprehensive characteristic value of the filling and packaging process is 60, the historical maximum value is 80, the minimum value is 50, the standardized value is (60-50)÷(80-50)≈0.333, and the contribution value is 0.2×0.333+0.05≈0.117;

[0109] The real-time comprehensive feature value of the packaging process is 85, the historical maximum value is 95, the minimum value is 75, the standardized value is (85-75)÷(95-75)=0.5, and the contribution value is 0.15×0.5+0.05=0.125;

[0110] Multiplying the contribution value of each step by the corresponding influence weight obtained in Example 2 and summing the results, we get:

[0111] 0.2×0.23+0.242×0.133+0.338×0.283+0.133×0.117+0.089×0.125≈0.046+0.032+0.096+0.015+0.011≈0.20, which, after processing with the Sigmoid activation function, is approximately 0.55. Converted to a percentage, the probability of water quality meeting standards is approximately 55%, corresponding to a high-risk level. If the real-time ozone concentration in the sterilization and disinfection process is lower than the historical acceptable range, this parameter will be output as a critical risk parameter.

[0112] This embodiment enables early prediction of water quality compliance and more precise identification of risk points. It can promptly detect potential quality problems during the production process, prevent unqualified products from entering the market, and provide a clear direction for process adjustments. It improves the foresight and accuracy of quality control and effectively reduces production losses.

[0113] Example 4

[0114] This embodiment achieves its operation through the collaborative work of the decision-making and resource allocation module and the equipment linkage control module. The process is as follows: After receiving the water quality impact weights and water quality compliance probabilities output by the deep learning analysis module, the decision-making and resource allocation module first dynamically allocates monitoring resources. For links with high impact weights, it increases the sampling frequency, deploys redundant sensors, and raises the data transmission priority. For links with medium impact weights, it maintains the normal sampling frequency and periodically calibrates the sensors. For links with low impact weights, it reduces the sampling frequency to conserve resources. Simultaneously, the process adjustment strategy generation unit formulates targeted treatment measures based on the water quality impact weights and water quality compliance probabilities. If abnormal water source parameters lead to a decrease in the compliance probability, it generates instructions to activate the backup water source and increase the number of pretreatment filtration stages. If abnormal purification parameters are generated, it generates instructions to switch to the backup filter membrane and adjust the filtration pressure. If abnormal sterilization parameters are generated, it generates instructions to adjust the sterilization time and ozone generator power. If abnormal filling parameters are generated, it generates instructions to stop for inspection and trigger secondary cleaning of the bottle neck, and transmits these control instructions to the equipment linkage control module. The water source parameter adjustment unit, purification adjustment unit, sterilization adjustment unit, and filling adjustment unit in the equipment linkage control module receive corresponding instructions and execute operations. Data during the operation process is fed back to the system in real time, realizing data closure. The working principle of this embodiment is to achieve reasonable allocation of monitoring resources based on quantified influence weights, so that resources are concentrated in key links. At the same time, based on accurate compliance prediction and risk positioning, production processes are adjusted in a targeted manner. Through automated linkage of equipment to execute adjustment instructions, abnormal situations are handled in a timely manner, forming a continuous management process of monitoring, analysis, decision-making, and execution.

[0115] The technical solution of this embodiment avoids the ineffective waste of monitoring resources, improves monitoring efficiency and the pertinence of process adjustments, realizes real-time control of the entire bottled water production process, effectively reduces quality problems caused by abnormal process parameters, ensures the stability and consistency of production control, and at the same time reduces manual intervention through the linkage of automated equipment, reduces the risk of human error, and improves production efficiency and quality control level.

[0116] Example 5

[0117] This embodiment implements bottled water production monitoring and management according to the entire process steps, such as... Figure 1As shown, specifically, the system is first deployed and initialized. Corresponding sensors are installed at the well water extraction port, the inlet and outlet of the purification equipment, the sterilization device, and the bottling line. An edge gateway, a local server, a cloud graphics processor server, and a distributed database cluster are deployed and connected to a programmable logic controller (PLC), which is connected to the production equipment. At the same time, data acquisition software and a deep learning framework are installed. Common deep learning frameworks include TensorFlow and PyTorch, as well as a visual management platform. User permissions and data encryption transmission channels are configured, and water quality standard thresholds for each stage are set. These thresholds can refer to GB19298-2014 "Packaged Drinking Water". It also includes setting the initial sampling frequency of the sensors, the initial parameters of the deep learning model, such as the number of convolutional kernels in the convolutional neural network, the number of hidden layer nodes in the long short-term memory network, and the warning threshold.

