Dynamic alarm threshold adaptive adjustment method based on reinforcement learning

Abstract

The invention discloses a dynamic alarm threshold adaptive adjustment method based on reinforcement learning, and relates to the technical field of industrial automation and intelligent operation and maintenance, and the method comprises the steps: obtaining the operation parameters of all devices in a system, the upstream and downstream correlation parameters between the devices and historical alarm data in real time, carrying out the feature extraction of the obtained data, and carrying out the feature extraction; respectively obtaining an operation parameter feature vector and an associated parameter feature vector; based on the feature vectors, generating an equipment state influence matrix through a parameter interaction analysis model; when the real-time operation parameter of any device is close to the current alarm threshold value, combining a preset device correlation model and a device state influence matrix, and analyzing the correlation and the influence degree between the current device and other devices; through step-by-step adjustment and model joint optimization, the reasonability of threshold adjustment can be continuously improved, long-term stable operation of the system is ensured, and unnecessary intervention of operation and maintenance personnel is reduced.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation and intelligent operation and maintenance technology, specifically a dynamic alarm threshold adaptive adjustment method based on reinforcement learning. Background Technology

[0002] In the operation monitoring of complex systems such as industrial control systems, data center server clusters, and smart grids, the rationality of equipment alarm thresholds directly determines the accuracy and timeliness of system fault warnings.

[0003] The current mainstream alarm threshold adjustment methods mostly adopt preset fixed thresholds or static adjustment strategies based on historical data of a single device. The former sets the threshold to a fixed value, which cannot adapt to scenarios where the operating parameters of the device change dynamically with the operating conditions. The latter updates the threshold based on the historical data of a single device, but does not consider the upstream and downstream correlation between devices in the system. When the operating parameters of a certain device are close to the threshold, only the threshold of that device is adjusted, ignoring its impact on related devices, which may lead to missed alarms for related device failures.

[0004] Meanwhile, the existing adjustment strategy lacks a closed-loop optimization mechanism: after the threshold is adjusted, the reasonableness of the threshold is judged only based on the alarm result of the device, without combining the fault handling result after the alarm to optimize the device correlation analysis model and threshold calculation logic in reverse, which makes it difficult to continuously improve the accuracy of threshold adjustment. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic alarm threshold adaptive adjustment method based on reinforcement learning to solve the problems mentioned in the background art.

[0006] A method for adaptive adjustment of dynamic alarm thresholds based on reinforcement learning, comprising: S1. Real-time acquisition of operating parameters of each device in the system, upstream and downstream correlation parameters between devices, and historical alarm data, and feature extraction of the acquired data to obtain operating parameter feature vectors and correlation parameter feature vectors respectively; S2. Based on the feature vector, generate a device status influence matrix through a parameter interaction analysis model; when the real-time operating parameters of any device are close to the current alarm threshold, combine the preset device correlation model and device status influence matrix to analyze the correlation and influence degree between the current device and other devices. S3. Based on the aforementioned correlation and degree of influence, construct a multi-objective optimization function, combine real-time operating parameters and historical alarm data, and perform step-by-step threshold adjustment on the current device and perform threshold pre-adjustment on the associated devices through reinforcement learning algorithm; S4. Apply the adjusted threshold to system monitoring. When the equipment operating parameters exceed the adjusted alarm threshold in multiple consecutive sampling periods, an alarm is triggered. Based on the alarm results, the equipment correlation model and parameter interaction analysis model are jointly optimized.

[0007] Furthermore, the feature extraction in step S1 includes: Dynamically filter out key operating parameters that are closely related to alarm events; The upstream and downstream correlation parameters are processed in time and space to extract the characteristics related to the connection and transmission between devices; Based on the processed data, a comprehensive parameter representation is formed that simultaneously includes the device's own operating status and the inter-device relationship status.

[0008] Furthermore, the generation of the device state influence matrix in step S2 includes: Through the parameter interaction analysis model, the operating parameter data is transformed into a set of device sensitivity data, and the associated parameter data is transformed into a set of associated transmission data. The device sensitivity data set and the associated transmission data set are fused together to generate a dynamically updatable device status influence matrix; The values ​​in the matrix are dynamically updated according to the system's operating mode, and are used to characterize the dynamic changes in the interaction relationships between parameters.

[0009] Furthermore, the device status influence matrix is ​​dynamically switched according to the system operating conditions, and the method further includes: Several typical system operating conditions are predefined, and a corresponding equipment state influence matrix is ​​trained for each operating condition. Real-time monitoring of system load and start / stop status parameters to determine the current operating condition; Call the equipment status influence matrix that matches the current operating conditions for the correlation and influence analysis in step S2.

[0010] Compared with the prior art, the beneficial effects of the present invention are: The innovation of this invention lies in its threshold adjustment method, which combines multi-dimensional data and algorithms. When equipment parameters approach alarm thresholds, it can analyze the impact of related devices, avoiding false alarms or missed alarms caused by traditional fixed thresholds. For example, when the inverter output voltage in the system fluctuates slightly, approaching but not reaching the alarm threshold, traditional methods might ignore this fluctuation. However, this invention, through correlation analysis, discovers the high correlation between the inverter and the water pump, and pre-adjusts the water pump threshold in advance. If the inverter voltage continues to drop and triggers an alarm, the water pump can respond quickly, preventing the fault from escalating due to the threshold not being adjusted. Simultaneously, step-by-step adjustment and model joint optimization continuously improve the rationality of threshold adjustments, ensuring long-term stable system operation and reducing unnecessary intervention by maintenance personnel. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the method framework structure of the present invention. Detailed Implementation

[0012] 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.

[0013] Please see Figure 1 This application provides a dynamic alarm threshold adaptive adjustment method based on reinforcement learning, including: S1. Real-time acquisition of operating parameters of each device in the system, upstream and downstream correlation parameters between devices, and historical alarm data. Feature extraction is performed on the acquired data to obtain feature vectors for operating parameters and correlation parameters. It should be noted that operating parameters refer to parameters that directly reflect the operating status of the device itself, such as pump speed, inverter output current, and sensor-detected temperature. These can be acquired in real-time through the device's built-in acquisition module, with the acquisition frequency set according to the device type. Upstream and downstream correlation parameters refer to parameters that reflect the dependency relationship between devices. For example, as an upstream device of the pump, the inverter's output voltage change directly affects the pump speed; therefore, the data on the correspondence between the inverter's output voltage and the pump speed belongs to upstream and downstream correlation parameters. Historical alarm data refers to the parameter records, alarm types, and processing results when devices triggered alarms over a past period.

