Intelligent auxiliary decision method and system for large language model driven net treatment system

Abstract

The present application relates to data learning technology, specifically to a large language model driven net treatment system intelligent auxiliary decision method and system, including data acquisition module, data processing module, large language model module and decision output module;When judging the periodic abnormal reason, through the accurate comparison of business process cycle and equipment maintenance cycle, the abnormal root positioning time is shortened, and the abnormal processing response speed is improved;When attributing to non-periodic abnormality, through dynamic analysis of frequency change value, differential response is realized;Through abnormal level comprehensive calculation, intelligent decision of maintenance opportunity is realized, immediate maintenance is triggered when the situation is judged to be urgent, and loss caused by excessive wear of equipment is avoided;In non-urgent situation, the advance maintenance strategy can prolong the service life of key equipment and reduce maintenance cost;From abnormal identification, attribution, decision to verification, a whole-process risk closed-loop management is formed, and the environment data linkage filtering in closed-loop management makes the system still maintain stable operation under extreme working conditions.

Description

Technical Field

[0001] This invention relates to data learning technology, specifically to a method and system for intelligent auxiliary decision-making driven by a large language model in a purification system. Background Technology

[0002] Traditional wastewater treatment systems rely on fixed thresholds for anomaly management. However, when the influent load of a wastewater treatment plant changes abruptly, these fixed thresholds can easily lead to numerous false alarms or missed alarms, resulting in low accuracy in anomaly identification. Furthermore, the time required for manual investigation of the root cause of anomalies is lengthy, severely impacting the system's response speed. Traditional maintenance strategies are mostly based on fixed cycles, without considering the actual operating status of the equipment. Performing maintenance according to fixed cycles can easily lead to over-maintenance or delayed maintenance, resulting in malfunctions. Traditional wastewater treatment systems do not systematically filter maintenance operation data, causing occasional valid data to be included in model training, resulting in slow adaptation of large language models to new anomaly scenarios.

[0003] To address the aforementioned technical problems, this application proposes a solution. Summary of the Invention

[0004] The purpose of this invention is to address the problems raised in the background section above, and to propose a large language model-driven intelligent auxiliary decision-making method and system for cleaning and treatment systems.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A large language model-driven intelligent auxiliary decision-making method for a purification system is disclosed. This method is implemented through an intelligent auxiliary decision-making system, which includes a data acquisition module, a data processing module, a large language model module, and a decision output module. The intelligent auxiliary decision-making method for the purification system includes the following steps:

[0007] S1: The data processing module analyzes outliers in the data transmitted from the data acquisition module. If the number of outliers exceeds a preset number, the fluctuation range of outliers is expanded. If the number of outliers still exceeds the preset number, the moment is marked as an outlier moment. Otherwise, outliers are removed and the mean is recalculated as valid data.

[0008] S2: Retrieve historical data, categorize by usage time, and calculate the frequency of abnormal moments in each category. If the frequencies are the same, the abnormality is determined to be periodic. If the abnormal period is the same as the business process period, further compare the overlap of time periods. If the abnormal period occupies more than 80% of the business process time period, it is determined to be caused by business operations. If the abnormal period is the same as the equipment maintenance period, compare the two time periods. If they are completely consistent, it is determined to be caused by maintenance operations. If the abnormal period is unrelated to the business process and maintenance period, generate an early maintenance signal and transmit it to the decision output module.

[0009] S3: If the frequencies are different, calculate the frequency change value; if the frequency change value continues to increase, check for manual operation: if there is no operation, determine that the system will automatically recover and record the conditions; if there is operation, confirm that the manual decision is effective, maintain the parameters and monitor; if the frequency change value continues to decrease, determine that the risk is increasing, and generate a warning signal when the threshold is reached; if the frequency change value continues to fluctuate, check for manual operation: if there is operation, wait for it to be completed and re-detect; if there is no operation, generate a warning signal and trigger an audible and visual alarm or information reminder according to the frequency change amplitude.

[0010] S4: The decision output module receives the signal and notifies the staff to carry out maintenance operations. During the maintenance operation, the data after maintenance is recorded and compared with the preset fluctuation range. The difference value is calculated. If the difference value is greater than 1, the operation is determined to be valid and recorded. Otherwise, it is ignored. The valid data is filtered to remove accidental factors. After filtering, the valid data is used for the optimization of the large language model.

[0011] In a preferred embodiment of the present invention, the data processing module performs the following steps for determining abnormal moments:

[0012] K1: Sort the collected data by time, and calculate the mean A1 and standard deviation B for multiple data of the corresponding item collected at the same time. Set the fluctuation range of the corresponding item detection data [A1-bB, A1+bB] based on the mean A1 and standard deviation B, where b is a positive integer.

