A method for early warning of EDI membrane stack failure and protection against thermal damage
By constructing a three-dimensional sensing network covering water, electricity, and heat, and using multi-dimensional sensor monitoring, combined with a PLC control unit and fault diagnosis model, the problem of lack of advance fault prediction in EDI membrane stack protection technology is solved. This enables timely early warning and protection against various types of faults in EDI membrane stacks, ensuring system safety and stability.
Patent Information
- Application Number
- CN202611381013.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
AI Technical Summary
Existing EDI membrane stack protection technologies lack the ability to predict potential faults in advance and cannot identify hidden faults such as flow rate attenuation, pressure difference abnormalities, and operating condition deviations, which makes the membrane stack prone to heat accumulation and dry burning failure.
A three-dimensional sensing network covering water, electricity, and heat is constructed. Key parameters of the EDI membrane stack are monitored in real time through multi-dimensional sensors. Real-time diagnosis is performed by combining the PLC control unit and fault diagnosis model to achieve multi-dimensional early warning and protection.
It achieves comprehensive coverage and timely early warning of various types of faults in EDI membrane stacks, avoids irreversible damage, reduces false alarm rate and false alarm rate, and ensures system safety and stability.
Smart Images

Figure CN123105808A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of electrodeionization technology and ultrapure water preparation technology, specifically to a method for early warning of EDI membrane stack failure and protection against thermal damage. Background Technology
[0002] Continuous electrodeionization (EDI) technology couples electrodialysis and ion exchange, using an electric field to drive ion migration and selective separation via ion exchange membranes to purify water. An EDI membrane stack consists of multiple unit pairs sandwiched between two electrodes. Each unit pair contains a desalination chamber and a concentrate chamber filled with mixed ion exchange resin, separated by anion and cation exchange membranes. This technology requires no acid or alkali regeneration and produces no wastewater discharge, and has been widely used in the preparation of ultrapure water in industries such as electronics, pharmaceuticals, and power.
[0003] Current EDI membrane stack protection technology systems are relatively simple. Equipment protection and fault diagnosis mostly rely on a few single-point fixed thresholds such as temperature and electrode polarity to achieve passive protection logic. They can only trigger protection actions after the equipment operating parameters exceed the hardware limit thresholds or a fault has occurred. They lack the ability to proactively perceive the equipment's operating conditions and predict faults. At the same time, traditional protection solutions have limited monitoring dimensions, covering only a few conventional operating parameters. They do not combine multi-dimensional core operating condition characteristics such as permeate flow rate, membrane stack pressure difference, heat load status, inlet and outlet water temperature difference, and electrical operating parameters. The parameter monitoring coverage is insufficient and the fault identification dimension is limited. In actual industrial operation scenarios, failures and damages of EDI membrane stacks are mostly caused by slowly accumulating latent abnormal operating conditions. Latent fault signs such as gradual flow attenuation, abnormal fluctuations in membrane stack pressure difference, system thermal load deviation, and unstable electrical conditions cannot be effectively identified through traditional single-point threshold monitoring methods. The long-term accumulation of latent abnormal operating conditions makes it impossible to provide timely warnings and interventions, which can easily lead to irreversible damage such as continuous heat accumulation, local overheating, and dry burning inside the membrane stack, reducing safety during use. Summary of the Invention
[0004] The purpose of this invention is to provide an EDI membrane stack fault early warning and thermal damage protection method to solve the problems of existing EDI membrane stack protection technology mentioned in the background art, which only relies on temperature and polarity single-point threshold for passive protection after the fact, lacks the ability to predict faults in advance, has single monitoring parameters, incomplete fault coverage, and cannot identify hidden faults such as flow attenuation, pressure difference abnormality, and operating condition deviation, which easily leads to cumulative thermal damage and dry burning failure of the membrane stack.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for EDI membrane stack fault early warning and thermal damage protection, comprising the following steps: S1. Hardware Equipment Configuration: Temperature sensors, flow sensors, pressure sensors, and voltage / current transformers are installed on the inlet, product water, and concentrate sides of the EDI membrane stack and on the power supply line. A double-insurance hard-wired shutdown circuit is constructed. The double-insurance hard-wired shutdown circuit connects the normally open contact of the flow switch and the normally closed contact of the temperature switch in series to the coil circuit of the DC contactor of the membrane stack, thus constructing a three-dimensional sensing network covering the three dimensions of water, electricity, and heat. All sensor signals are connected to the PLC control unit through shielded cables or fieldbus. S2. Real-time online monitoring of multi-dimensional operating parameters: Through the multi-dimensional sensors deployed in S1, key parameters such as the feed water temperature, permeate flow rate, concentrate flow rate, feed water pressure, permeate pressure, concentrate pressure, membrane power supply voltage, and membrane operating current of the EDI membrane stack are monitored online in milliseconds. The PLC control unit performs preprocessing such as filtering and noise reduction, bad value removal, and range normalization on the collected raw operating data to eliminate sensor noise and abnormal interference. S3. Data aggregation and analysis: The PLC control unit aggregates the pre-processed multi-dimensional operating data and constructs a fault diagnosis model that integrates temperature, flow, pressure and electrical characteristics. The fault diagnosis model consists of a basic rule layer and a parameter self-optimization layer. The two form a closed-loop mechanism of diagnosis-recording-training-calibration-optimization diagnosis. When the parameters have not yet reached the inherent danger threshold of the equipment, a predictive diagnostic conclusion is output by identifying abnormal trends. S4. Abnormal Early Warning Response and Tiered Protection: Construct a three-level early warning mechanism of yellow, orange and red, trigger different levels of response strategies based on the diagnostic results output by S3, and generate corresponding protection instructions; S5. Execution and Feedback Optimization: The PLC control unit sends the protection command generated in S4 to the corresponding actuator (contactor, frequency converter, electric regulating valve, etc.) to execute the graded linkage protection action. The system automatically records the change curves of each parameter before and after the warning is triggered, forming a complete fault event package (including trend chart, data snapshot, warning level and execution action). The system synchronously records each warning, fault and its processing result into the fault database to form a closed-loop feedback.
