Electrochemical descaling automatic control system based on autonomous prediction of pole reversal time sequence and operation mode

Abstract

The invention relates to the technical field of water treatment, and discloses an electrochemical descaling automatic control system based on autonomous prediction of a pole reversal time sequence and an operation mode, the system comprises a sensing layer used for collecting raw water quality parameters and system state parameters in real time, and the system state parameters comprise real-time scale thickness and real-time scale thickness; the prediction layer is connected with the sensing layer, the prediction layer is connected with the sensing layer, the acquired data are utilized, the future scale thickness and the descaling efficiency under different control strategies are predicted through an AI prediction model, and the decision-making layer is connected with the prediction layer and used for solving an optimization problem with constraints according to a prediction result, and an optimal reverse-pole power supply control strategy is generated. The system can sense the dynamic change of the water quality of industrial raw water and the equipment state in real time, dynamically predict the scaling trend through the AI model and generate an optimally matched inverse pole strategy, so that the system can automatically adjust the equipment descaling state to the optimal working point according to the water quality fluctuation, efficient descaling is realized, and the effect is ensured to be continuous and stable.

Description

Technical Field

[0001] This invention relates to the field of water treatment technology, specifically to an electrochemical descaling automatic control system based on autonomous prediction of polarity reversal timing and operating mode. Background Technology

[0002] During traditional symmetrical reverse polarity switching, the pH value of the water can jump by 1.2 units within 10 seconds, causing irreversible polymerization of colloidal fouling. At the physical structure level, the electric field distribution formed by fixed-spacing electrodes is uneven, and the edge effect of the electrodes results in a 30% descaling blind zone. Even after treatment with traditional equipment, a 0.3-0.5mm thick scale layer still remains on the inner wall of the heat exchange tubes. While existing electrochemical reverse polarity technology has environmental advantages, there is still significant room for improvement in its adaptability to complex water qualities. Furthermore, it needs to address the challenges posed by the dynamic characteristics of water quality to the descaling system, such as in industrial scenarios where the makeup water source needs to be switched. (For example, switching from municipal water to groundwater) can cause calcium hardness to increase fourfold within one hour. Sudden pH changes caused by process leaks (such as a drop in pH from 8.5 to 4.2 due to a leak in pickling waste liquid) can completely change the reaction balance of electrochemical descaling. Moreover, the influence weight of different water quality parameters on the reverse polarity effect varies significantly. This dynamic characteristic makes it difficult to maintain stable descaling efficiency based on empirically set reverse polarity parameters (such as cycle and voltage). However, it shows obvious shortcomings in scenarios such as excessively long descaling cycles in high-hardness water (greater than 450 mg / L) water quality and difficulty in removing scale in high-SiO2 environments such as geothermal water.

[0003] Currently, traditional electrochemical descaling systems still have shortcomings in adapting to complex water conditions, making it difficult to maintain stable descaling efficiency. Moreover, the electrode wear during the switching process is significant, failing to meet the technical requirements of industrial users for long cycles, low maintenance, and high stability. Summary of the Invention

[0004] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an electrochemical descaling automatic control system based on autonomous prediction of electrode reversal timing and operating mode. This system solves the problems of traditional electrochemical descaling systems, which still have shortcomings in adapting to complex water quality, making it difficult to maintain stable descaling efficiency. Moreover, the electrode wear during the electrode reversal process is relatively large, which cannot meet the technical requirements of industrial users for long cycle, low maintenance, and high stability.

[0005] Technical solution To achieve the above objectives, the present invention provides the following technical solution: an electrochemical descaling automatic control system based on autonomous prediction of reversal timing and operating mode, comprising: The sensing layer is used to collect raw water quality parameters and system status parameters in real time, including real-time scale thickness. The prediction layer, connected to the perception layer, uses the collected data to predict future scale thickness and descaling efficiency under different control strategies through an AI prediction model. The decision layer, connected to the prediction layer, is used to solve constrained optimization problems based on the prediction results and generate the optimal reversing power supply control strategy, including reversing frequency (f), duty cycle (d), and operating current (I). The execution layer, connected to the decision layer, is used to receive the control strategy and drive the reverse power supply to perform the corresponding reverse operation; The perception layer, prediction layer, decision-making layer, and execution layer together form a complete closed loop of perception-prediction-decision-execution-feedback.

[0006] As a further description of the above technical solution, the sensing layer is a data acquisition module composed of various sensors and instruments, including a pH sensor, a conductivity sensor, a calcium ion selective electrode, a temperature sensor, a flow meter, an ultrasonic thickness gauge, a voltage and current sensor, and a differential pressure sensor. The sensing layer is equipped with a signal conditioning and data preprocessing module for filtering and standardizing sensor signals, and a data aggregation and communication module for protocol conversion and data transmission. The pH sensor, conductivity sensor, calcium ion selective electrode, and temperature sensor in the data acquisition module output 4-20mA analog current signals, which are then connected to the signal conditioning and data preprocessing module via shielded twisted-pair cables. The ultrasonic thickness gauge communicates with the data aggregation and communication module via an RS-485 interface using the Modbus RTU protocol. The signal conditioning and data preprocessing module is implemented by the analog input module of the PLC, responsible for A / D conversion and preliminary filtering. The core hardware of the data aggregation and communication module is a PLC or industrial gateway, which runs a periodic scanning control program to perform data acquisition, protocol conversion, and data uploading. The PLC acquires digital sensor data through the RS-485 interface and communicates with the upper-level decision-making layer via an Ethernet interface using Modbus TCP / IP, OPC UA, or MQTT protocols.

