Intelligent AI data analysis system for power plant equipment condition monitoring

Abstract

This invention relates to the field of data analysis and discloses an intelligent AI data analysis system for power plant equipment status monitoring. This system enables intelligent monitoring of power plant equipment operating status, accurate fault diagnosis, dynamic early warning, and provides a scientific basis for operation and maintenance decisions. The system includes collecting and standardizing multi-sensor operating data, ensuring data quality through outlier identification, constructing a weighted causal graph based on the equipment's physical structure and data correlation, clarifying the dynamic correlation of sensor nodes, calculating anomaly scores by combining entropy values, node centrality, and entropy change trends, and determining early warning levels and anomaly correlations based on the scores and causal graph connections. Monitoring data is continuously stored to form a knowledge base, and information entropy reference baselines and dynamic early warning thresholds are updated periodically. This invention comprehensively considers multiple factors in equipment operation, achieving dynamic and accurate monitoring and early warning, effectively improving the operation and maintenance level and reliability of power plant equipment.

Description

Technical Field

[0001] This invention relates to the field of data analysis, and more particularly to an intelligent AI data analysis system for monitoring the condition of power plant equipment. Background Technology

[0002] In today's rapidly developing power industry, the stable operation of power plant equipment is crucial for ensuring the reliability, security, and economy of power supply. As power plant equipment becomes larger, more complex, and more automated, the connections between internal components are becoming increasingly close, and the amount of data generated during operation is exponentially increasing. Traditional power plant equipment condition monitoring methods mainly rely on regular manual inspections and simple threshold alarm mechanisms. This approach has many limitations and cannot meet the demands of efficient and precise operation and maintenance in modern power plants.

[0003] Manual inspections not only consume a lot of manpower, resources, and time, but also make it difficult to grasp the operating status of equipment in real time and comprehensively. Some potential faults may not be detected in time, which can easily lead to the expansion of equipment failures, thereby affecting the normal production of the entire power plant and even causing safety accidents.

[0004] Simple threshold alarm mechanisms compare data from a single sensor with a preset fixed threshold, ignoring the complex relationships between various components of the equipment and the dynamic changes in data under different operating conditions. This isolated data analysis approach is prone to false alarms and missed alarms, failing to accurately determine the true health status of the equipment and providing effective early warning information, making it difficult for maintenance personnel to take targeted maintenance measures in advance.

[0005] Therefore, we propose an intelligent AI data analysis system for power plant equipment condition monitoring to solve the above problems. Summary of the Invention

[0006] This invention provides an intelligent AI data analysis system for monitoring the status of power plant equipment, which enables intelligent monitoring of the operating status of power plant equipment, accurate fault diagnosis, dynamic early warning, and provides a scientific basis for operation and maintenance decisions.

[0007] The first aspect of the present invention provides an intelligent AI data analysis system for monitoring the status of power plant equipment. The intelligent AI data analysis system for monitoring the status of power plant equipment includes: a data acquisition module for acquiring and processing operational data from multiple sensors in the power plant equipment. The causal module is used to build a weighted causal graph based on the correlation between the physical structure of the equipment and standardized data. The processing module is used to calculate the information entropy value of each sensor based on the distribution characteristics of standardized data; The anomaly module is used to calculate anomaly scores by combining information entropy values, node centrality in the causal graph, and entropy change trends. The early warning module is used to determine the early warning level and the correlation between anomalies based on the anomaly score and the connection relationship in the cause-effect graph.

[0008] Optionally, in a first implementation of the first aspect of the present invention, the method includes: Based on the piping and instrumentation diagrams and functional connections of power plant equipment, determine the physical connections between sensor nodes; Calculate the time-delay cross-correlation sequence of different sensor data pairs within a set time window in standardized data; Extract the peak value and its corresponding time delay value from the time-delay cross-correlation sequence; Calculate the dynamic association weights between sensor nodes based on the peak value and the time delay value; The dynamic association weights are assigned to the connection edges between the corresponding nodes.

[0009] Optionally, in a second implementation of the first aspect of the present invention, the method includes: For each sensor's standardized data sequence, an adaptive binning scheme is determined based on the dispersion of its numerical distribution and the number of samples. The number of data points falling into each data interval is used to obtain the statistical frequency of each data interval; The probability distribution of data points falling into each data interval is calculated based on the statistical frequency. The information entropy value representing the uncertainty of the state is calculated based on the probability distribution.

[0010] Optionally, in a third implementation of the first aspect of the present invention, the method includes: Based on historical operating conditions, a reference baseline for the information entropy value of each sensor is determined, and the deviation of the current information entropy value from the reference baseline is calculated. Calculate the node centrality metric based on the weight of the connecting edges of each node in the weighted causal graph. Calculate the rate of change of entropy value based on the difference between the current entropy value and the historical entropy value in adjacent time windows; Anomaly scores are calculated by weighted summation based on the degree of deviation, the node centrality metric, and the rate of change of entropy.

