A remote calibration method and system for a smart water station

Abstract

The present application relates to the field of remote calibration technology, and in particular to a remote calibration method and system for a smart water station. The present application proposes the following scheme: by acquiring the status information of each liquid path, constructing a path state matrix, and calculating the pollution risk response value based on the calibration task, a pollution adaptability matrix is ​​generated, from which the most compatible path is screened to perform the calibration task; the modeling of the path state introduces a state evolution graph and a graph neural network deduction mechanism, and uses interrupt edges to control the propagation path, dynamically reflecting the impact of the cleaning operation on the state inheritance; after the task is completed, the operation information is written into the log and backfilled into the state graph to realize the continuous evolution and update of the path state.

Description

Technical Field

[0001] The present application relates to the field of remote calibration technology, and in particular to a remote calibration method and system for a smart water station. Background Art

[0002] With the increasing demand for water environment monitoring, smart water stations are widely deployed in scenarios such as watershed management, outlet supervision, and drinking water sources to achieve real-time online detection of multiple water quality indicators. In such water quality analysis systems, in order to ensure detection accuracy, it is often necessary to periodically introduce standard liquids or blind samples for automatic calibration. Since detection equipment usually adopts a multiple liquid path reuse design, including standard liquids, cleaning liquids, and blind sample liquids, which are all delivered to a shared measurement cavity after being switched and controlled by multiple valves, there are problems of liquid residue and cross-contamination when the paths are reused between different calibration tasks.

[0003] The existing technology has the problems raised in this background technology: in actual operation, due to factors such as differences in pipeline structure, valve control lag, and pump flow fluctuations, there are differences in cleaning effect. Even if some paths are cleaned, there may still be non-negligible residual effects. The lack of a dynamic and evolvable path state modeling mechanism can easily lead to inaccurate path judgment, resource waste, or calibration errors. To solve the above problems, this application designs a remote calibration method and system for smart water stations. Summary of the Invention

[0004] The technical problem to be solved by this application is to address the shortcomings of the existing technology and provide a remote calibration method and system for a smart water station. The method obtains the status information of each liquid path, constructs a path state matrix, calculates the pollution risk response value based on the calibration task, generates a pollution adaptability matrix, and selects the most compatible path to perform the calibration task. The modeling of the path state introduces the state evolution graph and the graph neural network deduction mechanism, and uses the interrupt edge to control the propagation path, dynamically reflecting the impact of the cleaning operation on the state inheritance. After the task is completed, the operation information is written to the log and backfilled into the state graph to achieve the continuous evolution and update of the path state.

[0005] To achieve the above objectives, this application provides the following technical solutions:

[0006] A remote calibration method for a smart water station is applied to a water quality analysis system having multiple liquid paths, wherein the liquid paths include a control valve corresponding to the calibration liquid, a sampling pump, and a liquid pipeline for conveying the calibration liquid to a measurement chamber. The method includes:

[0007] Acquiring state information of each liquid path and constructing a path state matrix, wherein the constructed path state matrix derives the state information according to a preset state update rule, the state update rule including interruption behavior modeling logic for simulating the impact of a cleaning operation on resetting the state information;

[0008] According to the calibration task, the contamination risk response value of each liquid path to the target calibration task is calculated through the path state matrix to generate a contamination adaptability matrix;

[0009] According to the pollution adaptability matrix, an optimal compatibility path is obtained, and a calibration task is performed through the optimal compatibility path.

[0010] The construction path state matrix includes:

[0011] Acquiring status information of each liquid path, the status information including the type of liquid last injected, whether a cleaning operation is performed, and the time interval since the last operation;

[0012] Deducing the state information according to a preset state update rule to generate a path state corresponding to the liquid path;

[0013] A current path state matrix is ​​constructed according to the path state.

[0014] The deducing the state information according to a preset state update rule to generate a path state corresponding to the liquid path includes:

[0015] The last injected liquid type of each liquid path is used as the initial node, and the time interval from the last operation is mapped to the evolution step to construct the path state evolution graph;

[0016] On the path state evolution graph, based on historical operation information, state deduction calculation is performed through the graph neural network to output the state evolution result.

[0017] The state deduction calculation is performed through the graph neural network, and the state evolution result is output, including:

[0018] According to the historical operation information, a semantic additional node is added to the original node structure of the path state evolution graph, wherein the semantic additional node includes operation frequency, path idle time, historical cleaning density, and liquid type switching amplitude;

[0019] Calculate the node sequence of the path state evolution graph, establish directed edges between adjacent nodes in the node sequence, and generate a time series graph, wherein if there is a cleaning operation between adjacent nodes, the corresponding directed edge is marked as a state interruption edge, and a state interruption identifier and a decay weight parameter are assigned;

[0020] The time series graph is used as the input of the graph neural network, and the state vector of the last node in the time series graph is extracted through preset rounds of propagation and updating as the state evolution result, wherein the graph neural network includes a multi-layer graph convolution unit and a state update unit.

[0021] The propagation and updating through the preset rounds include:

[0022] In each round of propagation, the state information of each node in the time series graph is aggregated through the graph convolution unit, where a node receives a state vector from its previous node. The state information aggregation is weighted based on the type and edge attributes of the directed edge. When a directed edge is marked as a state-interrupted edge, the directional information transmission is disabled according to the state interruption flag during the propagation process, and the incoming information strength is adjusted according to the corresponding attenuation weight parameter;

[0023] After completing information aggregation, the node is input into the state update unit, and the current node state is updated through a nonlinear conversion function. When a state interruption edge exists on the node, the state update unit blocks the inheritance of the historical state residual according to the state interruption flag.

[0024] The contamination risk response value of each liquid path to the target calibration task is calculated using the path state matrix to generate a contamination adaptability matrix, including:

[0025] Extracting the current state vector of each liquid path in the path state matrix;

[0026] For each liquid path, the contamination risk mapping function is called, with the current state vector as input and the target liquid corresponding to the calibration task as a parameter, to calculate the contamination risk response value of the corresponding liquid path relative to the target liquid;

[0027] The contamination risk response values ​​of the liquid paths were arranged according to the path numbers to construct a contamination adaptability matrix.

