Operation and maintenance system and method for double-membrane process water purification plant based on internet of things technology

Abstract

This invention belongs to the field of water treatment and data acquisition technology, specifically disclosing an operation and maintenance system and method for a dual-membrane process water purification plant based on Internet of Things (IoT) technology. The system includes a data sensing layer, an edge computing layer, a cloud platform analysis layer, and an operation and maintenance decision execution layer. Data is collected through sensors, the edge computing layer performs soft measurement and feature extraction, the cloud platform analysis layer uses deep learning models for trend prediction, anomaly detection, and fault diagnosis, and finally, the operation and maintenance decision execution layer generates maintenance or emergency commands. This solution achieves accurate prediction of membrane performance and early intelligent handling of faults, reducing operation and maintenance costs and improving system reliability.

Description

Technical Field

[0001] This invention belongs to the field of water treatment and data acquisition technology, specifically relating to the operation and maintenance system and method of a dual-membrane process water purification plant based on Internet of Things technology. Background Technology

[0002] In the field of water treatment technology, membrane separation technology has become an important component of modern water purification processes due to its advantages such as high efficiency, energy saving, and stable effluent quality. Dual-membrane technology, combining ultrafiltration and reverse osmosis, is widely used in the production of high-quality drinking water and the deep treatment and reuse of wastewater. The long-term stable operation and efficient maintenance of membrane modules ensure the economic benefits and water supply security of the entire water treatment plant.

[0003] In existing technologies, the operation and maintenance management of membrane systems heavily relies on monitoring the health status of membrane modules, particularly membrane flux, a key performance indicator. Instruments that directly monitor membrane flux online are not only expensive to purchase and maintain, but are also prone to damage in complex chemical and physical environments, leading to long-term data loss or unreliability. Traditional maintenance strategies based on fixed-cycle replacement or simple threshold alarms struggle to accurately model and predict the nonlinear decay trend of membrane flux over operating time, failing to provide a scientific basis for preventative maintenance. For sudden mechanical failures such as membrane fiber breakage and seal leakage, existing monitoring methods suffer from severe early warning delays, often only being detected after the fault has already affected permeate water quality or system pressure. This hinders early intervention, increasing the risk of unplanned downtime and the cost of membrane module replacement.

[0004] In the operation and maintenance of dual-membrane water purification plants, how to achieve accurate prediction of membrane flux decline trends and early intelligent warning of sudden failures at low cost and with high reliability has become a technical challenge that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an operation and maintenance system and method for a dual-membrane process water purification plant based on Internet of Things (IoT) technology, in order to solve the problems of high cost and poor reliability of membrane flux monitoring in the prior art, as well as the inability to accurately predict the nonlinear decay trend of membrane flux and the lag in early warning of sudden mechanical failures.

[0006] This invention provides an operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things (IoT) technology, comprising: The data sensing layer is used to collect multi-dimensional physical and chemical process parameters during the operation of the water purification plant. The data sensing layer includes pressure sensors, flow sensors and turbidity sensors deployed on the feed water side of the membrane unit, flow sensors and conductivity sensors deployed on the product water side of the membrane unit, and pressure sensors deployed on the concentrate side of the membrane unit. The edge computing layer is used to perform real-time preprocessing and feature extraction of raw high-frequency data from the data sensing layer. The edge computing layer performs soft measurement calculation of membrane flux, specifically by receiving instantaneous flow data from the permeate side flow sensor and combining it with preset membrane module filtration area parameters to calculate and output membrane flux values ​​in real time. The cloud platform analysis layer is used to receive and store time-series state feature vectors from the edge computing layer, and to perform in-depth analysis and prediction of membrane health status. The cloud platform analysis layer includes a trend prediction engine, an anomaly detection engine, and a fault diagnosis engine. The operation and maintenance decision execution layer is used to generate and issue specific operation and maintenance instructions based on the output results of the cloud platform analysis layer. The operation and maintenance decision execution layer includes a predictive maintenance decision module and a fault emergency decision module.

[0007] Preferably, the data sensing layer further includes a reagent flow meter and a concentration meter deployed on the membrane unit chemical cleaning system, and a vibration sensor array deployed on the membrane unit frame structure; All sensors aggregate and convert data through an industrial-grade IoT gateway, and upload real-time data streams at a preset sampling frequency.

