A method and system for online pH adjustment of drinking mineralized water based on a switching system

Abstract

This invention discloses a method and system for online pH adjustment of drinking mineralized water based on a switching system. The method includes: acquiring a global dataset, a local dataset, and optimal pre-trained weights; receiving production module replacement information and forming a replacement sequence; determining a first set of initial pre-trained weights when no production module replacement occurs, and determining the affected submatrix and generating a second set of initial pre-trained weights when a production module replacement occurs; performing central-side training based on the corresponding initial pre-trained weights to obtain the optimal central-side pre-trained weights and the system state transition matrix, and sending them to an edge intelligent control device for local training, parameter estimation, and rolling control. The system includes a central data center and multiple production modules. Using this invention can shorten the readjustment and debugging cycle after production module replacement, and improve the accuracy of online pH adjustment, control response speed, and operational stability.

Description

Technical Field

[0001] This application relates to the fields of drinking water treatment and industrial process control technology, and in particular to a method and system for online adjustment of pH value of drinking mineralized water based on a switching system. Background Technology

[0002] As industrial production moves towards flexibility, intelligence, and networking, production lines need to dynamically adjust their processes according to different production tasks. In the production of products such as mineral water and beverages, to adapt to changes in the formula, process parameters, or processing flow of the target product, it is often necessary to replace, switch, or reorganize some production modules in the production system to complete the production tasks of different products.

[0003] In the aforementioned production process, when a production module is switched, the structural relationships and operational state of the original production system change accordingly, leading to significant alterations in the dynamic characteristics of the entire production system. Due to the coupling relationships between different production modules, module switching not only affects local processing stages but also has a cascading impact on the overall system's state evolution, control input response, and output results. Therefore, after completing the module switch, the production system needs to be readjusted and optimized to ensure that the system output meets the target production requirements.

[0004] In existing technologies, for industrial production systems after module switching, the control parameters and operating status are typically tuned repeatedly through reference value setting, manual experience correction, and long-term on-site debugging to gradually bring the system output to the expected indicators. This approach is applicable in industrial scenarios with relatively simple structures, fixed control logic, or where a long debugging period is permissible. However, it has significant shortcomings for current intelligent industrial systems with frequent task switching, high system coupling, and high deployment efficiency requirements.

[0005] Specifically, on the one hand, the system dynamics change significantly after module switching, and traditional methods relying on manual experience and repeated debugging are insufficient to establish an effective control model for the system after the switch, resulting in a long time required for system redeployment and stable operation. On the other hand, the debugging process relies heavily on engineering experience, making it difficult to guarantee control accuracy and output consistency, which is not conducive to the rapid switching and efficient operation of complex industrial systems between different production tasks. Especially in production scenarios such as drinking mineral water, where both parameter adjustment accuracy and production efficiency are highly demanding, existing technologies struggle to balance system adjustment speed and output quality.

[0006] Therefore, in response to the problems of low system adjustment efficiency, long deployment cycle and insufficient output accuracy of flexible production lines when production tasks are updated or changed, there is an urgent need for a technical solution that can quickly adapt to changes in system dynamics after production module switching, shorten the debugging cycle and improve the accuracy of online parameter adjustment. Summary of the Invention

[0007] This application aims to solve the technical problem that changes in system dynamics after the replacement of production modules lead to the need for readjustment of online pH adjustment in drinking mineralized water, resulting in a long debugging cycle.

[0008] According to one aspect of this application, a method for online pH adjustment of drinking mineralized water based on a switching system is provided, executed by a central data center and applied to a production line communicatively connected to the central data center. The production line includes multiple production modules, each of which includes a sensing device, an adjustment device, and an edge intelligent control device, comprising: Obtain the global dataset corresponding to the m-th round of production, the local dataset corresponding to each production module, and the optimal pre-trained weights corresponding to the m-th round of production. Before the start of the (m+1)th round of production, receive the production module replacement information and form a replacement sequence S; Determine whether a production module replacement occurs in the (m+1)th round of production based on the replacement sequence S. If no production module is replaced, the optimal pre-trained weights corresponding to the m-th round of production are determined as the first pre-trained initial weights for the (m+1)-th round of production. If a production module is replaced, the affected submatrix in the system state transition matrix is ​​determined, and the second pre-trained initial weights for the (m+1)th round of production are generated. Training is performed based on the global dataset corresponding to the (m+1)th round of production and the corresponding first or second pre-training initial weights to obtain the optimal pre-training weights on the central side and the system state transition matrix. The optimal pre-trained weights on the central side and the system state transition matrix are sent to the edge intelligent control devices of each production module, so that each edge intelligent control device can perform local training, parameter estimation and rolling control to realize online adjustment of the pH value of drinking mineralized water.

[0009] Preferably, the system state transition matrix is ​​divided into multiple system state transition sub-matrices according to the correlation between production modules; Based on the replacement sequence S, the system state transition submatrix corresponding to the production module that has been replaced is determined from the plurality of system state transition submatrixes as the affected submatrix; When the replacement sequence S is represented as S=[E, ..., F], construct an index set. P = {(i, j) | E ≤ i ≤ F or E ≤ j ≤ F}, and the system state transition submatrix that satisfies (i, j) ∈ P is determined as the affected submatrix.

[0010] Preferably, when generating the second pre-training initial weights for the (m+1)th round of production, only the affected submatrices are initialized, while the remaining system state transition submatrices remain unchanged, to obtain the initialized system state transition matrix; The system state transition matrix, the global dataset corresponding to the m-th production round, and the optimal pre-trained weights corresponding to the m-th production round are used to train the second pre-trained initial weights for the (m+1)-th production round. The training employs a loss function that includes a fitting error term for the affected submatrix, which is used to constrain the difference between the trained affected submatrix and the initialized affected submatrix.

[0011] Preferably, when training based on the global dataset corresponding to the (m+1)th round of production and the first pre-trained initial weights or the second pre-trained initial weights, if the replacement sequence S represents that no production module replacement has occurred, then the first pre-trained initial weights are used as the initial training weights. If the replacement sequence S represents a change in the production module, then the second pre-training initial weights are used as the initial training weights. The training employs a loss function that includes a continuous learning regularization term to control the rate of weight change between adjacent production rounds, and estimates the system state transition matrix based on the optimal pre-trained weights obtained from the training on the central side.

[0012] Preferably, each of the edge intelligent control devices uses the optimal pre-trained weights on the central side as the initial weights for local training, and performs iterative training based on the local dataset corresponding to each of the production modules to obtain the optimal local training weights. When the difference between the training weights obtained from two adjacent iterations is less than the corresponding convergence threshold, the current training weight is determined to be the local optimal training weight.