[0013] K2: Mark the data for items outside the fluctuation range as outliers, and assign the number of outliers Y to each item. a Statistical analysis shows that the total number of data points detected simultaneously for the corresponding item is Z. a If Y a ≥k1*Z a If the detected data is abnormal, the count of value b is incremented by one to narrow the range of abnormal values. k1 is a preset proportional coefficient, and a represents the sorting number of the corresponding data item.

[0014] K3: If the value b is incremented by one and Y still satisfies the condition... a ≥k1*Z a Then mark the detection time as an abnormal time; if Y a <k1*Z a If outliers are removed, the mean A2 is calculated for the remaining corresponding item detection data, and the calculated mean A2 is used as the corresponding item data detected at that time.

[0015] In a preferred embodiment of the present invention, the data processing module performs periodic analysis at abnormal times as follows:

[0016] M1: Retrieve historical data for the corresponding item, classify the historical data for the corresponding item according to the length of time each time it is used, calculate the frequency of abnormal moments in the corresponding item data in each category, and if the frequency of abnormal moments is the same in each category, it is determined that the occurrence of abnormality is periodic.

[0017] M2: Compare the period of the abnormal moment with the business process cycle of the cleaning system and the maintenance cycle of the equipment. If the period of the abnormal moment is the same as the business process cycle of the cleaning system, it is initially determined to be a periodic abnormality caused by business operation. Then, compare the time period corresponding to the period of the abnormal moment with the time period corresponding to the business process cycle of the cleaning system. If the time period corresponding to the abnormal moment occupies more than 80% of the time period corresponding to the business process cycle of the cleaning system, it is determined to be a periodic abnormality caused by business operation.

[0018] M3: If the period of the abnormal moment is the same as the maintenance cycle of the equipment, it is initially determined to be a periodic abnormality caused by maintenance operation; then, the two periodic time periods of the abnormal moment and the equipment maintenance cycle are compared. If the two periodic time periods are the same, it is determined to be a periodic abnormality caused by maintenance operation; if the comparison finds that the period of the abnormal moment is not caused by the business process cycle of the cleaning system and the maintenance cycle of the equipment, an early maintenance signal is generated and transmitted to the decision output module.

[0019] In a preferred embodiment of the present invention, the data processing module performs the following non-periodic analysis steps at abnormal times:

[0020] N1: Calculate the frequency change value at abnormal times. If the frequency change value at abnormal times continues to increase, obtain the manual operation data of the corresponding item within the set time period before the current time point. If there is no manual operation data for the corresponding item, it is determined that the system has spontaneously recovered, and the recovery conditions are recorded to provide a reference for subsequent similar anomaly adjustments. If there is manual operation data for the corresponding item, it is determined that the manual decision has taken effect, and the current effective parameters are kept running for a period of time. The abnormal frequency is continuously monitored, and after the trend is confirmed to be stable, the investment of emergency resources is reduced in an orderly manner.

[0021] N2: If the frequency change value at the abnormal moment continues to decrease, it is determined that the risk of system failure or water quality deterioration is gradually increasing. When the frequency change value at the abnormal moment reaches the preset frequency change threshold, an early warning signal is generated and transmitted to the decision output module.

[0022] N3: If the frequency change value at the abnormal moment keeps fluctuating, then obtain the manual operation data of the corresponding item within the set time period before the current time point. If there is manual operation data for the corresponding item, then after the manual operation is completed, the frequency change value at the abnormal moment will be detected again. If there is no manual operation data for the corresponding item, an alarm signal will be generated and transmitted to the decision output module.

[0023] In a preferred embodiment of the present invention, the data processing module determines the validity of operational data as follows:

[0024] G1: Detect and record the changes in various data during the maintenance process, and record the values ​​of the changes in various data after the corresponding data is adjusted. Compare the recorded values ​​of the corresponding data changes with the preset fluctuation range of the corresponding data. The preset fluctuation range is between the minimum and maximum values ​​of the corresponding data.

[0025] G2: If the value of the corresponding data after the change is within the preset fluctuation range of the corresponding data, the operation adjustment is deemed effective, and the data change caused by the operation adjustment is recorded; if the value of the corresponding data after the change is not within the preset fluctuation range of the corresponding data, the absolute value of the difference between the value of the corresponding data after the change and the preset fluctuation range of the corresponding data is calculated, and the minimum absolute value of the calculated difference is the gap between the value of the corresponding data after the change and the preset fluctuation range of the corresponding data. The gap value is equal to the difference divided by the difference between adjacent values ​​within the fluctuation range of the corresponding data.