[0008] Preferably, the edge computing layer also performs the construction of the operating state feature vector, specifically by performing a moving average calculation of the influent pressure, product flow rate, concentrate pressure, transmembrane pressure difference, and vibration spectrum energy value with a time window of 5 minutes, and calculating the rate of change of transmembrane pressure difference relative to influent pressure, as well as the integral value of vibration spectrum energy in the 100 Hz to 500 Hz frequency band, and encapsulating these calculated feature values ​​together with the real-time membrane flux value into a standardized state feature vector; The edge computing layer uploads the state feature vector every 1 minute.

[0009] Preferably, the trend prediction engine is used to model and predict the long-term decline trend of membrane flux. It adopts a bidirectional long short-term memory network based on the attention mechanism as the prediction model. The input is a sequence of historical state feature vectors for the past 72 consecutive hours, and the output is a sequence of hourly predicted values ​​of membrane flux for the next 24 hours. The percentage of the predicted flux value to the current designed flux value is calculated and defined as the flux retention rate prediction value. The anomaly detection engine is used to identify sudden abnormal states of the membrane system in real time. It adopts an unsupervised learning algorithm based on isolated forest. In the real-time detection stage, it receives each latest state feature vector uploaded by the edge computing layer and inputs it into the trained isolated forest model for calculation. The isolated forest model outputs the anomaly score of the state feature vector. When the anomaly score is greater than the dynamic threshold set according to the normal sample distribution, the system state at the current moment is determined to be abnormal. The fault diagnosis engine is used to perform fine-grained identification of specific fault types after the anomaly detection engine issues an alarm. It constructs a multi-class gradient boosting decision tree model. After receiving an anomaly alarm, it immediately retrieves the state feature vector sequences one hour before and after the anomaly point, extracts 12 high-order statistical features including pressure fluctuation variance, peak frequency of vibration energy spectrum, and conductivity abrupt gradient, and inputs them into the trained multi-class gradient boosting decision tree model. The multi-class gradient boosting decision tree model outputs the probability distribution of various preset fault types, and outputs the fault type with the highest probability as the preliminary diagnosis result. After the initial diagnostic results of the fault diagnosis engine are output, the system starts the human-machine collaborative verification process, pushing the diagnostic results along with the relevant raw data and feature data to the operation and maintenance personnel's terminal. The operation and maintenance personnel need to check on-site and then confirm or correct the fault type on the terminal. Confirmed failure cases and their complete data chains will be automatically added to the historical failure case library for subsequent incremental learning of the gradient boosting decision tree model.

[0010] Preferably, the predictive maintenance decision module receives the flux retention rate prediction value from the trend prediction engine, and has three preset maintenance decision thresholds: an early warning threshold, a suggested cleaning threshold, and a forced replacement threshold. Based on the comparison results between the flux retention rate prediction value and these thresholds, it generates a first-level early warning notification, a chemical cleaning suggested work order, or a membrane module replacement early warning report. When the predictive maintenance decision module generates a chemical cleaning recommendation work order, the cleaning agent formula and cleaning duration parameters automatically recommended in the work order are calculated by matching the current flux decay rate with the historical cleaning effect database. The fault emergency decision-making module receives abnormal alarms from the abnormality detection engine and preliminary diagnostic results from the fault diagnosis engine. It has a pre-set emergency control strategy library that is linked to different fault types, generates alarms of different levels based on the diagnostic results, and executes corresponding emergency control operations through the control interface.

[0011] Preferably, the edge computing layer processes the vibration signal as follows: The vibration spectrum is obtained by performing a fast Fourier transform on the original vibration signal with a time window of 0.1 seconds. Calculate the sum of squares of the spectral amplitudes within the 100 Hz to 500 Hz frequency band as the vibration energy value for the time window; Finally, the vibration energy values ​​of 300 consecutive time windows are arithmetically averaged to obtain the vibration spectrum energy integral characteristic value uploaded once per minute.

[0012] Preferably, after each prediction is completed, the trend prediction engine calculates the root mean square error between the predicted sequence and the actual subsequent collected sequence; When the average root mean square error of 10 consecutive predictions is greater than twice the error of the validation set during the model training phase, the system automatically triggers the model retraining process, which uses historical data from the most recent 3 months.