[0118] After completing the above steps, the full-process data acquisition and preprocessing stage begins. The specific work process, parameter acquisition range, and data cleaning, standardization, feature extraction, and screening methods in this stage are all carried out according to the collaborative working method of the full-process data acquisition module and data preprocessing module in Example 1. All acquired data is processed and outputs a feature set that meets the requirements of deep learning analysis. Then, deep learning model training is implemented, constructing a convolutional neural network-long short-term memory network attention hybrid model. At least one year of historical production data is collected and labeled with compliance status and reasons for exceeding standards. The training set, validation set, and test set are divided in a 7:2:1 ratio. The Adam optimizer is used, with the sum of cross-entropy loss and mean squared error loss as the loss function for model training. Early stopping is used to prevent overfitting. Once the water quality compliance prediction accuracy on the test set reaches no less than 95% and the impact weight assessment error does not exceed 5%, the model is pruned, quantized, and deployed to edge nodes and the cloud.

[0119] After the system enters the real-time monitoring and analysis phase, the preprocessed feature data is input into the edge node model in real time. The model first calculates the impact weight of each link through the water quality impact assessment sub-model. The parameter input type, feature extraction logic, weight calculation formula and principle of this sub-model are all executed in the manner of Example 2. The specific calculation process can be referred to the numerical example in Example 2. At the same time, the water quality compliance prediction sub-model inputs real-time parameters and historical data. Its parameter mapping relationship, compliance probability calculation formula and risk point location method are all carried out in the manner of Example 3. The numerical calculation logic can be referred to the compliance probability derivation process in Example 3.

[0120] Finally, the system enters the decision-making and execution phase. The resource allocation logic of the decision-making and resource allocation module, the generation method of process adjustment strategies, and the instruction reception and execution operations of the equipment linkage control module are all performed according to the collaborative working method of the decision-making and resource allocation module and the equipment linkage control module in Embodiment 4. Data during the operation process is fed back to the system storage in real time, and this data will be used for subsequent model iteration and optimization. This embodiment integrates the core functions of the aforementioned embodiments through a closed-loop process of deployment, data acquisition and preprocessing, training, monitoring and analysis, and decision execution. It uses a deep learning model to quantify the impact of each stage on water quality and the probability of water quality compliance, and combines this with automated equipment linkage to achieve dynamic control of the production process. Its technical effect is that, based on integrating the advantages of the aforementioned embodiments, it solves the problems of isolated information and difficulty in quantifying the impact of each stage in existing production monitoring and management technologies. It achieves real-time controllability of the entire bottled water production process, effectively improves the water quality compliance rate and monitoring efficiency, reduces waste of monitoring resources and losses during the production process, and effectively ensures the stability of production control. At the same time, it forms a complete technical closed loop, avoiding the collaborative gaps that may occur when each module works independently.

[0121] To better implement the technical solution of this invention, the system hardware support for this invention can be configured as follows:

[0122] Sensor layer: pH sensor, turbidity sensor, TDS sensor, heavy metal sensor, microbial sensor, ozone concentration sensor, temperature sensor, pressure sensor, sealing tester, etc.;

[0123] Edge computing layer: edge gateways and local servers, used for real-time data preprocessing and rapid decision-making;

[0124] Cloud layer: GPU servers, used for model training and optimization, and distributed database clusters;

[0125] Control layer: PLC controller, frequency converter, solenoid valve, alarm device. The PLC controller can be a Siemens SIMATIC S7-1500 series PLC controller.

[0126] Specifically, the sensor models that can be selected are as follows:

[0127] pH sensor: The Bell T255 pH sensor is optional. This sensor is suitable for pH measurement in a variety of situations, including high temperature, high pressure and wastewater and sewage situations. It can meet the real-time pH monitoring needs of well water extraction and purification in bottled water production, and can withstand routine cleaning operations in the production process. It is also suitable for food-grade water quality monitoring scenarios.

