Intelligent operation and maintenance system and method for reverse osmosis system based on physical-AI hybrid drive

Abstract

The invention discloses an intelligent operation and maintenance system and method for a reverse osmosis system based on physical-AI hybrid drive, and relates to the technical field of water treatment and membrane separation engineering.The method comprises the following steps that data of changes of high-pressure pump frequency, water inlet pressure and water production flow along with time in the operation process of the reverse osmosis system are collected; all the operation data are uniformly mapped to the same time scale, and an operation rhythm reference time zone covering the whole reverse osmosis device is established. By establishing the running rhythm reference time zone with the unified time scale, synchronous alignment and fusion of the physical model and the artificial intelligence model are achieved, and it is ensured that the time sequence of operation and maintenance judgment is consistent and the decision is reliable; and a single decision chain and a ripple time regulation and control ring are constructed in a high-risk period, so that a control instruction is continuously and smoothly output, and the operation stability and the intelligent regulation and control capability of the reverse osmosis system are improved.

Description

Technical Field

[0001] This invention relates to the field of water treatment and membrane separation engineering technology, specifically to an intelligent operation and maintenance system and method for reverse osmosis systems based on a physical-AI hybrid drive. Background Technology

[0002] Intelligent operation and maintenance (O&M) of reverse osmosis (RO) systems, driven by a hybrid physics-AI approach, refers to an O&M technology model that moves away from relying solely on human experience or black-box AI models in the operation and management of RO systems and industrial control systems. Instead, it leverages the physical mechanisms and engineering constraints of the RO process itself, incorporating AI algorithms for collaborative modeling and decision-making. In essence, this model utilizes principles of mass conservation, hydraulic relationships, and engineering standards such as ASTM D4516 to physically verify and standardize collected operational data such as pressure, flow rate, and conductivity, ensuring that the data entering the model accurately reflects the membrane performance. Based on this, an interpretable performance baseline is established using a physical mechanism model. Then, a deep learning model is used to correct trends and make short-term predictions for nonlinear and time-varying characteristics such as membrane fouling, thereby enabling intelligent decision-making regarding cleaning timing, energy consumption control, and reagent dosing. This hybrid driving mode essentially uses physical laws to constrain the reasoning boundaries of AI, and uses AI to make up for the shortcomings of traditional mechanism models in responding to complex pollution evolution and operational disturbances. This transforms the operation and maintenance of reverse osmosis systems and industrial control systems from post-event response and experience-based judgment to a forward-looking, verifiable, and explainable intelligent operation and maintenance system based on real physical conditions and prediction results.

[0003] The existing technology has the following shortcomings: In existing technologies, when making operational and maintenance decisions for reverse osmosis systems based on the collaboration of physical and artificial intelligence (AI) models, the physical model typically updates continuously according to fixed time steps or event triggering methods, based on real-time collected operating parameters such as pressure, flow rate, and conductivity, to reflect the system's immediate operational status. Meanwhile, the AI ​​model primarily relies on historical time-series data, outputting trend judgments according to a preset inference cycle. In dynamic operating scenarios where the reverse osmosis system experiences rapid load changes, such as frequent start-ups and shutdowns of high-pressure pumps, adjustments to recovery rates, or sudden changes in feed water quality, the inconsistency in update frequency and inference rhythm between the physical and AI models can easily lead to multiple sets of mutually deviating state judgments existing within the system simultaneously. When these different model outputs are used simultaneously to generate operational control or maintenance recommendations without unified time-series alignment and consistency constraints, contradictory control decisions can easily arise in existing technologies. This can lead to frequent switching of maintenance commands, execution conflicts, and even misleading human intervention, thereby affecting the stability of the reverse osmosis system and the reliability of maintenance decisions.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent operation and maintenance system and method for reverse osmosis systems based on a physical-AI hybrid drive, so as to solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid approach, comprising the following steps: Data on the changes in high-pressure pump frequency, inlet water pressure, and permeate flow rate over time during the operation of the reverse osmosis system are collected. All operating data are mapped to the same time scale to establish a reference time band for the operating rhythm of the entire reverse osmosis unit, which provides a unified time benchmark for subsequent data comparison. Based on the reference time zone of the running rhythm, the real-time running data output by the physical model and the trend prediction results output by the artificial intelligence model are aligned point by point according to the time scale to generate a time difference map, which is used to characterize the offset relationship between the judgment results of different models within the same time scale. Extreme points of the offset segments are extracted along the time difference map to form a rhythm conflict list. The rhythm conflict list is then linked with the load change amplitude of the corresponding time period to identify high-risk decision periods and mark them on the running time axis. Based on the rhythm conflict list, the judgment results of the physical model and the judgment results of the artificial intelligence model are arranged into priority channels according to the degree of risk. During the high-risk decision period, the execution order of the control instructions is reorganized to construct a single decision chain, so that the judgment results of the previous moment are continuously fed into the single decision chain. A ripple-like time control loop is set up around a single decision chain to adjust the effective time of control commands during high-risk decision periods, insert short buffer windows and smooth rhythm fluctuations, so that the single decision chain maintains a continuous output trajectory on the time axis, thereby avoiding operation and maintenance command conflicts under dynamic load changes.

[0007] Preferably, the steps for generating the reference time band for the operating rhythm are as follows: During the operation of the reverse osmosis unit, real-time changes in high-pressure pump frequency, inlet water pressure, and product water flow rate are continuously collected and arranged in the order of sampling time to form an original time series. Using the sampling time series of the high-pressure pump frequency as the reference main time axis, the time records of inlet pressure and product water flow are synchronously extended, and the operating parameters from different sources are redistributed to a unified time node. The values ​​of high-pressure pump frequency, inlet pressure and product water flow rate after time mapping are combined point by point in time sequence to generate a reference time band for the operating rhythm covering the entire device. Using the operational rhythm reference time zone as a unified benchmark, the time nodes and trends of each operational parameter are recorded in the same time frame, forming an information structure containing a complete time index for subsequent data comparison and model analysis.