[0026] G3: If the difference value is greater than 1, the operation adjustment is deemed effective, and the data change caused by the operation adjustment is recorded; otherwise, the operation adjustment is deemed invalid, and the data change caused by the operation adjustment is not recorded.

[0027] In a preferred embodiment of the present invention, the data processing module performs the following filtering step for random factors:

[0028] P1: Obtain the operation data of the corresponding operation record in the historical data, retrieve the number of times the corresponding operation data was performed, filter the operation data with fewer than the preset number of times, and complete a data filtering and screening operation.

[0029] P2: Mark the operation data obtained after completing a data filtering and screening operation, mark the environmental data corresponding to the marked operation data again, then obtain the environmental data corresponding to the same operation data in the historical data, and compare the marked environmental data with the corresponding items in the obtained environmental data.

[0030] P3: If the fluctuation of the environmental data corresponding to the two operation data is within the preset fluctuation range of the corresponding item, then the two operation data will be retained; otherwise, both sets of operation data will be filtered out to complete the secondary data filtering and removal operation.

[0031] P4: Record the operation data after the secondary data filtering and screening operation, and substitute the operation data into the large language model to optimize the trained large language model.

[0032] As a preferred embodiment of the present invention, the decision output module executes the steps for the advance maintenance signal as follows:

[0033] Q1: Calculate the average value A3 of the abnormal data at the abnormal moment, obtain the corresponding item data at the adjacent moments of the abnormal moment, and calculate the average value A4. Divide the difference between the average values A3 and A4 by the preset fluctuation value to obtain the fluctuation ratio c. If (d - 1)*c1 < c < d*c1, then determine that the abnormal level is level d, where c1 is the preset comparison value;

[0034] Q2: Compare the cycle length d at the abnormal moment Y with the preset cycle length d S to determine that the abnormal level is level;

[0035] Q3: Calculate the sum value d1 of , and compare the sum value d1 with the preset comparison value d2. If d1 < d2, then determine to maintain the original maintenance cycle; if d2 < d1 < k1*d2, then determine to perform maintenance operation in advance , where k1 is the preset proportional coefficient and t1 is the time duration between the current time point and the next maintenance cycle; if d1 ≥ k1*d2, then determine to perform maintenance operation immediately.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. Through the setting of the dynamic standard deviation fluctuation range and the determination of the abnormal value level, automatic identification and cleaning of data anomalies are achieved, improving the accuracy of marking abnormal moments and avoiding decision misjudgments caused by data noise; when judging the reasons for periodic anomalies, through the precise comparison of the business process cycle and the equipment maintenance cycle, the time for positioning the root cause of anomalies is greatly shortened, and the response speed for anomaly handling is improved; when judging the reasons for non-periodic anomalies, through the dynamic analysis of the frequency change value, differential response is achieved;

[0038] 2. Through the comprehensive calculation of the abnormal level, intelligent decision-making for the maintenance timing is achieved: when it is determined that the situation is urgent, immediate maintenance is triggered to avoid downtime losses caused by excessive wear of the equipment; in non-emergency situations, the advance maintenance strategy can extend the life of key equipment and reduce maintenance costs; a full-process risk closed-loop management is formed from anomaly identification, attribution, decision-making to verification, and the environmental data linkage filtering in the closed-loop management enables the system to still operate stably under extreme working conditions;

[0039] 3. By recording the operation data during the maintenance process, the effective operation data is screened and the effective operation data is filtered for accidental factors. The filtered effective operation data is then injected into the model, and incremental learning is used to improve the model's adaptation speed to new abnormal scenarios. Attached Figure Description

[0040] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0041] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0042] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0043] Example:

[0044] Please see Figure 1 As shown, a method and system for intelligent auxiliary decision-making driven by a large language model for cleaning and treatment includes a data acquisition module, a data processing module, a large language model module, and a decision output module.

[0045] The data acquisition module collects water quality data, equipment operation data, and environmental data. Water quality data includes pH value, turbidity, and flow rate data; equipment operation data includes equipment temperature, speed, and vibration data; and environmental data includes ambient temperature and humidity data.

[0046] The data processing module sorts the collected data by time and calculates the mean A1 and standard deviation B for multiple data points of the same item collected at the same time. It then sets the fluctuation range [A1-bB, A1+bB] for the corresponding item's detection data based on the mean A1 and standard deviation B, where b is a positive integer. Data points outside this fluctuation range are marked as outliers, and the number of outliers for each item is Y. a Statistical analysis shows that the total number of data points detected simultaneously for the corresponding item is Z. a If Y a ≥k1*Z a If the detected data is deemed abnormal, the count of value b is incremented by one to narrow down the range of abnormal values. k1 is a preset proportional coefficient, and a represents the sorting number of the corresponding data item. If the count of value b, after being incremented, still satisfies Y... a ≥k1*Z a Then mark the detection time as an abnormal time; if Y a <k1*Za If outliers are removed, the mean A2 is calculated for the remaining corresponding item detection data, and the calculated mean A2 is used as the corresponding item data detected at that time.