[0007] As a further description of the above technical solution, the AI ​​prediction model is a multi-task spatiotemporal perception network, which includes, in sequence: Bidirectional Long Short-Term Memory (Bi-LSTM) layers are used to extract features from input time-series data and capture their long-term bidirectional dependencies. The attention mechanism layer is used to weight the feature sequence output by the Bi-LSTM layer to focus on key time steps and key features. A multi-task output layer is used to simultaneously calculate and output the predicted scale thickness and the predicted descaling efficiency based on the output of the attention mechanism layer.

[0008] As a further description of the above technical solution, the AI ​​prediction model is trained by minimizing a multi-task loss function, which is composed of a weighted sum of a first loss term for predicting scale thickness and a second loss term for predicting descaling efficiency. Both the first and second loss terms adopt robust loss functions that are insensitive to outliers.

[0009] As a further description of the above technical solution, the decision layer is configured to solve an optimization problem, the optimization objective of which is to minimize the system's operating energy consumption, and must simultaneously satisfy the following four types of constraints: 1) Descaling effect constraint to ensure that the predicted final scale thickness does not exceed the safety threshold; 2) Scaling trend constraints ensure that the adopted strategies can produce a clean descaling effect; 3) Hardware safety constraints to ensure that the operating current, reversing frequency, and duty cycle of the reverse-polarity power supply operate within their permissible physical limits; 4) Operational feasibility constraints to ensure that the electrode reversal time is sufficient to effectively remove scale.

[0010] As a further description of the above technical solution, the decision layer uses a Bayesian optimization algorithm to solve the optimization problem. The Bayesian optimization algorithm selects the next candidate strategy to be evaluated through a collection function. The collection function is configured to balance the trade-off between utilizing the currently known better region and exploring uncertain but potential regions.

[0011] As a further description of the above technical solution, the polarity reversal mode executed by the execution layer includes at least one of the following: standard balance mode, powerful descaling mode, energy-saving maintenance mode, cathode cleaning mode, and anode protection mode.

[0012] As a further description of the above technical solution, the system also includes a continuous learning module, used to automatically execute the following steps during system idle periods: Step 1: Synchronize recently run data from the database; Step 2: Using the current model weights as pre-training values, fine-tune the AI ​​prediction model on new data; Step 3: Verify the performance of the new model. If it is improved, deploy it online via shadow mode.

[0013] This invention also provides a self-cleaning electrochemical descaling method, comprising: 1. a sensing step, which involves real-time acquisition of raw water quality parameters and system state parameters, including scale thickness; 2. a prediction step, which involves predicting future scale thickness and descaling efficiency based on the acquired data using an AI prediction model; 3. a decision-making step, which involves generating an optimal reversing power supply control strategy based on the prediction results; and 4. an execution step, which involves receiving the control strategy and driving the reversing power supply to perform the corresponding reversing operation.

[0014] As a further description of the above technical solution, the decision-making step includes using a Bayesian optimization algorithm to minimize energy consumption and satisfy the requirements of descaling effect and hardware safety, to solve for the optimal reversal frequency (f), duty cycle (d), and operating current (I).

[0015] Beneficial effects Compared with existing technologies, this invention provides an electrochemical descaling automatic control system based on autonomous prediction of reversal timing and operating mode, which has the following beneficial effects: 1. This system can sense the dynamic changes in industrial raw water quality (such as hardness, pH, and flow rate) and equipment status (such as scale thickness and impedance) in real time, and dynamically predict the scaling trend through an AI model. This generates an inverted polarity strategy that is optimally matched with the dynamic changes in industrial raw water quality and equipment status, changing the "uniform" working mode of traditional descaling systems. The system can automatically adjust the descaling status of the equipment to the optimal working point according to water quality fluctuations, achieving efficient descaling and ensuring continuous and stable results.

[0016] 2. With minimizing energy consumption as the optimization objective, the decision algorithm can accurately calculate the optimal current, frequency, and duty cycle while ensuring descaling effectiveness, avoiding energy waste. Especially under low hardness or low load conditions, the system can automatically enter a low-energy-consumption maintenance mode. Compared with traditional systems that always operate at full load, this can significantly reduce power consumption and operating costs. 3. It integrates rule-based emergency response, optimization-based intelligent decision-making, and experience-based fuzzy backup strategies, ensuring that the system can still operate safely and without losing control under abnormal conditions (such as sudden changes in water quality or sensor malfunctions). The highest priority safety rule base can directly trigger emergency strategies, effectively preventing equipment blockage or damage, and the system has extremely high availability.