[0014] It should be understood that when extracting features from data, for operating parameters, the mean, variance, and trend are extracted. For example, the pump speed data collected 10 times consecutively (1450 rpm, 1448 rpm, 1452 rpm, etc.) are calculated to have a mean of 1450 rpm, a variance of 2, and a stable trend. These data combined constitute the operating parameter feature vector. For correlated parameters, the correlation coefficient between parameters of different devices is extracted. For example, for every 10V change in the inverter output voltage, the corresponding change in pump speed is 50 rpm, and the correlation coefficient between the two is 0.95, thus forming the correlated parameter feature vector. This transforms the massive amount of raw data into key information reflecting the equipment status, laying the foundation for subsequent analysis and avoiding interference from redundant information in the raw data.

[0015] S2. Based on feature vectors, a device status influence matrix is ​​generated through a parameter interaction analysis model. When the real-time operating parameters of any device approach the current alarm threshold, the correlation and influence degree between the current device and other devices are analyzed by combining a preset device correlation model and the device status influence matrix. It should be noted that the parameter interaction analysis model refers to training the feature vectors of operating parameters and related parameters using machine learning algorithms (such as random forest algorithms) to establish a model of the interaction and influence relationships between devices. The model is optimized by inputting historical data to improve the accuracy of the analysis. The device status influence matrix is ​​a dataset that presents the degree of mutual influence between devices in matrix form. The rows and columns in the matrix represent different devices, and the values ​​at the intersections represent the influence weight of the corresponding row device on the column device; the larger the value, the stronger the influence.

[0016] The preset device correlation model refers to a model that pre-determines the degree of correlation between devices based on their physical connections and functional dependencies within the system. For example, in a water supply system, a water pump is directly connected to a water pipe pressure sensor, and the pressure sensor needs to monitor the water pump's supply pressure in real time; therefore, their correlation is set to 0.9. However, although the water pump and a distant level sensor both belong to the water supply system, they are not directly physically connected, so their correlation is set to 0.3. When the real-time speed of the water pump approaches the alarm threshold, the system calls the device correlation model and finds that the correlation between the water pump and the pressure sensor is 0.9. Combining this with the water pump's influence weight of 0.7 on the pressure sensor in the device status influence matrix, the system comprehensively judges that if the water pump speed continues to decrease and triggers an alarm, the pressure sensor's detected pressure may drop from the normal 0.4 MPa to 0.3 MPa, indicating a significant impact, requiring close monitoring of the pressure sensor's status.

[0017] S3. Based on the correlation and degree of influence, a multi-objective optimization function is constructed. Combining real-time operating parameters and historical alarm data, a reinforcement learning algorithm is used to perform step-by-step threshold adjustment for the current device and pre-adjust the thresholds for associated devices. It should be noted that the multi-objective optimization function is a mathematical function constructed with the goals of reducing the false alarm rate of the current device, reducing cascading alarms of associated devices, and ensuring stable system operation. First, by setting the weight of each objective, the optimization values ​​of different threshold adjustment schemes are calculated, and the scheme with the highest optimization value is selected.

[0018] For example, in one specific embodiment, the initial threshold for the water pump speed alarm is 1200 rpm. Historical data shows that when the speed is between 1200-1250 rpm, 80% of the cases will not trigger subsequent faults, and only 20% will actually require an alarm, indicating a high false alarm rate. By combining real-time speed data of 1250 rpm and historical alarm data, and training with a reinforcement learning algorithm (such as Q-learning), the algorithm will try different threshold adjustment schemes. Option 1 lowers the threshold to 1180 rpm, which reduces the false alarm rate to 10%. However, the associated pressure sensor may trigger an alarm due to the lower pump speed, increasing the cascading alarm rate to 15%. Option 2 adopts a step-by-step adjustment. First, the threshold is reduced to 1220 rpm, and three sampling cycles are observed. If no false alarms occur and the equipment operates stably, the threshold is then reduced to 1200 rpm. At this point, the false alarm rate drops to 12%, and the cascading alarm rate is only 5%, which is a higher optimization value and becomes the final choice.

[0019] Furthermore, when performing threshold pre-adjustment on related devices, the pre-adjustment range is set according to the degree of influence. For example, the normal alarm threshold of the pressure sensor is 0.3MPa. Because the water pump has a high degree of influence on it, it is pre-adjusted to 0.32MPa. If the water pump speed triggers an alarm later, the pressure sensor can respond quickly without significantly adjusting the threshold, avoiding alarm lag caused by a fixed threshold. For the level sensor with a low correlation, its alarm threshold is only slightly adjusted from 1.5m to 1.48m, reducing excessive adjustment of non-critical devices and reducing the system burden.

[0020] S4. The adjusted threshold is applied to system monitoring. When the equipment operating parameters exceed the adjusted alarm threshold for multiple consecutive sampling periods, an alarm is triggered, and the equipment correlation model and parameter interaction analysis model are jointly optimized based on the alarm results. It should be noted that when the adjusted alarm threshold of 1220 rpm for the water pump is applied, if the water pump speed is 1210 rpm for three consecutive sampling periods (each period is 2 seconds), exceeding the adjusted threshold, the system will trigger an alarm and notify the maintenance personnel to handle it.

[0021] Meanwhile, the system records detailed information about this alarm: the water pump speed was 1210 rpm, the pressure sensor pressure was 0.35 MPa, and there were no related equipment cascading alarms. Based on this information, the system optimizes the equipment correlation model and finds that the water pump alarm did not have the expected high impact on the pressure sensor. The correlation between the two is then reduced from 0.9 to 0.8. When optimizing the parameter interaction analysis model, the alarm data is input into the model to retrain the equipment status influence matrix, adjusting the influence weight of the water pump on the pressure sensor from 0.7 to 0.6, thus improving the accuracy of subsequent analysis. The innovation of this invention lies in its threshold adjustment method, which combines multi-dimensional data and algorithms. When equipment parameters approach alarm thresholds, it can analyze the impact of related devices, avoiding false alarms or missed alarms caused by traditional fixed thresholds. For example, when the inverter output voltage in the system fluctuates slightly, approaching but not reaching the alarm threshold, traditional methods might ignore this fluctuation. However, this invention, through correlation analysis, discovers the high correlation between the inverter and the water pump, and pre-adjusts the water pump threshold in advance. If the inverter voltage continues to drop and triggers an alarm, the water pump can respond quickly, preventing the fault from escalating due to the threshold not being adjusted. Simultaneously, step-by-step adjustment and model joint optimization continuously improve the rationality of threshold adjustments, ensuring long-term stable system operation and reducing unnecessary intervention by maintenance personnel.

[0022] As a further embodiment, the feature extraction in step S1 includes: Dynamically filter out key operating parameters closely related to alarm events; it should be noted that key operating parameters refer to parameters that have a direct indicative effect on equipment alarms and frequently show anomalies in historical alarm events, rather than all operating data of the equipment.