[0006] Preferably, in S1, a double-safety hard-wired shutdown circuit is constructed and connected in series to the coil circuit of the membrane stack DC contactor: Flow switch (normally open contact): Closes when there is flow, and opens when the flow rate is below the hardware detection accuracy threshold; Temperature switch (normally closed contact): Closes when the temperature is normal, and opens when the temperature exceeds 60℃; If either the flow switch or the temperature switch is disconnected, the contactor coil will immediately lose power, and the membrane stack will be physically de-energized.
[0007] Preferably, in S2, the multidimensional sensor continuously collects data at a sampling frequency of 1 second / time, and uses a sliding window mid-range filter (taking the median value of the most recent 5 values) to filter out spike interference in one go.
[0008] Preferably, in S3, the specific diagnostic logic includes the following steps: S31. Establishment and Update of Reference Anchor Point: After the EDI membrane stack has been running stably for 24 hours, the PLC automatically calculates the average values of the following parameters as the initial reference anchor point and stores them in the power-off holding register: Reference membrane stack pressure difference. Benchmark production water flow rate Reference inlet and outlet water temperature difference Reference film stack voltage Reference membrane stack current Reference heat load index ; S32. Cascaded multidimensional diagnosis: The criteria priority is arranged from high to low (L0>L1>L2>L3>L4). When multiple criteria are triggered at the same time, only the diagnosis conclusion corresponding to the highest priority is output. The diagnosis results of the other low priority criteria are only recorded and stored and do not participate in the output decision of the current cycle. When all triggering conditions are eliminated, the system automatically removes the abnormal state and generates a recovery event record.
[0009] Preferably, the formula for calculating the baseline heat load index is: .
[0010] Preferably, the criteria for cascaded multidimensional diagnosis include: Level L0: Based on the on / off status of the series hard-wired circuit, it determines the emergency events of "water shortage and dry burning" or "overheating damage", and has the highest priority, which is higher than all PLC-level diagnostics. Level L1: When the water production flow rate is... ≤ ×0.75, heat load index ≥ ×1.15, Inlet and outlet water temperature difference ≥ When all three conditions of +2.5℃ are met simultaneously and last for 10 seconds, it is determined to be a "sign of impending dry burning"; Level L2: When the membrane stack pressure difference ≥ ×1.4 and continue for 5 seconds to determine "slight membrane blockage". ≥ If the value is ×1.7 and remains so for 5 seconds, it is determined to be "severe blockage of membrane stack / short circuit of channel"; Level L3: When the heat load index ≥ ×1.2 and ≥ A temperature of +2°C for 15 seconds is considered "mild heat accumulation," and the heat load index... ≥ ×1.5 and ≥ A temperature of +4°C for 10 seconds is considered "severe heat accumulation". Level L4: When the membrane stack voltage ≥ ×1.15 or current ≥ ×1.15 for 10 seconds is considered "voltage / current abnormally high". ≥ ×1.3 or ≥ ×1.3 is considered a "serious electrical abnormality" if it lasts for 5 seconds.
[0011] Preferably, in S4, the yellow warning is triggered by the PLC software. The triggering condition is that the membrane stack pressure difference is ≥ the reference value × 1.4 or the electrical parameter is ≥ the reference value × 1.15. The response strategy is an audible and visual alarm and a pop-up prompt from the host computer, without directly interfering with the operation of the equipment. Orange Alert: Triggered by PLC software. The triggering conditions are membrane stack pressure difference ≥ reference value × 1.7 or heat load index exceeding the standard and inlet / outlet water temperature difference exceeding the standard. The response strategy is to reduce the frequency of the inverter by 5% to 10% to reduce the membrane stack current, or to open the bypass regulating valve to increase the concentrate flow rate to remove heat. When performing automatic adjustment, the PLC synchronously monitors the feedback signal of the actuator. If there is no change in the feedback within 5 seconds after the command is issued, or the change is less than 50% of the command value, the actuator is judged to be faulty and automatically upgraded to a red alert. Red Alert: Triggered by hardware wiring. The triggering condition is that the flow switch or temperature switch is turned off. The response strategy is to directly cut off the power supply to the contactor coil through hardware contacts to achieve emergency shutdown. The PLC is only responsible for event recording.