[0028] The pollution risk mapping function includes state compatibility matching logic, where:

[0029] The state compatibility matching logic calculates a probability score for residual liquid interference in the liquid path in the current state based on the chemical interference level, cleaning behavior interruption frequency, and time interval indicators between the liquid type evolution trajectory recorded in the liquid path and the target liquid.

[0030] According to the pollution adaptability matrix, the optimal compatibility path is obtained, including:

[0031] Sorting the pollution risk response values ​​corresponding to the liquid paths in the pollution adaptability matrix to determine a path sequence from low to high pollution risk;

[0032] Performing availability checks on the paths in the path sequence in sequence, wherein the availability checks include: whether the paths are in operation, whether the minimum idle time has been reached, and whether there is a fault mark;

[0033] Among the liquid paths that have passed the availability verification, the liquid path with the lowest contamination risk response value is selected as the path with the best compatibility.

[0034] After the calibration task is completed, the method further includes:

[0035] Write the usage information of the liquid path used to perform the calibration task into the path operation log, and update the type of the most recently injected liquid, whether the cleaning operation was performed, and the completion time;

[0036] According to the updated content in the path operation log, the node information of the corresponding liquid path in the path state evolution map is calculated, and the calculation result is updated to the path state matrix.

[0037] A remote calibration system for a smart water station, the system comprising:

[0038] A status acquisition module is used to obtain status information of each liquid path, including the type of liquid injected last time, whether a cleaning operation is performed, and the time interval since the last operation;

[0039] a state deduction module that derives the state information according to preset state update rules to generate a path state of the liquid path, wherein the state update rules include interruption behavior modeling logic for simulating the impact of cleaning operations on the reset of the path state, and performs state deduction calculations through a graph neural network;

[0040] The risk calculation module calculates the contamination risk response value of each liquid path based on the path state matrix and the target liquid of the calibration task, and generates a contamination adaptability matrix;

[0041] The path scheduling module determines the most compatible path according to the pollution adaptability matrix, controls the path to execute a calibration task, and updates the path operation log and the path status matrix after the calibration task is completed.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] This application constructs a liquid path state evolution map and introduces a graph neural network for state deduction, achieving dynamic modeling and precise expression of the residual state of the liquid path, which can effectively reflect the actual impact of cleaning behavior on path state inheritance. Compared with the traditional method of judging path availability with Boolean logic, this application provides a path state representation mechanism with strong continuity, high expressiveness, and evolutionary capabilities. By combining the contamination risk mapping function with the task target liquid, the state compatibility matching between the path and the task is further achieved, which improves the pertinence and safety of path scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0045] Figure 1 This is a schematic diagram of an exemplary application scenario of an embodiment of the present application;

[0046] Figure 2 This is a flow chart of a remote calibration method for a smart water station according to an embodiment of the present application;

[0047] Figure 3 A schematic diagram of the process of constructing the path state matrix according to an embodiment of the present application;

[0048] Figure 4 This is a schematic diagram of the principle of state information derivation according to an embodiment of the present application;

[0049] Figure 5 This is a flow chart of the path matrix updating method according to an embodiment of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0051] References to "embodiments" herein mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of a phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It will be understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0052] The embodiments of this application are applicable to water quality analysis systems with a shared structure featuring multiple liquid pathways and a fixed, non-switchable measurement chamber. Typical application scenarios include multi-parameter detection equipment and automated online calibration devices, where multiple calibration liquids, blind sample solutions, cleaning solutions, and other liquids are reused through valve control using the same fluid path and injection assembly, ultimately entering the same measurement chamber.

[0053] Understandably, since the measurement cavity is shared by the entire system, the cleaning process encompasses not only the measurement cavity itself but also the flow of cleaning fluid through the path. Therefore, the cleaning effect is theoretically not perceptible in real time and is subject to control strategies, fluid resource constraints, or task scheduling policies. Some path switching operations may not complete the cleaning process, or system policies may allow for continuous reuse of paths to conserve cleaning resources.

[0054] It's important to note that under these conditions, the actual state of a path cannot be accurately determined by a single cleaning flag, nor can it be statically identified through structural partitioning or path numbering. Because cleaning is essentially a logical process event, its execution depends on dynamic factors such as liquid flow conditions, operation duration, and valve control accuracy. Traditional Boolean state management methods based on process records cannot effectively capture the actual evolution of residual interference on the path.

[0055] See also Figure 1 , Figure 1 A schematic diagram of an exemplary application scenario provided for an embodiment of the present application.

[0056] Figure 1 The control valve of the present application is shown, including a cleaning valve, a standard liquid valve and a blind sample valve. The standard liquid valve in an optional embodiment includes a standard liquid valve 1, a standard liquid valve 2 and a standard liquid valve 3. Each valve is connected to a corresponding liquid source. The liquid enters the sampling pump through the liquid pipeline and is then injected into the measuring chamber by the sampling pump to complete the detection and calibration of the target parameters. Figure 1 A liquid displacement pump is also shown for liquid displacement operations.

[0057] It can be understood that the structural feature of the present application is that after all liquid paths are dispatched by the control valve, they share the same set of liquid pipelines and measurement chambers to achieve sequential injection and reuse of different calibration liquids.

[0058] Those skilled in the art will appreciate that, for resource optimization purposes, cleaning operations are not mandatory for every path switch. During cleaning, uncertainties such as short cleaning times, low flow rates, and valve control response deviations can occur, leading to unstable cleaning results. If cleaning is inadequate, residual liquid from the previous path may enter the measurement chamber, causing cumulative errors in subsequent calibrations.

[0059] In one optional embodiment, the present application is applicable to a water quality testing device with automatic calibration of five parameters, including pH, conductivity, dissolved oxygen, turbidity, and temperature. This embodiment uses pH as an example, and other parameters can be implemented using the same control logic.