[0008] Preferably, in the process of generating the operating rhythm reference time zone, the time series of the high-pressure pump frequency is used as the core reference. By extending and weighting the time records of the inlet pressure and product water flow, the operating parameters form a continuous numerical sequence at a unified time node. The operating rhythm reference time zone is established in chronological order, thereby ensuring that the time alignment and changes of the operating parameters of the reverse osmosis unit are consistent in different operating stages.

[0009] Preferably, the steps for aligning the physical model output and the artificial intelligence model output point by point based on the runtime rhythm reference time band are as follows: The high-pressure pump frequency, inlet pressure, and product water flow rate are selected as alignment benchmarks. The real-time operating data output by the physical model and the trend prediction results output by the artificial intelligence model are matched one by one according to the time nodes of the operating rhythm reference time zone. The physical model output and the artificial intelligence model output at each time point are analyzed to extract the values ​​of the running variables and form a continuous difference dataset. The difference dataset is arranged in chronological order to form a time difference sequence, and the time difference sequence is mapped onto the running rhythm reference time band to generate a time difference map; Extended analysis of the offset trend in the time difference map is performed to form a time offset band that reflects the changing trend of the model's judgment difference, which is used for subsequent operation status analysis and rhythm conflict identification.

[0010] Preferably, in the process of generating the time difference map, the difference values ​​of adjacent time nodes in the time difference sequence are connected by a continuous time mapping method, so that the time difference map forms a continuously distributed difference trajectory band on the time axis, and the extreme points of the offset amplitude change are used as key feature points to characterize the offset pattern of the judgment results of the physical model and the artificial intelligence model at different operating stages.

[0011] Preferably, the steps for extracting extreme points of the offset segments along the time difference map and forming a rhythm conflict list are as follows: The offset data distributed along the time axis of the time difference map are continuously scanned to identify offset segments where the judgment results of the physical model and the artificial intelligence model differ. Extract the extreme points of the shift amplitude change along the time direction of each shift segment, and record the time position of the extreme points and the corresponding shift amplitude; A rhythm conflict list is constructed using extreme points as key indexes, and the time position, offset amplitude, offset direction and duration corresponding to the extreme points are recorded as a structured data sequence. The rhythm conflict list is linked to the load change range of the reverse osmosis unit within the corresponding time period to identify high-risk decision-making periods and mark them on the operating timeline.

[0012] Preferably, the step of linking the rhythm conflict list with the load change range is further defined as follows: by comparing the changes in influent pressure, the changes in permeable flow rate, and the changes in high-pressure pump frequency in adjacent time periods before and after the extreme point, the correlation between the load change range and the offset range is determined, and this is used as the basis for determining the high-risk decision period, so that the high-risk period forms a corresponding marked interval on the operating time axis.

[0013] Preferably, the steps for constructing a single decision chain based on a rhythm conflict list are as follows: Risk level analysis is performed on the conflict events recorded in the rhythm conflict list. The risk level is determined based on the offset amplitude, duration and load change amplitude, and the judgment results of the physical model and artificial intelligence model are extracted. The results of the physical model assessment and the results of the artificial intelligence model are arranged in a hierarchical manner according to the risk level, forming a risk priority channel, and maintaining the mapping relationship between the time sequence and the risk level. During high-risk decision-making periods, control instructions are reorganized, and control instructions from different models are arranged into a continuous execution sequence according to time order and risk level. The reorganized control logic is mapped onto the runtime timeline, forming a single decision chain that is time-continuous and logically unified, so that the judgment results of the previous moment are continuously incorporated into the decision process of the next moment.

[0014] Preferably, the steps for setting up a ripple-like time control loop around a single decision chain are as follows: Continuous scanning of the operational timeline covered by a single decision chain identifies time intervals containing high-risk decision-making periods and establishes time control intervals. A ripple-like time control loop is set up around the high-risk decision-making period to form a buffer zone with time decay characteristics before and after the control command takes effect, so that the command execution process has a flexible time span. Within the ripple-like time control loop, the effective time of control commands is reallocated based on risk level and execution priority, forming a time-layered structure that diffuses from the core to the periphery; The adjusted instruction effective time and buffer window are mapped back to the single decision chain, so that the control instructions form a continuous output trajectory on the time axis and maintain a stable operating rhythm.

[0015] The intelligent operation and maintenance system for reverse osmosis systems, driven by a hybrid physical-AI approach, includes a time baseline construction module, a time series alignment analysis module, a conflict identification and extraction module, a decision chain construction module, and a time series regulation and smoothing module. The time reference construction module collects data on the changes in high-pressure pump frequency, inlet water pressure, and product water flow rate over time during the operation of the reverse osmosis system. It maps all the operating data to the same time scale and establishes a reference time band for the operating rhythm of the entire reverse osmosis unit, which is used to provide a unified time reference for subsequent data comparison. The time-series alignment analysis module, based on the reference time band of the running rhythm, aligns the real-time running data output by the physical model with the trend prediction results output by the artificial intelligence model point by point according to the time scale, and generates a time difference map to characterize the offset relationship between the judgment results of different models within the same time scale. The conflict identification and extraction module extracts the extreme points of the offset segments along the time difference map to form a rhythm conflict list. It then binds the rhythm conflict list with the load change amplitude of the corresponding time period to identify high-risk decision periods and mark them on the running time axis. The decision chain construction module, based on the rhythm conflict list, arranges the judgment results of the physical model and the judgment results of the artificial intelligence model according to the risk level as priority channels, and reorganizes the execution order of control instructions during high-risk decision periods to construct a single decision chain, so that the judgment results of the previous moment are continuously fed into the single decision chain. The timing control and smoothing module sets up a ripple-like timing control loop around a single decision chain. During high-risk decision periods, it makes minor adjustments to the effective time of control commands, inserts short buffer windows, and smooths rhythm fluctuations, so that the single decision chain maintains a continuous output trajectory on the time axis, thereby avoiding operational command conflicts under dynamic load changes.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a reference time band for operational rhythm and achieves point-to-point alignment between the physical model and the artificial intelligence model under a unified time scale. This enables the two types of models to form a synchronous correlation in the time dimension, thereby ensuring that operational judgments from different sources can be compared and fused with a consistent time benchmark. This method allows the operating status of the reverse osmosis device under dynamic load conditions to be continuously expressed along the entire timeline, avoiding judgment deviations caused by time misalignment between model outputs. It also ensures stable temporal consistency and traceability in the generation process of control commands, thereby improving the coordination of system operation and the reliability of data-driven decisions.