[0013] Preferably, the local optimal training weights are used to estimate the system state transition matrix, control state transition matrix, and perception matrix of the corresponding production module; Based on the estimated system state transition matrix, control state transition matrix, and sensing matrix, an optimization problem corresponding to the production objective is constructed to solve for the optimal control sequence.

[0014] Preferably, the objective function of the optimization problem includes a tracking error term between the sensing sequence and the corresponding production target, and a control cost term for the control sequence, and is superimposed with control magnitude constraints and / or control rate of change constraints; The optimization problem is resolved in each control cycle based on the latest sensing data to continuously update the optimal control sequence.

[0015] Preferably, after the (m+1)th round of production is completed, the newly added local data uploaded by each of the production modules is received, and the newly added local data is added to the global dataset corresponding to the (m+1)th round of production for use in pre-training updates in subsequent production rounds. The data collected by each of the aforementioned sensing devices includes at least one or more of the following: pH value, water quality data, feeding status, valve status, and environmental data.

[0016] The present invention also provides an online pH adjustment system for drinking mineralized water based on a switching system, which is applied to the online pH adjustment method for drinking mineralized water based on a switching system as described in any of the above claims, including a central data center and multiple production modules, wherein each of the production modules is communicatively connected to the central data center; Each of the production modules includes sensing devices, adjustment devices, and edge intelligent control devices, and the central data center includes a data processing module and a communication module. The sensing device is used to collect production data from the corresponding production module and send the production data to the edge intelligent control device; The edge intelligent control device is used to receive the production data and send the production data and / or local processing results to the central data center through the communication module; The data processing module is used to receive production module replacement information and generate training results based on the production data and the production module replacement information; The communication module is also used to send the training results to the edge intelligent control devices of each of the production modules; The edge intelligent control device is also used to generate control commands based on the training results and the production data, and output the control commands to the adjustment device; The regulating device is used to perform regulating operations according to the control command.

[0017] Preferably, the data processing module includes a switching adaptation unit and a central training unit, wherein the switching adaptation unit is connected to the central training unit; The edge intelligent control device includes a local training unit and a scrolling control unit, wherein the local training unit is connected to the scrolling control unit; The switching adaptation unit is used to receive the production module replacement information, generate the initialization result of the affected submatrix based on the production data, and send the initialization result to the central training unit. The central training unit is used to receive the initialization result and generate the training result based on the initialization result and the production data; The local training unit is used to receive the training results and the production data, generate parameter estimation results, and send the parameter estimation results to the rolling control unit; The rolling control unit is used to receive the parameter estimation results and generate the control commands.

[0018] This application offers the following advantages: After a production module is replaced, the affected submatrix in the system state transition matrix is ​​determined based on the replacement information, and corresponding pre-trained initial weights are generated accordingly. Based on this, central-side training is performed using global data to obtain the optimal pre-trained weights and system state transition matrix, which are then sent to the edge intelligent control device for local training, parameter estimation, and rolling control. Through this technical solution, the adjustment model and control parameters can be quickly updated after the system dynamics change due to a production module replacement, reducing reliance on manual experience and repeated tuning, and shortening the debugging cycle for online pH adjustment of drinking mineralized water. Simultaneously, it improves the parameter estimation efficiency and online adjustment accuracy of the system after switching, enhancing the rapid deployment and stable operation capabilities of flexible production lines under different production tasks. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a logic block diagram of the online pH adjustment method for drinking mineralized water based on a switching system according to an embodiment of this application. Figure 2 This is a module connection diagram of the online pH adjustment system for drinking mineralized water based on a switching system, as described in one embodiment of this application.

[0021] The following are the diagram labels: 100, Production Module; 110, Sensing Device; 120, Edge Intelligent Control Device; 121, Rolling Control Unit; 122, Local Training Unit; 130, Adjustment Device; 200, Central Data Center; 210, Communication Module; 220, Data Processing Module; 221, Central Training Unit; 222, Switching Adaptor Unit. Detailed Implementation

[0022] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings. Preferred embodiments of this application are shown in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] Please refer to Figure 1 One embodiment of this application provides an online pH adjustment method for drinking mineralized water based on a switching system. This method is executed by a central data center and applied to a production line communicatively connected to the central data center. The production line includes multiple production modules, each comprising sensing devices, adjustment devices, and edge intelligent control devices. S10. Obtain the global dataset corresponding to the m-th production round, the local dataset corresponding to each production module, and the optimal pre-trained weights corresponding to the m-th production round. It should be noted that "obtaining" in this step includes the unified reading and organization of historical data from the previous production round, model parameters, and operational data from each production module. Preferably, the central data center receives process operation data from the sensing devices corresponding to each production module. This process operation data includes at least pH value, feeding status, valve status, and environmental status data, and may further include conductivity, turbidity, temperature, dissolved solids content, flow rate, and / or pressure data. The central data center performs time alignment, resampling, outlier removal, missing value completion, and noise filtering on the data uploaded by each production module to form a global dataset for training on the central side; simultaneously, each production module retains its corresponding local dataset for subsequent local training on the edge side.

[0025] Preferably, the mineralized water production process can be described using a discrete state-space model: x i (k)= A i,j x j (k 1)+B i u i (k 1)+w i (k 1) y i (k)=Ci x i (k)+v i (k-1) Where, x i (k) represents the state variable of the i-th production module at time k, u i (k) represents the control input, y i (k) represents the perception output, A i,j B represents the system state transition submatrix. i C represents the control state transition matrix. i Let w represent the perception matrix. i With v i This represents noise and disturbances during system operation and status observation. Multiple A's i,j These can be concatenated to form a system state transition matrix A at the production line level. The optimal pre-trained weights corresponding to the m-th production round can serve as a parameter representation of the common dynamics of the entire production line, providing an initial foundation for subsequent switching adaptation and continuous learning.

[0026] S20. Before the start of the (m+1)th round of production, receive production module replacement information and form a replacement sequence S. In this step, it should be noted that the production module replacement information is used to characterize the changes in module composition in the (m+1)th round of production compared to the mth round. The central data center can standardize this information as a replacement sequence S = [E, ..., F], where 1 ≤ E ≤ F ≤ a, a is the total number of production modules in the production line, and E and F represent the starting and ending module numbers of the replacement, respectively. The replacement sequence S can correspond to a single module replacement or multiple consecutive module replacements; for a mixed serial-parallel production line, the actual replacement modules can be mapped to numbering intervals before forming the replacement sequence S. By uniformly encoding the module replacement information as a replacement sequence S, it is easier to subsequently determine the affected submatrices and initialization range according to consistent rules.