[0060] For reference Figure 1 The water quality testing device includes multiple control valves corresponding to the calibration liquids. These control valves are connected to a sample pump via fluidic piping, which delivers the liquids to the measurement chamber for testing. The device is configured as a typical shared liquid path architecture, with sequential injection and switching between different parameters or calibration liquids using fluidic multiplexing.

[0061] In traditional control logic, calibration processes, such as blind sample calibration and standard solution 1 calibration, follow a sequence of draining, filling, measuring, and purging steps. For example, a pH low-standard calibration typically involves: turning on the drain pump to drain the foreline solution → filling standard solution 1 → checking the level → stabilizing the reading → filling standard solution 2 → purging → ending. Multiple steps in this process rely on a combination of delay settings, level determination signal feedback, and valve and pump action logic.

[0062] It is understandable that in the application environment concerned by this application, the cleaning operation has the following characteristics:

[0063] Cleaning is not mandatory every time, but is determined by the control strategy. Some paths may skip cleaning due to scheduling efficiency requirements;

[0064] The effectiveness of cleaning behavior cannot be perceived in real time;

[0065] Insufficient cleaning will result in liquid residue that cannot be removed, affecting subsequent calibration.

[0066] Therefore, it is logically insufficient to use only whether cleaning is performed as a criterion for path availability. The method of this application records the historical operation sequence of each liquid path, including information such as the type of liquid injected last, whether a cleaning operation was performed, and the time interval since the last operation, to construct a path state evolution map.

[0067] Furthermore, in the path state evolution graph, state interrupt edges are introduced as key modeling elements.

[0068] It's understandable that when a purge occurs in an operation sequence, state-breaking edges are established between adjacent operation nodes, assigned interruption flags and weight coefficients to control the state truncation effect of the purge during graph propagation. The presence of state-breaking edges enables the graph neural network to dynamically adjust the direction of edge information flow, propagation strength, and historical residual inheritance strategy during the state propagation phase, thereby simulating the logical reset effect of purges on path states.

[0069] Next, in conjunction with the accompanying drawings, a remote calibration method for a smart water station provided by an embodiment of the present application is introduced. The method is applied to a water quality analysis system having multiple liquid paths, wherein the liquid paths include a control valve corresponding to the calibration liquid, a sampling pump, and a liquid path pipeline for transporting the calibration liquid to the measurement chamber. Figure 2 The method shown includes the following steps S1-S3, and the specific steps are as follows:

[0070] S1: Obtain the status information of each liquid path and construct a path state matrix;

[0071] In this embodiment, the path status information is obtained from the operation control data, and the data content includes at least: the type of liquid injected last time, whether a cleaning operation was performed, and the time interval since the last operation.

[0072] It is understandable that state information has the advantages of clear operation sources and low collection costs, and can all be obtained through historical operation instructions recorded by the logic control system. Based on this state information, the state of each path is modeled and derived using preset state update rules. The update rules introduce interruption behavior modeling logic to model and abstract the path state reset effect caused by cleaning behavior, and use it as a regulatory factor in the information propagation mechanism in subsequent path state deduction. Each path state ultimately formed by the path state matrix is ​​a comprehensive state expression that includes time, behavior, and structure dimensions.

[0073] S2: According to the calibration task, the contamination risk response value of each liquid path to the target calibration task is calculated through the path state matrix to generate a contamination adaptability matrix;

[0074] In this embodiment, the calculation of the contamination risk response value is based on the aforementioned path state vector and the target liquid type required to be injected for the target calibration task, and is performed by the contamination risk mapping function. The core logic of the function is state compatibility matching. The matching relationship is determined by factors such as the chemical interference relationship between the liquid type change trajectory recorded in the path and the current target liquid, the interruption frequency of the cleaning behavior, and the length of time the path is idle. The contamination risk response value reflects the potential interference risk of the current path to the target liquid. The higher the response value, the greater the residual interference impact of the path. The contamination response values ​​of all paths are summarized to form the contamination adaptability matrix.

[0075] S3: Obtaining an optimal compatibility path according to the pollution adaptability matrix, and performing a calibration task through the optimal compatibility path;

[0076] In this embodiment, the optimal compatibility path is selected based on the lowest response value in the pollution adaptability matrix. During actual path scheduling, after sorting the response values, the path's current availability is also verified, including control constraints such as whether it is operational, whether it meets the minimum idle time, and whether there are logical fault flags. When multiple paths have similar response values, the optimal path is dynamically selected based on factors such as the calibration task priority weight and the current system scheduling load.

[0077] Before developing the specific technical content corresponding to the steps, the embodiments of this application need to be emphasized again.

[0078] Smart water stations are often used in watersheds or water source locations, requiring high-frequency online monitoring and calibration of multiple water quality indicators. Because multiple liquids (such as calibration solutions, blind samples, and cleaning solutions) share the same fluid path and measurement chamber, residual liquid in the path can lead to contamination and errors if not completely cleaned. Therefore, a refined and intelligent mechanism for path selection and status management is required.

[0079] This example does not attempt to directly model the residual liquid state, but rather transforms the problem into a path behavior evolution modeling problem. By constructing a state transition graph for a liquid path between different historical operation stages and introducing state interruption edges representing cleaning behavior into the graph structure as logical nodes for behavior reset, a graph neural network inference path with memory and truncation mechanisms is established.

[0080] The logic employed in this embodiment does not rely on observable states or a pre-set rule base. Instead, it infers path states from a causal perspective through the path's own behavioral trajectories, event types, and temporal evolution. State interruptions do not indicate whether a path has been cleaned. Instead, they serve as structural breakpoints in the information propagation process, controlling the propagation of node states during transmission, aggregation, and residual inheritance. This allows the graph neural network to express the behavioral semantics of how a cleansing can cause the previous state to be neither fully inherited nor completely discarded.

[0081] Next, the part of the method of the present application regarding constructing the path state matrix is ​​further expanded.