[0047] The fluctuation range [A1-bB, A1+bB] is set by using the mean A1 and standard deviation B. The core principle is to quantify the dispersion of data using the statistical concept of standard deviation. The standard deviation B reflects the amplitude of data fluctuation around the mean. When b is 1, the fluctuation range covers approximately 68% of the data (consistent with a normal distribution); when b is 2, it covers approximately 95%. In practice, if the number of outliers Y... a Exceeding the preset ratio k1*Z a This indicates that the current fluctuation range fails to accurately capture data anomalies. In this case, increasing b can narrow the range for anomaly detection (e.g., from b=2 to b=3), thereby improving the sensitivity of anomaly identification by dynamically shrinking the range. For example, the COD data of a wastewater treatment plant often fluctuates during the rainy season. When b=2 initially, the proportion of anomalies reached 30% (exceeding k1=0.2). After increasing b to 3, the proportion of anomalies dropped to 15%, effectively filtering out the noise data caused by the impact of the rainy season.

[0048] When Y a ≥k1*Z a The system identifies data anomalies in real time, avoiding misjudgments caused by a single data point anomaly. For example, a momentary sensor malfunction may produce a single outlier, but if the proportion of outliers does not exceed k1 (e.g., k1 = 0.2), the system will remove the outlier and recalculate the mean A2. Only when the proportion of outliers still exceeds the threshold after adjusting b is it marked as an anomaly. This is because the secondary verification mechanism ensures the authenticity of the anomaly. For instance, the proportion of anomalies in the vibration data of a water pump was 25% (k1 = 0.2) when b = 2. Even after increasing b to 3, the proportion of anomalies still reached 22%, which was marked as an anomaly. Subsequent inspection revealed bearing wear, verifying the accuracy of the mechanism.

[0049] Retrieve the historical data of the corresponding item, classify the historical data of the corresponding item according to the time length of each use, calculate the frequency of abnormal moments in the corresponding item data in each classification. If the frequency of abnormal moments in each classification is the same, it is determined that the occurrence of the abnormality is periodic; compare the period of the abnormal moment with the business process period of the purification system and the maintenance period of the equipment. If the period of the abnormal moment is the same as the business process period of the purification system, it is preliminarily determined that it is a periodic abnormality caused by business operations; then compare the corresponding period of the abnormal moment with the corresponding period of the business process period of the purification system. If the corresponding period of the abnormal moment occupies more than 80% of the corresponding period of the business process period of the purification system, it is determined that it is a periodic abnormality caused by business operations; if the period of the abnormal moment is the same as the maintenance period of the equipment, it is preliminarily determined that it is a periodic abnormality caused by maintenance operations; then compare the two period time segments of the abnormal moment period and the equipment maintenance period. If the time segments of the two periods are the same, it is determined that it is a periodic abnormality caused by maintenance operations; if it is found through comparison that the period of the abnormal moment is not caused by the business process period of the purification system and the maintenance period of the equipment, an early maintenance signal is generated and the early maintenance signal is transmitted to the decision output module;

[0050] The determination basis that the abnormal period occupies more than 80% of the business process period is that when the abnormality is strongly related to business operations, its occurrence time should highly coincide with the business cycle; for example, if the high-load production period of a certain factory is from 18:00 to 20:00 every day, and if 90% of the abnormal moments are concentrated in this period, it can be determined that it is caused by business operations; and the requirement for the maintenance period is "the time segments are exactly the same" because equipment maintenance is usually carried out within a fixed time window (such as from 8:00 to 10:00 on the 1st of each month). If the abnormality only appears during the maintenance period (such as vibration abnormality only occurs after the maintenance on the 1st of each month), it can be directly attributed to the maintenance operation; the maintenance period of the aerator in a sewage treatment plant is the 5th of each month, and the abnormal moments in three consecutive months all occur within 1 hour after the maintenance. Through the complete matching of the time segments, it is determined that it is caused by improper maintenance debugging;