[0023] Furthermore, filtering can be achieved by comparing historical alarm data with normal operating data. For example, first, count the frequency of abnormal occurrences of each operating parameter when the equipment triggers an alarm over the past year, then calculate the correlation probability between the parameter abnormality and the alarm event, and list parameters with a correlation probability exceeding a preset value (such as 70%) as key parameters. For another example, for water pump equipment, historical data shows that in alarm events, abnormal speed occurred 92 times, abnormal casing temperature occurred 65 times, and abnormal fan speed occurred only 18 times. When calculating the correlation probability, the correlation probability between abnormal speed and alarm is 92% (92 / 100 alarms), temperature is 65%, and fan speed is 18%. If the preset value is set to 70%, then only speed will be listed as a key operating parameter.

[0024] This approach avoids including parameters with low correlation to alarms, such as fan speed, in the analysis, reducing data redundancy and preventing frequent normal fluctuations from being misjudged as potential anomalies by the system, interfering with the analysis of truly critical speed parameters and affecting the accuracy of subsequent feature vectors.

[0025] The upstream and downstream related parameters are processed in both time and space dimensions to extract the characteristics related to the connection and transmission between devices. It should be noted that the time dimension processing refers to aligning the acquisition time of the upstream and downstream device parameters to ensure that the analysis focuses on the parameter interaction relationship within the same time period. The spatial dimension processing is based on the physical location and connection path of the devices in the system to filter out the related parameters that directly involve data or energy transmission and exclude irrelevant data that are indirectly related.

[0026] For example, in a water supply system, the output voltage data of the frequency converter (upstream equipment) is collected at the 1st, 3rd, and 5th seconds, while the speed data of the water pump (downstream equipment) is collected at the 2nd, 4th, and 6th seconds. When processing the time dimension, the voltage data of the frequency converter at the 1st second will be matched with the speed data of the water pump at the 2nd second (because the voltage change needs 1 second to be transmitted to the water pump and affect the speed), forming a time-aligned parameter pair. In terms of spatial dimension, the frequency converter and the water pump are directly connected by a cable, and the voltage signal is directly transmitted. Their voltage-speed parameters are directly related. However, the frequency converter and the water tank level sensor at a distance are not directly connected. The level data needs to be transferred through multiple devices, which is an indirect relationship and will be excluded.

[0027] This processing method allows for the extraction of the true transmission characteristics between devices. For example, after time alignment, it can be observed that one second after the inverter voltage drops from 380V to 360V, the water pump speed drops from 1450 rpm to 1380 rpm, clearly demonstrating the transmission effect of voltage on speed. If time alignment is not performed and the inverter voltage in the first second is directly compared with the water pump speed in the first second, the transmission delay may lead to the mistaken belief that there is no obvious correlation between the two, making it impossible to extract accurate transmission characteristics.

[0028] Based on the processed data, a comprehensive parameter representation is formed that simultaneously includes the device's own operating status and the inter-device relationship status. It should be noted that the comprehensive parameter representation integrates the characteristics of key operating parameters with the processed upstream and downstream relationship parameters to form a dataset that comprehensively reflects the device's own status and its interaction with other devices, rather than a single-dimensional combination of parameters. It should be understood that the construction of the comprehensive parameter representation first standardizes the characteristics such as the mean and variance of key operating parameters and the characteristics such as the propagation delay and correlation coefficient of relationship parameters, and then performs a weighted fusion according to preset weights.

[0029] This representation can simultaneously reflect the status of the equipment itself and its associated status, reflecting both the pump's own abnormalities and the uninterrupted transmission of signals. It provides a comprehensive basis for the subsequent generation of operating parameter feature vectors and associated parameter feature vectors, avoiding the problem of focusing only on the equipment's own status while ignoring whether the associated equipment is transmitting signals normally.

[0030] The innovation of this application lies in the fact that, from the perspective of operation and maintenance efficiency, dynamic screening of key operating parameters can eliminate redundant data with low correlation, reduce the amount of data processing, reduce system computing power consumption, and enable operation and maintenance personnel to quickly locate core parameters, thereby improving the speed of operation and maintenance response. In terms of accuracy, processing the time and space dimensions of the associated parameters can solve the judgment bias caused by parameter time misalignment and fuzzy association path in traditional analysis, making the transmission characteristics between devices more realistic. When generating the device status influence matrix, the influence weight between devices can be accurately set, providing an accurate basis for judging the risk of device abnormal chain reaction. From the perspective of system stability, the comprehensive parameter representation covers both the device's own status and the associated status, avoiding misjudgments caused by relying on only a single parameter. When constructing a multi-objective optimization function, it can take into account the status of multiple devices, formulate a more reasonable threshold adjustment scheme, reduce the frequency of system alarms or missed alarms, maintain stable equipment operation, and reduce unnecessary operation and maintenance interventions.

[0031] It should be understood that in industrial control systems, the parameter interactions between devices often change dynamically with changes in operating load and environmental conditions. If the device status influence matrix remains fixed, it can easily lead to subsequent correlation analysis deviating from reality.

[0032] As a further embodiment, the generation of the device state influence matrix in step S2 includes: Through the parameter interaction analysis model, the operating parameter data is converted into a device sensitivity data set, and the associated parameter data is converted into an associated transmission data set. It should be noted that the device sensitivity data set refers to a dataset that reflects the sensitivity of the device to changes in its own operating parameters. The more obvious the device's state response after parameter changes, the higher the sensitivity value.

[0033] During the conversion, first define the normal fluctuation range of the parameters, then calculate the probability of equipment malfunction when the parameters exceed the range. For example, if the normal voltage range of the frequency converter is 370V-390V, the probability of triggering overcurrent protection when the voltage drops to 360V is 85%, so the sensitivity value is set to 0.85; when it rises to 400V, the probability of overvoltage protection is 60%, so the value is 0.6. This conversion can clearly reflect the differences in the sensitivity of the equipment to changes in different parameters, avoiding the bias of judgment based on experience.

[0034] The correlation propagation dataset reflects the impact of upstream parameter changes on the status of downstream equipment. By combining time-aligned correlation parameters, the probability of downstream anomalies is calculated for each unit change in upstream parameters. For example, if the inverter voltage drops by 10V, the water pump speed drops by 70 rpm with a 30% probability of abnormal noise, and the correlation propagation value is set to 0.3; if the voltage increases by 10V, the probability of abnormal water pump noise is only 5%, and the value is 0.05. This distinguishes the differences in the impact of parameter propagation direction, making subsequent matrices more accurate.

[0035] The equipment sensitivity data set and the associated transmission data set are fused to generate a dynamically updated equipment status influence matrix. It should be noted that the fusion process first determines the matrix dimensions. For example, for a water supply system with 4 devices, the matrix is ​​set to 4 rows and 4 columns. The rows and columns represent the source of influence and the affected device, respectively, and the intersection is the influence weight.

[0036] During fusion, the source sensitivity data and the associated transmission data are multiplied together, and then adjusted according to the functional dependency coefficient. For example, for the inverter to the water pump, the sensitivity is 0.85, the association transmission is 0.3, and the water pump is completely dependent on the inverter (coefficient 1), so the weight is 0.85×0.3×1=0.255; for the water pump to the pressure sensor, the sensitivity is 0.7, the association transmission is 0.6, and the sensor is partially dependent (coefficient 0.8), so the weight is 0.7×0.6×0.8=0.336. This method takes into account multiple factors and avoids weight bias caused by a single data point.