[0017] 4. The optimized reversal strategy can effectively avoid electrode coating corrosion and wear caused by unnecessary frequent reversal or excessively long reversal time, reduce the replacement frequency and maintenance cost of core components such as electrodes, extend the service life of the entire descaling system, and improve the return on investment. Attached Figure Description

[0018] Figure 1 This is a block diagram of the overall architecture of the system of the present invention; Figure 2 This is a schematic diagram showing the module composition and connection relationships of the perception layer; Figure 3 This is a schematic diagram of the structure of an AI prediction model (multi-task spatiotemporal perception network); Figure 4 Flowchart of a hybrid decision engine for the intelligent decision-making layer; Figure 5 This is a schematic diagram of the output waveforms of the execution layer system and the five inverted polarity modes. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments; Overall architecture of this system: refer to Figure 1 This invention, a self-cleaning electrochemical descaling system that autonomously predicts reversal timing and operating modes, is a cyber-physical system (CPS). Data flows from the sensing layer through the prediction, decision, and execution layers, ultimately impacting the physical reactor. Reactor state changes are captured by the sensing layer, forming a closed loop. Simultaneously, historical data feedback is used for model retraining and updating the prediction layer, enabling continuous system evolution. The core architecture of the above technical solution includes: The perception layer, acting as the system's sensory organ, is responsible for collecting all relevant data from the physical world, used to collect raw water quality parameters and system status parameters in real time.

[0021] The prediction layer, acting as the brain of the system, is used to predict future scaling trends and the effectiveness of different control strategies based on data from the perception layer through AI models.

[0022] The decision-making layer, as the central hub of the system, is used to generate the optimal inverted power supply control strategy based on the prediction results, while satisfying the system constraints.

[0023] The execution layer acts as the system's two hands, receiving instructions from the decision layer and driving the reverse power supply to perform the corresponding reverse operation.

[0024] The feedback learning module collects the results from the execution layer again from the perception layer to evaluate system performance and initiate continuous learning of the model.

[0025] Through the design of the above technical solution, compared with the bottlenecks of traditional descaling systems such as rapid electrode wear and poor adaptability to working conditions, this system controls a high-performance descaling power supply and has multiple descaling modes to cope with changes in industrial water quality. This enables the system to accurately, flexibly and safely cope with various complex working conditions, and ultimately achieve intelligent, adaptive, efficient and energy-saving descaling operations.

[0026] Detailed design of the perception layer: refer to Figure 2 The perception layer is the data entry point, and its design directly affects system performance.

[0027] The hardware components of the data acquisition module include a pH sensor, a conductivity (TDS) sensor, a calcium ion selective electrode, a temperature sensor, an ultrasonic thickness gauge, a voltage / current sensor, and a differential pressure sensor. The wiring and communication method involves the pH, conductivity, calcium ion, and temperature sensors outputting 4-20mA analog current signals, which are hardwired to the PLC's analog input (AI) module via shielded twisted-pair cables. It is powered by 24V DC. The ultrasonic thickness gauge outputs a 4-20mA signal or communicates with the PLC via an RS-485 interface using the Modbus RTU protocol. The voltage / current sensor outputs a 0-10V or 4-20mA signal, which is connected to the PLC's AI module.

[0028] The hardware for the signal conditioning and data preprocessing module is integrated into the analog input module of the PLC (such as Siemens SM331). The control logic is a periodic scan (e.g., 100ms) executed by the PLC's internal program. During the program execution phase, the raw digital quantities are converted into engineering quantities and filtered by moving average, and data alignment and timestamping are performed.

[0029] The core hardware of the data aggregation and communication module is an industrial PLC (such as a Siemens S7-1500) or an industrial gateway. The connection and control logic involves the PLC polling Modbus RTU device data via its RS-485 port and simultaneously reading data from the AI ​​module. Subsequently, the internal program packages all data and uploads it to the decision-making layer's local server via an Ethernet port using MQTT, OPC UA, or Modbus TCP / IP protocols. The system's data storage module is deployed on the decision-making layer's local server, employing a hybrid architecture combining time-series databases, object storage, and relational databases to store real-time sensor data, model files, and system metadata, respectively. This design ensures low latency and high availability for system data processing while meeting the security requirement of preventing industrial data from leaving the factory. To further ensure data security and maintainability, the module also includes data lifecycle management, a local high-availability cluster, and optional bidirectional data synchronization.

[0030] Prediction layer: This layer is the intelligent core of the system, employing a multi-task spatiotemporal perception network model, the structure of which includes: Bi-LSTM (Long Short-Term Memory) layers are used to extract features from input temporal data and capture their long-term bidirectional dependencies. Their computation is defined by the following formula: The computation process of the forward LSTM is defined by the following formula: Forgotten Gate:

[0031] Input Gate:

[0032] Candidate cell status:

[0033] Cell status update:

[0034] Output gate:

[0035] Hidden state output:

[0036] The computation process of backward LSTM is defined by the following formula: Forgotten Gate:

[0037] Input Gate:

[0038] Candidate cell status:

[0039] Cell status update:

[0040] Output gate:

[0041] Hidden state output:

[0042] Ultimately, Bi-LSTM at each time step Output It is a concatenation of the forward and backward hidden states:

[0043] in Represents the Sigmoid activation function. Represents the hyperbolic tangent activation function. element-wise multiplication , , , It is the weight matrix of the corresponding gate. , , , These are the bias vectors for the corresponding gates, while the weight matrix and bias vectors are the parameters that the model needs to learn.

[0044] The attention mechanism layer is used to weight the feature sequence output by the Bi-LSTM layer to focus on key time steps and key features, processing all hidden state sequences output by the Bi-LSTM layer. The process involves calculating the weights for each time step and generating a context vector. The calculation formula is as follows:

[0045]

[0046]

[0047] Here, the score function is a learnable feedforward neural network, for example... ,in , , These are learnable parameters; The multi-task output layer is used to simultaneously calculate and output the predicted scale thickness and the predicted descaling efficiency based on the output of the attention mechanism layer, and receives the context vector output by the attention mechanism layer. It outputs two predicted targets through two independent fully connected layers:

[0048]

[0049] in, It is the predicted future scale thickness. This is the predicted descaling efficiency. , It is a weight matrix. , It is the bias vector.