[0082] See Figure 3 , Figure 3 A schematic diagram of the process of constructing the path state matrix in an embodiment of the present application.

[0083] In an example, the specific steps of S1 are as follows:

[0084] S1.1: Obtaining status information for each liquid path, including the type of liquid last injected, whether a cleaning operation is performed, and the time interval since the last operation;

[0085] S1.2: Deducing the state information according to a preset state update rule to generate a path state corresponding to the liquid path;

[0086] Understandably, the actual effectiveness of the cleaning action is affected by factors such as valve response delay, pump pressure change, and path flow resistance differences. Whether or not a cleaning action has been performed before is not logically sufficient for a path to be safe and usable, and therefore cannot be represented using a Boolean state. This application establishes a state derivation mechanism that can compensate for the unobservable path state through logical reasoning without relying on sensor monitoring.

[0087] Specifically, the derivation of liquid path state is an evolutionary modeling method established under the premise that cleaning behavior is unreliable and the path cannot be observed in real time. In this embodiment, the path state is not directly equivalent to whether cleaning is currently in progress. Instead, it represents an estimate of the residual state of the liquid path driven by a series of historical operations. This includes the evolution trajectory of historical liquid types, the impact of cleaning behavior on path memory, and the correction coefficient for the residual weakening effect of time intervals.

[0088] Taking status update as an example, you can refer to Figure 4 To understand, Figure 4 This is a schematic diagram of the principle of state information derivation in an embodiment of the present application. Figure 4 The diagram shows the nodes composed of the last injection operation. Each node represents an injection operation performed on the liquid path at different times, including standard liquid 1 injection, standard liquid 2 injection, blind sample injection, and cleaning liquid injection. The nodes are connected by directed edges, indicating the state evolution relationship in time sequence.

[0089] Figure 4 It shows that there is no injection of cleaning liquid between the injection of standard liquid 2 and the injection of blind sample. That is, the edge marked with a dotted line in the figure represents a state interruption edge, indicating that there is an interruption behavior in the propagation of the path state.

[0090] Figure 4 Five moments from t1 to t5 are shown, wherein the boundary between t1 and t2 and between t4 and t5 is the injection of cleaning liquid, and the specific lengths of the corresponding lines can be regarded as time intervals.

[0091] Figure 4 It is further shown that on the basis of the state node path structure, historical operation information nodes are also included as semantic control inputs, the content of which includes but is not limited to the behavioral characteristics of operation frequency, cleaning density, path vacancy time and liquid switching amplitude, which are input into the graph neural network module in the form of structured attributes and combined with the state path diagram for state deduction.

[0092] Figure 4It further shows that the graph neural network takes the graph structure of the current path as input, combines the interrupt edge control mechanism to perform multiple rounds of state propagation and node state updates, and finally outputs the state evolution result representing the current path state.

[0093] In one example, the deducing the state information according to a preset state update rule to generate a path state corresponding to the liquid path includes:

[0094] S1.2.1: Take the last injected liquid type of each liquid path as the initial node, map the time interval from the last operation to the evolution step, and construct the path state evolution graph;

[0095] Specifically, this step aims to construct a structured graph model representing the historical operational behavior of the liquid path and its temporal evolution, enabling reasonable prediction and modeling of the path's current state. Because liquid paths lack real-time status monitoring capabilities, the traditional approach of relying on a single Boolean flag (cleaning or not) to determine path reliability fails to capture the evolving characteristics of actual residual conditions. Therefore, it is necessary to model the state evolution process within the path's behavioral chain itself.

[0096] In this embodiment, each liquid path is considered as a state evolution process. Its state node is first composed of the information of the last operation. Each injection event is abstracted as a graph node. The node attributes include but are not limited to:

[0097] Control parameters such as the type of injected liquid, whether cleaning is performed, operation timestamp, task type corresponding to the operation, and whether liquid level stability is achieved.

[0098] Furthermore, nodes are connected by directed edges, which represent the temporal order and carry an evolutionary step attribute. The evolutionary step is not a fixed setting; instead, the time interval between two adjacent operations is discretized or normalized and mapped to a propagation distance parameter on the edge, which is used to influence the intensity of information transmission during the subsequent graph neural network propagation stage. Longer time intervals increase the likelihood of physical attenuation or chemical neutralization of residual substances in the path, and therefore should be given a higher attenuation weight during propagation.

[0099] It is understandable that the path state evolution graph here is just an initial form and only contains the last operation information.

[0100] S1.2.2: Based on the path state evolution graph, perform state deduction and calculation using a graph neural network based on historical operation information, and output the state evolution results;

[0101] Specifically, the goal of this step is to perform node state propagation and reasoning operations on the aforementioned path state evolution graph through a graph neural network model, thereby obtaining an estimate of the current path state. Since the internal residue of the liquid path cannot be directly observed, and the path reuse and cleaning behaviors are highly dynamic in the control strategy, it is difficult to capture the changing trend of the true state of the path by relying solely on static rules or empirical functions. This embodiment performs end-to-end state reasoning on the graph structure based on a graph neural network, which can integrate multi-step historical information and consider the temporal cumulative effect of behavioral influencing factors, thereby realizing semantic modeling and logical deduction of the path state.

[0102] In one example, the state deduction calculation is performed through the graph neural network, and the state evolution result is output, including:

[0103] S1.2.2.1: Based on the historical operation information, add semantic additional nodes to the original node structure of the path state evolution graph, wherein the semantic additional nodes include operation frequency, path idle time, historical cleaning density, and liquid type switching amplitude;

[0104] Specifically, this step introduces a structured representation of auxiliary behavior information on the state node graph to enhance the node's contextual awareness during graph neural network propagation. During the path state evolution process, although the operational behaviors carried by different nodes are clear in the temporal structure, there is still information missing about the background environment in which their state evolves. This is especially true in scenarios where paths are frequently reused, cleaning intervals are not fixed, and liquid switching strategies are dynamically adjusted. Simply relying on time intervals and liquid type labels cannot fully characterize the true evolution trend of the path. Therefore, it is necessary to structurally express a series of feature quantities that are strongly related to operational behaviors but were not originally included in the graph structure in the graph.