[0037] The values ​​in the matrix are dynamically updated according to the system's operating mode, representing the dynamic changes in the interaction relationships between parameters. It should be noted that the system operating modes include full load, half load, standby maintenance, etc. The normal range of parameters and transmission efficiency differ in different modes. Real-time data is collected every 10 minutes to update the matrix, thereby avoiding the analysis lag problem caused by matrix fixation. The innovation of this invention lies in generating a device status influence matrix through this process. First, it provides real-time and accurate interaction relationship basis for subsequent device correlation analysis, avoiding misjudgment of the degree of influence due to data deviation. Second, it can dynamically adapt weights according to the system operation mode, ensuring accurate identification of key related devices under different load scenarios and reducing the risk of missed detection. Third, it can balance monitoring accuracy and system computing power consumption, focusing on core related devices to improve monitoring effectiveness under high load and reasonably reducing monitoring frequency under low load, making the entire alarm threshold adjustment scheme more adaptable to actual operation and maintenance needs and reducing unnecessary intervention and resource waste.

[0038] As a further embodiment, the device status influence matrix is ​​dynamically switched according to the system operating conditions, and the method further includes: Several typical system operating conditions are predefined, and corresponding equipment state influence matrices are trained for each condition. It should be noted that typical system operating conditions refer to common operating states categorized based on industrial production needs, equipment load intensity, and environmental conditions; they do not cover all extreme cases but focus on routine scenarios that account for over 90%. The parameter characteristics of each condition differ significantly. Specifically, during peak morning water supply, the system operates at full load, resulting in frequent and large fluctuations in equipment parameters; during low-load nighttime water supply, the equipment load is only 30% of full load, and parameter changes are gradual; during equipment maintenance and standby, most equipment is in a low-power state, with only a few monitored devices operating. It should be understood that when training a dedicated matrix for each operating condition, at least three months of historical operating data for that condition must be collected, including changes in equipment parameters, correlation transmission patterns, and alarm records. This data is then trained separately through parameter interaction analysis models. For example, for the morning peak water supply condition, data on inverter voltage, pump speed, and pressure sensor data are collected from 6:00 AM to 9:00 AM. The equipment sensitivity values ​​(the inverter's sensitivity to voltage fluctuations rises to 0.92) and correlation transmission values ​​(the correlation transmission value between the inverter and the pump reaches 0.45) are calculated during this period. These values ​​are then fused to generate a dedicated equipment status influence matrix for that operating condition. The influence weight of the inverter on the pump in the matrix is ​​set to 0.92 × 0.45 × 1 = 0.414. In the training data for low-load water supply at night, the inverter sensitivity value dropped to 0.65, the correlation transmission value dropped to 0.2, and the corresponding influence weight was set to 0.65×0.2×1=0.13, which is much lower than that of the morning peak condition. This method of training under different operating conditions can make the matrix parameters more in line with the actual characteristics of different operating conditions and avoid the adaptation deviation of a single matrix under extreme operating conditions.

[0039] Real-time monitoring of system load and start / stop status parameters determines the current operating condition. It should be noted that system load parameters include overall power consumption and the output power of key equipment; start / stop status parameters include the number of devices in operation and the status of maintenance switches. For example, if the system's overall power consumption reaches 90% of its rated power, all six water pumps are running, and no maintenance switch is triggered, combined with the current time being 7:00 AM, it can be determined that the system is currently in peak morning water supply operation. To ensure accurate identification of operating conditions, the system first performs a preliminary matching of operating conditions by using load and start / stop parameters. Then, it compares the real-time parameter fluctuation patterns with the historical characteristics of the operating condition. If the matching degree exceeds 85%, the current operating condition is confirmed. If the matching degree is insufficient, the system will trigger a temporary monitoring mode to increase the parameter collection frequency until the operating condition is clearly identified, thus avoiding misjudgment of operating conditions due to temporary parameter fluctuations.

[0040] The system retrieves the equipment status influence matrix that matches the current operating condition for the correlation and influence analysis in step S2. It should be noted that after the operating condition is confirmed, the system automatically retrieves the specific matrix for that condition from the database, replacing the general dynamically updated matrix. For example, if the system determines that it is a morning peak water supply condition, it retrieves the specific matrix for that condition. In this matrix, the influence weight of the frequency converter on the water pump is 0.414, and the influence weight of the water pump on the pressure sensor is 0.38. Subsequent analysis of equipment correlation is based on these specific weights for that condition. If the operating condition switches to nighttime low-load water supply, the corresponding specific matrix is ​​retrieved, and the weights are adjusted to 0.13 and 0.15.

[0041] This method of matching operating conditions allows the correlation analysis to better reflect the current operating status: for example, during the morning peak hours, when the pump speed is close to the alarm threshold, the system will quickly identify the high impact risk of inverter voltage fluctuations based on the high weight of the dedicated matrix and prioritize monitoring the inverter status; while during nighttime operating conditions, even if the pump speed fluctuates by the same magnitude, the system will determine that the impact range is small based on the low weight, and no excessive intervention is required, only routine monitoring is needed.

[0042] By dynamically switching matrices based on operating conditions, the adaptability and accuracy of the matrix influencing equipment status can be further improved. Compared to relying solely on dynamic updates of a single matrix, condition-specific matrices can better handle parameter characteristics under extreme conditions, avoiding misjudgments of the degree of impact due to differences in operating conditions. Simultaneously, the automated process of real-time condition identification and matrix retrieval eliminates the need for manual intervention, reducing the workload of maintenance personnel and making the entire correlation analysis process more efficient and accurate. This provides a more realistic basis for subsequent alarm threshold adjustments, further reducing the risk of false alarms and missed alarms, and ensuring the stable operation of the industrial control system under different operating conditions.

[0043] As a further embodiment, the multi-objective optimization in step S3 includes: Consistency in threshold adjustments for associated devices is one of the optimization objectives. It's important to note that threshold adjustment consistency means that the adjustment direction and magnitude of thresholds for devices with parameter transfer relationships with the current device must match those of the current device to avoid breakage of the associated logic due to excessive adjustment differences. For example, if the inverter voltage threshold is lowered by 5V from 370V to 365V, but the water pump speed threshold is increased by 50 rpm, the water pump will still frequently alarm even when the inverter voltage decreases. After incorporating consistency into the optimization, the water pump adjustment magnitude must be controlled within ±30% of the inverter adjustment magnitude, i.e., a decrease of 1.5-8.5 rpm, to ensure coordination between the two.

[0044] A reinforcement learning algorithm is used to generate threshold adjustment schemes to simultaneously improve false positive rate, false negative rate, system stability, and threshold adjustment consistency. It should be noted that in the reinforcement learning algorithm, the adjustment schemes are generated synchronously in the system running state environment, and scores are obtained through a reward function.