[0017] This invention constructs a single decision chain and sets up a ripple-like time control loop during high-risk decision-making periods, transforming the execution of control commands from instantaneous to continuous and smooth output, forming a time transition zone with buffering characteristics. This approach effectively reduces execution conflicts and oscillations caused by the superposition of multiple commands, enabling the reverse osmosis system to maintain a smooth transition of its operating rhythm when the load changes or operating conditions fluctuate. It ensures the continuity of operation and maintenance decisions over time and the flexibility and consistency of control response, thereby improving the overall operational stability and intelligent control capabilities of the system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive, as described in this invention.

[0020] Figure 2 This is a schematic diagram of the module of the intelligent operation and maintenance system for reverse osmosis systems based on physical-AI hybrid drive of the present invention. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] This invention provides, for example Figure 1 The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid approach, as shown, includes the following steps: Data on the changes in high-pressure pump frequency, inlet water pressure, and permeate flow rate over time during the operation of the reverse osmosis system are collected. All operating data are mapped to the same time scale to establish a reference time band for the operating rhythm of the entire reverse osmosis unit, which provides a unified time benchmark for subsequent data comparison. To address the dynamic variations in high-pressure pump frequency, inlet water pressure, and permeate flow rate during different operating stages of the reverse osmosis unit, a reference time band covering the entire unit's operating rhythm is established using a unified time scale. This ensures a consistent time base for subsequent data comparison and analysis. The specific steps are as follows: When the reverse osmosis unit is in stable operation, real-time data on the high-pressure pump frequency, inlet water pressure, and permeate flow rate are continuously collected. These operating parameters are then arranged chronologically according to the sampling time sequence from the physical sensors to form a raw time series. During the data acquisition process, to ensure consistency across different parameters in the time dimension, a corresponding sampling timestamp is recorded for each operating parameter. This ensures that the high-pressure pump frequency signal, inlet water pressure signal, and permeate flow rate signal form a continuous numerical sequence within the same time reference frame. In this way, the collected raw operating data not only reflects the physical trends of each parameter but also possesses traceability and mappability in the time dimension, laying the foundation for subsequent time scale mapping and rhythm unification.

[0023] After obtaining the complete original time series, the sampling time series of the high-pressure pump frequency is used as the reference master time axis to synchronously extend the time records of influent pressure and permeate flow rate, so that operating parameters from different sources can be rearranged at the same time intervals. In this step, data with inconsistent original sampling intervals are redistributed to a unified time node through time interpolation, time extension, or time weighting, so that the high-pressure pump frequency, influent pressure, and permeate flow rate have corresponding operating values ​​at each time node. Through this unified mapping, the operating data of the entire reverse osmosis unit forms a continuous and equal-step time scale band in the time dimension, thereby realizing a time alignment structure centered on the high-pressure pump. This structure can truly reflect the coordinated changes between key operating parameters of the unit within any time period, giving the time series data of each operating parameter a unified reference significance.

[0024] After establishing a unified time scale, the mapped values ​​of high-pressure pump frequency, inlet pressure, and permeate flow rate are combined point-by-point in chronological order to generate a continuous time series set reflecting the overall operating rhythm of the unit. This time series set not only reflects the variation of individual parameters over time but also includes the interrelationships between various operating parameters at the same moment. In this way, the resulting operating rhythm reference time belt can fully cover all key operating variables from the inlet to the permeate end, presenting the rhythmic change trajectory of the reverse osmosis unit under different loads and operating conditions. This operating rhythm reference time belt has full continuity in the time dimension and covers the core operating units of the entire unit in the spatial dimension, enabling it to reflect the operating dynamics of the unit and provide a unified time scale basis for subsequent model comparison and trend correlation.

[0025] After generating the operating rhythm reference time zone, this time zone serves as a unified benchmark, simultaneously recording the time nodes, numerical amplitudes, and trends of all operating parameters within the same time frame, forming an information structure with a complete time index. This information structure maintains consistency with the sampling interval in terms of time resolution and a one-to-one correspondence with the operating variables of the reverse osmosis unit in terms of parameter dimensions, ensuring comparability and synchronization of different parameters at the same time scale. In subsequent intelligent operation and maintenance processes, whether for real-time judgment of the physical model or trend analysis of the artificial intelligence model, this time zone can be directly invoked as a unified operating rhythm reference, enabling cross-model and cross-time period data alignment and correlation analysis. In this way, the operating data of the entire reverse osmosis unit is uniformly incorporated into a continuous and scalable time scale system on the time axis, allowing the changes in high-pressure pump frequency, influent pressure, and permeate flow rate to be accurately described at the same time benchmark. This provides stable, unified, and globally consistent time reference conditions for subsequent intelligent analysis, operation prediction, and control decisions.