[0105] In this embodiment, a combination structure of a main node and a semantic node is constructed based on the original operation node by introducing a semantic additional node into the state node.

[0106] In this embodiment, semantic nodes include but are not limited to the following dimensions:

[0107] Operation frequency, used to express the usage density per unit time of the path where the current node is located;

[0108] The length of time the pathway is idle reflects the stable dormant period of the liquid pathway before the operation;

[0109] Historical cleaning density is the frequency of cleaning behaviors within a unit of time in the sliding window;

[0110] Liquid type switching amplitude quantifies the level of physical and chemical interference between the current injected liquid and the previously injected liquid.

[0111] Furthermore, each semantic node is connected to the corresponding state node through an edge, and its attribute value is used as a high-dimensional auxiliary feature vector in the input stage of the graph neural network to be spliced ​​and fused with the main node features to construct the node initialization state.

[0112] It is understandable that, in an optional embodiment, the historical operation information will also include multiple injection events to add state nodes, and the time step of the injection event is determined by a person skilled in the art through experiments. It is understandable that the time step here refers to the time between the injection event and the last operation.

[0113] Furthermore, during the network propagation phase, these semantic nodes do not serve as relay nodes for message transmission, but rather as static weight biases to guide the directionality of information during the aggregation process. Specifically, through the attention mechanism, the master node's absorption of information from different neighboring nodes when aggregating adjacent states is regulated, thereby achieving semantically driven refinement of state representation. This enables the model to effectively distinguish path operations that are similar in form but have vastly different contexts, improving the discriminability and historical interpretability of state modeling.

[0114] S1.2.2.2: Calculate the node sequence of the path state evolution graph, establish directed edges between adjacent nodes in the node sequence, and generate a time series graph. If a cleaning operation occurs between adjacent nodes, mark the corresponding directed edge as a state interruption edge and assign a state interruption identifier and a decay weight parameter.

[0115] Specifically, the purpose of this step is to construct a series of operation events in the path into a linearized, clearly structured directed graph sequence according to the time sequence, and at the same time introduce an edge structure with conditional propagation control capabilities, namely the state interruption edge, to express the behavioral interference effect of the cleaning operation on the path state inheritance chain.

[0116] Understandably, since the calibration liquid, blind sample liquid, and cleaning liquid share the same fluid path, cleaning should logically be a state-clearing action. However, due to the influence of pump pressure, valve response, and residual conductivity under real-world conditions, this action does not always completely eliminate the previous state of the path. The path state after cleaning may still retain some historical residue. Failure to consider the inheritance of residual risks can easily lead to over-idealization of path state judgments.

[0117] In this embodiment, when constructing historical operation information as a node sequence, a directed edge is established between every two adjacent nodes to indicate their temporal sequence and state evolution relationship. Classification is then performed based on whether cleaning behavior occurs between the two nodes connected by the edge:

[0118] If there is cleaning behavior, the edge is explicitly marked as a state-breaking edge; otherwise, it is a normal evolving edge.

[0119] It's easy to understand that the essential difference between state interruption edges and ordinary evolving edges lies in their role in constraining propagation paths during graph neural network propagation. State interruption edges possess two key properties: an interruption flag and a decay weight parameter. The interruption flag triggers the gating module that controls message transmission during graph propagation, blocking, delaying, or redirecting information propagation in a certain direction to prevent the previous state in the cleaned path from strongly interfering with the subsequent state. The decay weight parameter is set based on dynamic conditions such as the operation time interval, cleaning frequency, and cleaning type or intensity. While allowing some information from the previous state to remain, a reduction function is applied to the propagation intensity to achieve fuzzy decay-like inheritance of state information.

[0120] S1.2.2.3: Using the time series graph as input to the graph neural network, extracting the state vector of the last node in the time series graph through a preset number of rounds of propagation and updates as a state evolution result, wherein the graph neural network includes a multi-layer graph convolution unit and a state update unit;

[0121] Specifically, after the structure is constructed, the model execution phase begins. The constructed time series graph is input into a graph-based neural network, where multiple rounds of state information propagation and node state updates are performed. Ultimately, the state vector of the latest node on the current path is output as a semantic representation of the path's current usage state. Because path operation histories often exhibit significant behavioral sequence variations, incomplete cleaning, and complex chemical switching interference, traditional methods based on rule trees or flowcharts struggle to adapt to residual interference judgment in actual operating conditions. Therefore, this embodiment employs a graph neural network approach, directly performing vector propagation and aggregation on the structure graph, enabling end-to-end, adaptive feature integration.

[0122] In one example, the processing of the graph neural network includes:

[0123] In each round of propagation, the state information of each node in the time series graph is aggregated through the graph convolution unit, where a node receives a state vector from its previous node. The state information aggregation is weighted based on the type and edge attributes of the directed edge. When a directed edge is marked as a state-interrupted edge, the directional information transmission is disabled according to the state interruption flag during the propagation process, and the incoming information strength is adjusted according to the corresponding attenuation weight parameter;

[0124] Specifically, the node state aggregation operation is performed in rounds within the graph neural network. The core is to complete the directed transmission of the upstream node state through the edge connection relationship in the structure graph, and to regulate the influence intensity of different path sources in the information aggregation stage. In the path state evolution graph, each node represents a specific liquid operation behavior, and each edge represents the time evolution relationship of the path state. In traditional graph neural networks, the information of all adjacent nodes will be equally weighted or merged into the current node based on static weights during the propagation process. However, for evolutionary systems such as liquid paths that have the characteristics of state cleaning interruption, behavior reset, and physical inheritance inaccessibility, if logical filtering at the propagation rule level is not introduced in the propagation stage, it will not be possible to effectively express the truncation effect of the cleaning operation on the path state inheritance chain.