[0051] After receiving the early maintenance signal, the decision output module calculates the average value A3 of the abnormal data at the abnormal moment, obtains the corresponding item data at the adjacent moments of the abnormal moment, and calculates the average value A4. Divide the difference between the average values A3 and A4 by the preset fluctuation value to obtain the fluctuation ratio c. If (d - 1)*c1 < c < d*c1, it is determined that the abnormal level is level d, and c1 is the preset comparison value; the period length d of the abnormal moment Y is compared with the preset period length d S to determine the abnormal level as level; then calculate The sum value d1 is obtained, and the sum value d1 is compared with a preset comparison value d2. If d1 < d2, it is determined to maintain the original maintenance cycle; if d2 < d1 < k1 * d2, it is determined to perform maintenance in advance. Here, k1 is a preset proportionality coefficient, and t1 is the time duration between the current time point and the next maintenance cycle. If d1 ≥ k1 * d2, it is determined to perform maintenance immediately;

[0052] Fluctuation ratio reflects the current abnormal amplitude, and the cycle length ratio reflects the deviation degree of the abnormal duration from the preset cycle. The design of adding the two is based on a risk quantification model: the greater the amplitude and the shorter the cycle, the higher the risk. For example, for a certain abnormality, c = 1.5 (medium amplitude), d Y = 3 days, d S = 7 days, the cycle ratio is 0, and the comprehensive grade d1 = 1.5 + 0 = 1.5. If the preset comparison value d2 = 2, it is determined to maintain the original maintenance cycle; when c = 2.0, d Y = 2 days, d S = 7 days (cycle ratio 1), d1 = 3.0 ≥ k1 * d2 (k1 = 1.5, d2 = 2), then immediate maintenance is triggered. When a water pump vibrates abnormally, bearing wear is detected 2 days in advance through this calculation, avoiding a shutdown accident;

[0053] Early maintenance time The setting of the early maintenance time is based on the equipment wear theory. For example, the remaining life prediction of a certain equipment is t1 = 48 hours, and early hour maintenance can balance the maintenance cost and the failure risk. The k1 in the immediate maintenance threshold d1 ≥ k1 * d2 usually takes 1.5 - 2.0 to ensure that emergency abnormalities are processed first. For example, for a certain equipment with an abnormal vibration grade d1 = 3.0, d2 = 2.0, k1 = 1.5, 3.0 ≥ 1.5 × 2.0, the system immediately triggers shutdown maintenance, and it is found through inspection that the bearing balls are broken, avoiding equipment scrapping;

[0054] If the anomaly is determined to be non-periodic, the frequency change value at the time of the anomaly is calculated. If the frequency change value at the time of the anomaly continuously increases, the manual operation data of the corresponding item within a set time period prior to the current time point is obtained. If there is no manual operation data for the corresponding item, it is determined that the system has spontaneously recovered, and the recovery conditions are recorded to provide a reference for subsequent adjustments to similar anomalies. If there is manual operation data for the corresponding item, the manual decision is determined to be effective, and the current effective parameters are maintained for a period of time. The anomaly frequency is continuously monitored, and after confirming that the trend is stable, the investment of emergency resources is reduced in an orderly manner. If the frequency change value at the time of the anomaly continuously decreases, it is determined that the risk of system failure or water quality deterioration is gradually increasing. When the frequency change value at the time of the anomaly reaches the preset frequency change threshold, an early warning signal is generated and transmitted to the decision output module. If the frequency change value at the time of the anomaly continuously fluctuates, the manual operation data of the corresponding item within a set time period prior to the current time point is obtained. If there is manual operation data for the corresponding item, the frequency change value at the time of the anomaly is detected again after the manual operation is completed. If there is no manual operation data for the corresponding item, an alarm signal is generated and transmitted to the decision output module.

[0055] The criterion for determining frequency change is based on a sliding window algorithm (e.g., a 5-minute window), which calculates the rate of change of the frequency at abnormal times; "continuously increasing" means that the rate of change is consistently positive and exceeds a threshold (e.g., 0.1 times / hour). 2 "Continuously decreasing" means the rate of change is consistently negative and its absolute value exceeds the threshold; "fluctuation" means the rate of change oscillates between positive and negative. The preset frequency change threshold is usually taken as the 3σ value of historical data. For example, the threshold for the abnormal frequency change rate of a water pump is set to 0.2 times / hour. 2 When the rate of change is detected to be 0.25 times / hour for three consecutive windows. 2 When the risk increases, an early warning is generated; this dynamic analysis can capture the development trend of anomalies, such as when the frequency of COD anomalies in a certain water quality increases from 0.5 times / day to 1.2 times / day (a change rate of 0.3 times / day). 2 The system provides 24-hour advance warnings, preventing water quality from exceeding standards.