[0013] Preferably, when the fault emergency decision-making module diagnoses a membrane fiber breakage or sealing ring leakage, it immediately generates a high-level alarm and sends instructions to the PLC to which the membrane unit belongs via the control interface. The instructions sequentially execute the operations of closing the corresponding membrane group inlet valve, opening the drain valve, and isolating the membrane group from the product water sequence, while automatically starting the commissioning procedure of the standby membrane group.

[0014] Preferably, the operation and maintenance decision execution layer also includes a comprehensive health assessment module, which periodically calculates a comprehensive health index of the membrane system between 0 and 100 by combining the long-term decay rate of the trend prediction engine, the monthly anomaly frequency of the anomaly detection engine, and the historical diagnostic accuracy of the fault diagnosis engine.

[0015] The present invention also provides an operation and maintenance method for a dual-membrane process water purification plant based on Internet of Things (IoT) technology, which uses the above-mentioned operation and maintenance system for a dual-membrane process water purification plant based on IoT technology to realize the operation and maintenance of the dual-membrane process water purification plant.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an IoT-based soft measurement and multi-sensor fusion system, which eliminates the expensive and easily damaged direct membrane flux monitoring instruments. It indirectly and reliably calculates membrane flux using conventional pressure and flow sensor data, and introduces vibration sensing to capture mechanical state. While reducing hardware investment and maintenance costs, it improves the dimensionality, continuity and reliability of condition monitoring data, laying a solid data foundation for advanced analysis.

[0017] 2. This invention deploys a trend prediction engine and uses a bidirectional long short-term memory network based on an attention mechanism to perform deep learning on massive time-series data. It can capture and predict the complex nonlinear decay law of membrane flux as it changes with operating time and cleaning cycle. This realizes the transformation from passive response maintenance to proactive preventive maintenance based on accurate prediction, scientifically guides the timing of chemical cleaning and membrane replacement, extends membrane life and reduces operation and maintenance costs.

[0018] 3. This invention employs a cascaded design of an anomaly detection engine and a fault diagnosis engine. It utilizes unsupervised learning algorithms to quickly detect any abnormal states deviating from the normal pattern at the system level, and supervised learning models to perform refined classification and diagnosis of anomalies. This two-tiered mechanism enables early and rapid identification and location of sudden mechanical faults such as membrane filament rupture and seal leakage. Furthermore, the fault emergency decision-making module automatically executes pre-set isolation and emergency procedures, minimizing the impact of the fault and improving the reliability and safety of system operation, thus avoiding unplanned downtime and water quality safety risks. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of membrane flux trend prediction based on the attention mechanism of the bidirectional long short-term memory network in this invention. Figure 3 This is a flowchart of the cascaded fault handling logic of the anomaly detection engine and the fault diagnosis engine in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the data perception layer, the edge computing layer and the cloud platform analysis layer in this invention; Figure 5 This is a logical flowchart of the operation and maintenance decision execution layer in this invention, which generates decision instructions based on prediction and diagnosis results. Detailed Implementation

[0020] The overall architecture of the dual-membrane process water purification plant operation and maintenance system based on Internet of Things technology proposed in this invention is as follows: Figure 1 As shown, the system consists of four main parts: a data perception layer, an edge computing layer, a cloud platform analysis layer, and an operation and maintenance decision execution layer. These layers communicate with each other via industrial-grade communication protocols to achieve highly reliable, low-latency data interaction and command loops. Deployed in a dual-membrane water treatment plant employing ultrafiltration and reverse osmosis membrane units operating in series, the system is deeply integrated with on-site membrane treatment equipment, chemical cleaning systems, programmable logic controllers (PLCs), and human-machine interface terminals, forming a comprehensive intelligent operation and maintenance system covering perception, computation, analysis, decision-making, and execution.

[0021] Please refer to the attached document. Figure 4The data sensing layer, serving as the physical information entry point for the entire system, is responsible for collecting multi-dimensional physical and chemical process parameters closely related to membrane performance during the operation of the water purification plant. This data sensing layer is equipped with pressure sensors, flow sensors, and turbidity sensors on the feed water side of the ultrafiltration and reverse osmosis membrane units, respectively, to monitor the raw water pressure, instantaneous flow rate, and suspended solids concentration entering the membrane modules in real time. High-precision flow sensors and conductivity sensors are configured on the product water side to obtain the effective permeate flow rate and desalination effect. A pressure sensor is installed on the concentrate side to reflect changes in fluid resistance within the membrane modules.