[0012] Preferably, in S5, based on the accumulated fault case library, the system continuously calibrates and optimizes the threshold coefficients of the basic rule layer through an offline training mechanism on the host computer and an online parameter injection mechanism on the PLC. Specific steps include: S51. Model structure definition: The CART classification model is used as the optimization engine. The six-dimensional input features are permeate flow rate, membrane stack pressure difference, heat load index, inlet and outlet water temperature difference, membrane stack voltage and membrane stack current. The seven types of faults are used as output labels. The model splitting threshold corresponds to the multiplier coefficient and temperature difference increment of the basic rule layer. The leaf node weights do not participate in the threshold mapping. S52. Training data labeling: Operation and maintenance personnel manually label each diagnosis result as "fault confirmed / false alarm / cause to be investigated". The labeling results that are consistent with the diagnosis conclusion are used as positive samples, and those labeled as false alarms are used as negative samples. When the number of valid fault cases reaches 50, the model retraining is automatically triggered. S53. Training Execution: The PLC uploads the desensitized data to the host computer, calls the Scikit-learn library to use the ten-fold cross-validation method with F1-Score as the evaluation index to retrain the model, outputs the optimized splitting threshold and generates a comparison report for review. S54, Safety Download: After the new threshold is manually reviewed and confirmed, it is written into the PLC power-off retention register. The single adjustment range shall not exceed ±10% of the original baseline value. The L0 hardware threshold does not participate in software optimization. The interval between two adjacent automatic optimizations shall not be less than 72 hours. S55. Closed-loop verification: After the new threshold is deployed, the false alarm rate, false negative rate and protection failure rate are statistically analyzed on a weekly basis. When the false alarm rate increases by more than 5 percentage points, the false negative rate increases by more than 3 percentage points, or a failure to activate event occurs within 72 hours, the system will automatically roll back to the previous version. S56. Regular Evaluation: The system automatically generates a performance evaluation report every quarter, which engineers can use to manually trigger retraining or manually adjust thresholds.
[0013] Preferably, in S53, during 10-fold cross-validation, the labeled dataset is randomly divided into 10 equal parts. One part is used as the validation set and the remaining 9 parts are used as the training set for 10 independent training and evaluations. The average of the 10 F1-Scores is used as the basis for the final performance evaluation of the model.
[0014] Preferably, in S55, the false alarm rate = number of diagnoses marked as false alarms / total number of diagnoses in the period × 100%; The false negative rate = (Number of events confirmed as false negatives / Total number of failures actually occurring in the period) × 100%; Protection failure rate = number of protection failure records / total number of protection commands issued in the current period × 100%.
[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In this invention, the L1-level dry-burning warning criterion is based on the simultaneous fulfillment of three conditions: the permeate flow rate drops to 75% of the baseline, the heat load index rises to 115% of the baseline, and the inlet and outlet water temperature difference exceeds 2.5℃ of the baseline. This can provide an early warning 10-15 minutes before the hardware flow switch is activated, allowing maintenance personnel to intervene in a timely manner and eliminate potential hazards before thermal damage occurs, fundamentally preventing membrane stack dry-burning failure. Multi-dimensional sensors are deployed on the membrane stack feed water side, permeate water side, concentrate side, and power supply line to construct a three-dimensional sensing network covering water, electricity, and heat, achieving comprehensive coverage of multiple types of faults such as flow rate attenuation, membrane stack blockage, electrical abnormalities, and heat accumulation. 2. In this invention, by using L2-level membrane stack blockage diagnosis, L3-level thermal accumulation diagnosis, and L4-level electrical anomaly diagnosis, the invention achieves hierarchical identification and timely early warning of latent abnormal signs, effectively preventing irreversible damage. Through a parameter self-optimization layer, using the CART classification model as the engine, six-dimensional operating parameters as input, and seven types of faults as output, the diagnostic threshold is continuously trained and optimized based on the fault database. The optimized threshold is manually reviewed and then written into the PLC. The single adjustment range does not exceed ±10%, ensuring that the threshold always adapts to the current operating state of the membrane stack, effectively reducing the false alarm rate and the missed alarm rate. 3. In this invention, a double-insurance hard-wired shutdown loop is constructed. Its response time is less than a preset value and is completely independent of the PLC scanning cycle. The hardware-level criteria have the highest priority over all software-level diagnoses. If any condition is not met, the contactor will immediately physically disconnect. The software-level criteria output a predictive diagnosis when the parameters do not reach the hard-wired threshold, realizing the coordinated protection of hard-wired fallback and software prediction. The diagnostic conclusions are manually annotated and entered into the fault database. When the accumulated value reaches the preset threshold, the model retraining is automatically triggered. After the new threshold is deployed, the false alarm rate, false negative rate, and failure to operate rate are periodically monitored. When the values are exceeded, the system automatically rolls back to the previous version to ensure that the system always operates in an optimal and safe state. Attached Figure Description
[0016] Figure 1 This is a flowchart of an EDI membrane stack fault early warning and thermal damage protection method according to the present invention. Detailed Implementation
[0017] 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.