[0026] Based on the reference time zone of the running rhythm, the real-time running data output by the physical model and the trend prediction results output by the artificial intelligence model are aligned point by point according to the time scale to generate a time difference map, which is used to characterize the offset relationship between the judgment results of different models within the same time scale. Based on the established operational rhythm reference time band, a time difference map is constructed by aligning the real-time operational data output by the physical model with the trend prediction results output by the artificial intelligence model point by point. This map describes the differences in judgment between the two types of models at the same time scale, revealing the judgment offset relationship between different models in the time dimension. The specific steps are as follows: After obtaining the reference timeframe for the operating rhythm, the main operating variables of the reverse osmosis unit are selected as the alignment benchmark, including physical output data such as high-pressure pump frequency, feed water pressure, and permeate flow rate. Simultaneously, trend prediction results within the corresponding time range are extracted from the artificial intelligence model, and the two types of data are matched one-to-one according to the time nodes of the operating rhythm reference timeframe. During this process, to ensure that data from different sources can be compared at the same time scale, the time labels of the physical output data and the time labels of the artificial intelligence prediction results are paired at completely consistent time intervals, so that each time node simultaneously contains corresponding operating status descriptions from both the physical and artificial intelligence sides. This pairing method aligns the data structures of different models in the time dimension, providing the foundation for subsequent point-by-point comparisons and differential calculations. This step ensures that the physical model and the artificial intelligence model have the same reference starting point in the same timeframe, making the descriptions of the reverse osmosis unit's operating status by different models comparable on the time axis.

[0027] After pairing time tags, the real-time output data of the physical model and the trend prediction results of the artificial intelligence model are analyzed at each time point, and the corresponding operating variable values ​​are extracted, such as the high-pressure pump frequency, inlet pressure, and product water flow rate at the same time. By calculating the difference between the outputs of the physical model and the artificial intelligence model at the same time point, a preliminary differential dataset reflecting the difference in the judgments of the two models on the operating status at the same time scale can be formed. The key to this process is to ensure that each time point contains results from both models simultaneously, so as to achieve point-by-point alignment in time. The resulting differential dataset can be continuously distributed in the time dimension and reflects the changing trend of the judgment results between the models. In this way, the time scale is no longer just used for data arrangement of a single model, but becomes a continuous reference frame connecting the judgment results of the two models, so that the differences in operating status can be continuously tracked in time.

[0028] After obtaining a continuous difference dataset, the difference values ​​at each time point are arranged chronologically to form a complete time difference sequence. This time difference sequence is then mapped onto a reference time band for the operating rhythm, creating a time difference map covering the entire operation process. This time difference map, with time as the horizontal axis and the differences in judgments from different models as the vertical axis, can comprehensively present the judgment shifts between the physical model and the artificial intelligence model throughout the entire operating cycle at a visual or data structure level. Through this mapping, each operating moment of the reverse osmosis unit corresponds to a set of clear offset values, and these offset values ​​are continuously arranged along the time direction, forming a holistic map that reflects the changes in the consistency of model judgments at different operating stages. The formation of the time difference map allows the operating status of the reverse osmosis unit to no longer be based on a single model result, but rather to achieve parallel comparison of physical and intelligent judgments at a unified time scale, clearly demonstrating the differences within the same time scale.

[0029] After the time difference map is generated, the offset trend of each time period in the map is analyzed. The offset fluctuations within continuous time periods are extended based on the time axis, and a complete difference trajectory band is formed by linking each point in the time direction. This difference trajectory band not only reflects the magnitude of the difference in model judgments at each time node, but also shows the trend and persistence of offset changes over time, thereby identifying the offset patterns of the model under different load change stages and different operating rhythm cycles. In this process, the time difference map, as a unified data carrier, continuously expresses the judgment offset relationship between the physical model and the artificial intelligence model across the entire time scale, ensuring that each time scale node corresponds to a specific physical operating state, thus forming a time offset band with full correlation. The existence of this time offset band allows subsequent operating state analysis, rhythm conflict identification, and priority decision-making to be extended based on continuous time difference information, achieving a connection from local judgment to full-time domain coordination.

[0030] Extreme points of the offset segments are extracted along the time difference map to form a rhythm conflict list. The rhythm conflict list is then linked with the load change amplitude of the corresponding time period to identify high-risk decision periods and mark them on the running time axis. Based on the generated time difference map, an in-depth analysis of the offset relationship between the judgment results of the physical model and the artificial intelligence model at different time scales is conducted. Offset segments are identified along the time difference map, and extreme points of offset amplitude changes are extracted. This further forms a rhythm conflict list, which is then linked to the load change amplitude of the reverse osmosis unit within the corresponding time period to identify high-risk decision-making periods and visualize them on the operating timeline. The specific steps are as follows: After the time-difference map is formed, the offset data distributed along the time axis in the map is continuously scanned to identify offset segments where the judgments of the physical model and the artificial intelligence model differ significantly in the time dimension. This scanning process uses time as the main thread, analyzing the offset amplitude at each time scale point one by one, and identifying offset segments with stable extension characteristics by the changing trends of offset values ​​within continuous time periods. At this point, each offset segment corresponds to a time interval, within which the offset direction and amplitude are relatively concentrated, representing the continuous difference between physical and intelligent judgments during a certain operational phase. Through this offset identification along the time axis, a series of discrete offset segments can be formed throughout the entire reverse osmosis operation cycle. These segments constitute the key analytical units of the time-difference map, providing a basic data source for subsequent extreme value extraction.