[0045] For example, Option A: Reduce the inverter voltage by 3V and the water pump speed by 2 RPM. After application, the false alarm rate decreases by 2%, the missed alarm rate decreases by 1%, the stability improves by 1.5%, and the consistency meets the standard, with a total reward of 17.75 points. Option B: Lowering the inverter voltage by 5V and the water pump speed by 1 RPM reduces the false alarm rate by 3%, but the consistency fails to meet the standard. The total reward is 7.5 points, and the algorithm prioritizes Option A. This approach avoids the deterioration of other indicators due to the optimization of a single indicator.

[0046] During the optimization process, it is determined in real time whether the adjustment scheme will cause the system state to exceed the allowable range, and schemes that exceed the range are eliminated. It should be noted that the allowable range of the system state is the parameter boundary for the safe operation of the equipment, such as the minimum voltage of the frequency converter of 350V and the minimum speed of the water pump of 1000 rpm. For example, if scheme C proposes to lower the frequency converter threshold to 355V and the water pump to 1100 rpm, simulation shows that the frequency converter voltage may drop to 348V, exceeding the allowable range, and it is directly eliminated; scheme A only enters the evaluation after simulation because all parameters are within the safe range. If the scheme meets the simulation target but historical data shows a 30% probability of exceeding the range, the adjustment range will also be adjusted. The innovation of this invention lies in its multi-objective optimization design. On the one hand, consistent adjustments to related equipment prevent operational chaos; for example, synchronized downsizing of the frequency converter and water pump leads to more stable parameter transmission. On the other hand, reinforcement learning enables simultaneous improvement of multiple indicators, allowing for real-time judgment and prevention of unsafe solutions. This system-level collaborative optimization further enhances the reliability of the dynamic alarm threshold adjustment scheme.

[0047] As a further embodiment, step S3, which involves performing a step-by-step threshold adjustment on the current device, includes: Based on the parameter change trend, preliminary adjustments are made to parameters that may exceed the threshold. It should be noted that the parameter change trend needs to be determined by real-time data from 5-8 consecutive sampling cycles. For example, if the current voltage threshold of the frequency converter is 370V, and the voltage of 6 consecutive sampling cycles is 372V, 369V, 367V, 365V, 363V, and 361V respectively, showing a continuous downward trend and approaching the threshold, it is predicted that the voltage may fall below the threshold and trigger a false alarm. In this case, the preliminary adjustment should be set according to the trend slope, with the voltage decreasing by an average of 2V per sampling cycle. If it is desired to maintain a safe distance of more than 5V between the threshold and the current voltage, the threshold is initially lowered from 370V to 356V to avoid frequent alarms when the voltage continues to drop. This trend-based preliminary adjustment can respond to parameter changes in advance and avoid the response lag caused by passive adjustment.

[0048] The threshold adjustment is compensated by considering the reverse impact of related devices on the current device. It should be noted that the reverse impact of related devices refers to the degree to which the parameters of the current device are affected by the changes in the state of the related devices. For example, if a water pump is a downstream device of a frequency converter, a sudden increase in the load of the water pump will cause the output current of the frequency converter to increase, indirectly accelerating the voltage drop. In this case, it is necessary to calculate the incremental parameter change caused by the reverse impact. The increase in load causes an additional voltage drop of 1V / cycle. If the threshold is still adjusted to the original initial value of 356V, it may quickly fall below the threshold due to the accelerated voltage drop. Therefore, the adjustment needs to be compensated by lowering the threshold by another 4V to 352V to offset the parameter fluctuation risk caused by the reverse impact of related devices.

[0049] The adjusted system state is evaluated through parameter interaction simulation, and the threshold is recalibrated based on the state deviation. It should be noted that parameter interaction simulation must simulate the parameter interaction process between the current device and related devices after adjustment. For example, after adjusting the inverter threshold to 352V, the simulation simulates the interaction changes of inverter voltage and current with water pump speed and load over the next 10 sampling cycles. The simulation shows that the voltage is stable at 355-358V for the first 5 cycles, with a reasonable distance from the threshold of 352V. However, in the 6th cycle, due to a brief drop in water pump load, the inverter voltage rises to 365V. At this point, the gap between the threshold of 352V and the actual voltage is too large, which may lead to missed alarms in the event of a real anomaly. Based on this state deviation, the threshold needs to be recalibrated, increasing it from 352V to 360V, so that the safe distance remains at around 5V when the voltage rises, avoiding both false alarms and missed alarms. If the system state does not deviate significantly in the simulation, no calibration is needed, and the compensated and modified threshold can be used directly.

[0050] The innovations of this application are as follows: First, it enables advance adjustment through trend prediction, avoiding frequent alarms caused by parameters passively exceeding thresholds and improving the foresight of adjustments; second, it compensates for the reverse influence of related devices, reducing adjustment deviations caused by interactions between devices and ensuring the adaptability of thresholds to changes in related parameters; third, it corrects state deviations through simulation calibration, eliminating the risk of missed or false alarms caused by parameter fluctuations, which helps ensure threshold accuracy and avoids the blindness of one-time adjustments. This ensures that the final threshold can adapt to the current operating state and cope with dynamic changes, providing a reliable guarantee for the stable operation of the system.

[0051] As a further embodiment, the magnitudes of the initial adjustment and secondary calibration are determined by an adaptive algorithm; The adaptive algorithm is generated based on a comprehensive calculation of the deviation of the real-time parameter value from the threshold, the slope of the parameter change trend, and the fluctuation range of the parameter in the same historical period. It should be noted that the deviation of the real-time parameter value from the threshold refers to the difference between the actual value of the device parameter in the current sampling period and the current threshold. If the parameter shows a downward trend, the deviation is negative; if it shows an upward trend, the deviation is positive. For example, if the current real-time voltage value of the frequency converter is 361V and the current threshold is 370V, the deviation is -9V. The slope of the parameter change trend refers to the average rate of change of the parameter value over time in 5-8 consecutive sampling periods. In the calculation, the sampling period is used as the horizontal axis and the parameter value is used as the vertical axis. The slope is obtained through linear fitting. For example, if the voltage of the frequency converter is sampled for 6 consecutive times as 372V, 369V, 367V, 365V, 363V, and 361V, with each sampling period interval of 2 seconds, the slope is (361-372) / (6-1)÷2=-1.1V / second. The negative sign indicates that the voltage is decreasing. The historical range of parameter fluctuations refers to the difference between the highest and lowest values ​​of the parameter within the past three months, under the same conditions and operating conditions as the current period.