[0056] To determine if the system will spontaneously recover without human intervention, it is necessary to compare environmental parameters (such as temperature and flow rate) and equipment status parameters (such as speed and voltage) before and after the anomaly. The recovery condition record should include a complete data sequence for one hour before and after the anomaly, such as "2024-04-10 10:00, pH anomaly recovered, temperature at recovery time 25℃, influent flow rate 500m³ / h". 3 / h, dissolved oxygen 2.5mg / L; The pH anomaly at a wastewater treatment plant automatically recovered after the rainstorm stopped. The system recorded "Recovery conditions: rainfall stopped, influent flow rate from 800m³ / h, dissolved oxygen 2.5mg / L"; 3 / h dropped to 500m 3The system uses the " / h" parameter to provide a reference for subsequent heavy rain response. When manual intervention is involved, the system confirms the effectiveness of the decision by monitoring the stability time of parameters (e.g., no fluctuation for 30 consecutive minutes), thus avoiding frequent adjustments to resource allocation.

[0057] Upon receiving an early warning signal, the decision output module identifies the corresponding data. If the warning signal is for pH, it notifies staff via SMS to adjust the pH level within a set timeframe. If the warning signal is for turbidity or flow rate, it records the turbidity / flow rate data, marks the corresponding abnormal moment, and sends the abnormal information to staff to draw their attention. If the warning signal is for vibration, it sends the abnormal information to staff and immediately triggers shutdown protection when the equipment vibration amplitude exceeds 200% of the preset rated value. If the warning signal is for speed or equipment temperature, it reduces the equipment load. If the warning signal is for ambient temperature / humidity, it implements insulation measures for the purification system and strengthens ventilation and dehumidification operations. Upon receiving an alarm signal, the decision output module detects the frequency change amplitude at the abnormal moment. If the frequency change amplitude exceeds a preset comparison threshold, it triggers an audible and visual alarm to notify staff to perform timely maintenance. If it does not exceed the preset comparison threshold, it sends the abnormal information to staff to draw their attention.

[0058] When staff perform maintenance operations, changes in various data during the maintenance process are detected and recorded. The values ​​of each data point after adjustment are also recorded. These changed values ​​are compared to the preset fluctuation range of the corresponding data (the preset fluctuation range is between the minimum and maximum values ​​of the corresponding data). If the changed value is outside the preset fluctuation range, the absolute value of the difference between the changed value and the preset fluctuation range is calculated. The minimum absolute value of this difference is taken as the gap between the changed value and the preset fluctuation range. The difference is calculated as the difference divided by the difference between adjacent values ​​within the fluctuation range of the corresponding data. If the difference is greater than 1, the operation adjustment is considered effective, and the resulting data change is recorded. Conversely, if the difference is less than 1, the operation adjustment is considered ineffective, and the resulting data change is not recorded.

[0059] The difference value is defined as "difference / difference between adjacent values ​​within the fluctuation range," where "difference" is the minimum distance between the adjusted value and the preset fluctuation range, and "difference between adjacent values" reflects the data accuracy (e.g., the accuracy of pH is 0.1). When the difference value > 1, it means that the deviation between the adjusted value and the target range exceeds one accuracy unit, and the operation has a substantial improvement on the system. For example, if the adjusted pH value is 5.5, the preset range is 6.0-8.0, the difference is 0.5, and the difference between adjacent values ​​is 0.1, the difference value = 0.5 / 0.1 = 5 > 1, and the operation is deemed effective. If the adjusted value is 5.9, the difference is 0.1, and the difference value = 1.0, then it just meets the standard. This quantification method avoids the vague judgment of "being close to meeting the standard is effective," ensuring the quality of operation.

[0060] The system retrieves operation data from historical data records corresponding to the operations, and calculates the number of times each operation was performed. Operation data with fewer than a preset number of performances is filtered out, completing the first data filtering operation. The resulting operation data is then marked, and the corresponding environmental data is marked again. The system then retrieves the environmental data corresponding to the same operation data from historical data and compares the marked environmental data with the corresponding items in the retrieved environmental data. If the fluctuation of the environmental data corresponding to the two operation data is within a preset fluctuation range, both operation data are retained; otherwise, both sets of operation data are filtered out, completing the second data filtering operation. The operation data after the second data filtering operation is recorded and then fed into a large language model to optimize the already trained large language model.

[0061] The first filter removes data that has been performed less than the preset number of times (e.g., the preset number is 5 times), as a small amount of data lacks statistical significance and may be only occasionally effective. The second filter compares environmental data fluctuations. For example, if an operation is effective at an ambient temperature of 25℃, but another identical operation at a temperature of 35℃ has an environmental fluctuation exceeding ±5℃, then that data is removed to avoid environmental interference causing the model to mislearn. A dosing operation at a wastewater treatment plant was occasionally effective during the rainy season (performed 3 times), and was removed during the first filter. Another dosing operation was effective 5 times under the same temperature and flow conditions, with an environmental data fluctuation of <±2℃, and was retained for model training, improving the model's adaptability to normal operating conditions.