[0031] After obtaining the offset segments, extreme points reflecting the turning points or peaks of the offset trend are extracted based on the changes in the offset amplitude within each segment. Specifically, along the time direction of each offset segment, the offset amplitude at continuous time scales is compared point by point to identify the time nodes when the offset value reaches its maximum or minimum, and the time position and corresponding offset amplitude of these nodes are recorded. In this way, the extraction of extreme points not only reveals the peak and trough characteristics of offset changes but also characterizes the moment when the difference between the physical model and the artificial intelligence model reaches a critical state. The formation of each extreme point signifies that the model difference at that moment has reached a critical level, which is highly likely to lead to inconsistent judgment tendencies in subsequent operational decisions within that period. Therefore, these extreme points are considered core indicators reflecting the potential risks of operational rhythm conflicts. By extracting extreme points from all offset segments, a time-series-based set of offset extreme points can be formed, covering the key moment distribution of the entire operational timeline.

[0032] After extracting the extreme points of the offset segments, a rhythmic conflict list is constructed using these extreme points as key indices. This rhythmic conflict list forms a structured data sequence by recording the time position, offset magnitude, offset direction, and duration of the offset segment corresponding to each extreme point. In this way, the rhythmic conflict list not only contains offset information at a single time point but also associates the offset persistence and offset evolution trend of the time periods before and after it, thus providing a complete description of the model conflict behavior of the reverse osmosis unit at different operating stages in the time dimension. To give the rhythmic conflict list a relational meaning related to the operating state, the operating time corresponding to each extreme point is associated with the load change magnitude of the reverse osmosis unit during the construction process. This load change magnitude can be reflected by comparing the changes in feed water pressure, permeate flow rate, and high-pressure pump frequency in adjacent time periods before and after the extreme point, giving each rhythmic conflict event a clear operating context. Through this binding relationship, the rhythmic conflict list not only reflects the magnitude of the model's judgment offset but also clarifies the dynamic state of the unit's operating load when the offset occurs, thus giving the conflict information a physical meaning of operating context.

[0033] After the rhythm conflict list is constructed and linked to the load change magnitude, each high-risk conflict event recorded in the list is mapped back to the operating timeline, and high-risk decision periods are marked by their positions on the timeline. In this process, events with large offsets and corresponding load change magnitudes exceeding a set threshold are selected as representative events of high-risk decision periods and arranged chronologically on the operating timeline, forming continuous or discrete marked intervals for each high-risk period. This method visually presents the time periods of concentrated potential conflicts during the entire reverse osmosis unit's operation, enabling maintenance personnel or subsequent decision-making logic to identify time areas where model judgments show significant discrepancies. Each marked high-risk decision period corresponds not only to a specific offset segment but also to a specific stage of operating load change, making the operating timeline a comprehensive mapping carrier connecting model judgment offsets with the physical operating state. This visual marking on the timeline allows subsequent decision reconstruction and timing control processes to be directly based on high-risk periods, thus achieving a logical connection from conflict identification to risk prevention and control.

[0034] Based on the rhythm conflict list, the judgment results of the physical model and the judgment results of the artificial intelligence model are arranged into priority channels according to the degree of risk. During the high-risk decision period, the execution order of the control instructions is reorganized to construct a single decision chain, so that the judgment results of the previous moment are continuously fed into the single decision chain. Based on a rhythmic conflict list, this study analyzes the differences between the judgment results of the physical model and the artificial intelligence model at different risk levels. The two types of judgment results are then ordered according to their risk level. During high-risk decision-making periods, the execution order of control instructions is reorganized to form a temporally continuous and logically unified single decision chain. This ensures that the judgment result of the previous moment naturally continues into the decision-making process of the next moment, thereby achieving temporal coherence of control logic and integrated connection of decision output. The specific steps are as follows: After the rhythmic conflict list is formed, the risk level of each conflict event recorded in the list is analyzed. The risk level of each conflict event is determined based on the comprehensive relationship between the offset magnitude, offset duration, and corresponding load change magnitude. During this process, decision-making periods marked as high-risk are distinguished from ordinary operating periods of medium and low risk on the timeline, giving each time interval a clear risk identification attribute. Subsequently, based on each risk-identified interval, the judgment results of the physical model within that period are extracted, including operational pressure adjustment suggestions, flow balance judgments, and energy consumption trend predictions. Simultaneously, the trend judgment results of the artificial intelligence model at the same time scale are extracted. In this way, judgment outputs from both models are obtained simultaneously at a unified time scale, providing a data foundation for subsequent risk ranking and decision reorganization. The completion of this step ensures that each period on the entire operating timeline has a clear risk label and corresponding dual-model judgment results, thus laying the foundation for establishing priority channels.

[0035] After clarifying the risk levels for each time period, for each high-risk decision-making period, the judgment results of the physical model and the artificial intelligence model are hierarchically arranged according to risk level, forming a risk priority channel. Specifically, within the same time interval, if the rhythm conflict list shows that the offset of that period is large and the duration is long, then that interval is classified as a high-risk channel, and the judgment results reflecting the constraints of physical laws are placed first within the channel to ensure that the control logic first refers to the decision basis with stronger physical characteristics under high-risk conditions. Conversely, in intervals with smaller offsets, the trend judgment results of the artificial intelligence model can be used first to leverage its predictive ability under nonlinear changes. Through this hierarchical arrangement, the decision sources in different risk intervals have a differentiated priority order, thus forming a risk-driven hierarchical structure in the decision-making logic. At the same time, during the channel construction process, the mapping relationship between time sequence and risk level is maintained, so that the judgment results of each time segment can maintain continuity with adjacent time periods in terms of risk level, thereby avoiding sudden jumps in the decision-making logic.

[0036] After the risk priority channel is established, the control instructions for high-risk decision-making periods are reorganized. Multiple control instructions originally output independently by different models are rearranged according to time sequence and risk level, forming a continuous execution sequence. This step uses the time axis as the main thread and risk level as a constraint to merge various control instructions within the same high-risk period. For example, when the physical model suggests reducing the high-pressure pump frequency to balance flux, while the artificial intelligence model simultaneously suggests delaying the cleaning cycle to maintain operational stability, the execution order of the two is reorganized so that the high-pressure pump frequency adjustment instruction takes priority, while the judgment result of delaying cleaning is executed in the next time scale, thus forming a logically coherent and sequentially reasonable decision chain. In this process, the time position and execution order of each control instruction are constrained by the judgment result of the previous moment, making the entire control decision present a continuous logical trajectory in time. Through this reorganization method, not only is the control instruction rearranged in the time dimension, but the judgment results output by different models are also ensured to have a smooth connection in execution logic, transforming the operational decision-making in high-risk periods from a previously dispersed output into a single, orderly decision-making process.