[0027] S30. Determine whether a production module replacement occurred in the (m+1)th round of production based on the replacement sequence S. In this step, it should be noted that the central data center can determine whether a production module replacement occurred in the (m+1)th round of production based on whether the replacement sequence S is empty or meets the preset valid interval conditions. When the replacement sequence S is empty or does not contain a valid module number, it indicates that no module replacement occurred in this round of production, and the model knowledge obtained from the previous training round can be directly inherited. When the replacement sequence S is a valid interval, it indicates that at least one production module has been replaced, and the local dynamic changes caused by the module replacement need to be adapted. The judgment result of this step is used to determine whether to use the first or second pre-training initial weights subsequently, and to decide whether to trigger the initialization and constraint training process of the affected submatrix.

[0028] S40. If no production module replacement occurs, the optimal pre-trained weights corresponding to the m-th production round are determined as the initial pre-trained weights for the (m+1)-th production round. In this step, it should be noted that when S30 determines that no production module replacement occurred in the (m+1)-th production round, the central data center directly uses the optimal pre-trained weights corresponding to the m-th production round as the initial pre-trained weights for the (m+1)-th production round, thus letting θ... m+1 (1) =θ m * Since the production line module composition remains unchanged, the main structure of the system dynamics has not changed significantly, so there is no need to reinitialize the local parameters. This approach allows for the inheritance of global dynamic knowledge learned from the previous production run, while also reducing the number of training iterations at the start of a new production run, enabling central-side training to proceed as a continuation of training rather than restarting from random initial values.

[0029] S50. If a production module replacement occurs, determine the affected submatrix in the system state transition matrix and generate the second pre-training initial weights for the (m+1)th round of production. In this step, it should be noted that when S30 determines that a production module replacement has occurred in the (m+1)th round of production, the central data center constructs an affected index set based on the replacement sequence S=[E, ..., F]. P = {(i,j) | E ≤ i ≤ F or E ≤ j ≤ F}, and the system state transition submatrix A satisfying (i,j) ∈ P is... i,j The affected submatrices are identified. Subsequently, initialization is performed only on the affected submatrices, while the estimation results from the previous round remain unchanged for the remaining unaffected submatrices, thus obtaining the initialized system state transition matrix Ã. Preferably, the initialization process can be performed by assigning preset initial matrix values ​​or by assigning values ​​randomly according to a preset distribution.

[0030] Based on this, the central data center generates the second pre-training initial weights for the (m+1)th round of production. Preferably, the following objective function can be constructed: L c (θ)=L(θ,D m )+Σ (i,j)∈P ||Â i,j (θ)-Ã i,j || F 2 +μR(θ,D m ) Where L(θ,D) m ) represents the global dataset D based on the m-th round. m Basic training loss, i,j (θ) represents the (i,j)th submatrix predicted by the model, R(θ,D) m) represents the regularization term, μ is the regularization coefficient, and the subscript F denotes the Frobenius norm. Further, the following iterative method can be used: θc (n) =θc (n-1) -η θcL(θc (n-1) D m )+Σ (i,j)∈P ||Â i,j (θc (n-1) )-Ã i,j || F 2 +μR(θc (n-1) D m ) When ||θc (n) -θc (n-1) When || < ε, the iteration stops, and the current weights are determined as the second pre-training initial weights produced in the (m+1)th round. This process allows the model to focus on adapting to the affected regions, while the unaffected parts continue to inherit historical knowledge.

[0031] S60. Based on the global dataset corresponding to the (m+1)th production round and the corresponding first or second pre-training initial weights, training is performed to obtain the optimal pre-training weights and system state transition matrix on the central side. In this step, it should be noted that the central data center determines the initial training values ​​for the (m+1)th production round based on the processing results of S40 or S50. If no production module change occurs, the first pre-training initial weights are used as the initial training weights; if a production module change occurs, the second pre-training initial weights are used as the initial training weights. Subsequently, the central data center combines the global dataset D corresponding to the (m+1)th production round... m+1 Perform continuous learning training on the central side to obtain the optimal pre-trained weights θ on the central side. m+1 * .

[0032] Preferably, the central-side training can adopt the following objective function: θ m+1 * =arg min θ m+1 L(θ m+1 D m+1 )+λΩ(θ,θ m * ) Where, Ω(θ, θ) m * ) represents the continuous learning regularization term used to control the rate of change of weights between adjacent production rounds, and λ is a hyperparameter. Further, the following update method can be adopted: θ m+1 (n) =θm+1 (n-1) -η θ m+1 [L(θ m+1 (n-1) D m+1 )+λΩ(θ m+1 (n-1) D m+1 )] When ||θ m+1 (n) -θ m+1 (n-1) When || < ε, training stops and the optimal pre-trained weights on the central side are output. The central data center further estimates the system state transition matrix A at the production line level based on the optimal pre-trained weights on the central side. m+1 The purpose of this step is to connect the local switching adaptation with the global model update, so that the central model can both inherit knowledge from previous rounds and adapt to the new round of production data.

[0033] S70. The optimal pre-trained weights and system state transition matrix from the central data center are sent to the edge intelligent control devices of each production module for local training, parameter estimation, and rolling control, thereby achieving online pH adjustment of the drinking mineralized water. In this step, it should be noted that after obtaining the optimal pre-trained weights and system state transition matrix from the central data center, they are distributed to the corresponding edge intelligent control devices of each production module. Upon receiving these, each edge intelligent control device uses the optimal pre-trained weights from the central data center as the initial weights for local training, and performs local training using its local dataset to obtain the local optimal training weights φ. i,m+1 * Then, based on the locally optimal training weights, estimate the system state transition matrix A corresponding to the production module. i，m+1 Control state transition matrix B i,m+1 and perception matrix C i,m+1 Preferably, local training can be represented as: Φ i,m+1 * =argminφ i,m+1 L i (φ i,m+1 D i,m+1 ) Among them, D i,m+1 This represents the local dataset of the i-th production module in the (m+1)-th round of production.