[0125] In this embodiment, during each round of propagation, the graph convolution unit traverses each node and collects the state vectors of all preceding nodes as the candidate input set for the current node. For each edge, the network detects its edge type label. If it is a normal state edge, a weighted merge operation is performed using the default propagation path. If it is a state-disconnected edge, the propagation control sub-logic is activated to apply directional gating to the message input process of the state-disconnected edge.

[0126] It can be understood that directional gating consists of two aspects:

[0127] The first aspect is to determine whether the edge allows incoming information;

[0128] Secondly, the vector amplitude of the incoming information is reduced based on the attenuation factor attached to the edge. This attenuation factor is not a fixed constant, but is dynamically generated based on the differences in the properties of the nodes at both ends of the edge. This includes, but is not limited to, the magnitude of the change in liquid type, whether it is marked as the first injection after cleaning, and the length of time the path has been idle. This is used to simulate the actual intensity of the impact of cleaning behavior on the path state. By using propagation constraints, the graph network no longer aggregates information unconditionally during state aggregation. Instead, it constructs selective information transmission pathways based on cleaning behavior, achieving structural mapping of residual suppression in the physical flow path.

[0129] After completing information aggregation, the node is input into the state update unit, and the current node state is updated through a nonlinear conversion function. When a state interruption edge exists on the node, the state update unit blocks the inheritance of the historical state residual according to the state interruption flag.

[0130] Specifically, state updates correspond to the node state upgrade phase after each round of propagation. Their purpose is to fuse the multidimensional information aggregated from previous nodes with the current node's own state input to generate a semantic state representation for the node at the next stage. In common graph neural network architectures, state updates typically use a simple activation function to perform a weighted combination of the aggregation results and the current state. However, for path state modeling, due to the critical decision point of whether to retain the influence of the previous task, a more behaviorally sensitive state update structure is required. This structure can determine whether the current node inherits the state residual of the previous node based on whether cleaning behavior has occurred.

[0131] In this embodiment, after receiving the aggregation result, the state update unit does not directly perform nonlinear transformation, but first determines whether the current node has an information path introduced by the state interrupt edge;

[0132] If present, the residual blocking logic is activated. This logic includes a state inheritance gating factor, which is used to adjust the proportion of historical memory vectors incorporated into node state updates. When the cleaning interrupt flag is activated, the gating factor is forced to zero. The current node state is determined solely by the results of the current round of propagation in the graph and the current node's own semantic vector, completely removing any traces of the previous state. If the interrupt flag is not activated, the factor proportionally incorporates the previous node state vector to express the continuity of state evolution.

[0133] It can be understood that the residual blocking logic is essentially an equivalent modeling of the behavior of physical cleaning operations in the mathematical propagation model, allowing the node state to be logically cut off in the propagation chain, simulating a two-state scenario in which the path state is reset to zero and the path residue is brought in.

[0134] Furthermore, to avoid complete loss of state information due to forced blocking, this embodiment introduces a retained residual channel in the state update unit. This allows a certain amount of information to penetrate under conditions of boundary uncertainty, such as insufficient cleaning duration or extremely high liquid affinity before and after cleaning, adding flexible control capabilities to state expression. This enables the model to maintain stable state estimation and behavioral consistency in scenarios such as highly uncertain path behavior, uneven cleaning quality, and irregular operation intervals, thereby improving the reliability of the final path availability determination.

[0135] S1.3: Constructing a current path state matrix according to the path state;

[0136] In an example, the specific steps of S2 are as follows:

[0137] S2.1: extract the current state vector of each liquid path in the path state matrix;

[0138] Specifically, the core goal of this step is to perform a structured extraction of the pathway representations in the pathway state matrix, processed during the state modeling phase, to provide quantitative input for the pollution risk response calculation. Each row of the pathway state matrix corresponds to the state representation of a liquid pathway, and the vector representation of each entry in the matrix represents a state encoding containing complex information such as time, behavior, attenuation, and contextual history residuals.

[0139] In this embodiment, the state vector of the final node of the current round is extracted path by path from the path state matrix. Each vector is the result of multiple rounds of propagation and nonlinear updates by the graph neural network. The current state vector is essentially a high-dimensional feature representation, capable of complex expression of multiple dimensions of information, including residual liquid identification codes, behavioral evolution trends, the impact of cleaning breakpoints, and path vacancy stability.

[0140] S2.2: For each liquid path, call a contamination risk mapping function, using the current state vector as input and the target liquid corresponding to the calibration task as a parameter, to calculate the contamination risk response value of the corresponding liquid path relative to the target liquid. The contamination risk mapping function includes state compatibility matching logic, where:

[0141] The state compatibility matching logic calculates a probability score of residual liquid interference in the liquid path in the current state based on the chemical interference level, cleaning behavior interruption frequency, and time interval indicators between the liquid type evolution trajectory recorded in the liquid path and the target liquid;

[0142] Specifically, this step is the core process of converting the path state vector into a contamination risk value. The projection operation from the path state to the contamination risk space is achieved by establishing a contamination risk mapping function. Although the current state vector corresponding to the liquid path itself contains rich behavioral and residual characteristic information, for calibration tasks, whether it constitutes contamination interference depends on whether the state is compatible with the target liquid. Therefore, it is necessary to introduce matching logic with the target liquid based on the state vector expression to determine whether there is a risk of chemical interference between the residual behavior implied by the current state and the target liquid.

[0143] In this embodiment, the pollution risk mapping function consists of two parts:

[0144] The first part is the input channel and the matching logic body. The input channel receives the current state vector of the path and the target liquid identification code. The target liquid code is used as a compatibility reference value in the matching logic body to perform a difference comparison with the historical liquid type code, liquid switching frequency, and liquid chemical category label in the state vector. The matching logic body performs the following judgment logic: First, an interference weight table is established based on the evolution trajectory of the liquid type and the chemical interference level between the target liquid. The chemical interference level can be predefined as the reaction intensity, solubility, or adsorption tendency between liquid pairs. Second, the interruption frequency of the historical cleaning behavior recorded in the state vector is counted. If the path interruption is discontinuous or the cleaning is insufficient, the weight coefficient of the interference level of the cleaning behavior will be increased accordingly; third, the operation interval information recorded in the path is called, and the path idle time is mapped to the residual attenuation correction factor.