[0018] Reference Figure 1 As shown: A method for early warning of EDI membrane stack failure and protection against thermal damage includes the following steps: Step 1: Hardware Equipment Configuration: Temperature sensors, flow sensors, pressure sensors, and voltage / current transformers are installed on the feed water side, product water side, concentrate side, and power supply line of the EDI membrane stack to build a three-dimensional sensing network covering the three dimensions of water, electricity, and heat. All sensor signals are connected to the PLC control unit through shielded cables or fieldbus.
[0019] Construct a double-safety hard-wired shutdown circuit, connected in series with the coil circuit of the membrane stack DC contactor: Flow switch (normally open contact): Closes when there is flow, and opens when the flow rate is below the hardware detection accuracy threshold; Temperature switch (normally closed contact): Closes when the temperature is normal, and opens when the temperature exceeds 60°C.
[0020] The contactor coil can only be energized and the membrane stack can only be energized when both the flow switch and the temperature switch are closed. If either condition is not met, the contactor will immediately physically disconnect. The circuit response time is less than 10ms and the PLC program scan cycle is completed independently.
[0021] All analog signals (4-20mA) are connected to the PLC analog input module via shielded cables, and all digital signals are connected to the PLC digital input module. Meanwhile, the contacts of the flow switch and temperature switch are hardwired separately and connected in parallel to the contactor coil.
[0022] Step 2: Real-time online monitoring of multi-dimensional operating parameters: The multi-dimensional sensors deployed in S1 perform millisecond-level real-time online monitoring of key parameters of the EDI membrane stack, such as influent temperature, product water temperature, product water flow rate, influent pressure, membrane stack inlet and outlet pressure difference, membrane stack voltage, and membrane stack current; The PLC control unit performs preprocessing such as filtering and noise reduction, bad value removal, and range normalization on the collected raw data to eliminate sensor noise and abnormal interference; The multidimensional sensor continuously collects data at a sampling frequency of 1 second / time. It uses a sliding window mid-value filter (taking the median value of the 5 most recent values) to filter out spike interference in one go, thereby achieving noise reduction and bad value removal. The pre-processed valid data is used for subsequent steps.
[0023] Step 3: Data aggregation and analysis: The PLC control unit aggregates the pre-processed multi-dimensional operating data and constructs a fault diagnosis model that integrates temperature, flow rate, pressure, and electrical characteristics. When the parameters have not yet reached the inherent danger threshold of the equipment, the system outputs a predictive diagnostic conclusion by identifying abnormal trends.
[0024] The fault diagnosis model consists of a basic rule layer and a parameter self-optimization layer, which together form a closed loop of "diagnosis → recording → training → calibration → optimized diagnosis".
[0025] The specific diagnostic logic includes the following steps: S31. Establishment and Update of Reference Anchor Point: After the EDI membrane stack has been running stably for 24 hours, the PLC automatically calculates the average values of the following parameters as the initial reference anchor point and stores them in the power-off holding register: Reference membrane stack pressure difference. Benchmark production water flow rate Reference inlet and outlet water temperature difference Reference film stack voltage Reference membrane stack current Reference heat load index .
[0026] The calculation formula is: .
[0027] Heat load index The electrical power input of the membrane stack corresponding to a unit product water flow rate is characterized by the heat source of EDI (Electrical Discharge Irrigation), which is mainly Joule heating, while the product water flow rate... The ability of water flow to carry away heat is determined when A rise in temperature means that a unit volume of water needs to remove more Joule heat, increasing the risk of heat accumulation inside the membrane stack. Therefore... Temperature difference between inlet and outlet water The two ends of the heat balance equation are: Characterizing the intensity of the heat source, Characterizing heat accumulation results—the combined judgment of both can effectively identify the thermal imbalance state of the membrane stack.
[0028] The baseline anchor point is updated only in the following two cases: (1) After chemical cleaning (CIP) of the membrane stack or resin regeneration, a new round of 24-hour calibration is automatically triggered; (2) Engineers evaluate every quarter, and when a certain reference parameter deviates from the initial value by more than ±15%, it is manually recalibrated and written into the PLC after manual confirmation.
[0029] S32. Cascaded multidimensional diagnosis: The criteria priority is arranged from high to low (L0>L1>L2>L3>L4). When multiple criteria are triggered at the same time, only the diagnosis conclusion corresponding to the highest priority is output. The diagnosis results of the other low priority criteria are only recorded and stored and do not participate in the output decision of the current cycle. When all triggering conditions are eliminated, the system automatically removes the abnormal state and generates a recovery event record.
[0030] L0 (hardware level, independent of PLC) - Based on the on / off state of the series hard-wired circuit in step one: When the circuit is disconnected, the system directly determines it as an emergency event of "dry burning due to lack of water" or "overheating damage". This determination result has the highest priority (higher than all PLC-level diagnostics). The PLC only records the event by capturing the falling edge after the hard-wired action.