[0034] Furthermore, the edge intelligent control device is based on the estimated A i,m+1 B i,m+1 C i,m+1This involves constructing an optimization problem based on the corresponding production target to find the optimal control sequence. Preferably, the following objective function can be used: J i = (y i (k)-r i (k)) T Q(y i (k)-r i (k))+u i (k) T Ru i (k) Where, r i (k) represents the target observation sequence, Q represents the error weight matrix, R represents the control weight matrix, and N p This represents the prediction time domain length, and the solution is obtained by combining control magnitude constraints and / or control rate of change constraints. For nonlinear processes of pH change in mineralized water bodies, a linearized prediction model can also be constructed near the local operating point. Δx k+1 =A k Δx k +B k Δu k Δy k =C k Δx k And based on the deviation between the latest pH measurement and the predicted value e k =pH k meas -pH k pred Generate feedback correction amount δu k To compensate for linearization errors, the edge intelligent control device only outputs the control quantity corresponding to the current moment as the control command in the current control cycle, and re-solves based on the latest data in the next control cycle, thereby realizing online closed-loop regulation of pH value.

[0035] The technical solution implemented in this embodiment enables branching of the central training process before and after the production module replacement: when no replacement occurs, the optimal pre-trained weights from the previous round are directly inherited, reducing redundant training; when a replacement occurs, initialization and adaptation training are performed only for the affected submatrices, avoiding full retuning of the unaffected parts. Furthermore, the central data center completes continuous learning training based on the corresponding first or second pre-trained initial weights and sends the training results to the edge intelligent control devices for local training, parameter estimation, and rolling control, thereby enabling the system dynamics after module switching to complete learning and adaptation more quickly. This shortens the debugging cycle for online pH adjustment after production module replacement and improves the parameter estimation accuracy, control response speed, and operational stability of the system after switching.

[0036] In one specific embodiment, the system state transition matrix is ​​divided into multiple system state transition sub-matrices according to the association relationships between production modules. It should be noted that a production line containing multiple production modules can be abstracted as a discrete switching system, whose system state transition matrix is ​​denoted as A, and divided into a block matrix form A=[A...] according to the numbering order of each production module. i,j ], where A i,j This represents the influence of the state change of the j-th production module on the state evolution of the i-th production module. If the state dimension of the i-th production module is n... i The state dimension corresponding to the j-th production module is n. j ,but Matrix A i,j The dimension is n i ×n j Among them, the diagonal submatrix A i,i The off-diagonal submatrix A is used to characterize the state continuity and inherent dynamic characteristics of the corresponding production module. i,j (i≠j) is used to characterize the coupling effect between different production modules.

[0037] Based on the replacement sequence S, the system state transition submatrix corresponding to the production module that has been replaced is determined from multiple system state transition submatrixes as the affected submatrix. It should be noted that the replacement sequence S characterizes the location and scope of the module replacement in the (m+1)th production round relative to the mth production round. After receiving the production task adjustment information, the central data center maps the replaced module range to the replacement sequence S. The affected submatrix is ​​not limited to the state transition submatrix of the replaced module itself, but also includes other system state transition submatrixes that have input-output coupling relationships with the replaced module.

[0038] When the sequence S is changed to S=[E, ...,F], an index set is constructed. P = {(i, j) | E ≤ i ≤ F or E ≤ j ≤ F}, and the system state transition submatrices satisfying (i, j) ∈ P are determined as the affected submatrices. It should be noted that when the replacement sequence S is represented as S = [E, ..., F], by constructing the index set P = {(i, j) | E ≤ i ≤ F or E ≤ j ≤ F}, all system state transition submatrices that have row or column associations with the replacement module interval can be filtered out at once. Furthermore, the set of affected submatrices can be represented as A. aff =A i,j|(i,j)∈P. For example, when the production line includes 6 production modules and the replacement sequence S=[2,3], then all system state transition submatrices satisfying i∈2,3 or j∈2,3 can be determined as affected submatrices. It should be noted that the total number of production modules is greater than F. The replacement of modules is only a local replacement.

[0039] The technical solution implemented in this embodiment can limit the impact of module replacement on system dynamics to the level of local sub-matrices in the system state transition matrix, thus narrowing the parameter adjustment target after the production module replacement from the entire system to the set of affected sub-matrices. Therefore, dynamic model information corresponding to the parts that have not been replaced can be preserved, and an accurate range can be provided for subsequent local initialization and local adaptation training.

[0040] In one specific embodiment, when generating the second pre-training initial weights for the (m+1)th round of production, only the affected sub-matrices are initialized, while the remaining system state transition sub-matrices remain unchanged, to obtain the initialized system state transition matrix. It should be noted that: when determining the set A of affected sub-matrices... aff =A i,j After |(i,j)∈P, the central data center only applies to set A. aff Each affected submatrix in the set undergoes initialization processing, while the remaining system state transition submatrixes not belonging to the set retain the matrix estimation results corresponding to the end of the m-th production round. Preferably, the initialization method for the affected submatrix can be a preset initial matrix assignment, zero-mean random initialization, or prior assignment based on historical parameters of similar modules. For example, the affected submatrix can be initialized to A. i,j init For the unaffected submatrix A i,j ((i,j) If P), then retain the estimated result A obtained from the previous round of training. i,j (m) The state remains unchanged. Therefore, the initialized system state transition matrix can be formed as à = [Ã]. i,j ].

[0041] The system is trained based on the initialized system state transition matrix, the global dataset corresponding to the m-th production round, and the optimal pre-trained weights corresponding to the m-th production round, to generate the second pre-trained initial weights for the (m+1)-th production round. It should be noted that the central data center uses the optimal pre-trained weights corresponding to the m-th production round as the training starting point, combined with the initialized system state transition matrix à and the global dataset D corresponding to the m-th production round. m Iterative training is performed on the central pre-trained model to generate a second pre-trained initial weight suitable for the (m+1)th round of production.

[0042] Training employs a loss function that includes a fitting error term for the affected submatrix. This affected submatrix fitting error term constrains the difference between the trained affected submatrix and the initialized affected submatrix. It should be noted that, to ensure the generated second pre-trained initial weights both retain historical training results and accurately reflect the local dynamics after module replacement, a loss function including the affected submatrix fitting error term is constructed during training. Preferably, the total loss function can be expressed as: L c (θ)=L(θ,D m )+αL aff (θ,Ã)+μR(θ,D m ) Among them, L aff (θ,Ã)=Σ (i,j)∈P |Â i,j (θ)-Ã i,j || F 2 .

[0043] By introducing a fitting error term for the affected submatrix, the training process can focus on constraining the affected submatrix to converge toward the initialized target structure.