[0022] To detect potential mechanical anomalies in the membrane system, a triaxial vibration sensor array is installed at each of the flange connections at both ends of the ultrafiltration membrane module and the reverse osmosis membrane pressure vessel, forming a vibration sensor array. This array can simultaneously acquire vibration acceleration signals in the X, Y, and Z directions at a sampling frequency of no less than 10 kHz. Simultaneously, the chemical cleaning system is equipped with a chemical flow meter and a concentration meter to record the type, flow rate, and actual concentration of the chemicals used in each cleaning operation, ensuring the traceability and quantifiability of the cleaning process. All of the above sensors are locally aggregated through an industrial-grade IoT gateway. The gateway supports multiple industrial communication protocols such as Modbus TCP, OPC UA, and MQTT. After protocol conversion and data encapsulation, the raw data stream is pushed to the edge computing layer at a sampling frequency of once per second.

[0023] The edge computing layer consists of industrial edge computers deployed in the water treatment plant's control room. It performs real-time preprocessing, feature extraction, and state vector construction on high-frequency raw data from the data sensing layer, significantly reducing the amount of data uploaded to the cloud and improving analysis efficiency. This edge computing layer first performs soft-sensor calculations of membrane flux. Specifically, the edge computing layer receives the instantaneous volumetric flow rate values ​​output by the permeate flow sensor. (Unit: cubic meters per hour), and combined with the membrane module filtration area preset in the local configuration file. (Unit: square meters), based on the formula: ; Real-time calculation of membrane flux (Unit: cubic meters / square meter·hour). The above formula shows that membrane flux is the amount of water produced per unit membrane area per unit time, and is an important indicator for measuring membrane performance. This calculation process does not rely on a dedicated flux sensor; it can achieve high-precision indirect measurement using only conventional flow rates and known geometric parameters, reducing hardware costs and maintenance complexity.

[0024] After completing the membrane flux calculation, the edge computing layer further constructs an operational status feature vector. This operational status feature vector contains several key features, and its generation process is as follows: For the feedwater pressure, permeate flow rate, concentrate pressure, transmembrane pressure difference (i.e., the difference between feedwater pressure and permeate pressure), and vibration spectrum energy value, a 5-minute moving average is applied to filter out high-frequency noise and retain trend information. The rate of change of transmembrane pressure difference relative to feedwater pressure is calculated. This rate of change reflects the rate of resistance growth caused by membrane fouling. It is calculated by taking the derivative of the average transmembrane pressure difference with respect to the average feedwater pressure within the current 5-minute window, and the result is smoothed before being incorporated into the feature vector.

[0025] For vibration signals, the edge computing layer performs a Fast Fourier Transform (FFT) on the raw time-domain signal of each triaxial vibration sensor with a time window of 0.1 seconds to obtain the corresponding vibration spectrum. It then calculates the sum of squares of the amplitudes at all frequency points within the 100 Hz to 500 Hz band as the vibration energy value for each time window. Finally, it performs an arithmetic average of the vibration energy values ​​over 300 consecutive time windows (i.e., 30 seconds) to obtain the vibration spectrum energy integral characteristic value, updated every minute. The selection of this frequency band is based on engineering experience; vibration energy within the 100 Hz to 500 Hz range primarily reflects characteristic frequencies of mechanical anomalies such as loosening of the membrane module's internal structure, seal failure, or fluid impact.

[0026] The edge computing layer encapsulates the seven features—the average transmembrane pressure difference, average feed water pressure, average permeate flow rate, average concentrate pressure, rate of change of transmembrane pressure difference, vibration spectrum energy integral value, and real-time membrane flux value—into a standardized state feature vector according to a predefined order, and uploads it to the cloud platform analysis layer through a secure encrypted channel at a 1-minute interval.

[0027] Please refer to the attached document. Figure 2 The cloud platform analytics layer is deployed on a cloud server cluster and is responsible for receiving, storing, and deeply analyzing time-series state feature vectors from multiple edge nodes. This cloud platform analytics layer comprises three collaborative analytics engines: a trend prediction engine, an anomaly detection engine, and a fault diagnosis engine.