[0052] Calculate the product of deviation and slope, and map this product to an adjustment range referenced to historical fluctuations to determine the specific adjustment range for this operation. It's important to note that the calculation process first requires calculating the product of deviation and slope. This product reflects the combined effect of the degree and rate of change of the parameter deviating from the threshold. Larger absolute values ​​of both deviation and slope, along with a larger absolute value of the product, indicate that the parameter is more likely to exceed the threshold quickly, requiring a larger adjustment range; conversely, smaller values ​​require a smaller adjustment range. Furthermore, the upper and lower limits of the adjustment range need to be set based on the historical fluctuation range. Typically, the upper limit is 1 / 3 of the historical fluctuation range, and the lower limit is 1 / 6 of the historical fluctuation range. This avoids excessive adjustment range leading to sudden threshold changes, and also prevents the adjustment range from being too small to cope with parameter changes. For example, if the historical fluctuation range of the frequency converter is 27V, the adjustment range is 9V (27×1 / 3) to 4.5V (27×1 / 6). When mapping, it is necessary to establish a correspondence between the product and the interval. The larger the product, the closer the adjustment range after mapping is to the upper limit of the interval; the smaller the product, the closer it is to the lower limit of the interval. The product of the above frequency converter is 9.9V·s. Under this operating condition, the historical product range is 3V·s to 12V·s. Through linear mapping calculation, the adjustment range corresponding to 9.9V·s is (9.9-3) / (12-3)×(9-4.5)+4.5=8.4V. Therefore, during the initial adjustment, the threshold is lowered from 370V by 8.4V, and determined to be 361.6V (in actual applications, one decimal place is retained according to the precision of the equipment parameters). Furthermore, the adaptive algorithm logic in the secondary calibration stage is consistent with that in the initial adjustment, but the input data needs to be updated to the parameter state before calibration. For example, after the inverter's initial adjustment, the threshold is set to 361.6V. Before the secondary calibration, the continuously sampled voltages are 358V, 356V, 355V, 357V, 359V, and 365V. At this time, the real-time parameter value of 365V deviates from the current threshold of 361.6V by 3.4V. The slope is obtained by fitting as 0.5V / second (voltage recovery). The historical fluctuation range is still 27V. First, the product of the deviation and the slope is calculated: 3.4 × 0.5 = 1.7V·second. Then, it is mapped to the adjustment range of 4.5V-9V. Because the product is small, the adjustment range after mapping is 5.2V. Finally, the threshold is increased by 5.2V from 361.6V to 366.8V, ensuring that the threshold still maintains a reasonable and safe distance from the actual value when the voltage recovers.

[0053] This invention addresses two main issues. First, it avoids adaptation deviations caused by fixed amplitudes. For example, it automatically increases the amplitude when parameters change rapidly to prevent parameters from exceeding the threshold, and automatically decreases the amplitude when parameters fluctuate slowly to avoid frequent threshold adjustments affecting system stability. Second, it combines historical data from the same period to ensure that the amplitude is within a reasonable range, preventing over-adjustment due to a single abnormal fluctuation. This further improves the accuracy of initial adjustment and secondary calibration, making step-by-step threshold adjustment more closely match the dynamic changes in equipment parameters, and providing more accurate threshold support for stable system operation.

[0054] As a further embodiment, step S3, performing threshold pre-adjustment on the associated device, includes: Based on the equipment correlation model, determine the material or energy transfer path from the current device to related devices. It's important to note that the equipment correlation model stores the physical connections and functional dependencies between devices. The transfer path should be selected based on this, focusing on links directly involved in the material or energy transfer of the current device, not all related links. For example, if the current device is a frequency converter that outputs electrical energy, related devices include a water pump that receives the electrical energy, a pressure sensor that monitors the water pump's supply pressure, a water pipe that transmits the water flow, and a water tank that stores the water flow. The correlation model determines the transfer path as frequency converter to water pump to water pipe to water tank. This path includes both energy and material transfer; energy is converted from electrical energy to mechanical energy, and the material is the transfer of water flow. While the pressure sensor is associated with the water pump, it does not directly participate in the material or energy transfer and is not included in this path. When determining the path, verify the directness of the transfer. For example, the frequency converter's electrical energy must be directly delivered to the water pump via cable, and the water pump's mechanical energy must directly drive the impeller to deliver water to the water pipe. Ensure there are no indirect relay devices in the path to avoid errors in subsequent node classification.

[0055] The equipment along the transmission path is categorized into critical and non-critical nodes. It's important to note that critical nodes are those that play a decisive role in the continuity and stability of the transmission path; abnormal parameters will directly interrupt transmission or trigger a cascading failure. Non-critical nodes, on the other hand, have a minor impact on transmission; parameter fluctuations only have a localized effect and can be quickly recovered from. The criteria for judgment should consider both the equipment's function and its role in the path. In the path from the frequency converter to the water pump to the water pipe to the water tank, the water pump is the core of energy conversion and material transfer, responsible for converting electrical energy into mechanical energy to drive water flow. Abnormal water pump parameters will directly cause the water flow transmission to be interrupted, therefore it is classified as a critical node. Although the water pipe transmits water flow, slight parameter fluctuations will not immediately interrupt the overall transmission, and can be compensated for by adjusting the water pump output; therefore, it is classified as a non-critical node. The water tank only stores water flow and has no active impact on the transmission process; similarly, it is classified as a non-critical node. After classification, the core parameters of critical nodes need to be marked, such as the water pump's speed and output power, to provide a basis for subsequent high-priority adjustments.

[0056] High-priority pre-adjustment operations are performed on critical node equipment, while non-critical node equipment is adjusted in stages according to its importance, and the adjustment range changes dynamically based on path attenuation.

[0057] Specifically, high-priority pre-adjustments must prioritize responding to the parameter requirements of critical nodes, with the adjustment timing synchronized with the current equipment adjustment and the adjustment direction consistent with the current equipment. For example, if the current equipment's frequency converter adjusts its threshold due to a voltage drop, the critical node's water pump must simultaneously pre-adjust its speed threshold, and the adjustment direction should also be downward. This prevents the water pump from frequently alarming due to the threshold not being adjusted after the frequency converter voltage drops. The adjustment range needs to be calculated in conjunction with path attenuation. Path attenuation refers to the degree of loss of material or energy during the process of transferring from the current equipment to related equipment. The farther the transmission distance and the more links, the more obvious the attenuation. The adjustment range should decrease as attenuation increases. For example, the energy attenuation from the frequency converter to the water pump is the power transmission loss. The current equipment adjustment range is 8.4V. As the critical node, the water pump is pre-adjusted to 90% of the current equipment adjustment range to deduct the attenuation effect, resulting in 7.56V. As a non-critical node, the water pipe has a 15% attenuation in its water flow transmission to the water pump. The adjustment range is 60% of the water pump's pre-adjustment range. Since the water pipe's effect is relatively low, the result is 4.54V. The water tank has the lowest effect, so the adjustment range is 30% of the water pipe's adjustment range, resulting in 1.36V.