[0044] The technical solution implemented in this embodiment enables initialization and adaptation training to be performed only on the affected submatrix after a production module is replaced, while keeping the remaining system state transition submatrix unchanged, thereby avoiding a complete reset of all dynamic parameters of the entire production line. Furthermore, by introducing a fitting error term for the affected submatrix, the training focus can be concentrated on the affected region, improving the matching degree of the second pre-trained initial weights to the switched system.

[0045] In one specific embodiment, when training based on the global dataset corresponding to the (m+1)th round of production and the first pre-trained initial weights or the second pre-trained initial weights, if the production module is not changed when the sequence S is replaced, the first pre-trained initial weights are used as the initial weights for training.

[0046] If the production module is changed due to the replacement of sequence S, the second pre-training initial weights are used as the initial training weights.

[0047] The training employs a loss function that includes a continuous learning regularization term to control the rate of weight change between adjacent production rounds, and estimates the system state transition matrix based on the optimal pre-trained weights obtained from the training on the central side.

[0048] It should be noted that when the replacement sequence S represents that no production module has been changed in the (m+1)th round of production, the central data center can directly inherit the optimal pre-trained weights corresponding to the mth round of production as the first pre-trained initial weights and use them as the starting weights for training on the central side; when the replacement sequence S represents that a production module has been changed in the (m+1)th round of production, the central data center uses the second pre-trained initial weights generated in the aforementioned steps as the starting point for training.

[0049] Regardless of whether the first or second pre-training initial weights are used, the central data center can construct a loss function containing a continuously learned regularization term based on the global dataset produced in the (m+1)th round to obtain the optimal pre-training weights on the central side. Preferably, the total loss function can be expressed as: L m+1 (θ)=L(θ,D m+1 )+λΩ(θ,θ m * ) Where, L(θ, D) m+1 ) represents the global dataset D corresponding to the (m+1)th production round. m+1 The basic training loss term, Ω(θ, θ) m * ) represents the continuous learning regularization term, and λ represents the weight coefficient of the regularization term. The continuous learning regularization term is used to constrain the degree of deviation between the weights obtained in the current training round and the optimal pre-trained weights in the previous round, so as to avoid excessive parameter drift of the central side model under the drive of the new round of production data.

[0050] The central data center can update parameters in the following ways: θ (n) =θ (n-1) -η θ[L(θ (n-1) D m+1 )+λΩ(θ (n-1) θ m * )] When two consecutive iterations satisfy ||θ (n) -θ (n-1) When || < ε, the current weight can be determined as the optimal pre-trained weight θ on the central side. m+1 * After obtaining θ m+1 * Subsequently, the central data center can output the system state transition matrix A based on the trained model. m+1 .

[0051] The technical solution implemented in this embodiment can select either the first pre-training initial weight or the second pre-training initial weight as the training starting point during the central-side training phase, depending on whether the production module has been changed. This allows the model training when no module has been changed to inherit the global knowledge from the previous round of production, while the model training when a module has been changed can complete a global update based on local adaptation. Furthermore, by introducing a continuous learning regularization term during the training process, the drastic fluctuations of the model under the drive of a new round of production data can be reduced, improving the continuity and stability of the central-side training.

[0052] In one specific embodiment, each edge intelligent control device uses the optimal pre-trained weights on the central side as the initial weights for local training, and performs iterative training based on the local datasets corresponding to each production module to obtain the optimal local training weights.

[0053] When the difference between the training weights obtained from two adjacent iterations is less than the corresponding convergence threshold, the current training weight is determined to be the local optimal training weight.

[0054] It should be noted that after completing the central-side training, the central data center distributes the optimal pre-trained weights to the edge intelligent control devices corresponding to each production module. Each edge intelligent control device uses these optimal pre-trained weights as the initial parameter values ​​for its local training model. Let φ be the local training weight of the i-th production module in the (m+1)-th round of production. i,m+1 Then its initial value can be set to satisfy φ i,m+1 (0) =θ m+1 * .

[0055] Each edge intelligent control device further bases its data on the local dataset D of the corresponding production module. i，m+1 Iterative training is performed to make the model parameters better fit the local dynamic characteristics and operating conditions of this module. Preferably, the local dataset D i,m+1 This can include at least the pH value, feeding status, valve status, and environmental status data corresponding to this module. Edge intelligent control devices can update locally trained weights using gradient descent, stochastic gradient descent, mini-batch gradient descent, or adaptive learning rate optimization methods, for example: φ i,m+1 (n) =φ i,m+1 (n-1) -η i φ i L i (φ i,m+1 (n-1) D i，m+1 ).

[0056] When two consecutive iterations satisfy ||φi,m+1 (n) -φ i,m+1 (n-1) ||<ε i When this happens, the current training weights can be determined as the local optimal training weights φ. i,m+1 * .

[0057] The technical solution implemented in this embodiment enables each edge intelligent control device to further learn local operating condition features using local data from the corresponding production module, based on the global dynamic knowledge contained in the optimal pre-trained weights of the central side. This improves the adaptability of the local training results to the actual operating state of the module. Furthermore, by setting a convergence threshold, the local training time can be controlled while ensuring sufficient training, thereby improving the efficiency of edge-side model updates.

[0058] In one specific embodiment, the locally optimal training weights are used to estimate the system state transition matrix, control state transition matrix, and perception matrix of the corresponding production module.

[0059] Based on the estimated system state transition matrix, control state transition matrix, and sensing matrix, an optimization problem corresponding to the production objective is constructed to solve for the optimal control sequence.

[0060] It should be noted that after obtaining the local optimal training weights for the corresponding production module, each edge intelligent control device can identify and update the local dynamic model parameters of its module based on these weights to obtain the system state transition matrix, control state transition matrix, and perception matrix that match the current production task. Preferably, for the i-th production module, its local dynamic process can be represented as: x i (k+1)=A i x i (k)+B i u i (k)+w i (k) y i (k)=C i x i (k)+v i (k) Among them, A i B represents the system state transition matrix. i C represents the control state transition matrix. i This represents the perception matrix.

[0061] After obtaining the corresponding production module A i B i and C iSubsequently, the edge intelligent control device can establish a local predictive control model oriented towards production goals based on the matrix, and construct an optimization problem accordingly. Preferably, it can be based on the state variable x at the current moment. i (k) and control input u i (k) predicts the state trajectory and output trajectory over several future control cycles, thereby obtaining information about the future control input sequence U. i =[u i (k),u i (k+1),…,u i (k+N c The prediction model of -1), where N c This represents the control time domain length. Based on production targets, the edge intelligent control device constructs an optimization problem about the future control input sequence and solves it to obtain the optimal control sequence U for the corresponding production module. i * .