[0050] Model Training: The AI ​​prediction model is trained by minimizing a multi-task loss function. This multi-task loss function is a weighted sum of a first loss term for predicting scale thickness and a second loss term for predicting descaling efficiency. Both the first and second loss terms employ robust loss functions that are insensitive to outliers. The formula for the loss function is as follows: ; Huber Loss uses multi-task smoothing L1 loss as follows:

[0051] in, It is a hyperparameter that balances the weights of the two tasks. and The loss value for a single task. For the sample size, For the first in the dataset The true label value of each sample, For AI models to the first Prediction results for each sample For threshold; In the model training described above, TotalLoss represents the overall validation error of the model across all samples in the entire training dataset. It is also known as the objective function or cost function. The ultimate goal of the training process is to minimize the value of TotalLoss using an optimization algorithm (such as gradient descent). The lower the TotalLoss value, the closer the model's predicted scale thickness and descaling efficiency are to their true values ​​overall, and the higher the model accuracy.

[0052] (Lambda, weight hyperparameter) A pre-defined constant between 0 and 1, used to balance the importance and contribution of Task 1 (scale thickness prediction) and Task 2 (descaling efficiency prediction) to the total loss. Since the two tasks may have different dimensions and numerical ranges, simply adding them together will cause the model optimization to favor one task. An adjustment knob is provided. When When the value approaches 1, model training will focus on improving the accuracy of scale thickness prediction.

[0053] when When the value is close to 0, model training will focus on improving the accuracy of descaling efficiency prediction.

[0054] It is usually set according to business needs, for example, it can be set to = 0.7 indicates a greater focus on the prediction accuracy of the core parameter (scale thickness).

[0055] and The loss value for a single task; The loss value for Task 1 is the sum of the differences between the scale thickness predicted by the model and the actual scale thickness measured.

[0056] The loss value for Task 2, which is the sum of the differences between the descaling efficiency predicted by the model and the actual calculated descaling efficiency, is mainly used to quantify the model's performance on each specific task.

[0057] It is a specific function for calculating the loss of a single task, namely the smoothed L1 loss. It combines the advantages of mean squared error (MSE) and mean absolute error (MAE).

[0058] (Number of samples) is the total number of training samples in the batch used to calculate the loss value. The loss value is averaged over all samples in the batch (1 / N). Σ(...)) is used to obtain the average loss for that batch. This avoids fluctuations in the loss value due to different batch sizes, ensuring the stability of the training process.

[0059] (True value) in the dataset, the th The true label value of each sample. Specific meaning in the task: exist middle, This represents the actual thickness of the scale measured by an ultrasonic thickness gauge at a future time Δt.

[0060] exist middle, This represents the actual descaling efficiency calculated based on the actual reduction in scale thickness and energy consumption within a cycle T.

[0061] (Predicted value) is the AI ​​model's prediction for the first... The prediction results for each sample.

[0062] δ (threshold hyperparameter) is a pre-defined constant greater than 0. It is the threshold in the definition of Huber Loss, used to determine whether the prediction error of a sample is "small" or "large". It determines the critical point at which Huber Loss switches between mean squared error (MSE) and mean absolute error (MAE).

[0063] when When the error is small, the loss function is expressed as mean squared error (MSE). When the error is close to zero, its gradient is also close to zero. This allows the model to converge quickly and stably when it is close to the optimal solution, and the model parameters can be finely adjusted to approach the optimal solution.

[0064] when When the error is large, the loss is linearly related to the absolute value of the error, and its gradient magnitude remains constant (approximately ±δ). Compared to MSE (where the gradient increases linearly with the error), it is significantly less sensitive to outliers, effectively suppressing the excessive influence of outliers on model training, enhancing the robustness of the model, handling outliers or noise in the data, and preventing the model from deviating from the overall trend in order to fit a few outliers.

[0065] Intelligent decision-making layer: refer to Figure 4 The decision-making level employs a hybrid decision engine.

[0066] The optimization objective of the decision-making layer is to minimize energy consumption J:

[0067] min J means that the goal of the optimization algorithm is to find a set of operation variables that minimizes the objective function J. ; J is the objective function, which physically represents the electrical energy consumed by the system within a complete reversal period T. (Selection) This is because the energy loss (Joule heating) in the electrochemical descaling system is proportional to the square of the operating current; The operating current is the DC current output by the reverse-pole power supply (unit: ampere, A), which is one of the decision variables to be optimized. This is the system equivalent impedance (unit: ohms, Ω). It is a dynamic parameter determined by factors such as the electrolyte conductivity, electrode plate condition, and scale deposition thickness within the electrochemical reactor, and can be calculated from real-time monitored voltage U and current I. ; T is the reciprocal period (unit: seconds), which is the reciprocal of the reciprocal frequency f (T = 1 / f). f is one of the decision variables to be optimized; The following constraints must be met:

[0068] The constraints define the feasible solution space of the decision variable A(f, d, I), ensuring that the optimization strategy is within a safe and effective range.

[0069] Constraint (a): Endpoint fouling thickness constraint. The scale thickness predicted by the AI ​​model at time T, one cycle after the execution of strategy A (unit: mm).