[0028] The trend prediction engine models and predicts the long-term nonlinear decay trend of membrane flux. This engine employs a bidirectional long short-term memory network based on an attention mechanism as the prediction model. The input to the prediction model is a sequence of historical state feature vectors for the past 72 consecutive hours (4320 time steps), with state feature vectors for each minute. The output is a membrane flux prediction value for each hour within the next 24 hours, totaling 24 prediction points. During the model training phase, the system extracts long-term operational data containing complete cleaning-running-decaying-re-cleaning cycles from the historical database to form a training set.

[0029] The training data is first normalized, scaling each feature to the 0-1 range, and then divided into a 72-hour input sequence and its corresponding 24-hour target sequence. The prediction model uses mean squared error as the loss function and is trained end-to-end using backpropagation until the prediction error on the validation set stabilizes and converges. During online operation, the trend prediction engine automatically triggers a prediction task every 6 hours, using the latest 72-hour data to predict the membrane flux curve for the next 24 hours. After the prediction is completed, the system calculates the percentage of the predicted flux value to the current design flux value, i.e., the initial flux of the new membrane, defined as the flux retention rate prediction value. For example, if the design flux is 30 cubic meters per square meter per hour, and the predicted flux after 24 hours is 24, then the flux retention rate prediction value is 80%.

[0030] The system also calculates the root mean square error between the predicted sequence and the actual subsequent collected sequences. When the average root mean square error of 10 consecutive predictions is greater than twice the error of the validation set during the model training phase, the system automatically triggers the model retraining process, retraining the model using historical data from the most recent 3 months to adapt to the long-term drift of the membrane module performance degradation curve.

[0031] An anomaly detection engine is used to identify sudden abnormal states in the membrane system that deviate from normal operating mode in real time. This engine employs an unsupervised learning algorithm based on Isolation Forest. After initial system commissioning or major overhaul, the system operates in a stable and efficient state at the water treatment plant. Two weeks of continuous data collection of state feature vectors forms a normal sample set. This normal sample set is used to train the Isolation Forest model. The Isolation Forest model constructs multiple isolation trees by randomly selecting features and segmentation values. Abnormal samples are isolated more quickly and through shorter paths due to their sparse distribution.

[0032] During the real-time detection phase, whenever a new state feature vector is uploaded from the edge computing layer, this vector is input into the trained Isolation Forest model. The Isolation Forest model outputs an anomaly score, which is inversely proportional to the average path length of the sample within the isolation trees. The system sets a dynamic threshold based on the 95th percentile of the anomaly scores in the normal sample set. When the anomaly score of a new sample exceeds this dynamic threshold, the system is determined to be in an abnormal state. This dynamic threshold is automatically updated quarterly to incorporate new normal operation data, ensuring the Isolation Forest model's adaptability to slowly changing operating conditions.

[0033] Please refer to the attached document. Figure 3When the anomaly detection engine issues an anomaly alert, the fault diagnosis engine immediately activates to perform refined classification of the anomaly's cause. This anomaly detection engine constructs a multi-classification Gradient Boosting Decision Tree (GBDT) model. The GBDT model's training relies on a historical fault case database. Each record in this database contains a sequence of state feature vectors for the two hours preceding and following the fault, along with a fault type label ultimately confirmed by operations personnel. The preset fault types include four categories: membrane filament rupture, sealing ring leakage, water ingress blockage, and chemical contamination.

[0034] During the diagnostic phase, the fault diagnosis engine retrieves the state feature vector sequences one hour before and after the anomaly point, extracting 12 high-order statistical features, including but not limited to: the standard deviation of inlet pressure within a 10-minute window, the peak frequency of the vibration energy spectrum in the 200 Hz to 400 Hz band, and the maximum gradient change in conductivity within 5 minutes. These high-order features are input into the trained GBDT model, which outputs a four-dimensional probability distribution, corresponding to the confidence levels of four fault types. The system outputs the fault type with the highest probability as the preliminary diagnostic result. To improve diagnostic reliability, the system then initiates a human-machine collaborative verification process: the preliminary diagnostic result, along with the relevant raw sensor data, feature vectors, and high-order statistical features, is pushed to the mobile terminal of the on-site maintenance personnel. The maintenance personnel must complete the on-site inspection within 2 hours and confirm or correct the fault type on their terminal. All confirmed fault cases and their complete data chains are automatically added to the historical fault case library for subsequent incremental learning of the GBDT model, thus forming a closed loop of continuously optimized diagnostic capabilities.