[0062] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for intelligent auxiliary decision-making in a purification system driven by a large language model, characterized in that, The intelligent decision-making assistance method is implemented through an intelligent decision-making assistance system, which includes a data acquisition module, a data processing module, a large language model module, and a decision output module. The intelligent decision-making assistance method of the purification system includes the following steps: S1: The data processing module performs outlier analysis on the collected data. If the number of outliers exceeds the preset number, the range of outlier fluctuations is expanded. If it is still greater than the preset number, this moment is marked as an outlier moment. Otherwise, after removing outliers, the mean is recalculated as valid data. S2: Retrieve historical data and classify it by usage time. Calculate the frequency of abnormal moments in each category. If the frequencies are the same, the abnormality is determined to be periodic. Further compare the abnormality period with the business process period or equipment maintenance period. Based on the overlap of time periods or whether the time periods are consistent, determine whether the abnormality is caused by business operations or maintenance operations. If the abnormality period is unrelated to either, generate an early maintenance signal and pass it to the decision output module. S3: If the frequencies are different, calculate the frequency change value; if the frequency change value continues to increase, check if there is any manual operation. If there is no operation, determine that the system will automatically recover and record the conditions. If there is operation, confirm that the manual decision is effective and maintain parameter monitoring; if the frequency change value continues to decrease, determine that the risk is increasing and generate an early warning signal when the threshold is reached; if the frequency change value continues to fluctuate, check for manual operation. If there is operation, wait for it to be completed and re-detect. If there is no operation, generate an alarm signal and trigger an audible and visual alarm or information reminder according to the frequency change amplitude. S4: After receiving the signal, the decision output module calculates the fluctuation ratio of the average abnormal data at the abnormal time to the average of adjacent times to determine the abnormal level. It also calculates the comprehensive abnormal level by combining the ratio of the abnormal period length to the preset period length. The maintenance timing is determined based on the comparison between the comprehensive abnormal level and the preset threshold. If the signal is a warning or alert, the module notifies the staff to carry out maintenance. During maintenance, the data changes are recorded and compared with the preset fluctuation range. The difference value is calculated. If the difference value is greater than 1, the operation is deemed valid and recorded; otherwise, it is ignored. After filtering out accidental factors from the valid data, it is used for large language model optimization.

2. The intelligent auxiliary decision-making method for a large language model-driven purification system according to claim 1, characterized in that, The data processing module performs the following steps to determine abnormal situations: K1: Sort the collected data by time and average the data for the corresponding item collected at the same time. The calculation of the mean and standard deviation B, and the expression for the mean. The corresponding item test data fluctuation range was performed with the standard deviation B. In this setting, b is a positive integer; K2: Mark the data for items outside the fluctuation range as outliers, and specify the number of outliers for each item. Statistical analysis showed that the total number of data detected for the corresponding item at the same time was: ,like If the detected data is deemed abnormal, the count for value b is incremented by one to narrow down the range of abnormal values. This is a preset proportional coefficient. Indicates the sorting number of the corresponding data item; K3: If the value b is incremented by one and still satisfies the condition... If this detection time is not specified, then mark this detection time as an abnormal time; if If outliers are removed, the remaining corresponding test data are averaged. The calculation, using the calculated mean This is the corresponding data detected at this moment.

3. The intelligent auxiliary decision-making method for a large language model-driven purification system according to claim 2, characterized in that, The data processing module performs periodic analysis at abnormal times using the following steps: M1: Retrieve historical data for the corresponding item, classify the historical data for the corresponding item according to the length of time each time it is used, calculate the frequency of abnormal moments in the corresponding item data in each category, and if the frequency of abnormal moments is the same in each category, it is determined that the occurrence of abnormality is periodic. M2: Compare the period of the abnormal moment with the business process cycle of the cleaning system and the maintenance cycle of the equipment. If the period of the abnormal moment is the same as the business process cycle of the cleaning system, it is initially determined to be a periodic abnormality caused by business operation. Then, the corresponding time period of the abnormal moment is compared with the corresponding time period of the business process cycle of the purification system. If the corresponding time period of the abnormal moment accounts for more than 80% of the corresponding time period of the business process cycle of the purification system, it is determined to be a periodic abnormality caused by business operation. M3: If the period of the abnormal time is the same as the maintenance cycle of the equipment, it is initially determined to be a periodic abnormality caused by maintenance operation; then compare the two periodic time periods of the abnormal time with the equipment maintenance cycle. If the two periodic time periods are the same, it is determined to be a periodic abnormality caused by maintenance operation. If the comparison reveals that the abnormal time period is not caused by the business process cycle of the cleaning system and the maintenance cycle of the equipment, an early maintenance signal is generated and transmitted to the decision output module.