[0037] After the execution sequence of control commands is reorganized, the reorganized continuous control logic is mapped onto the operating timeline, forming a complete single decision chain. This single decision chain, guided by a time scale, connects control commands for each high-risk period according to temporal continuity and risk priority, allowing the judgment result of the previous moment to naturally transition to the execution process of the next moment, thus constructing a cross-time-extended decision structure. In this way, the single decision chain is no longer limited to single-point control, but forms a continuous decision logic throughout the entire operating cycle, achieving temporal fusion and logical integration of physical model judgments and artificial intelligence model judgments. This single decision chain maintains a correspondence with the rhythm conflict list at each time scale, ensuring that each decision output matches the risk level and operating status, thereby eliminating decision conflicts and duplicate execution problems that easily occur under traditional independent model outputs. Through continuous mapping on the timeline, the single decision chain forms a logical closed loop from past judgments naturally continuing to the current decision and continuously influencing subsequent periods, making the operating commands of the reverse osmosis unit exhibit temporal continuity, logical stability, and unified execution characteristics under dynamic load changes.

[0038] A ripple-like time control loop is set up around a single decision chain to make minor adjustments to the effective time of control commands during high-risk decision periods, insert short buffer windows and smooth rhythm fluctuations, so that the single decision chain maintains a continuous output trajectory on the time axis to avoid operation and maintenance command conflicts under dynamic load changes. Based on the established single decision chain, a ripple-like time control loop with buffering and extension characteristics is introduced on the time axis to slightly adjust the effective time of control commands during high-risk decision periods. Furthermore, a time buffer band with brief delay characteristics is inserted during continuous control to achieve a smooth transition of operating commands in the time series. This ensures that the entire single decision chain maintains a continuous and stable output trajectory under dynamic load changes in the reverse osmosis unit, thereby avoiding abrupt changes in control logic or decision conflicts. The specific steps are as follows: After a single decision chain is formed, the entire operational timeline covered by the chain is continuously scanned to identify high-risk decision periods. Using the time intervals corresponding to the rhythm conflict list as a reference, the start and end boundaries of these high-risk periods are distinguished from the normal operation segments. Within the identified high-risk periods, the control instructions to be executed in the single decision chain and their effective times are extracted, and the time intervals, execution sequences, and scopes of action between adjacent instructions are analyzed. At this point, for periods with short time intervals or insufficient execution sequence continuity, time adjustment intervals are established, providing a positional basis for subsequent effective time adjustments and buffer settings. In this way, the time layout of the entire decision chain has an operable time window, providing spatial conditions for the insertion of subsequent control loops. This step, by identifying decision segments with high time concentration, defines "time fluctuation source areas" for the subsequent adjustment process, thereby clarifying the specific location range where fine-tuning needs to be implemented.

[0039] Based on clearly defined high-risk periods and their corresponding control commands, a ripple-like time control loop with progressive extension characteristics is established around these time segments. This control loop extends bidirectionally forward and backward from the command activation point, with time as its core axis. By constructing a continuous buffer zone with time decay characteristics, it allows control commands to have a certain flexible time span before and after activation. Specifically, a slightly advanced response zone is set at the leading edge of the control loop, allowing commands about to take effect to enter the pre-response phase before formal triggering, thus providing time for state transition within the system. A delay buffer is set at the trailing edge of the control loop, allowing the execution effect of the command to gradually dissipate over time after activation, avoiding instantaneous superposition of command effects that could cause operational fluctuations. Through this bidirectional time ripple design, the activation process of each control command changes from instantaneous triggering to gradual connection, creating a flexible transition layer in the time dimension for continuous control behaviors, thereby reducing direct conflicts between control commands during high-risk decision-making periods.

[0040] After setting up the ripple-like time control loop, the time structure within the loop is refined. The effective times of different control commands are rearranged within the control loop according to their risk level and execution priority. High-risk control commands are located in the central area of ​​the control loop, while low-risk control commands are located in the outer area, thus creating an execution rhythm that gradually spreads from the core to the periphery. At this time, by making minor adjustments to the time interval between adjacent control commands, high-risk decision events are given a larger time buffer, while low-risk events are executed with a shorter delay in the outer ring. Through this time-layered structure, a hierarchical time sequence is formed between different control commands, avoiding operational fluctuations caused by simultaneous triggering. Simultaneously, a time connection zone is set in the outer area of ​​the control loop, allowing the end time of the control loop to naturally connect to the pre-response zone of the next command, achieving a time transition between continuous rings and ensuring the output trajectory of the entire single decision chain remains coherent on the time axis. This step, through time-differentiated configuration, achieves "time-layered execution" between commands, which manifests as a gradual and smooth transition of the device's operating state in physical operation.

[0041] After the time structure of the ripple-like time control loop is determined, the adjusted command activation time and buffer window are mapped back to the single decision chain. This causes each control command to form a time action band with extended boundaries on the time axis, thus constituting a continuous time output trajectory. At this point, each decision node in the single decision chain no longer appears as a discrete time point, but rather as a strip of time slices distributed on the operating time axis. Multiple adjacent time slices are naturally connected through buffers, making the entire decision chain exhibit a continuous fluctuation pattern in output. Through this temporal connection, the reverse osmosis unit can maintain a stable operating rhythm under complex dynamic conditions such as load changes, flow fluctuations, and pressure disturbances during execution. Each control command execution is accompanied by a brief buffer transition period, thus avoiding operational conflicts and system oscillations caused by the superposition of multiple commands. Furthermore, through the repeated nesting of the time control loop, the overall output of the decision chain forms a periodic and smooth ripple structure in the time dimension, making the trajectory of the unit's operating state exhibit continuous and balanced time distribution characteristics. In this way, the entire single decision chain no longer responds to external disturbances in a sudden manner during high-risk periods, but maintains continuous output through slight extension and time transition, thereby achieving stable control of operation and maintenance instructions in the time domain.