[0145] In the second part, the matching logic weights and combines the above multiple judgment dimensions and outputs a contamination risk response value, which indicates the possibility of contamination of the current state of the path relative to the target liquid.

[0146] Understandably, the contamination risk response value is not an absolute residual indicator, but rather a scored expression of the likelihood of contamination from the perspective of structural evolution and behavioral logic. Its construction allows for risk estimation without relying on actual sensor detection, relying solely on the comprehensive relationship between structure, time, behavior, and tasks.

[0147] In one example, the contamination risk mapping function in this application is used to calculate the compatibility score between the current path state vector S and the calibration target liquid code L, and output the contamination risk response value R. Its mathematical form is as follows:

[0148] ;

[0149] in, 、 and Indicates the weight coefficient, which satisfies the sum of 1 and can be set according to the empirical value. represents the liquid chemical interference subfunction, that is, the chemical compatibility with the target liquid, which is specifically expressed as the weighted average of the two liquids after mixing. represents the sub-function of the frequency of interruption of cleaning behavior, represents the path idle time decay sub-function;

[0150] The mathematical form of the path idle time decay sub-function is as follows:

[0151] ;

[0152] in, represents the coefficient that controls time decay, Indicates the time the path has been idle since it was last filled.

[0153] S2.3: Arrange the contamination risk response values ​​of the liquid pathways by pathway number and construct a contamination adaptability matrix;

[0154] In this embodiment, the pollution adaptability matrix is ​​a one-dimensional structure, wherein the value of each position is the pollution risk response value of the corresponding path.

[0155] As a preferred embodiment, the contamination adaptability matrix is ​​expandable into a two-dimensional structure to support concurrent multi-task or multi-target liquid assessment scenarios. The horizontal axis represents the path number, the vertical axis represents the target liquid type, and the matrix elements represent the corresponding contamination response scores. This supports vectorized scheduling decisions, compatibility screening, and resource reuse strategy generation, effectively reducing path decision latency and improving the accuracy and flexibility of calibration resource scheduling.

[0156] In an example, the specific steps for S3 are as follows:

[0157] S3.1: Sort the pollution risk response values ​​corresponding to the liquid paths in the pollution adaptability matrix to determine a path sequence from low to high pollution risk;

[0158] S3.2: Perform availability checks on the paths in the path sequence in sequence, where the availability checks include: whether the paths are in operation, whether the minimum idle time has been reached, and whether there is a fault mark;

[0159] S3.3: Among the liquid paths that have passed the availability check, select the liquid path with the lowest contamination risk response value as the optimal compatibility path.

[0160] It's understandable that after the multi-path structure in a water quality analysis system undergoes a calibration task and completes the injection operation, the internal state of the path has actually changed due to this round of injection operations. Failure to promptly update its state information will cause the path state evolution graph to remain in its historical state for a long time, affecting the accuracy of the next round of state deduction. This is especially true in scenarios where the system lacks real-time state perception sensors. Path states primarily rely on evolutionary mechanisms for indirect modeling, making the operation log mechanism a necessary part of the information closure. The introduction of operation logs is not solely for recordkeeping and archiving; rather, it serves to structurally maintain the path state graph and logically backfill the state matrix, enabling the state model to form a causal closed loop after each round of tasks, thus possessing self-evolution and self-updating capabilities.

[0161] In this embodiment, after the calibration task is completed, the scheduling logic records the key operating parameters of the liquid path used in this round and writes them to the path operation log. The log records include, but are not limited to: the type of injected liquid, whether the cleaning operation was performed, the completion timestamp, the duration of the injection, the execution failure flag, etc. All information is stored in the form of structured key-value pairs and associated with the unique identification number of the path. The operation log can be viewed as a time-rolling window structure in the system, with append-write, historical backtracking, and data recovery mechanisms. Its main purpose is to provide raw behavioral data for generating new state nodes in the path state evolution graph.

[0162] Furthermore, after the operation log is updated, the system parses the latest record through a preset log parsing module, extracts the core fields, generates a new node on the path state evolution graph, and establishes a connection relationship with the last node in the graph. The attribute fields of the newly created node are constructed based on the original log information, including liquid type code, operation time, cleaning flag, task type, etc. The edge attributes of the newly created edge are determined by the difference between the last completion time and the start time of the current round of tasks. At the same time, it is determined whether it needs to be set as a state interruption edge, and the corresponding interruption flag and propagation attenuation factor are attached. The supplementary operation of the graph structure provides the necessary structural input for the next round of state deduction of the graph neural network, avoiding inference errors in the path state due to incomplete information or graph breaks.

[0163] Finally, after the path state evolution graph completes its structural update, the latest node states in the graph are synchronously updated to the path state matrix via a state propagation mechanism. This update operation can be divided into two strategies: replacement update, in which the new node state is directly used as the latest state vector of the path in the state matrix; and fusion update, in which the new node state is weightedly merged with the original state vector to construct a long-term state representation that integrates multiple rounds of behavioral residuals. The specific strategy depends on the system configuration or model settings.

[0164] by Figure 5 For example, Figure 5 This is a flow chart of the path matrix updating method according to an embodiment of the present application, which specifically includes:

[0165] S4: writing the usage information of the liquid path used to perform the calibration task into the path operation log, updating the type of the most recently injected liquid, whether the cleaning operation was performed, and the completion time;

[0166] S5: Calculate the node information of the corresponding liquid path in the path state evolution graph according to the updated content in the path operation log, and update the calculation result to the path state matrix.