4. The intelligent auxiliary decision-making method for a large language model-driven purification system according to claim 3, characterized in that, The data processing module performs the following steps for non-periodic analysis during abnormal moments: N1: Calculate the frequency change value at abnormal times. If the frequency change value at abnormal times continues to increase, obtain the manual operation data of the corresponding item within the set time period before the current time point. If there is no manual operation data for the corresponding item, it is determined that the system has spontaneously recovered, and the recovery conditions are recorded to provide a reference for subsequent similar anomaly adjustments. If there is manual operation data for the corresponding item, it is determined that the manual decision has taken effect, and the current effective parameters are kept running for a period of time. The abnormal frequency is continuously monitored, and after the trend is confirmed to be stable, the investment of emergency resources is reduced in an orderly manner. N2: If the frequency change value at the abnormal moment continues to decrease, it is determined that the risk of system failure or water quality deterioration is gradually increasing. When the frequency change value at the abnormal moment reaches the preset frequency change threshold, an early warning signal is generated and transmitted to the decision output module. N3: If the frequency change value at the abnormal moment keeps fluctuating, then obtain the manual operation data of the corresponding item within the set time period before the current time point. If there is manual operation data for the corresponding item, then after the manual operation is completed, the frequency change value at the abnormal moment will be detected again. If there is no manual operation data for the corresponding item, an alarm signal will be generated and transmitted to the decision output module.

5. The intelligent auxiliary decision-making method for a large language model-driven purification system according to claim 2, characterized in that, The data processing module determines valid operation data as follows: G1: Detect and record the changes in various data during the maintenance process, and record the values ​​of the changes in various data after the corresponding data is adjusted. Compare the recorded values ​​of the corresponding data changes with the preset fluctuation range of the corresponding data. The preset fluctuation range is between the minimum and maximum values ​​of the corresponding data. G2: If the value of the corresponding data after the change is within the preset fluctuation range of the corresponding data, the operation adjustment is deemed effective, and the data change caused by the operation adjustment is recorded. If the value of the corresponding data after the change is not within the preset fluctuation range of the corresponding data, then calculate the absolute value of the difference between the value of the corresponding data after the change and the preset fluctuation range of the corresponding data. The minimum absolute value of the calculated difference is the gap between the value of the corresponding data after the change and the preset fluctuation range of the corresponding data. The gap value is equal to the gap divided by the difference between adjacent values ​​within the fluctuation range of the corresponding data. G3: If the difference value is greater than 1, the operation adjustment is deemed effective, and the data changes caused by the operation adjustment are recorded; Conversely, if the operation adjustment is invalid, the data changes caused by the operation adjustment will not be recorded.

6. The intelligent auxiliary decision-making method for a large language model-driven purification system according to claim 2, characterized in that, The data processing module performs the following steps to filter out random factors: P1: Obtain the operation data of the corresponding operation record in the historical data, retrieve the number of times the corresponding operation data was performed, filter the operation data with fewer than the preset number of times, and complete a data filtering and screening operation. P2: Mark the operation data obtained after completing a data filtering and screening operation, mark the environmental data corresponding to the marked operation data again, then obtain the environmental data corresponding to the operation data that is the same as the marked operation data in the historical data, and compare the marked environmental data with the corresponding items in the obtained environmental data. P3: If the fluctuation of the environmental data corresponding to the two operation data is within the preset fluctuation range of the corresponding item, then the two operation data will be retained. Conversely, both sets of operational data are filtered out to complete the secondary data filtering operation; P4: Record the operation data after the secondary data filtering and elimination operation, and substitute the operation data into the large language model to optimize the trained large language model.

7. The intelligent auxiliary decision-making method for a large language model-driven purification system according to claim 4, characterized in that, The decision output module executes the following steps for the advance maintenance signal: Q1: Mean of outlier data at outlier times Perform calculations, obtain corresponding data for time intervals adjacent to the abnormal time, and calculate the mean. , the mean and The difference is divided by the preset volatility value to obtain the volatility ratio c. If so, the anomaly level is determined to be level d. This is the preset comparison value; Q2: What is the period length at the time of the anomaly? With preset cycle length The comparison was made to determine the anomaly level. class; Q3: Calculation The sum of and sum value Comparison with preset value If a comparison is made, If so, the original maintenance cycle will be maintained; if Then it is determined to be early. Perform maintenance operations. This is a preset proportional coefficient. This refers to the duration between the current time and the next maintenance cycle. like If so, it is determined that maintenance operations should be performed immediately.

Images