[0058] If there are multiple critical nodes in the transmission path, such as in a path from boiler to steam pipe to heat exchanger to circulating water pump, where both the boiler and heat exchanger are critical nodes, then they need to be pre-adjusted sequentially according to the transmission order. First, adjust the boiler, which is closest to the current equipment, then adjust the downstream heat exchanger to ensure that upstream adjustments do not cause downstream parameter imbalances due to attenuation. For non-critical nodes, hierarchical adjustments should avoid excessive intervention. For example, steam pipes, as non-critical nodes, only require slight adjustments to their pressure thresholds; they do not need to be adjusted at the same magnitude as critical nodes to reduce system computational power consumption.

[0059] This pre-adjustment strategy ensures that the thresholds of associated devices and current devices are coordinated and adapted. High-priority adjustments of critical nodes can avoid transmission link interruptions, while hierarchical adjustments of non-critical nodes can balance adjustment accuracy and system load. The magnitude of path attenuation adaptation prevents remote devices from under-adjusting or over-adjusting due to attenuation. The overall operation makes threshold pre-adjustment no longer a one-size-fits-all synchronous operation, but a precise control that fits the transmission characteristics, reducing the risk of parameter conflicts for subsequent stable system operation and further improving the synergy of the entire threshold adjustment scheme.

[0060] As a further embodiment, the path attenuation is dynamically varied; The method also includes: Real-time monitoring of the actual operating efficiency of key nodes along the transmission path is crucial. It should be noted that the actual operating efficiency of a key node refers to the effectiveness with which the equipment converts input matter or energy into output. This efficiency is calculated by collecting the input and output parameters of the equipment, and the calculation method differs for different types of key nodes. For example, when the key node is a water pump, the input parameter is the electrical energy delivered by the frequency converter, and the output parameters are the actual water supply and pressure of the water pump. The actual operating efficiency is the ratio of output power to input power. The monitoring frequency must be synchronized with the parameter sampling period to ensure real-time capture of efficiency changes. For instance, input and output parameters should be collected every 2 seconds, and the actual operating efficiency calculated every 10 seconds to avoid data lag leading to untimely attenuation correction.

[0061] The efficiency deviation coefficient is obtained by comparing the actual operating efficiency with the rated efficiency. It should be noted that the rated efficiency refers to the standard efficiency value of the equipment under design conditions and normal aging range, which is usually provided by the equipment manufacturer or obtained through long-term normal operation data statistics. The efficiency deviation coefficient is calculated by subtracting the rated efficiency from the actual operating efficiency and then dividing by the rated efficiency. The result is rounded to two decimal places. If the actual efficiency is lower than the rated efficiency, the coefficient is negative; if it is higher than the rated efficiency, the coefficient is positive.

[0062] Furthermore, this coefficient can intuitively reflect the efficiency loss of key nodes. The larger the negative value, the lower the equipment efficiency, the greater the loss in the energy or material transfer process, and the higher the degree of path attenuation.

[0063] The pre-adjustment range is dynamically corrected based on the efficiency deviation coefficient. When the actual operating efficiency is lower than the rated efficiency, the attenuation level is increased, thereby reducing the adjustment range for downstream equipment. It should be noted that the correspondence between the efficiency deviation coefficient and the attenuation level must be established before correction: the larger the absolute value of the negative coefficient, the higher the proportion of increase in attenuation; when the coefficient is positive or close to zero, the attenuation level remains at the original proportion or is slightly reduced. For example, if the original setting for energy attenuation from the frequency converter to the water pump is 5%, and the water pump efficiency deviation coefficient is -0.08 (actual efficiency is lower than rated), the attenuation increase ratio is set to 1.5 times the absolute value of the coefficient, i.e., an increase of 0.08 × 1.5 = 12%. The corrected attenuation is 5% × (1 + 12%) = 5.6%. At this time, the current frequency converter adjustment range is 8.4V. The pre-adjustment range of the water pump as a critical node should originally be 8.4V × 90% = 7.56V (the original 5% attenuation corresponds to a 95% retention, and 90% is taken as a safety redundancy). After correction, it needs to be calculated based on the new attenuation of 5.6%, and the retention ratio is adjusted to 95% - (5.6% - 5%) = 94.4%. The pre-adjustment range becomes 8.4V × 94.4% × (90% ÷ 95%) ≈ 7.32V, which is 0.24V less than the original range. This avoids parameter imbalance caused by the downstream equipment still adjusting according to the original range due to the low efficiency and high energy loss of the water pump.

[0064] If the efficiency deviation coefficient is positive, for example, the actual efficiency of the heat exchanger is slightly higher than the rated efficiency, with a coefficient of 0.02, then the attenuation level can be slightly reduced. The original steam transfer attenuation is 10%, and after correction, the attenuation is 10% × (1 - 0.02 × 1) = 9.8%. The pre-adjustment range of the downstream circulating water pump can be increased from the original 3.2V to 3.24V to ensure sufficient adaptation to the transfer characteristics after the efficiency improvement. If the coefficient is close to zero (e.g., -0.01 or 0.01), it indicates that the equipment efficiency is basically normal, and the attenuation level does not need to be significantly adjusted; only a slight adjustment of 0.5%-1% is needed to avoid excessive correction leading to frequent changes in the amplitude.

[0065] The innovation of this application lies in the fact that, through this dynamic correction method, the path attenuation level can always match the actual operating status of the key nodes, avoiding the adjustment deviation caused by fixed attenuation values. Furthermore, the overall operation further improves the accuracy of pre-adjustment of associated devices, making the entire threshold adjustment scheme more capable of responding to dynamic changes in the operating status of devices and ensuring the long-term stable operation of the system.

[0066] As a further embodiment, the joint optimization in step S4 includes: By analyzing the interactions between parameters, key parameter combinations that trigger alarms can be identified. It's important to note that parameter interactions must be analyzed using real-time data from 30 minutes before and after the alarm to observe the correlation between changes in multiple parameters. For example, when an inverter triggers an overcurrent alarm, the current or voltage parameters alone may appear only slightly abnormal. However, simultaneous monitoring reveals that when the current exceeds 14A and the voltage is below 362V, the probability of the overcurrent alarm triggering increases sharply, forming a "current-voltage" key parameter combination. Similarly, in water pump noise alarms, an alarm is guaranteed to trigger when the speed is below 1180 rpm and the bearing temperature exceeds 65℃; these two parameters constitute another key combination. During identification, cases of single-parameter anomalies must be excluded; only combinations of multiple parameters that change synergistically and are strongly correlated with the alarm should be retained to avoid mistakenly including irrelevant parameters.

[0067] Based on the key parameter combinations, adjust the key factors in the parameter interaction analysis. It should be noted that key factors in the parameter interaction analysis model include parameter weights and interaction coefficients, which need to be optimized for key combinations. For example, for the "current-voltage" combination, the original model had a current weight of 0.6, a voltage weight of 0.4, and an interaction coefficient of 0.3, resulting in the model not fully capturing the synergistic effect between the two. After adjustment, the interaction coefficient between current and voltage is increased to 0.7, and the overall weight of the combination is increased by 15%, ensuring that the model prioritizes identifying the synergistic changes of this combination when the parameters approach the threshold, providing early warning of overcurrent risks. Similarly, for the "speed-temperature" combination, the original model did not set a separate interaction term for the two. After adjustment, an interaction factor is added, allowing automatic association with temperature monitoring when the speed is below the threshold, avoiding alarm delays due to neglecting temperature parameters.