[0070] This is the safe threshold for scale thickness, which is preset by the system process requirements. This constraint ensures that the optimized strategy can control future scale thickness within the allowable range.

[0071] Constraint (b): Scaling Tendency Constraint S(t) is the scale thickness (in mm) measured in real time by an ultrasonic thickness gauge at the current moment. This constraint is an inequality that requires the predicted endpoint scale thickness to be less than the current scale thickness. This ensures that the strategy adopted must produce a clean scale removal effect (scale thickness reduction) rather than slowing down or accelerating scale formation.

[0072] Constraints (c), (d), and (e) are hardware safety and operational constraints that ensure the reverse-polarity power supply operates within a safe range.

[0073] , The minimum and maximum allowable current (in amperes, A) that a reverse-polarity power supply can output are determined by the power supply hardware specifications. , The minimum and maximum permissible polarity reversal frequencies (in Hertz, Hz or times / second) that a reversible power supply can perform are determined by the polarity switching speed of the power supply and the control logic. , The minimum and maximum permissible values ​​(dimensionless) of duty cycle are usually... >0.5, <1, to ensure the existence of reverse time and the dominant proportion of forward time; Constraint (f) is the lower limit constraint of the reverse time, which ensures that the reverse time is long enough to effectively clean the electrode.

[0074] It is the reverse time (unit: seconds), which is the time during which the current flows in the reverse direction within one cycle. Its value is calculated from the duty cycle d and the frequency f.

[0075] This is the minimum reverse time (in seconds). This constraint ensures that the reverse time is long enough to effectively remove scale from the cathode plate, preventing scale buildup that can lead to efficiency degradation and malfunctions. Simultaneously, it also prevents damage to the electrodes from arcing caused by excessively short reverse times.

[0076] Decision-making process: Security alert rule base: highest priority, IF S(t) ≥ THEN emergency shutdown; IF Ŝ(t+Δt)≥S_alert, THEN activate the most powerful preset strategy.

[0077] Bayesian optimization (main path): using a Gaussian process as a surrogate model and the expected improvement (EI) as the acquisition function for iterative optimization.

[0078] in This is the current optimal energy consumption value.

[0079] 1. Initialization: Randomly measure or manually specify several initial strategies A to obtain their actual energy consumption. and determine the initial .

[0080] 2. Construct a surrogate model: Using these initial data points, fit a Gaussian process surrogate model that can predict any... of (Including mean and variance).

[0081] 3. Optimize the acquisition function: Within the entire policy space, use optimization algorithms (such as gradient descent) to find the optimal acquisition function. The candidate strategy A_next with the largest value.

[0082] 4. Evaluation: Input A_next into the AI ​​prediction model, calculate its predicted descaling effect, and determine whether all constraints are met. If they are met, calculate its estimated energy consumption J(A_next).

[0083] 5. Update: Add the new data point (A_next, J(A_next)) to the observation dataset, and update the surrogate model and the currently known optimal value. .

[0084] 6. Repeat steps 3-5 until the optimal solution that satisfies all constraints is found, or the maximum number of iterations is reached.

[0085] Value: This The formula enables the system to efficiently find the global optimum with as few evaluations as possible (since each call to the AI ​​model is computationally expensive). It intelligently balances "developing currently known good regions" and "exploring unknown but potentially better regions," and is the core algorithm that enables the entire system to achieve intelligent and adaptive optimization.

[0086] Hybrid architecture: The decision engine also integrates a security alert rule base: a set of IF-THEN rules for handling emergency situations such as exceeding thresholds, with the highest priority.

[0087] Fuzzy controller: When AI predictions are unreliable, it serves as a fallback solution by controlling based on expert experience rules.

[0088] Execution layer refer to Figure 5 After the execution layer receives strategy A(f, d, I), the PLC controls the reverse power supply to execute various modes, including but not limited to: Standard balance mode: parameters are medium f, high d, and medium I, used to balance efficiency and energy consumption when water quality is stable.

[0089] Powerful descaling mode: Parameters are high f, high d, and high I, used to maximize the descaling rate when the scaling trend accelerates rapidly.

[0090] Energy-saving maintenance mode: parameters are low f, medium d, and low I, used to minimize energy consumption when operating under low load.

[0091] Cathode cleaning mode: Parameters are very high f, low d, and medium I, used for powerful scale removal when the cathode plate has abnormally thick scale.

[0092] Anode protection mode: Parameters are low f, extremely high d, and on-demand I, used to protect the anode when the concentration of corrosive ions is high.

[0093] Feedback and Continuous Learning Phase 1: Data Preparation. The system automatically collects recent operational data, cleans and validates it to ensure data quality. Then, it performs feature processing using predefined rules and generates a versioned training dataset to prepare for training. The process is automatically started by a scheduler (such as Apache Airflow) during periods of low system load (such as 2 AM every Sunday). If the system detects that the prediction error of the online model continues to exceed the limit, emergency retraining can be automatically triggered. The workflow script extracts all recent (e.g., the past 90 days) historical running data from a time-series database (such as InfluxDB). It retrieves the exact version code of the feature engineering pipeline that is compatible with the current production model from a model repository (such as MLflow) or object storage (such as S3). Perform rigorous automated data quality checks to handle missing data issues, remove outliers that exceed physically reasonable ranges (e.g., pH=100), verify the logical relationships between relevant sensor readings (e.g., whether conductivity and ion concentration match), and automatically discard or repair problematic data segments.