[0042] This invention constructs a reference time band for operational rhythm and achieves point-to-point alignment between the physical model and the artificial intelligence model under a unified time scale. This enables the two types of models to form a synchronous correlation in the time dimension, thereby ensuring that operational judgments from different sources can be compared and fused with a consistent time benchmark. This method allows the operating status of the reverse osmosis device under dynamic load conditions to be continuously expressed along the entire timeline, avoiding judgment deviations caused by time misalignment between model outputs. It also ensures stable temporal consistency and traceability in the generation process of control commands, thereby improving the coordination of system operation and the reliability of data-driven decisions.

[0043] This invention constructs a single decision chain and sets up a ripple-like time control loop during high-risk decision-making periods, transforming the execution of control commands from instantaneous to continuous and smooth output, forming a time transition zone with buffering characteristics. This approach effectively reduces execution conflicts and oscillations caused by the superposition of multiple commands, enabling the reverse osmosis system to maintain a smooth transition of its operating rhythm when the load changes or operating conditions fluctuate. It ensures the continuity of operation and maintenance decisions over time and the flexibility and consistency of control response, thereby improving the overall operational stability and intelligent control capabilities of the system.

[0044] This invention provides, for example Figure 2 The intelligent operation and maintenance system for reverse osmosis systems based on a physics-AI hybrid approach, as shown, includes a time baseline construction module, a time series alignment analysis module, a conflict identification and extraction module, a decision chain construction module, and a time series control and smoothing module. The time reference construction module collects data on the changes in high-pressure pump frequency, inlet water pressure, and product water flow rate over time during the operation of the reverse osmosis system. It maps all the operating data to the same time scale and establishes a reference time band for the operating rhythm of the entire reverse osmosis unit, which is used to provide a unified time reference for subsequent data comparison. The time-series alignment analysis module, based on the reference time band of the running rhythm, aligns the real-time running data output by the physical model with the trend prediction results output by the artificial intelligence model point by point according to the time scale, and generates a time difference map to characterize the offset relationship between the judgment results of different models within the same time scale. The conflict identification and extraction module extracts the extreme points of the offset segments along the time difference map to form a rhythm conflict list. It then binds the rhythm conflict list with the load change amplitude of the corresponding time period to identify high-risk decision periods and mark them on the running time axis. The decision chain construction module, based on the rhythm conflict list, arranges the judgment results of the physical model and the judgment results of the artificial intelligence model according to the risk level as priority channels, and reorganizes the execution order of control instructions during high-risk decision periods to construct a single decision chain, so that the judgment results of the previous moment are continuously fed into the single decision chain. The timing control and smoothing module sets up a ripple-like timing control loop around a single decision chain. During high-risk decision periods, it makes minor adjustments to the effective time of control commands, inserts short buffer windows, and smooths rhythm fluctuations, so that the single decision chain maintains a continuous output trajectory on the time axis, thereby avoiding operational command conflicts under dynamic load changes.

[0045] The intelligent operation and maintenance method for reverse osmosis systems based on physical-AI hybrid drive provided in this embodiment of the invention is implemented through the aforementioned intelligent operation and maintenance system for reverse osmosis systems based on physical-AI hybrid drive. For details of the specific methods and processes of the intelligent operation and maintenance system for reverse osmosis systems based on physical-AI hybrid drive, please refer to the embodiments of the intelligent operation and maintenance method for reverse osmosis systems based on physical-AI hybrid drive, which will not be repeated here.

[0046] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent operation and maintenance of reverse osmosis systems based on a physical-AI hybrid approach, characterized in that, Includes the following steps: Data on the changes in high-pressure pump frequency, inlet water pressure, and product water flow rate over time during the operation of the reverse osmosis system are collected. All operating data are mapped to the same time scale to establish a reference time band for the operating rhythm of the entire reverse osmosis unit. Based on the reference time band of the operating rhythm, the real-time operating data output by the physical model and the trend prediction results output by the artificial intelligence model are aligned point by point according to the time scale to generate a time difference map. Extreme points of the offset segments are extracted along the time difference map to form a rhythm conflict list, and the rhythm conflict list is linked with the load change amplitude of the corresponding time period. Based on the rhythm conflict list, the judgment results of the physical model and the judgment results of the artificial intelligence model are arranged into priority channels according to the degree of risk, and the execution order of control instructions is reorganized during high-risk decision periods to construct a single decision chain. A ripple-like time control loop is set up around a single decision chain to adjust the effective time of control commands during high-risk decision periods, insert short buffer windows and smooth rhythm fluctuations, so that the single decision chain maintains a continuous output trajectory on the time axis.

2. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive as described in claim 1, characterized in that, The steps for generating the rhythm reference time band are as follows: During the operation of the reverse osmosis unit, real-time changes in high-pressure pump frequency, inlet water pressure, and product water flow rate are continuously collected and arranged in the order of sampling time to form an original time series. Using the sampling time series of the high-pressure pump frequency as the reference main time axis, the time records of inlet pressure and product water flow are synchronously extended, and the operating parameters from different sources are redistributed to a unified time node. The values ​​of high-pressure pump frequency, inlet pressure and product water flow rate after time mapping are combined point by point in time sequence to generate a reference time band for the operating rhythm covering the entire device. Using the operating rhythm reference time zone as a unified benchmark, the time nodes and changing trends of each operating parameter are recorded in the same time frame, forming an information structure containing a complete time index.

3. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 2, characterized in that, In the process of generating the reference time zone for the operating rhythm, the time series of the high-pressure pump frequency is used as the core reference. By extending and weighting the time records of the inlet pressure and the product flow rate, the operating parameters form a continuous numerical sequence at a unified time node, and the operating rhythm reference time zone is established in chronological order.

4. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 2, characterized in that, The steps for point-by-point alignment of the physical model output and the artificial intelligence model output based on the runtime rhythm reference time band are as follows: The high-pressure pump frequency, inlet pressure, and product water flow rate are selected as alignment benchmarks. The real-time operating data output by the physical model and the trend prediction results output by the artificial intelligence model are matched one by one according to the time nodes of the operating rhythm reference time zone. The physical model output and the artificial intelligence model output at each time point are analyzed to extract the values ​​of the running variables and form a continuous difference dataset. The difference dataset is arranged in chronological order to form a time difference sequence, and the time difference sequence is mapped onto the running rhythm reference time band to generate a time difference map; Extended analysis of the offset trend in the time difference plot is performed to form a time offset band that reflects the trend of difference changes in the model's judgment.

5. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 4, characterized in that, In the process of generating the time difference map, the difference values ​​of adjacent time nodes in the time difference sequence are connected by a continuous time mapping method, so that the time difference map forms a continuously distributed difference trajectory band on the time axis, and the extreme points of the offset amplitude change are used as key feature points.

6. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 4, characterized in that, The steps for extracting extreme points of the offset segments along the time difference map and generating a rhythm conflict list are as follows: The offset data distributed along the time axis of the time difference map are continuously scanned to identify offset segments where the judgment results of the physical model and the artificial intelligence model differ. Extract the extreme points of the shift amplitude change along the time direction of each shift segment, and record the time position of the extreme points and the corresponding shift amplitude; A rhythm conflict list is constructed using extreme points as key indexes, and the time position, offset amplitude, offset direction and duration corresponding to the extreme points are recorded as a structured data sequence. The rhythm conflict list is linked to the load change range of the reverse osmosis unit within the corresponding time period to identify high-risk decision-making periods and mark them on the operating timeline.

7. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 6, characterized in that, The steps of linking the rhythm conflict list with the load change range are further defined as follows: by comparing the changes in influent pressure, permeate flow rate and high-pressure pump frequency in adjacent time periods before and after the extreme point, the correlation between the load change range and the offset range is determined, and this is used as the basis for determining high-risk decision periods.

8. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 6, characterized in that, The steps to construct a single decision chain based on a rhythm conflict list are as follows: Risk level analysis is performed on the conflict events recorded in the rhythm conflict list. The risk level is determined based on the offset amplitude, duration and load change amplitude, and the judgment results of the physical model and artificial intelligence model are extracted. The results of the physical model assessment and the results of the artificial intelligence model are arranged in a hierarchical manner according to the risk level, forming a risk priority channel, and maintaining the mapping relationship between the time sequence and the risk level. During high-risk decision-making periods, control instructions are reorganized, and control instructions from different models are arranged into a continuous execution sequence according to time order and risk level. The reorganized control logic is mapped onto the runtime timeline, forming a single decision chain that is time-continuous and logically unified, so that the judgment results of the previous moment are continuously incorporated into the decision process of the next moment.

9. The intelligent operation and maintenance method for reverse osmosis systems based on a physical-AI hybrid drive according to claim 8, characterized in that, The steps for setting up a ripple-like time control loop around a single decision chain are as follows: Continuous scanning of the operational timeline covered by a single decision chain identifies time intervals containing high-risk decision-making periods and establishes time control intervals. A ripple-like time control loop is set up around the high-risk decision-making period to form a buffer zone with time decay characteristics before and after the control command takes effect, so that the command execution process has a flexible time span. Within the ripple-like time control loop, the effective time of control commands is reallocated based on risk level and execution priority, forming a time-layered structure that diffuses from the core to the periphery; The adjusted instruction effective time and buffer window are mapped back to the single decision chain, so that the control instructions form a continuous output trajectory on the time axis and maintain a stable operating rhythm.

10. A reverse osmosis system intelligent operation and maintenance system based on a physical-AI hybrid drive, used to implement the intelligent operation and maintenance method for a reverse osmosis system based on a physical-AI hybrid drive as described in any one of claims 1-9, characterized in that, It includes a time baseline construction module, a time series alignment analysis module, a conflict identification and extraction module, a decision chain construction module, and a time series regulation and smoothing module: The time reference construction module collects data on the changes in high-pressure pump frequency, inlet water pressure, and product water flow rate over time during the operation of the reverse osmosis system, maps all operating data to the same time scale, and establishes a reference time band for the operating rhythm covering the entire reverse osmosis unit. The time-series alignment analysis module, based on the reference time band of the running rhythm, aligns the real-time running data output by the physical model with the trend prediction results output by the artificial intelligence model point by point according to the time scale, and generates a time difference map. The conflict identification and extraction module extracts the extreme points of the offset segments along the time difference map to form a rhythm conflict list, and binds the rhythm conflict list with the load change amplitude of the corresponding time period. The decision chain construction module, based on the rhythm conflict list, prioritizes the judgment results of the physical model and the judgment results of the artificial intelligence model according to the degree of risk, and reorganizes the execution order of control instructions during high-risk decision periods to construct a single decision chain. The timing control and smoothing module sets up a ripple-like timing control loop around a single decision chain. During high-risk decision periods, it makes minor adjustments to the effective time of control commands, inserts short buffer windows, and smooths rhythm fluctuations, so that the single decision chain maintains a continuous output trajectory on the time axis.

Images