[0031] L1 (PLC level, dry-burning warning sign recognition, lasts for 10 seconds) – when ① production water flow rate ≤ ×0.75、②Heat load index ≥ ×1.15、③ Inlet and outlet water temperature difference ≥ When all three conditions (+2.5℃) are met simultaneously, it is considered a precursor to dry burning. This criterion is based on the steady-state thermal balance parameter (…). and It replaces the traditional temperature rise rate differential calculation and can provide an early warning 10 to 15 minutes before the hardware flow switch is activated.
[0032] L2 (PLC level, membrane stack blockage diagnosis, lasts 5 seconds) – when the membrane stack pressure difference... ≥ When ×1.4, the diagnosis is "mild membrane stack blockage"; when ≥ When the value is ×1.7, the diagnosis is "severe blockage of membrane stack / short circuit of channel".
[0033] L3 (PLC level, thermal damage auxiliary diagnosis in the basic rule layer) — when ≥ ×1.2 and ≥ When +2℃ is maintained for 15 seconds, the diagnosis is "mild heat accumulation"; when ≥ ×1.5 and ≥ When the temperature is maintained at +4℃ for 10 seconds, the diagnosis is "severe heat accumulation".
[0034] L4 (PLC level, electrical anomaly diagnosis) – membrane stack voltage ≥ ×1.15 or current ≥ When ×1.15 lasts for 10 seconds, the diagnosis is "voltage / current is abnormally high"; ≥ ×1.3 or ≥ The pulse of ×1.3 lasted for 5 seconds, and the diagnosis was "serious electrical abnormality".
[0035] The initial threshold coefficients (such as 0.75, 1.15, 1.2, 1.4, 1.7, etc.) and temperature difference increments (such as 2℃, 2.5℃, 4℃, etc.) of each diagnostic criterion in the EDI membrane stack basic rule layer are determined comprehensively from three aspects: First, referencing the allowable operating range of the original manufacturer's technical specifications; second, based on statistical fitting of normal / fault measured data of the same model of membrane stack, with the parameter mean ±3σ as the normal range; and third, based on the heat balance theory, with the Joule heat generation rate exceeding the water flow heat removal rate by 20% as the heat accumulation risk criterion. The above initial thresholds are fine-tuned and solidified according to the membrane stack model during commissioning, but this set of coefficients is not permanently fixed. As the initial value of the "parameter self-optimization layer", it is periodically dynamically calibrated based on the actual fault database.
[0036] Step 4, Anomaly Warning Response and Tiered Protection: Construct a three-level warning mechanism of yellow, orange and red, trigger different levels of response strategies based on the diagnostic results output by S3, and generate corresponding protection instructions.
[0037] Yellow Alert: Triggered by PLC software, the triggering conditions are differential pressure ≥ 1.4 times the reference or electrical parameters ≥ 1.15 times the reference; Response strategy: audible and visual alarm plus supervisory control computer pop-up prompt, without directly interfering with equipment operation, the operation and maintenance personnel determine whether offline cleaning is required, and the warning prompt is realized by outputting the PLC contact to the alarm light and buzzer.
[0038] Orange Alert: Triggered by PLC software. Triggering conditions are differential pressure ≥ 1.7 times the baseline or heat load index exceeding the standard and temperature difference exceeding the standard. Response strategy: The PLC outputs a 4-20mA signal to adjust the frequency converter, reducing the membrane stack current by 5% to 10% frequency reduction, or opening the bypass regulating valve to increase the concentrate flow to remove heat. If the operating condition fails to recover within 5 minutes, it is reported as a serious fault. Automatic adjustment and control are completed by relying on the PLC's PID / slope function output to the frequency converter or regulating valve. When performing automatic adjustment actions, the PLC simultaneously monitors the frequency feedback signal from the inverter and the opening feedback signal from the regulating valve. If the feedback signal does not change or the change is less than 50% of the command value within 5 seconds after the execution command is issued, it is determined that the actuator is faulty. The system automatically upgrades the orange warning to a red warning, triggers a hardware hard-wired emergency shutdown, and records the "actuator failure" fault code on the host computer, prompting maintenance personnel to conduct on-site troubleshooting.
[0039] Red Alert: Triggered by hardware hard-wired connection, triggered when the flow switch or temperature switch is disconnected; Response Strategy: The hardware contacts directly cut off the power supply to the contactor coil, causing the membrane stack to immediately shut down in an emergency. At the same time, the PLC captures the falling edge of the switch signal to complete the event recording. Hard interlocking protection is achieved using pure relay hard-wired connection, and the PLC is only responsible for event recording.
[0040] Step 5, Execution and Feedback Optimization: The PLC control unit sends the protection command generated in Step 4 to the corresponding actuators (contactors, frequency converters, electric regulating valves, etc.) to execute the graded linkage protection actions. The system automatically records the change curves of each parameter before and after the warning is triggered, forming a complete fault event package (including trend charts, data snapshots, warning levels and execution actions). The system synchronously records each warning, fault and its handling results into the fault database to form a closed-loop feedback.