[0128] Turbidity sensor: Optional Mettler Toledo InPro 8100 / S / 297 / 3.1 turbidity sensor. This sensor is a single-fiber design, specially designed for measuring high particle concentration samples. It has a wide linear measurement range, can withstand corrosive media, and supports disinfection and high-temperature sterilization, meeting the needs of bottled water purification treatment and water turbidity monitoring before bottling.

[0129] TDS sensor: Optional Thermo Fisher Scientific CTS pocket measuring instrument, IP-67 waterproof rating suitable for the humid environment of bottled water production, suitable for real-time monitoring of total dissolved solids in the purification process, providing more accurate data for judging the water purification effect.

[0130] Heavy metal sensor: Optional Hach HQ411d heavy metal sensor, which can detect multiple heavy metal ions such as lead (Pb), cadmium (Cd), and arsenic (As) in water. The measurement range covers the low concentration range of ppb level, supports automatic temperature compensation function to reduce the impact of temperature changes on measurement results, is suitable for monitoring heavy metal content in the bottled water and well water extraction process, and can be compatible with water quality monitoring systems to achieve data linkage.

[0131] Microbial sensor: Hanwei Technology MWM-01 microbial sensor is optional. This sensor adopts the principle of bioelectrochemical and can monitor microbial indicators such as total bacterial count and coliform bacteria in water in real time. It is suitable for detecting microbial residues after sterilization and disinfection of bottled water and before bottling. It can quickly provide feedback on sterilization effect, avoid water quality problems caused by secondary pollution, and has a long maintenance cycle, making it suitable for continuous industrial production scenarios.

[0132] Ozone concentration sensor: Optional Alphasense O3-A2 ozone concentration sensor. This sensor adopts the electrochemical detection principle, has strong stability, and is suitable for real-time monitoring of ozone concentration in the sterilization and disinfection process of bottled water. It can more accurately capture ozone concentration fluctuations and ensure that the ozone concentration is maintained within the effective sterilization range during the sterilization process. It also supports linkage with PLC control system to realize automatic adjustment of ozone generator.

[0133] Temperature sensor: Siemens 7ML5033-1BA00-1AA1 temperature sensor is optional. This sensor is a platinum resistance thermometer (RTD) type with an IP67 protection rating. It is suitable for temperature monitoring in the sterilization and disinfection process of bottled water production (such as temperature control for ultraviolet sterilization and ozone sterilization) and water temperature monitoring in the well water extraction process. It can stably withstand vibration, humidity and other conditions in the production environment to ensure accurate transmission of temperature data.

[0134] Pressure sensor: Optional Hydac HDA4744-A-016-000 pressure sensor, which is suitable for monitoring the filtration pressure in the bottled water purification process and the pressure inside the bottle in the filling process. It can provide real-time feedback on changes in equipment operating pressure and trigger an alarm when the pressure exceeds the normal range, thus avoiding equipment failure or water quality problems caused by abnormal pressure.

[0135] Sealing tester: The Heres HRS-8800 sealing tester is optional. This device uses the negative pressure leak detection method and is suitable for testing the sealing of bottle mouths after bottled water is filled and packaged. It can quickly identify problems such as poor bottle mouth sealing and loose caps, avoiding secondary contamination or leakage during transportation and storage. It also supports data storage and traceability and can generate test reports, meeting the needs of production quality control and traceability.

[0136] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A bottled water full-process production monitoring and management system, characterized in that, include: The full-process data acquisition module is used to collect water quality parameters, equipment operating parameters, and environmental parameters of the production process, and output the raw collected data to the pretreatment module. The data preprocessing module receives data collected by the full-process data acquisition module, performs data cleaning, standardization, feature extraction and filtering, and outputs preprocessed feature data. The deep learning analysis module receives the feature data output by the data preprocessing module and establishes a water quality impact assessment sub-model and a water quality compliance prediction sub-model through a CNN-LSTM attention hybrid model. It inputs feature data to perform water quality impact assessment, water quality compliance prediction and risk point location, and outputs impact weight values ​​and prediction results. The decision-making and resource allocation module receives the analysis results from the deep learning analysis module, generates monitoring resource allocation and process parameter adjustment strategies, and outputs control commands and resource allocation schemes. The equipment linkage control module is used to receive control commands from the decision-making and resource allocation module, and output equipment operation signals to perform process adjustments, abnormal shutdowns, and alarm triggers.