[0035] Please refer to the attached document. Figure 5 The operations and maintenance decision execution layer, serving as the system's execution exit, automatically generates and issues specific operations and maintenance instructions based on the output of the cloud platform's analysis layer, achieving seamless integration from analysis to action. This layer includes a predictive maintenance decision module and a fault emergency decision module.

[0036] The predictive maintenance decision-making module receives flux retention rate predictions from the trend prediction engine. Internally, the module has three key thresholds: an early warning threshold of 85%, a recommended cleaning threshold of 75%, and a forced replacement threshold of 60%. When the predicted flux retention rate falls below 85% but above 75% within the next 24 hours, the module generates a Level 1 early warning notification, pushed to the operations team via WeChat or SMS, reminding them to pay attention to the membrane performance degradation trend and recommending increased daily inspections. When the predicted minimum is below 75%, the module automatically generates a chemical cleaning recommendation work order. The work order includes the recommended cleaning agent formulation, cleaning circulation flow rate, cleaning temperature, and cleaning duration. These parameters are not fixed values ​​but are calculated by matching the current flux degradation rate (the slope of flux decline over the past 72 hours) with a historical cleaning effect database.

[0037] If the current decay rate is 2% per day, and historical data shows that circulating a 0.5% citric acid solution for 2 hours can restore flux to over 90% under similar decay conditions, then the system recommends this approach. When the predicted flux retention rate is less than 60% for 48 consecutive hours, and the flux recovery rate after the most recent chemical cleaning (i.e., the ratio of flux after cleaning to flux before cleaning) is less than 60%, the predictive maintenance decision module determines that the membrane module has entered the irreversible aging stage, generates a membrane module replacement early warning report, and submits it to the equipment management department as the basis for spare parts procurement and downtime planning.

[0038] The fault emergency decision-making module receives anomaly alarms from the anomaly detection engine and preliminary diagnostic results from the fault diagnosis engine, and executes corresponding operations based on the preset emergency control strategy library. When the diagnostic result indicates membrane filament rupture or sealing ring leakage, the system determines it as a serious fault, immediately generates a high-level alarm (audible and visual alarm + SMS notification), and sends control commands to the PLC of the corresponding membrane module via the OPC UA interface to perform the following operations in sequence: first, close the inlet electric valve of the membrane module to cut off the water supply; second, open the concentrate side drain valve to drain the residual liquid in the membrane housing; finally, logically isolate the membrane module from the permeate sequence through PLC logic to prevent contamination spread.

[0039] The system automatically initiates the standby membrane module commissioning procedure, including opening the standby membrane module inlet valve, starting the matching booster pump, and adjusting the permeate regulating valve opening to ensure stable total permeate flow. When the diagnosis indicates inlet blockage, the system generates a medium-level alarm and recommends performing a backflushing procedure. The backflushing parameters are automatically calculated based on the rate of increase of the inlet pressure over the past 10 minutes; the faster the rate of increase, the stronger the backflushing. When the diagnosis indicates chemical fouling, the system generates a primary alarm and recommends adjusting the inlet pH or adding a specific dispersant. Simultaneously, the system records the characteristic data of this fouling event to the case database for optimizing future fouling prevention strategies.

[0040] In addition, the operation and maintenance decision execution layer also includes a comprehensive health assessment module. This module runs automatically at 00:00 on the 1st of each month, combining three indicators: the monthly average flux decay rate calculated by the trend prediction engine, the monthly frequency of abnormal events counted by the anomaly detection engine, and the accuracy rate of the fault diagnosis engine in all diagnostic tasks of the previous month (i.e., the proportion of correct diagnoses after human-machine collaboration). The system weights and fuses these three indicators with weights of 0.5, 0.3, and 0.2, respectively, to calculate a comprehensive health index of the membrane system between 0 and 100. For example, if the monthly decay rate is 0.8% / day, corresponding to a score of 85; the anomaly frequency is 2 times, corresponding to a score of 90; and the diagnostic accuracy rate is 92%, corresponding to a score of 92, then the comprehensive health index is 0.5×85+0.3×90+0.2×92=87.9. This comprehensive health index is presented to management through a web-based visual dashboard, supporting multi-dimensional drill-down analysis by membrane module, by month, and by fault type, providing quantitative decision-making basis for long-term asset management, membrane replacement budget preparation, and technological transformation investment.