[0041] Based on the accumulated fault case library, the system continuously calibrates and optimizes the threshold coefficients of the basic rule layer through an offline training mechanism on the host computer and an online parameter injection mechanism on the PLC. The specific steps include: S51. Structural Definition of the Diagnostic Model: The parameter self-optimization layer of the fault diagnosis model uses the interpretable classification model of the CART algorithm as the optimization engine, based on the real-time monitored water production flow rate in step two. membrane stack pressure difference Heat load index Inlet and outlet water temperature difference membrane stack voltage and membrane stack current As a six-dimensional input feature vector, “normal / precursor to dry burning / minor blockage / severe blockage / mild heat accumulation / severe heat accumulation / electrical abnormality” are used as seven-category output labels. The numerical thresholds used for node splitting in this model directly correspond to the multiplier coefficients (such as 1.15, 1.4, etc.) and temperature increments (such as 2℃, 4℃, etc.) of each criterion in the basic rule layer. The leaf node weights are only used for probability calculation of the output category and do not participate in the threshold mapping of the basic rule layer. The multiplier coefficients of 1.2 and 1.5 and the temperature increments of 2℃ and 4℃ involved in the L3 level criteria are used as specific splitting thresholds corresponding to the thermodynamic dimension in this model and are included in the scope of this optimization and calibration.
[0042] S52. Training Data Source and Labeling: The training set is derived from the fault database. After each diagnosis is triggered and a complete on-site handling is completed, the maintenance personnel must confirm the "true result" of the diagnosis on the host computer interface, select one of the three states: "fault confirmed / false alarm / cause under investigation," and enter the actual fault cause and handling measures to form a standard sample that has been manually labeled. Only samples whose diagnostic conclusions are consistent with the manually labeled results are included in the training set as "positive samples." Samples labeled as "false alarms" are used as "negative samples" to optimize the algorithm and reduce the false alarm rate. When the cumulative number of newly labeled valid fault cases in the database reaches the preset trigger threshold (recommended value is 50), the system automatically triggers a model retraining.
[0043] S53. Optimize the training execution process: The PLC uploads the de-identified feature data package to the host computer. The host computer calls standard machine learning libraries such as Scikit-learn and uses the ten-fold cross-validation method to retrain the model. The F1-Score (harmonic mean of precision and recall) is used as the model performance evaluation index. During model training, the CART algorithm aims to minimize the Gini coefficient, freely searching for optimal splitting features and splitting thresholds in the six-dimensional feature space without pre-setting any hard constraints on the splitting positions. After training, the splitting thresholds corresponding to the six-dimensional features at each internal node of the decision tree (such as the splitting value of the permeable flow branch node, the splitting value of the heat load index branch node, etc.) are extracted and mapped to the multiplier coefficients and temperature difference increments of the corresponding criteria in the basic rule layer. For example, the ratio of the threshold of the heat load index-related splitting node to the baseline heat load index is used as the optimized update value of the L1 level 1.15 coefficient or the L3 level 1.2 and 1.5 coefficients; the difference between the threshold of the inlet and outlet water temperature difference-related splitting node and the baseline temperature difference is used as the optimized update value of the corresponding temperature difference increment (2℃, 4℃). After training is completed, the optimized threshold values for each criterion are output (i.e., the specific values corresponding to coefficients such as 1.15, 1.2, and 1.4, and temperature difference increments such as 2℃ and 4℃), and a comparison report of the threshold values before and after optimization is generated for engineers to review.
[0044] In the 10-fold cross-validation, the labeled dataset is randomly divided into 10 parts. One part is used as the validation set and the other 9 parts are used as the training set for 10 independent training and evaluations. The average F1-Score of the 10 tests is used as the basis for the final performance evaluation of the model.
[0045] S54. Downloading and Safety Constraints of Optimization Results: After manual review and confirmation via a pop-up prompt on the engineer station interface, the optimized new threshold coefficients are written to the PLC's power-off holding register via industrial Ethernet, replacing the original coefficients. This allows for online dynamic calibration of the diagnostic model thresholds without modifying the core logic of the PLC ladder diagram. The system forcibly sets that the single threshold adjustment range must not exceed ±10% of the original benchmark value. Furthermore, the physical contact thresholds (60℃ and low flow setting values) of the L0 level (hardware hard-wired) criterion are inherent hardware parameters and do not participate in any software-level automatic optimization. They always maintain independent physical protection, and the time interval between two adjacent automatic optimization triggers must not be less than 72 hours.
[0046] S55. Closed-loop verification of optimization effect: After the new threshold is deployed, the host computer has a built-in performance evaluation module to monitor the optimization effect in real time by statistically analyzing the false alarm rate, false alarm rate and protection action rate.
[0047] False alarm rate: Calculated as the percentage of "false alarms" confirmed by maintenance personnel on the host computer out of the total number of diagnoses in the period, i.e., false alarm rate = number of diagnoses marked as "false alarms" / total number of diagnoses in the period × 100%; False alarm rate: Calculated by linking maintenance work orders with the historical trend database, the false alarm rate is the percentage of events confirmed as "false alarms" in the past period relative to the total number of actual faults in that period. False alarm rate = Number of events confirmed as "false alarms" / Total number of actual faults in that period × 100%. Specifically, when an unexpected shutdown occurs that is not a system-initiated shutdown or when equipment performance significantly degrades and requires maintenance, the host computer's backtracking analysis module automatically retrieves the parameter trends from the 30 minutes prior to the fault occurrence and replays the triggering states of criteria at levels L0 to L4. If the parameters have met or are close to the triggering conditions of a certain criterion but the system has not output a diagnosis, it is considered a false alarm.