[0062] The technical solution implemented in this embodiment enables edge intelligent control devices to make targeted estimates of the system state transition matrix, control state transition matrix, and perception matrix of the corresponding production module based on local optimal training weights, thereby establishing a dynamic prediction model adapted to the current local operating conditions. Furthermore, by constructing an optimization problem for the corresponding production target based on the estimated model parameters and solving for the optimal control sequence, control decisions can be directly based on the local dynamics of this module, improving the targeting of control.

[0063] In one specific embodiment, the objective function of the optimization problem includes a tracking error term between the sensing sequence and the corresponding production target, and a control cost term for the control sequence, plus control magnitude constraints and / or control rate of change constraints.

[0064] The optimization problem is resolved in each control cycle based on the latest sensing data to continuously update the optimal control sequence.

[0065] It should be noted that after each edge intelligent control device estimates the system state transition matrix, control state transition matrix, and perception matrix of its corresponding production module based on the local optimal training weights, it can further construct an optimization problem in the prediction time domain around the current production task. For the control task of the i-th production module in the (m+1)-th round of production, a production target sequence r can be constructed. i，m+1 =[r i，m+1 (1) r i，m+1 (2) ,...,r i，m+1 (Np) Correspondingly, a perceptual sequence y can be constructed. i，m+1 and control sequence u i，m+1 Among them, Np This indicates the prediction time domain length, which can be set to any integer value between 5 and 50 control cycles, depending on the system response speed and real-time requirements. For online pH adjustment scenarios in drinking mineralized water, production objectives may include maintaining the pH value within a target range.

[0066] The objective function of the optimization problem can be expressed as: J i,m+1 = (y i,m+1 (k)-r i,m+1 (k)) T Q(y i,m+1 (k)-r i,m+1 (k))+u i,m+1 T(k)Ru i,m+1 (k) The first term is the tracking error term, and the second term is the control cost term.

[0067] While constructing the objective function, each edge intelligent control device also applies control magnitude constraints and / or control rate of change constraints to the control sequence. For scenarios using incremental control input, Δu can also be... i,m+1 (k)=u i,m+1 (k)-u i,m+1 (k-1) is used as the constraint object.

[0068] During operation, each edge intelligent control device employs a rolling time-domain optimization approach. After receiving the latest sensing data in each control cycle, it updates or corrects the current system state and re-solves the optimization problem to obtain a new optimal control sequence. In the current control cycle, only the control quantity corresponding to the current moment in the optimal control sequence is output as the actual control command. For the nonlinear process of pH change in mineralized water, a linearized prediction model can be constructed near the local operating point, and a feedback correction quantity can be generated based on the deviation between the latest obtained pH measurement value and the predicted value to compensate for model mismatch and disturbance effects.

[0069] The technical solution implemented in this embodiment enables the corresponding production module to meet production targets while simultaneously considering the economy and stability of control inputs by simultaneously setting tracking error and control cost terms in the objective function. Furthermore, by superimposing control amplitude constraints and / or control rate of change constraints, the optimal control sequence obtained satisfies the actual equipment capacity boundaries and process safety requirements. Moreover, by resolving the optimization problem based on the latest sensing data in each control cycle and continuously updating the optimal control sequence, the dynamic response speed and steady-state control accuracy of online pH adjustment for drinking mineralized water can be improved.

[0070] In one specific embodiment, after the (m+1)th round of production is completed, the newly added local data uploaded by each production module is received, and the newly added local data is added to the global dataset corresponding to the (m+1)th round of production for use in pre-training updates in subsequent production rounds.

[0071] The data collected by each sensing device includes at least one or more of the following: pH value, water quality data, feeding status, valve status, and environmental data.

[0072] It should be noted that after the (m+1)th round of production, the edge intelligent control devices corresponding to each production module can upload the newly generated local data during this round of production to the central data center. This newly generated local data may include the sensing sequences, control input sequences, local parameter estimation results, and / or local processing results from this round of production. Upon receiving this data, the central data center can add it to the global dataset corresponding to the (m+1)th round of production, forming an extended data foundation for subsequent production rounds. Through this data feedback method, the central data center can continuously accumulate production data from multiple modules and operating conditions during the production rounds, thereby providing richer data support for subsequent pre-training updates.

[0073] The data collected by each sensing device includes at least one or more of the following: pH value, water quality data, feed status, valve status, and environmental data. Preferably, water quality data may include one or more of the following: conductivity, turbidity, temperature, and dissolved solids content; environmental data may include ambient temperature, ambient humidity, and / or external disturbance information. The data types collected by the sensing devices can be the same for different production modules, or they may differ depending on the module's function. Before adding new local data to the global dataset, the central data center can also perform time alignment, noise reduction, normalization, and label association processing on it to improve the consistency and usability of subsequent training data.

[0074] The technical solution implemented in this embodiment enables each production module to continuously transmit newly added local data back to the central data center after completing the current round of production, and incorporate it into the global dataset. This provides a continuously expanding data foundation for pre-training updates in subsequent production rounds. This improves the central model's coverage and learning capabilities across different operating conditions and task switching scenarios. Furthermore, by collecting multi-dimensional information such as pH value, water quality data, feeding status, valve status, and environmental data through sensing devices, the data integrity for subsequent model training and parameter estimation can be improved, providing more comprehensive data support for online pH adjustment of drinking mineralized water.

[0075] like Figure 2As shown, the present invention also provides an online pH adjustment system for drinking mineralized water based on a switching system, which is applied to the online pH adjustment method for drinking mineralized water based on a switching system according to any of the above claims, including a central data center 200 and multiple production modules 100, each of which is communicatively connected to the central data center 200.

[0076] Each production module 100 includes a sensing device 110, a regulating device 130, and an edge intelligent control device 120, while the central data center 200 includes a data processing module 220 and a communication module 210.

[0077] The sensing device 110 is used to collect production data from the corresponding production module 100 and send the production data to the edge intelligent control device 120.

[0078] The edge intelligent control device 120 is used to receive production data and send the production data and / or local processing results to the central data center 200 through the communication module 210.

[0079] The data processing module 220 is used to receive the replacement information of the production module 100 and generate training results based on the production data and the replacement information of the production module 100.

[0080] The communication module 210 is also used to send the training results to the edge intelligent control device 120 of each production module 100.

[0081] The edge intelligent control device 120 is also used to generate control commands based on training results and production data, and output the control commands to the regulating device 130.

[0082] The regulating device 130 is used to perform regulating operations according to control commands.