[0094] Feature engineering and dataset creation: The loaded feature engineering pipeline is used to process the cleaned raw data (such as standardization, generating lagged features, and calculating statistics) to ensure that the features are completely consistent with those used during model training. The processed data is divided into training and test sets in chronological order (e.g., the first 80 days for training and the last 10 days for testing), and random shuffling is strictly prohibited. The final generated dataset is versioned and named (e.g., dataset_20231015_v1.parquet) and stored in the storage system to achieve complete reproducibility. The second stage, model fine-tuning, does not involve training from scratch. Instead, it loads the currently used production model and uses it as a foundation for fine-tuning on a new dataset. This approach offers fast training speed, effectively absorbs new knowledge, and avoids forgetting already learned patterns. Based on the new data, a more powerful candidate model is generated in an efficient and safe manner. Environment setup and model loading occur in an independent computing environment (such as a Docker container or Kubernetes Pod), instantiating a pre-configured training image and downloading the weight file and structure definition of the model (Model_vN) currently used in the production environment from the model repository.

[0095] The model is fine-tuned using a transfer learning approach: the weights of the production model Model_vN are used as pre-training weights, and training is performed on a new training dataset. A small learning rate and fewer training epochs are used to focus on fine-tuning, avoiding the destruction of general features already learned by the model (to prevent catastrophic forgetting). The validation set loss is monitored during the training process, and an early stopping mechanism is used to prevent overfitting.

[0096] Phase 3: Performance Evaluation. The new model must be fairly compared with the old model on a completely new, untrained dataset. Only when its key metrics (such as prediction accuracy) improve by more than a preset threshold (e.g., 5%) will it be approved to proceed to the next phase. Otherwise, the process will automatically terminate without affecting the existing system. Offline Performance Evaluation. On the test set prepared in Phase 3 that has never been trained, a comprehensive benchmark test is performed on both the candidate model and the current production model, using a set of pre-defined rigorous evaluation metrics, such as mean absolute error (MAE) and root mean square error (RMSE).

[0097] Model Comparison & Decision calculates the percentage improvement in key performance indicators. For example, if "the new model's MAE is reduced by 5% compared to the old model," the decision rule is: the new model is considered acceptable only if the performance improvement exceeds a preset threshold (e.g., a reduction in MAE ≥ 3%). If the decision result is satisfactory: register the new model (Model_vN+1), its performance report, and all metadata (hyperparameters, code version, dataset version) into the model repository and proceed to the next stage.

[0098] If the decision result is satisfactory: trigger an alarm to notify the engineer, record the log and stop the process, and the system continues to use the original production model.

[0099] Phase Four: Secure Deployment. The new model, having passed the evaluation, will not immediately take over the system but will first enter "shadow mode." In this mode, it runs in parallel, receiving real data and making predictions, but its predictions are not used for actual control; they are only used for final verification. After a period of monitoring to confirm its stable performance, it will be manually approved by operations personnel before being finally switched to live, fully replacing the old model. The deployment system starts the new model Model_vN+1 in "shadow mode," running it in parallel with the current production model Model_vN. Real-time production data is sent to both models simultaneously for prediction, but only the prediction results of Model_vN are used for actual control. The prediction results of Model_vN+1 are recorded and compared with subsequent real data, but its output is not linked to the execution layer. After running in shadow mode for a period of time (e.g., 24 hours), the system automatically compares the online prediction accuracy of the two models, generates a detailed performance comparison report, and confirms that the new model performs stably and excellently. The system then sends a report and a request for approval to go live to the operations team. This is a critical safety valve that requires final manual confirmation. After manual approval, the deployed system performs a blue-green deployment, switching all traffic to Model_vN+1, making it the new production model. During the monitoring period after the new model goes live, the system continuously monitors its performance. If any anomaly is detected, the system can automatically switch back to the stable Model_vN within minutes, ensuring absolute production safety.

[0100] Experimental example: Verification of the technical effectiveness of the dynamic operation control strategy: The experiment was conducted in a 500L circulating water simulation system. A flow-through reaction chamber (effective cross-sectional area 0.12 m²) was constructed using quartz glass and equipped with a constant temperature control system (±0.5°C) and an online water quality analysis unit. Artificially prepared hard water (a mixture of CaCl₂·2H₂O and MgSO₄·7H₂O in a specific ratio) was used, with water quality parameters dynamically adjusted via a precision metering pump. The control group used a mainstream timed polarity reversal device (6-hour cycle, symmetrical polarity switching), while the experimental group used the system described above. Each experiment was repeated three times, and the average value was taken. The significance level was set at p<0.05.

[0101] Descaling efficiency and energy consumption optimization: The results of a 120-hour continuous test under standard water quality conditions of TDS=800mg / L (calcium ions=120mg / L, magnesium ions=80mg / L, pH=7.8) are as follows:

[0102] The experimental group experienced a brief drop in descaling rate (to 87%) at the 48th hour. Source analysis revealed this to be an adaptive adjustment process of the LSTM model to a sudden turbidity disturbance (turbidity jumped from 15 NTU to 32 NTU). The system recovered to its optimal state within 2 hours by temporarily activating a strong eddy current mode (increasing eddy current velocity to 1.2 m / s). Energy consumption monitoring showed that the dynamic pole reversal strategy reduced the proportion of ineffective energy consumption from 38% in the traditional system to 12%, mainly due to the reduction in reactive power loss caused by the asymmetric pole reversal mode (average duty cycle 3.2:1).