[0048] Protection failure rate: Based on the real-time comparison of feedback signals from the PLC to the actuators (actual frequency of the frequency converter, actual opening degree of the regulating valve, and status of the auxiliary contacts of the contactor) in step four, the number of times the feedback fails to meet the standard (allowable error ±5%) within a specified time after the execution command is issued is counted, and the percentage of the total number of protection commands issued in the cycle is calculated; Protection failure rate = Number of protection failure records / Total number of protection commands issued in the cycle × 100%.
[0049] The above indicators are updated on a weekly cycle. When any of the following conditions are met, the system will automatically trigger a threshold rollback to the previous optimized version and lock the optimized parameters for engineers to analyze: ① The false alarm rate increases by more than 5 percentage points compared to before optimization; ② A manually confirmed protection failure event occurs within 72 hours of operation (actuator failure or software-level failure); ③ The false alarm rate increases by more than 3 percentage points compared to before optimization.
[0050] S56. Regular evaluation and manual intervention: In addition to automatic retraining, the system automatically generates a diagnostic model performance evaluation report every quarter (including the trigger frequency of each level criterion, false alarm / false negative distribution, the offset between the current threshold value and the initial value, etc.) for engineers to conduct a comprehensive evaluation. Engineers can manually trigger model retraining based on the evaluation results, or manually adjust specific threshold coefficients and write them into the PLC.
[0051] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An EDI membrane stack failure early warning and thermal damage protection method, characterized in that: Includes the following steps: S1. Hardware equipment configuration: Multi-dimensional sensors are installed on the feed water side, product water side, concentrate side and power supply line of the EDI membrane stack, and a double-insurance hard-wired shutdown circuit is constructed. The double-insurance hard-wired shutdown circuit connects the normally open contact of the flow switch and the normally closed contact of the temperature switch in series to the coil circuit of the DC contactor of the membrane stack. All sensor signals are connected to the PLC control unit through shielded cables or fieldbus. S2. Real-time online monitoring of multi-dimensional operating parameters: The key parameters of the EDI membrane stack are monitored in real time by multi-dimensional sensors deployed in S1, and the PLC control unit preprocesses the collected raw operating data. S3. Data aggregation and analysis: The PLC control unit aggregates the pre-processed multi-dimensional operating data and constructs a fault diagnosis model that integrates temperature, flow, pressure and electrical characteristics. The fault diagnosis model consists of a basic rule layer and a parameter self-optimization layer. The two form a closed-loop mechanism of diagnosis-recording-training-calibration-optimization diagnosis. When the parameters do not reach the inherent danger threshold of the equipment, a predictive diagnostic conclusion is output by identifying abnormal trends. S4. Abnormal Early Warning Response and Tiered Protection: Construct a three-level early warning mechanism of yellow, orange, and red, trigger different levels of response strategies based on the diagnostic results output by S3, and generate corresponding protection instructions; S5. Execution and Feedback Optimization: The PLC control unit sends the protection command generated in S4 to the corresponding actuator to execute the hierarchical linkage protection action. The system automatically records the change curves of each parameter before and after the warning is triggered, forming a complete fault event package. Each warning, fault and its handling result are synchronously entered into the fault database to form a closed-loop feedback.
2. The method of claim 1, wherein: In S1, a double-safety hard-wired shutdown circuit is constructed, including: Flow switch: Closes when there is flow, and opens when the flow rate is below the hardware detection accuracy threshold; Temperature switch: Closed when the temperature is normal, and opened when the temperature exceeds 60℃; If either the flow switch or the temperature switch is disconnected, the contactor coil will immediately lose power, and the membrane stack will be physically de-energized.
3. The method of claim 1, wherein: In S2, the key parameters are inlet water temperature, product water temperature, product water flow rate, inlet water pressure, membrane stack inlet and outlet pressure difference, membrane stack voltage, and membrane stack current. The multi-dimensional sensor continuously collects data at a sampling frequency of 1 second / time.
4. The method of claim 1, wherein: In S3, the specific diagnostic logic includes the following steps: S31, benchmark anchor establishment and update: after EDI membrane stack debugging stable operation for 24 hours, PLC automatically calculates the average of each parameter as the initial benchmark anchor and solidifies in the power-off holding register: benchmark membrane stack pressure difference , benchmark water production flow , benchmark water inlet and outlet temperature difference , benchmark membrane stack voltage , benchmark membrane stack current , benchmark heat load index ; S32. Cascaded multidimensional diagnosis: The criteria priority is arranged from high to low, i.e., L0>L1>L2>L3>L4. When multiple criteria are triggered at the same time, only the diagnosis conclusion corresponding to the highest priority is output. The diagnosis results of the other low priority criteria are only recorded and stored, and do not participate in the output decision of the current cycle. After all triggering conditions are eliminated, the system automatically removes the abnormal state and generates a recovery event record.