[0041] Throughout the system's operation, the integrity and security of information flow are ensured through strictly defined data interfaces and communication protocols between each layer. The data perception layer and the edge computing layer use the MQTT protocol based on TLS 1.3 encryption to ensure secure transmission of sensor data within the local network. The edge computing layer and the cloud platform analysis layer interact through an API gateway, with all uploaded data undergoing digital signatures and integrity verification. The cloud platform analysis layer and the operation and maintenance decision execution layer achieve asynchronous decoupling through message queues, ensuring reliable command delivery under high concurrency. The system also incorporates a complete fault recovery mechanism: if an edge computing node fails, the IoT gateway can cache up to 72 hours of raw data and automatically retransmit it after network recovery; if the cloud platform analysis layer service is interrupted, the edge computing layer can switch to a local lightweight rule engine, execute basic alarms based on preset thresholds, and synchronize the status again after cloud recovery.

[0042] In summary, this embodiment achieves comprehensive, accurate, and real-time monitoring of the operating status of the membrane system in a dual-membrane water purification plant by constructing a multi-level, multi-modal intelligent sensing and analysis system. Based on this, it forms a closed-loop operation and maintenance capability from prediction and early warning to fault diagnosis and automatic execution, solving the problems of high monitoring costs, inaccurate predictions, and delayed response in traditional operation and maintenance models.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 process, method, article, or apparatus.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An operation and maintenance system for a dual-membrane water purification plant based on Internet of Things (IoT) technology, characterized in that: include: The data sensing layer is used to collect multi-dimensional physical and chemical process parameters during the operation of the water purification plant. The data sensing layer includes pressure sensors, flow sensors and turbidity sensors deployed on the feed water side of the membrane unit, flow sensors and conductivity sensors deployed on the product water side of the membrane unit, and pressure sensors deployed on the concentrate side of the membrane unit. The edge computing layer is used to perform real-time preprocessing and feature extraction of raw high-frequency data from the data sensing layer. The edge computing layer performs soft measurement calculation of membrane flux, specifically by receiving instantaneous flow data from the permeate side flow sensor and combining it with preset membrane module filtration area parameters to calculate and output membrane flux values ​​in real time. The cloud platform analysis layer is used to receive and store time-series state feature vectors from the edge computing layer, and to perform in-depth analysis and prediction of membrane health status. The cloud platform analysis layer includes a trend prediction engine, an anomaly detection engine, and a fault diagnosis engine. The operation and maintenance decision execution layer is used to generate and issue specific operation and maintenance instructions based on the output results of the cloud platform analysis layer. The operation and maintenance decision execution layer includes a predictive maintenance decision module and a fault emergency decision module.

2. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 1, characterized in that, The data sensing layer also includes a reagent flow meter and a concentration meter deployed on the membrane unit chemical cleaning system, and a vibration sensor array deployed on the membrane unit frame structure. All sensors aggregate and convert data through an industrial-grade IoT gateway, and upload real-time data streams at a preset sampling frequency.

3. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 2, characterized in that, The edge computing layer also performs the construction of the operating state feature vector. Specifically, it performs a moving average calculation of the influent pressure, product flow rate, concentrate pressure, transmembrane pressure difference, and vibration spectrum energy value with a time window of 5 minutes. It also calculates the rate of change of transmembrane pressure difference relative to influent pressure and the integral value of vibration spectrum energy in the 100 Hz to 500 Hz frequency band. These calculated feature values ​​are then encapsulated together with the real-time membrane flux value into a standardized state feature vector. The edge computing layer uploads the state feature vector every 1 minute.

4. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 3, characterized in that, The trend prediction engine is used to model and predict the long-term decline trend of membrane flux. It adopts a bidirectional long short-term memory network based on the attention mechanism as the prediction model. The input is the historical state feature vector sequence of the past 72 consecutive hours, and the output is the hourly predicted value sequence of membrane flux in the next 24 hours. It also calculates the percentage of the predicted flux value to the current designed flux value, which is defined as the flux retention rate prediction value. The anomaly detection engine is used to identify sudden abnormal states of the membrane system in real time. It adopts an unsupervised learning algorithm based on isolated forest. In the real-time detection stage, it receives each latest state feature vector uploaded by the edge computing layer and inputs it into the trained isolated forest model for calculation. The isolated forest model outputs the anomaly score of the state feature vector. When the anomaly score is greater than the dynamic threshold set according to the normal sample distribution, the system state at the current moment is determined to be abnormal. The fault diagnosis engine is used to perform fine-grained identification of specific fault types after the anomaly detection engine issues an alarm. It constructs a multi-class gradient boosting decision tree model. After receiving an anomaly alarm, it immediately retrieves the state feature vector sequences one hour before and after the anomaly point, extracts 12 high-order statistical features including pressure fluctuation variance, peak frequency of vibration energy spectrum, and conductivity abrupt gradient, and inputs them into the trained multi-class gradient boosting decision tree model. The multi-class gradient boosting decision tree model outputs the probability distribution of various preset fault types, and outputs the fault type with the highest probability as the preliminary diagnosis result. After the initial diagnostic results of the fault diagnosis engine are output, the system starts the human-machine collaborative verification process, pushing the diagnostic results along with the relevant raw data and feature data to the operation and maintenance personnel's terminal. The operation and maintenance personnel need to check on-site and then confirm or correct the fault type on the terminal. Confirmed failure cases and their complete data chains will be automatically added to the historical failure case library for subsequent incremental learning of the gradient boosting decision tree model.

5. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 4, characterized in that, The predictive maintenance decision module receives the flux retention rate prediction value from the trend prediction engine. It has three preset maintenance decision thresholds: early warning threshold, recommended cleaning threshold, and forced replacement threshold. Based on the comparison between the flux retention rate prediction value and these thresholds, it generates a first-level early warning notification, a chemical cleaning recommendation work order, or a membrane module replacement early warning report. When the predictive maintenance decision module generates a chemical cleaning recommendation work order, the cleaning agent formula and cleaning duration parameters automatically recommended in the work order are calculated by matching the current flux decay rate with the historical cleaning effect database. The fault emergency decision-making module receives abnormal alarms from the abnormality detection engine and preliminary diagnostic results from the fault diagnosis engine. It has a pre-set emergency control strategy library that is linked to different fault types, generates alarms of different levels based on the diagnostic results, and executes corresponding emergency control operations through the control interface.

6. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 5, characterized in that, The edge computing layer processes vibration signals as follows: The vibration spectrum is obtained by performing a fast Fourier transform on the original vibration signal with a time window of 0.1 seconds. Calculate the sum of squares of the spectral amplitudes within the 100 Hz to 500 Hz frequency band as the vibration energy value for the time window; Finally, the vibration energy values ​​of 300 consecutive time windows are arithmetically averaged to obtain the vibration spectrum energy integral characteristic value uploaded once per minute.

7. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 6, characterized in that, After each prediction is completed, the trend prediction engine calculates the root mean square error between the predicted sequence and the actual subsequent collected sequence. When the average root mean square error of 10 consecutive predictions is greater than twice the error of the validation set during the model training phase, the system automatically triggers the model retraining process, which uses historical data from the most recent 3 months.

8. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 7, characterized in that, When the fault emergency decision-making module diagnoses a membrane fiber breakage or sealing ring leakage, it immediately generates a high-level alarm and sends instructions to the PLC of the membrane unit through the control interface. The PLC sequentially executes operations such as closing the corresponding membrane group inlet valve, opening the drain valve, and isolating the membrane group from the product water sequence. At the same time, it automatically starts the commissioning procedure of the standby membrane group.

9. The operation and maintenance system for a dual-membrane process water purification plant based on Internet of Things technology according to claim 8, characterized in that, The operation and maintenance decision execution layer also includes a comprehensive health assessment module. This module periodically combines the long-term decay rate of the trend prediction engine, the monthly anomaly frequency of the anomaly detection engine, and the historical diagnostic accuracy of the fault diagnosis engine to calculate a comprehensive health index of the membrane system between 0 and 100.

10. A method for operation and maintenance of a dual-membrane water purification plant based on Internet of Things (IoT) technology, characterized in that: The operation and maintenance of a dual-membrane process water purification plant is achieved using the IoT-based dual-membrane process water purification plant operation and maintenance system as described in any one of claims 1 to 9.

Images