[0103] Complex water quality adaptability test: In the calcium hardness impact test, the hardness increased rapidly from 200 mg / L to 800 mg / L (300% increase) by instantaneously injecting CaCl2 solution. Traditional systems experienced descaling failure for 4 hours (efficiency dropped to 51%) due to exceeding the processing capacity of the preset reversal cycle. However, the AI-ReFlux™ system triggered an emergency reversal procedure within 180 seconds: the electrode spacing was reduced to 12 mm (initial value 15 mm), the current density was increased to 25 A per square meter, and with the activation of the eddy current distribution chamber, the descaling efficiency was maintained above 85%. The pH value cycle fluctuation experiment (4→11→4, switching every 2 hours) showed that the system achieved descaling stability under acid and alkaline environments through ORP collaborative monitoring (control window 200-600mV), and the efficiency fluctuation range was controlled within ±5%. In a 3000ppm high-chlorine environment test lasting 72 hours, the IrO2 coating of the comb electrode showed excellent corrosion resistance, with the polarization rate increasing by only 0.005Ω / h compared to standard water quality, far lower than the 0.031Ω / h of the traditional titanium electrode.

[0104] The system automatically matches five operating modes based on water quality characteristics: standard balancing mode (normal water quality), powerful descaling mode (high hardness), cathode cleaning mode (routine cleaning of cathode plates), anode protection mode (pH fluctuation, high chlorine environment), and energy-saving maintenance mode (low scaling risk). The mode switching response time is less than 15 seconds, and there is no overshooting of water quality parameters during the switching process, verifying the robustness of the "prediction-decision-execution" closed loop.

[0105] Pilot-scale data from a food factory showed that the equipment maintained a descaling efficiency of over 90% even during 3-5 incidents of process wastewater mixing per month.

[0106] An application case study using a circulating cooling water system (flow rate 200 cubic meters / liter, hardness 450 mg / L, pH 8.5-9.0) in a steel plant illustrates the following:

[0107] After adopting the system of this invention, the electrode reversal cycle is shortened from the traditional fixed 30 min to an adaptive 9-15 min (adjusted according to real-time scale thickness), the number of electrode replacements per year is reduced from 4 times to 1 time, saving 120,000 yuan in replacement costs annually. The scale particle recovery rate is increased to 87%, the turbidity of circulating water is stabilized below 5 NTU, the heat exchanger cleaning cycle is extended from 3 months to 12 months, and the annual downtime losses are reduced by more than 500,000 yuan.

[0108] To improve the cleaning of scale on the cathode plate surface, this system can also work in conjunction with an ultrasonic-cavitation module to achieve a synergistic self-cleaning and scale removal effect of "electrochemical reverse dissolution + physical stripping". The specific process is as follows. The module consists of 4-8 20-40 kHz piezoelectric ceramic transducers (power 50-100 W / each) and a cavitation generator. The transducers are evenly arranged around the circumference of the reaction chamber (spacing 150-200 mm), with a sound pressure level of 120-150 dB. The cavitation generator uses a venturi tube, which releases 0.3-0.5 MPa compressed air instantaneously through a solenoid valve during descaling, generating a 50-200 ms cavitation pulse with a peak power density of 100-300 W / cm².

[0109] The synergistic mechanism is manifested as follows: ultrasonic vibration induces fatigue cracks in the scale layer (amplitude 5-10μm), and the micro-jets generated by cavitation bubble collapse (velocity 100-200m / s) propagate the cracks into macroscopic ablation. Through fiber optic synchronous triggering, the delay between ultrasonic and cavitation actions is ≤100 milliseconds, and the energy utilization rate is improved by 50% compared with asynchronous control.

[0110] The specific steps for descaling are as follows: During system startup, the system first enters constant current mode for 5-10 minutes, with an initial current density J0 of 50-100 A / m² (set according to the raw water hardness gradient: 50 A / m² for hardness ≤300 mg / L, 80 A / m² for 300-500 mg / L, and 100 A / m² for >500 mg / L). The current is gradually increased (by 10 A / m² per minute) to avoid impacting the electrode coating with instantaneous high current. Conductivity changes are monitored during this phase; startup is considered complete when the system conductivity stabilizes within ±5%. Constant current startup rapidly establishes an initial passivation film on the electrode surface, improving the current efficiency of the subsequent pulse phase by 15%.

[0111] transition phase

[0112] After startup, it automatically switches to pulse mode, selecting PWM or PFM modulation based on the raw water hardness: PWM mode (duty cycle 30%-50%) is used for low-hardness water (≤500 mg / L), controlling energy input by adjusting the conduction time; PFM mode (frequency 500-1000 Hz) is used for high-hardness water (>500 mg / L), achieving precise dosing by adjusting the pulse density. The transition phase continues until the scale thickness L reaches the set threshold (0.05-0.12 mm). During this process, the ORP value needs to be maintained between -200 and -100 mV to ensure that CaCO3 precipitates in the aragonite crystal form (easier to peel than calcite).

[0113] During the reversal phase, the ultrasonic-cavitation module starts 50 ms before reversal, with the cavitation pulse precisely synchronized with the current reversal. Utilizing the scale loosening effect caused by the instantaneous change in electrode polarity, it achieves a synergistic effect of "electrochemical dissolution + physical stripping". Through timing optimization, the amount of scale stripped in a single reversal is 2.3 times that of traditional methods.