5. The method of claim 4, wherein: The calculation formula of the reference thermal load index is: .
6. The method of claim 4, wherein: The criteria for S32's cascaded multidimensional diagnosis include: Level L0: Based on the on / off status of the series hard-wired circuit, it determines emergency events such as water shortage, dry burning, or overheating damage, and has the highest priority, which is higher than all PLC-level diagnostics. L1 level: when the water production flow rate ≤ × 0.75, the heat load index ≥ × 1.15, the temperature difference between the inlet and outlet water ≥ + 2.5℃, all three conditions are met and last for 10 seconds, it is determined as a dry burning precursor; L2 level: when the membrane stack pressure difference ≥ x 1.4 for 5 seconds, the membrane stack is determined to be slightly clogged, ≥ x 1.7 for 5 seconds, the membrane stack is determined to be severely clogged / short circuit; L3 level: when the heat load index ≥ x 1.2 and ≥ + 2°C for 15 seconds is determined as mild heat accumulation, when the heat load index ≥ x 1.5 and ≥ + 4°C for 10 seconds is determined as severe heat accumulation; L4: When the membrane stack voltage ≥ × 1.15 or the current ≥ × 1.15 for 10 seconds, it is determined that the voltage / current is abnormally high, and when ≥ × 1.3 or ≥ × 1.3 for 5 seconds, it is determined that there is a serious electrical abnormality.
7. The method of claim 1, wherein: In S4, the yellow warning is triggered by the PLC software. The triggering conditions are that the membrane stack pressure difference is ≥ the reference value × 1.4 or the electrical parameter is ≥ the reference value × 1.
15. The response strategy is an audible and visual alarm and a pop-up prompt on the host computer. It does not directly interfere with the operation of the equipment. Orange Alert: Triggered by PLC software. The triggering conditions are membrane stack pressure difference ≥ reference value × 1.7 or heat load index exceeding the standard and inlet / outlet water temperature difference exceeding the standard. The response strategy is to reduce the membrane stack current by adjusting the frequency converter to reduce the frequency by 5% to 10%, or to open the bypass regulating valve to increase the concentrate flow rate to remove heat. When performing automatic adjustment, the PLC synchronously monitors the feedback signal of the actuator. If there is no change in the feedback within 5 seconds after the command is issued, or the change is less than 50% of the command value, the actuator is judged to be faulty and automatically upgraded to a red alert. Red Alert: Triggered by hardware wiring. The triggering condition is that the flow switch or temperature switch is turned off. The response strategy is to directly cut off the power supply to the contactor coil through hardware contacts to achieve emergency shutdown. The PLC is only responsible for event recording.
8. The method of claim 1, wherein: In S5, based on the accumulated fault case library, the system continuously calibrates and optimizes the threshold coefficients of the basic rule layer through offline training on the host computer and online parameter injection into the PLC. Specific steps include: S51. Model structure definition: The CART classification model is used as the optimization engine, with multi-dimensional parameters of temperature-flow-pressure-electrical characteristics as input features and multiple fault states as output labels. The model splitting threshold is mapped and correlated with the diagnostic criterion coefficients of the basic rule layer. The weights of leaf nodes do not participate in the threshold mapping. S52. Training data annotation: Operation and maintenance personnel manually annotate the diagnostic results to reflect the true state. The annotation results that are consistent with the diagnostic conclusions are included in the training set as positive samples, and those labeled as false alarms are included as negative samples. When the cumulative number of valid fault cases reaches the preset threshold, the system automatically triggers model retraining. S53. Training Execution: The PLC uploads the desensitized data to the host computer, calls the Scikit-learn library to use the ten-fold cross-validation method with F1-Score as the evaluation index to retrain the model, and outputs the optimized splitting threshold. S54, Safety Download: The new threshold is written to the PLC register after manual confirmation. The single adjustment range shall not exceed ±10% of the original baseline value. The L0 hardware threshold does not participate in software optimization. The interval between two adjacent automatic optimizations shall not be less than 72 hours. S55. Closed-loop verification: After the new threshold is deployed, the false alarm rate, false negative rate and protection failure rate are statistically analyzed on a weekly basis. When the false alarm rate increases by more than 5 percentage points, the false negative rate increases by more than 3 percentage points, or a failure to activate event occurs within 72 hours, the system will automatically roll back to the previous version. S56. Regular Evaluation: The system automatically generates a performance evaluation report every quarter, and engineers can manually trigger retraining or manually adjust the thresholds.
9. The method of claim 8, wherein: In S53, during 10-fold cross-validation, the labeled dataset is randomly divided into 10 equal parts. One part is used as the validation set, and the other 9 parts are used as the training set for 10 independent training and evaluations. The average of the 10 F1-Scores is used as the basis for the final performance evaluation of the model.
10. The EDI stack failure early warning and thermal damage prevention method of claim 8, wherein: In S55, the false alarm rate = number of diagnoses marked as false alarms / total number of diagnoses in the period × 100%; The false negative rate = (Number of events confirmed as false negatives / Total number of failures actually occurring in the period) × 100%; Protection failure rate = number of protection failure records / total number of protection commands issued in the current period × 100%.