[0167] In one example, the present application provides a remote calibration system for a smart water station, the system comprising:

[0168] A status acquisition module is used to obtain status information of each liquid path, including the type of liquid injected last time, whether a cleaning operation is performed, and the time interval since the last operation;

[0169] a state deduction module that derives the state information according to preset state update rules to generate a path state of the liquid path, wherein the state update rules include interruption behavior modeling logic for simulating the impact of cleaning operations on the reset of the path state, and performs state deduction calculations through a graph neural network;

[0170] The risk calculation module calculates the contamination risk response value of each liquid path based on the path state matrix and the target liquid of the calibration task, and generates a contamination adaptability matrix;

[0171] The path scheduling module determines the most compatible path according to the pollution adaptability matrix, controls the path to execute a calibration task, and updates the path operation log and the path status matrix after the calibration task is completed.

[0172] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A remote calibration method for a smart water station, applied to a water quality analysis system having multiple liquid paths, wherein the liquid paths include a control valve corresponding to the calibration liquid, a sampling pump, and a liquid pipeline for conveying the calibration liquid to a measurement chamber, characterized in that: The method comprises: Acquiring state information of each liquid path and constructing a path state matrix, wherein the constructed path state matrix derives the state information according to a preset state update rule, the state update rule including interruption behavior modeling logic for simulating the impact of a cleaning operation on resetting the state information; According to the calibration task, the contamination risk response value of each liquid path to the target calibration task is calculated through the path state matrix to generate a contamination adaptability matrix; Obtaining an optimal compatibility path according to the pollution adaptability matrix, and performing a calibration task through the optimal compatibility path; The construction path state matrix includes: Acquiring status information of each liquid path, the status information including the type of liquid last injected, whether a cleaning operation is performed, and the time interval since the last operation; The state information is deduced according to a preset state update rule to generate a path state corresponding to the liquid path, including: The last injected liquid type of each liquid path is used as the initial node, and the time interval from the last operation is mapped to the evolution step to construct the path state evolution graph; On the path state evolution graph, based on historical operation information, the graph neural network performs state deduction and calculation, and outputs the state evolution results, including: According to the historical operation information, a semantic additional node is added to the original node structure of the path state evolution graph, wherein the semantic additional node includes operation frequency, path idle time, historical cleaning density, and liquid type switching amplitude; Calculate the node sequence of the path state evolution graph, establish directed edges between adjacent nodes in the node sequence, and generate a time series graph, wherein if there is a cleaning operation between adjacent nodes, the corresponding directed edge is marked as a state interruption edge, and a state interruption identifier and a decay weight parameter are assigned; The time series graph is used as the input of the graph neural network. Through a preset round of propagation and update, the state vector of the last node in the time series graph is extracted as the state evolution result, wherein the graph neural network includes a multi-layer graph convolution unit and a state update unit, including: In each round of propagation, the state information of each node in the time series graph is aggregated through the graph convolution unit, where a node receives a state vector from its previous node. The state information aggregation is weighted based on the type and edge attributes of the directed edge. When a directed edge is marked as a state-interrupted edge, the directional information transmission is disabled according to the state interruption flag during the propagation process, and the incoming information strength is adjusted according to the corresponding attenuation weight parameter; After completing information aggregation, the node is input into the state update unit, and the current node state is updated through a nonlinear conversion function. When a state interruption edge exists on the node, the state update unit blocks the inheritance of the historical state residual according to the state interruption flag. constructing a current path state matrix according to the path state; The contamination risk response value of each liquid path to the target calibration task is calculated using the path state matrix to generate a contamination adaptability matrix, including: Extracting the current state vector of each liquid path in the path state matrix; For each liquid path, the contamination risk mapping function is called, with the current state vector as input and the target liquid corresponding to the calibration task as a parameter, to calculate the contamination risk response value of the corresponding liquid path relative to the target liquid; Arrange the contamination risk response values ​​of the liquid paths according to the path numbers to construct a contamination adaptability matrix; According to the pollution adaptability matrix, the optimal compatibility path is obtained, including: Sorting the pollution risk response values ​​corresponding to the liquid paths in the pollution adaptability matrix to determine a path sequence from low to high pollution risk; Performing availability checks on the paths in the path sequence in sequence, wherein the availability checks include: whether the paths are in operation, whether the minimum idle time has been reached, and whether there is a fault mark; Among the liquid paths that have passed the availability verification, the liquid path with the lowest contamination risk response value is selected as the path with the best compatibility.

2. A remote calibration method for a smart water station according to claim 1, characterized in that: The pollution risk mapping function includes state compatibility matching logic, where: The state compatibility matching logic calculates a probability score for residual liquid interference in the liquid path in the current state based on the chemical interference level, cleaning behavior interruption frequency, and time interval indicators between the liquid type evolution trajectory recorded in the liquid path and the target liquid.

3. The remote calibration method of a smart water station according to claim 1, characterized in that: After the calibration task is completed, the method further includes: Write the usage information of the liquid path used to perform the calibration task into the path operation log, and update the type of the most recently injected liquid, whether the cleaning operation was performed, and the completion time; According to the updated content in the path operation log, the node information of the corresponding liquid path in the path state evolution map is calculated, and the calculation result is updated to the path state matrix.

4. A remote calibration system for a smart water station, used to implement a remote calibration method for a smart water station according to any one of claims 1 to 3, characterized in that: The system comprises: A status acquisition module is used to obtain status information of each liquid path, including the type of liquid injected last time, whether a cleaning operation is performed, and the time interval since the last operation; a state deduction module that derives the state information according to preset state update rules to generate a path state of the liquid path, wherein the state update rules include interruption behavior modeling logic for simulating the impact of cleaning operations on the reset of the path state, and performs state deduction calculations through a graph neural network; The risk calculation module calculates the contamination risk response value of each liquid path based on the path state matrix and the target liquid of the calibration task, and generates a contamination adaptability matrix; The path scheduling module determines the most compatible path according to the pollution adaptability matrix, controls the path to execute a calibration task, and updates the path operation log and the path status matrix after the calibration task is completed.

Images