[0114] It should be noted that the term "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An electrochemical descaling automatic control system based on autonomous prediction of reverse polarity timing and operating mode, characterized in that: The system comprises: a perception layer for collecting real-time raw water quality parameters and system state parameters, the system state parameters including real-time scale thickness; a prediction layer connected with the perception layer for predicting future scale thickness and descaling efficiency under different control strategies by an AI prediction model using the collected data; a decision layer connected with the prediction layer for solving a constrained optimization problem based on the prediction results to generate an optimal reverse polarity power supply control strategy, including reverse polarity frequency (f), duty cycle (d) and working current (I); an execution layer connected with the decision layer for receiving the control strategy and driving the reverse polarity power supply to perform corresponding reverse polarity operation; The perception layer, prediction layer, decision layer and execution layer together constitute a complete closed loop of perception-prediction-decision-execution-feedback.

2. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode in accordance with claim 1, wherein: The perception layer is a data collection module composed of multiple sensors and instruments, including pH sensor, conductivity sensor, calcium ion selective electrode, temperature sensor, flow meter, ultrasonic thickness gauge, voltage and current sensor and differential pressure sensor; The perception layer is provided with a signal conditioning and data preprocessing module for filtering and standardizing sensor signals, and a data aggregation and communication module for protocol conversion and data transmission; The pH sensor, conductivity sensor, calcium ion selective electrode and temperature sensor in the data collection module output 4-20mA analog current signals and are connected to the signal conditioning and data preprocessing module through shielded twisted pair hardware, the ultrasonic thickness gauge is connected to the data aggregation and communication module in communication through RS-485 interface and Modbus RTU protocol, the function of the signal conditioning and data preprocessing module is realized by the analog input module of PLC, which is responsible for A / D conversion and preliminary filtering, and the core hardware of the data aggregation and communication module is PLC or industrial gateway, which runs a periodic scanning control program internally to perform data collection, protocol conversion and data upload, the PLC collects digital sensor data through RS-485 interface and communicates with the upper decision layer through Ethernet interface in Modbus TCP / IP, OPC UA or MQTT protocol.

3. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode in accordance with claim 1, wherein: The AI prediction model is a multi-task spatio-temporal perception network, which comprises in sequence: a bidirectional long short-term memory network layer (Bi-LSTM layer) for feature extraction and capturing long-term bidirectional dependencies of input time series data; an attention mechanism layer for weighting processing of feature sequences output by the Bi-LSTM layer to focus on key time steps and key features; a multi-task output layer for simultaneously calculating and outputting predicted scale thickness and predicted descaling efficiency based on the output of the attention mechanism layer.

4. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode of claim 1 or 3, wherein: The AI prediction model is trained by minimizing a multi-task loss function, which is composed of a weighted sum of a first loss term for predicting scale thickness and a second loss term for predicting descaling efficiency, wherein both the first loss term and the second loss term adopt a robust loss function insensitive to outliers.

5. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode in accordance with claim 1, wherein: The decision layer is configured to solve an optimization problem, with the optimization objective being to minimize the running energy consumption of the system while satisfying the following four types of constraints: 1) a descaling effect constraint, ensuring that the predicted end-point scale thickness does not exceed a safety threshold; 2) a scaling trend constraint, ensuring that the adopted strategy produces a net descaling effect; 3) a hardware safety constraint, ensuring that the working current, reverse polarity frequency, and duty cycle of the reverse polarity power supply operate within their allowed physical limit ranges; and 4) an operation feasibility constraint, ensuring that the reverse polarity time is sufficient to effectively strip the scale.

6. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode in accordance with claim 1, wherein: The decision layer uses a Bayesian optimization algorithm to solve the optimization problem, which selects the next candidate strategy to be evaluated through a collection function configured to balance the trade-off between exploiting the currently known superior regions and exploring uncertain but potentially superior regions.

7. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode in accordance with claim 1, wherein: The execution layer executes a reverse polarity mode including at least one of a standard balance mode, a strong descaling mode, an energy-saving maintenance mode, a cathode cleaning mode, and an anode protection mode.

8. The electro-chemical descaling system based on autonomous predictive reverse polarity timing and operational mode in accordance with claim 1, wherein: The system further includes a continuous learning module for automatically performing the following steps during idle periods of the system: Step one, synchronizing recent running data from the database; Step two, fine-tuning the AI prediction model on new data using the current model weight as the pre-training value; Step three, verifying the performance of the new model, and if it is improved, deploying it online through a shadow mode.

9. An electrochemical descaling method for the electrochemical descaling self-control system according to any one of claims 1-8, characterized in that, The method includes: one, a sensing step, real-time acquisition of raw water quality parameters and system state parameters including scale thickness; two, a prediction step, based on the collected data, predicting future scale thickness and descaling efficiency through an AI prediction model; three, a decision step, generating an optimal reverse polarity power supply control strategy based on the prediction results; four, an execution step, receiving the control strategy and driving the reverse polarity power supply to perform corresponding reverse polarity operations.

10. The self-cleaning electrochemical de-scaling method based on autonomous predictive reverse polarity timing and operational mode of claim 9, wherein: The decision step includes using a Bayesian optimization algorithm to minimize energy consumption as the objective, and to satisfy the descaling effect and hardware safety constraints, to solve the optimal reverse polarity frequency (f), duty cycle (d), and working current (I).

Images