[0083] In this embodiment, it should be noted that the online pH adjustment system for drinking mineralized water based on the switching system adopts a collaborative architecture of "central data center 200 - production module 100 - edge intelligent control device 120". Multiple production modules 100 are arranged in series and / or parallel along the production line to achieve continuous treatment of drinking mineralized water at different process stages. The sensing devices 110 in each production module 100 are used to collect real-time data on the current process status. The collected production data may include at least one or more of the following: pH value, water quality data, feeding status, valve status, and environmental data. Water quality data may include conductivity, turbidity, temperature, dissolved solids content, flow rate, and / or pressure data. The sensing devices 110 send the collected production data to the corresponding edge intelligent control device 120. The edge intelligent control device 120 receives, caches, and preprocesses the production data locally, and uploads the production data and / or local processing results to the central data center 200 via the communication module 210.

[0084] The data processing module 220 in the central data center 200 is mainly used to receive replacement information from production modules 100 and perform unified processing and training on the production data uploaded by each production module 100 to generate training results. The replacement information from production modules 100 can be sent to the central data center 200 by the production task management system, scheduling system, manual configuration terminal, and / or the edge intelligent control device 120 corresponding to the production module 100, representing the changes in module composition from the (m+1)th round of production to the mth round. Preferably, the data processing module 220 can perform time alignment, outlier removal, missing value completion, noise filtering, and normalization processing on the production data from different production modules 100 to form a global dataset for training the central-side model. Simultaneously, the data processing module 220 can also standardize the replacement information from production modules 100 into a replacement sequence and identify the impact range of module switching on system dynamics based on the replacement sequence. Based on this, the data processing module 220 further outputs training results. The training results may include at least one or more of the central-side optimal pre-trained weights and the system state transition matrix.

[0085] After receiving the training results, the edge intelligent control device 120 can perform local training, parameter estimation, and control decisions based on the training results and the corresponding production data of this module. The regulating device 130 can be a dosing device, a feeding device, a valve actuator, a pump control device, or a combination thereof, used to perform feeding regulation, opening regulation, flow regulation, or other process regulation operations according to control commands, thereby intervening in the operating status of the corresponding production module 100. Through the above structure, the system as a whole forms a closed-loop information flow and control flow of "sensing device 110 acquisition—edge device reception and uploading—central side processing and training—central side issuing results—edge side generating control commands—regulating device 130 executing regulation".

[0086] The technical solution implemented in this embodiment enables integrated processing of production data aggregation, training result distribution, and local control execution through the collaborative cooperation between the central data center 200 and multiple production modules 100. Because the functional division and connection relationships among the sensing device 110, edge intelligent control device 120, data processing module 220, communication module 210, and adjustment device 130 are clearly defined, the system can quickly complete data feedback, training result generation, and control command output after a production module 100 is replaced, thereby improving the real-time performance and operational stability of online pH adjustment for drinking mineralized water.

[0087] In one specific embodiment, the data processing module 220 includes a switching adaptation unit 222 and a central training unit 221, with the switching adaptation unit 222 connected to the central training unit 221.

[0088] The edge intelligent control device 120 includes a local training unit 122 and a scrolling control unit 121, with the local training unit 122 connected to the scrolling control unit 121.

[0089] The switching adaptation unit 222 is used to receive the replacement information of the production module 100, generate the initialization result of the affected submatrix based on the production data, and send the initialization result to the central training unit 221.

[0090] The central training unit 221 is used to receive the initialization results and generate training results based on the initialization results and production data.

[0091] The local training unit 122 is used to receive training results and production data, generate parameter estimation results, and send the parameter estimation results to the rolling control unit 121.

[0092] The rolling control unit 121 is used to receive parameter estimation results and generate control commands.

[0093] In this embodiment, it should be noted that: the data processing module 220 includes a switching adaptation unit 222 and a central training unit 221, wherein the switching adaptation unit 222 is connected to the central training unit 221 and is used to complete the conversion from module replacement information to model initialization results when the production module 100 is replaced; the edge intelligent control device 120 includes a local training unit 122 and a rolling control unit 121, wherein the local training unit 122 is connected to the rolling control unit 121 and is used to complete local parameter estimation and control command generation after receiving the training results from the central side.

[0094] The switching adaptation unit 222 receives the replacement information of the production module 100, generates the initialization result of the affected sub-matrices based on the production data, and sends the initialization result to the central training unit 221. Preferably, the switching adaptation unit 222 can first standardize the replacement information of the production module 100 into a replacement sequence S=[E, ..., F], and then determine the set of affected sub-matrices in the system state transition matrix according to the replacement sequence. For the affected sub-matrices, the switching adaptation unit 222 can generate the initialization result by using a preset initial matrix assignment, random initialization, or migration based on historical parameters, while keeping the original estimated value unchanged for the unaffected system state transition sub-matrices.

[0095] The central training unit 221 receives the initialization results and generates training results based on the initialization results and production data. Preferably, the central training unit 221 can perform model training based on the optimal pre-trained weights corresponding to the previous round of production, combined with the global production data corresponding to the current round; when the production module 100 is not changed, the central training unit 221 can directly use the inherited pre-trained initial weights as the training starting point; when the production module 100 is changed, the central training unit 221 can first complete local adaptation based on the initialization results and historical production data provided by the switching adaptation unit 222, and then perform continuous learning and training on the central side in combination with the production data of the current round.

[0096] The local training unit 122 receives training results and production data, generates parameter estimation results, and sends the parameter estimation results to the rolling control unit 121. Preferably, after receiving the training results output by the central training unit 221, the local training unit 122 uses the optimal pre-trained weights from the central side as the initial parameter values ​​for local training, and then performs local training by combining them with the local production data corresponding to the current production module 100. Based on the local training results, the local training unit 122 can output parameter estimation results, which include at least one or more of the following: the system state transition matrix, the control state transition matrix, and the perception matrix corresponding to the production module 100.

[0097] The rolling control unit 121 receives parameter estimation results and generates control commands. Preferably, the rolling control unit 121 can construct a discrete state-space prediction model based on the parameter estimation results and establish an optimization problem around the corresponding production target; in the current control cycle, it only outputs the control quantity corresponding to the current moment in the optimal control sequence as a control command and sends it to the regulating device 130 to perform the regulating operation. The regulating device 130 can be a dosing device, a feeding device, a valve actuator, a pump control device, or a combination thereof, and the control commands are used to control the regulating device 130 to perform one or more operations among feeding regulation, valve opening regulation, and flow regulation.