2. The bottled water full-process production monitoring and management system according to claim 1, characterized in that, The full-process data acquisition module includes: a well water extraction acquisition unit, a purification treatment acquisition unit, a sterilization and disinfection acquisition unit, and a filling and packaging acquisition unit, all of which are connected to the data preprocessing module via signal transmission. The well water extraction and acquisition unit is used to collect water temperature, pH value, turbidity, total hardness, heavy metal content, total bacterial count, and coliform bacteria through sensors. The purification and acquisition unit is used to collect turbidity, TDS value, filtration pressure, number of regeneration cycles of softening resin, and filter membrane pressure difference through sensors. The sterilization and disinfection data collection unit is used to collect sterilization time, sterilization temperature, ozone concentration or ultraviolet intensity, and microbial residue through sensors. The filling and packaging data acquisition unit is used to collect data on barrel pressure, filling volume, sealing performance, bottle mouth cleanliness, and secondary contamination indicators through sensors.

3. The bottled water full-process production monitoring and management system according to claim 1, characterized in that, The water quality impact assessment sub-model of the deep learning analysis module takes into account the standardized time-series data of parameters from each production stage, extracts the spatial correlation features of parameters from each stage using CNN, captures the temporal variation features using LSTM, and strengthens the weights of key parameters based on an attention mechanism to establish the following model: in, : No. The weight of water quality impact at each production stage; : No. Step, First The spatial feature contribution value of each link is calculated based on the correlation of various parameters of the corresponding production link, through the output of the CNN layer. : No. Step, First The attention weights of each step are output through the attention mechanism; , : No. Step, First The temporal stability coefficient of the process, output through the LSTM layer, reflects the impact of temporal fluctuations in parameters on water quality; the smaller the fluctuation, the better. The closer to 1; Total number of production stages; Production process number; The water quality impact assessment sub-model outputs the impact weights of each production stage on the final water quality.

4. The bottled water full-process production monitoring and management system according to claim 3, characterized in that, The water quality compliance prediction sub-model of the deep learning analysis module takes preprocessed real-time parameter data and historical compliance data as input. Based on the deep learning model learning the mapping relationship between parameters and water quality compliance, and combined with quantitative calculations, it predicts in real time the probability of water quality compliance and potential risk points of exceeding standards in the current production process, and establishes the following model: in: Probability of water quality meeting standards; Sigmoid activation function; : No. The weight of water quality impact in each stage; : No. Real-time comprehensive characteristic values ​​of the process; : No. Historical feature dataset of the process; : No. The historical compliance mapping coefficient of each stage is obtained by fitting the training set data and is used to reflect the correlation strength between the characteristics of the corresponding production stage and the compliance status. : No. The deviation correction term for the process is obtained through optimization during model training and is used to compensate for the effects of sensor errors and process fluctuations. Its value range is [-0.1, 0.1]. The water quality compliance prediction sub-model outputs the probability of water quality compliance, the risk level of exceeding the standard, and key risk parameters.

5. The bottled water full-process production monitoring and management system according to claim 4, characterized in that, The decision-making and resource allocation module includes: a process adjustment strategy generation unit, used for generating strategies based on water quality impact weights. and the probability of water quality meeting standards Develop targeted solutions; The equipment linkage control module includes: The water source parameter adjustment unit is used to activate the backup water source and increase the number of pretreatment filtration stages; The purification adjustment unit is used to switch to the backup filter membrane and adjust the filtration pressure. The sterilization adjustment unit is used to adjust the sterilization time and ozone generator power. The filling adjustment unit is used to perform shutdown checks and trigger secondary cleaning of the bottle neck; The signal output terminal of the process adjustment strategy generation unit is connected to the signal input terminals of the water source parameter adjustment unit, the purification adjustment unit, the sterilization adjustment unit, and the filling adjustment unit, respectively.