[0068] The correlation relationships corresponding to the key parameter combinations are updated in the equipment correlation model, achieving the coordinated evolution of the parameter interaction analysis model and the equipment correlation model. It should be noted that the correlation update needs to clarify the correlation strength of the equipment to which the parameters in the key combinations belong. For example, the "current-voltage" combination relates to the electrical parameters of the frequency converter itself, increasing the correlation degree of the "current module-voltage module" within the frequency converter from 0.75 to 0.9; the "speed-temperature" combination involves the mechanical and temperature parameters of the water pump, increasing the correlation degree of the water pump's "impeller assembly-temperature sensor" from 0.6 to 0.85. After the update, the adjusted interaction coefficients of the parameter interaction analysis model and the updated correlation degrees of the equipment correlation model correspond: for example, the increased interaction coefficient of the "current-voltage" combination corresponds to an increased correlation degree within the equipment, ensuring that subsequent analyses consider both parameter interactions and the correlations between equipment components, avoiding analytical biases caused by a disconnect between the two models.

[0069] The advantages of this invention are that it allows the model to continuously adapt to the system's operating characteristics, improves the model's sensitivity to coordinated parameter changes through key parameter combination identification and factor adjustment, and achieves coordinated evolution of the two models, avoiding analysis disconnect and ensuring consistency in subsequent correlation judgments and parameter interaction analysis. Finally, it promotes the model to shift from "passive response to alarms" to "proactive prevention of anomalies," further improving the accuracy and foresight of system alarms, reducing the recurrence of similar alarms, and providing more reliable model support for the stable operation of industrial control systems.

[0070] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A dynamic alarm threshold adaptive adjustment method based on reinforcement learning, characterized in that, include: S1. Real-time acquisition of operating parameters of each device in the system, upstream and downstream correlation parameters between devices, and historical alarm data, and feature extraction of the acquired data to obtain operating parameter feature vectors and correlation parameter feature vectors respectively; S2. Based on the feature vector, generate a device status influence matrix through a parameter interaction analysis model; when the real-time operating parameters of any device are close to the current alarm threshold, combine the preset device correlation model and device status influence matrix to analyze the correlation and influence degree between the current device and other devices. S3. Based on the aforementioned correlation and degree of influence, construct a multi-objective optimization function, combine real-time operating parameters and historical alarm data, and perform step-by-step threshold adjustment on the current device and perform threshold pre-adjustment on the associated devices through reinforcement learning algorithm; S4. Apply the adjusted threshold to system monitoring. When the equipment operating parameters exceed the adjusted alarm threshold in multiple consecutive sampling periods, an alarm is triggered. Based on the alarm results, the equipment correlation model and parameter interaction analysis model are jointly optimized.

2. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 1, characterized in that, The feature extraction in step S1 includes: Dynamically filter out key operating parameters that are closely related to alarm events; The upstream and downstream correlation parameters are processed in time and space to extract the characteristics related to the connection and transmission between devices; Based on the processed data, a comprehensive parameter representation is formed that simultaneously includes the device's own operating status and the inter-device relationship status.

3. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 1, characterized in that, The generation of the device state influence matrix in step S2 includes: Through the parameter interaction analysis model, the operating parameter data is transformed into a set of device sensitivity data, and the associated parameter data is transformed into a set of associated transmission data. The device sensitivity data set and the associated transmission data set are fused together to generate a dynamically updatable device status influence matrix; The values ​​in the matrix are dynamically updated according to the system's operating mode, and are used to characterize the dynamic changes in the interaction relationships between parameters.

4. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 3, characterized in that, The device status influence matrix is ​​dynamically switched according to the system operating conditions. The method also includes: Several typical system operating conditions are predefined, and a corresponding equipment state influence matrix is ​​trained for each operating condition. Real-time monitoring of system load and start / stop status parameters to determine the current operating condition; Call the equipment status influence matrix that matches the current operating conditions for the correlation and influence analysis in step S2.

5. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 1, characterized in that, The multi-objective optimization in step S3 includes: Consistency in threshold adjustment of associated devices is taken as one of the optimization objectives; A reinforcement learning algorithm is used to generate a threshold adjustment scheme to simultaneously improve the false positive rate, false negative rate, system stability, and threshold adjustment consistency. During the optimization process, it is determined in real time whether the adjustment scheme will cause the system state to exceed the allowable range, and the scheme that exceeds the range is eliminated.

6. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 1, characterized in that, The step-by-step threshold adjustment for the current device in step S3 includes: Based on the parameter change trend, make preliminary adjustments to parameters that may exceed the threshold; The threshold adjustment amount is compensated by taking into account the reverse influence of related devices on the current device; The adjusted system state is evaluated through parameter interaction simulation, and the threshold is recalibrated based on the deviation of the state.

7. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 6, characterized in that, The magnitudes of the initial adjustment and secondary calibration are determined by an adaptive algorithm; The adaptive algorithm is generated based on a comprehensive calculation of the deviation of the real-time parameter value from the threshold, the slope of the parameter change trend, and the fluctuation range of the parameter in the same period in history. Calculate the product of the deviation and the slope, and map the product to an adjustment range that is referenced to the historical fluctuation range, in order to determine the specific magnitude of this adjustment.

8. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 1, characterized in that, The threshold pre-adjustment of the associated device in step S3 includes: Based on the equipment correlation model, determine the material or energy transfer path from the current equipment to the associated equipment; The devices along the transmission path are divided into critical nodes and non-critical nodes; High-priority pre-adjustment operations are performed on critical node equipment, while non-critical node equipment is adjusted in stages according to its importance, and the adjustment range changes dynamically based on path attenuation.

9. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 8, characterized in that, The path attenuation level is dynamically changing; The method further includes: Real-time monitoring of the actual operating efficiency of key nodes on the transmission path; The actual operating efficiency is compared with the rated efficiency to obtain the efficiency deviation coefficient; The pre-adjustment range is dynamically corrected based on the efficiency deviation coefficient. When the actual operating efficiency is lower than the rated efficiency, the attenuation level is increased, thereby reducing the adjustment range for downstream equipment.

10. The method for adaptive adjustment of dynamic alarm threshold based on reinforcement learning according to claim 1, characterized in that, The joint optimization in step S4 includes: By analyzing the interactions between parameters, the key parameter combinations that trigger alarms can be identified. Based on the aforementioned key parameter combinations, adjust the key factors in the parameter interaction analysis; The correlation relationships corresponding to the key parameter combinations are updated in the device correlation model, thereby realizing the coordinated evolution of the parameter interaction analysis model and the device correlation model.

Images