[0098] The technical solution implemented in this embodiment enables the system to structurally possess the capabilities of module replacement and adaptation, central-side training update, edge-side local parameter estimation, and rolling control execution by setting up a switching adaptation unit 222 and a central training unit 221 within the central data center 200, and a local training unit 122 and a rolling control unit 121 within the edge intelligent control device 120. This facilitates the step-by-step completion of model updates and control adjustments after module replacement along a clear internal unit chain, further improving the data connection efficiency between central-side training and edge-side control, as well as the response speed of online pH adjustment.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for online pH adjustment of drinking mineralized water based on a switching system, executed by a central data center, applied to a production line communicatively connected to the central data center, the production line comprising multiple production modules, each production module including sensing devices, adjustment devices, and edge intelligent control devices, characterized in that... include: Obtain the global dataset corresponding to the m-th round of production, the local dataset corresponding to each production module, and the optimal pre-trained weights corresponding to the m-th round of production. Before the start of the (m+1)th round of production, receive the production module replacement information and form a replacement sequence S; Determine whether a production module replacement occurs in the (m+1)th round of production based on the replacement sequence S. If no production module is replaced, the optimal pre-trained weights corresponding to the m-th round of production are determined as the first pre-trained initial weights for the (m+1)-th round of production. If a production module is replaced, the affected submatrix in the system state transition matrix is ​​determined, and the second pre-trained initial weights for the (m+1)th round of production are generated. Training is performed based on the global dataset corresponding to the (m+1)th round of production and the corresponding first or second pre-training initial weights to obtain the optimal pre-training weights on the central side and the system state transition matrix. The optimal pre-trained weights on the central side and the system state transition matrix are sent to the edge intelligent control devices of each production module, so that each edge intelligent control device can perform local training, parameter estimation and rolling control to realize online adjustment of the pH value of drinking mineralized water.

2. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 1, characterized in that, The system state transition matrix is ​​divided into multiple system state transition sub-matrices according to the correlation between production modules; Based on the replacement sequence S, the system state transition submatrix corresponding to the production module that has been replaced is determined from the plurality of system state transition submatrixes as the affected submatrix; When the replacement sequence S is represented as S=[E, ..., F], construct an index set. P = {(i, j) | E ≤ i ≤ F or E ≤ j ≤ F}, and the system state transition submatrix that satisfies (i, j) ∈ P is determined as the affected submatrix.

3. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 2, characterized in that, When generating the second pre-training initial weights for the (m+1)th round of production, only the affected submatrices are initialized, while the remaining system state transition submatrices remain unchanged, to obtain the initialized system state transition matrix. The system state transition matrix, the global dataset corresponding to the m-th production round, and the optimal pre-trained weights corresponding to the m-th production round are used to train the second pre-trained initial weights for the (m+1)-th production round. The training employs a loss function that includes a fitting error term for the affected submatrix, which is used to constrain the difference between the trained affected submatrix and the initialized affected submatrix.

4. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 1, characterized in that, When training based on the global dataset corresponding to the (m+1)th round of production and the first pre-trained initial weights or the second pre-trained initial weights, if the replacement sequence S represents that no production module replacement has occurred, then the first pre-trained initial weights are used as the initial weights for training. If the replacement sequence S represents a change in the production module, then the second pre-training initial weights are used as the initial training weights. The training employs a loss function that includes a continuous learning regularization term to control the rate of weight change between adjacent production rounds, and estimates the system state transition matrix based on the optimal pre-trained weights obtained from the training on the central side.

5. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 1, characterized in that, Each of the edge intelligent control devices uses the optimal pre-trained weights on the central side as the initial weights for local training, and performs iterative training based on the local datasets corresponding to each of the production modules to obtain the optimal local training weights. When the difference between the training weights obtained from two adjacent iterations is less than the corresponding convergence threshold, the current training weight is determined to be the local optimal training weight.

6. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 5, characterized in that, The local optimal training weights are used to estimate the system state transition matrix, control state transition matrix, and perception matrix of the corresponding production module. Based on the estimated system state transition matrix, control state transition matrix, and sensing matrix, an optimization problem corresponding to the production objective is constructed to solve for the optimal control sequence.

7. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 6, characterized in that, The objective function of the optimization problem includes a tracking error term between the sensing sequence and the corresponding production target, and a control cost term for the control sequence, plus control magnitude constraints and / or control rate of change constraints. The optimization problem is resolved in each control cycle based on the latest sensing data to continuously update the optimal control sequence.

8. The method for online pH adjustment of drinking mineralized water based on a switching system according to claim 1, characterized in that, After the (m+1)th round of production is completed, the newly added local data uploaded by each of the production modules is received, and the newly added local data is added to the global dataset corresponding to the (m+1)th round of production for use in pre-training updates in subsequent production rounds. The data collected by each of the aforementioned sensing devices includes at least one or more of the following: pH value, water quality data, feeding status, valve status, and environmental data.

9. An online pH adjustment system for drinking mineralized water based on a switching system, applied to the online pH adjustment method for drinking mineralized water based on a switching system as described in any one of claims 1-8, characterized in that, It includes a central data center and multiple production modules, each of which is communicatively connected to the central data center; Each of the production modules includes sensing devices, adjustment devices, and edge intelligent control devices, and the central data center includes a data processing module and a communication module. The sensing device is used to collect production data from the corresponding production module and send the production data to the edge intelligent control device; The edge intelligent control device is used to receive the production data and send the production data and / or local processing results to the central data center through the communication module; The data processing module is used to receive production module replacement information and generate training results based on the production data and the production module replacement information; The communication module is also used to send the training results to the edge intelligent control devices of each of the production modules; The edge intelligent control device is also used to generate control commands based on the training results and the production data, and output the control commands to the adjustment device; The regulating device is used to perform regulating operations according to the control command.

10. The online pH adjustment system for drinking mineralized water based on a switching system according to claim 9, characterized in that, The data processing module includes a switching adaptation unit and a central training unit, and the switching adaptation unit is connected to the central training unit. The edge intelligent control device includes a local training unit and a scrolling control unit, wherein the local training unit is connected to the scrolling control unit; The switching adaptation unit is used to receive the production module replacement information, generate the initialization result of the affected submatrix based on the production data, and send the initialization result to the central training unit. The central training unit is used to receive the initialization result and generate the training result based on the initialization result and the production data; The local training unit is used to receive the training results and the production data, generate parameter estimation results, and send the parameter estimation results to the rolling control unit; The rolling control unit is used to receive the parameter estimation results and generate the control commands.

Images