6. A method for monitoring and managing the entire production process of bottled water, comprising the bottled water production monitoring and management system according to any one of claims 1-5, characterized in that, Includes the following steps: S100, System Deployment and Initialization; S200, full-process data acquisition; S300, data preprocessing and feature engineering; S400, deep learning model training; S500, real-time monitoring and analysis; S600, Decision-making and Execution.

7. The method for monitoring and managing the entire production process of bottled water according to claim 6, characterized in that, The S300 process specifically includes the following steps: S301, Data Cleaning, including: Outliers are removed by using the 3σ criterion to eliminate extreme data caused by sensor malfunctions; Missing value imputation: linear interpolation is used to imput a small number of missing data. A sensor calibration warning is triggered when the missing value rate exceeds 10%. Data deduplication involves removing redundant data that has been collected repeatedly. S302. Data standardization, including Z-score standardization of parameters with different dimensions; S303, Feature extraction, including: Temporal feature extraction, calculating the sliding window mean, variance, maximum value, and minimum value of each parameter; Feature extraction is performed, correlation coefficients between parameters in each stage are calculated, and spatial feature support for F_i calculation is strengthened. Engineering feature extraction and derivative feature construction; S304. Feature selection, including using mutual information to select features that are strongly correlated with water quality compliance and removing irrelevant features; outputting the selected feature set for deep learning model training and real-time analysis.

8. The method for monitoring and managing the entire production process of bottled water according to claim 6, characterized in that, The S400 process specifically includes the following steps: S401, Dataset Construction, including: Collect historical production data, including full-process parameter data for both qualified and unqualified products; The data should be labeled, indicating whether the water quality meets the standards, does not meet the standards, and the reasons for exceeding the standards; The training set, validation set, and test set are divided in a 7:2:1 ratio. S402. Model Architecture Construction: Construct a CNN-LSTM attention hybrid model, where the CNN layer includes three convolutional and pooling layers to extract spatial correlation features of parameters at each stage and output... ; The LSTM layer consists of two LSTM layers, capturing temporal variation features and outputting... ; The attention layer employs the Bahdanau attention mechanism to enhance the weights of key parameters and output... ; The output layer outputs both the weight of the impact on water quality and the probability of water quality meeting standards.

9. The method for monitoring and managing the entire production process of bottled water according to claim 7, characterized in that, S400 also includes the following steps: S403, Model Training and Validation, including: The Adam optimizer is used, and the loss function is the sum of cross-entropy loss and mean squared error loss. Early stopping is used during training to prevent overfitting, and training is stopped if the accuracy on the validation set does not improve for 5 consecutive epochs. Test set validation was conducted, with validation requirements including a water quality compliance prediction accuracy of ≥95% and an impact weight assessment error of ≤5%. S404, Model Optimization and Deployment, including: Pruning and quantization of the trained model improves the real-time calculation speed of water quality impact weights and water quality compliance probability. The model is deployed to edge computing nodes and cloud servers to calculate the water quality impact weight and the probability of water quality compliance in real time. Configure the model call interface to achieve linkage with the data preprocessing module and the decision and resource allocation module.

10. A method for monitoring and managing the entire production process of bottled water according to claim 6, characterized in that, S500 real-time monitoring and analysis specifically includes the following steps: S501, Real-time data input: The preprocessed feature data is input into the deep learning model deployed on the edge node in real time to calculate the real-time comprehensive feature value of each stage; The cloud server synchronously receives data for the subsequent iteration of the model's real-time monitoring and historical target mapping coefficients and deviation correction terms. S502, Water quality impact analysis: The model calculates the real-time impact weight of each production link and dynamically updates the weight distribution. S503, Water Quality Compliance Prediction: The model calculates the probability of water quality compliance and the risk level of exceeding the standard in real time, locates key risk points, and outputs the parameters that lead to the risk. S504, data association storage, synchronizes and stores real-time parameters, water quality impact weights, water quality compliance probability and risk point information to a distributed database, and associates timestamps with production batch numbers.

Images