[0011] Optionally, in a fourth implementation of the first aspect of the present invention, basic weight coefficients are assigned to the degree of deviation, the node centrality measure, and the entropy change rate according to the device type and the importance of the sensor function. Based on the subgraph structure where the sensor nodes are located in the weighted causal graph, adjust the weight coefficients corresponding to the node centrality metric. The adjusted weight coefficients are multiplied by the corresponding deviation degree, node centrality measure and entropy change rate, and then summed to obtain the preliminary anomaly score; The preliminary anomaly score is corrected based on the preliminary anomaly scores of other nodes that have strong connections with the sensor node.

[0012] Optionally, in a fifth implementation of the first aspect of the present invention, the method includes: The current operating conditions of the equipment are matched with a variety of pre-set standard operating conditions to determine the corresponding dynamic early warning thresholds; The anomaly score of each sensor is compared with the dynamic early warning threshold to determine the status level; In the weighted causal graph, path search is performed starting from sensor nodes whose state level exceeds the threshold to identify critical paths for anomaly propagation. Based on the status level and the critical path, a comprehensive early warning information is generated, which includes the warning level and the associated impact range.

[0013] Optionally, in a sixth implementation of the first aspect of the present invention, a monitoring module is also included: Continuously store standardized data sequences, dynamic cause-effect graphs, information entropy values, and anomaly scores generated during the monitoring process to form a historical monitoring knowledge base; Based on the historical monitoring knowledge base, the reference baseline of information entropy value and the dynamic early warning threshold under various operating conditions are updated regularly through statistical analysis and manual verification.

[0014] Optionally, in the seventh implementation of the first aspect of the present invention, abnormal scoring sequences and early warning records under each operating condition within the historical monitoring period are extracted; Analyze the statistical distribution of the abnormal score sequence to determine the typical normal fluctuation range of abnormal scores under different operating conditions, and update the dynamic early warning threshold. Identify abnormal patterns that frequently trigger early warnings in historical monitoring, and associate them with the corresponding set of sensor nodes and the abnormal propagation path; Based on the abnormal pattern, adjust the connection weights between associated sensor nodes in the dynamic causal graph.

[0015] Beneficial effects: Various time-series data are obtained from the power plant's distributed control system and plant-level monitoring information system. Through time alignment and interpolation, a fused data sequence with a unified time reference is formed. Then, through outlier identification, stationarity testing, and standardization using moving time window statistical methods, a solid foundation is laid for subsequent accurate analysis. Based on the physical structure of the equipment and the association of standardized data, not only is the physical connection relationship of sensor nodes determined, but also the time-delay cross-correlation sequence is calculated, the peak value and time delay value are extracted to determine the dynamic association weight and assign connection edges, and a weighted causal graph that can reflect the actual dynamic characteristics of the equipment is constructed. By combining information entropy value, causal graph node centrality, and entropy value change trend, anomaly score is calculated, comprehensively considering multiple factors of equipment status, avoiding the one-sidedness of single-factor evaluation, and achieving accurate quantification of the degree of equipment anomaly. Dynamic early warning thresholds are matched based on the equipment's operating conditions. Starting from nodes where the status level exceeds the threshold, the critical path for anomaly propagation is searched to generate comprehensive early warning information. Monitoring data is continuously stored to form a knowledge base, and the information entropy value reference baseline and dynamic early warning thresholds are updated regularly. Furthermore, the connection weights of the causal graph can be adjusted according to the anomaly pattern, giving the monitoring method adaptive and self-optimizing capabilities. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an embodiment of the intelligent AI data analysis system for monitoring the status of power plant equipment in this invention. Figure 2 This is a schematic diagram of another embodiment of the intelligent AI data analysis system for monitoring the status of power plant equipment, as described in this invention. Detailed Implementation

[0017] This invention provides an intelligent AI data analysis system for monitoring the status of power plant equipment, enabling intelligent monitoring of the operating status of power plant equipment, accurate fault diagnosis, dynamic early warning, and providing a scientific basis for operation and maintenance decisions. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent AI data analysis system for monitoring the status of power plant equipment in this invention includes: 101. Acquisition module, used to collect operating data from multiple sensors in power plant equipment and perform standardized processing.

[0019] It is understood that the executing entity of this invention can be an intelligent AI data analysis device for monitoring the status of power plant equipment, or it can be a terminal or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0020] Specifically, various types of sensor time-series data are acquired from the power plant's distributed control system and plant-level monitoring information system; sensor data with different sampling frequencies are time-aligned and interpolated to form a fused data sequence with a unified time reference; outlier identification and stationarity testing are performed on the fused data sequence; and a statistical method based on a moving time window is used to standardize the tested data sequence.

[0021] It should be noted that the monitored object is assumed to be the main steam system of the boiler. Real-time time-series data from three key sensors are collected from the power plant's distributed control system (DCS): main steam temperature (sensor T, sampling frequency 1Hz), main steam pressure (sensor P, sampling frequency 1Hz), and feedwater flow rate (sensor F, sampling frequency 1 / 60Hz, i.e., one point per minute). Simultaneously, the calculated steam enthalpy value is obtained from the plant-level monitoring information system (SIS) (sensor H, sampling frequency 1 / 300Hz, i.e., one point every 5 minutes).

[0022] Because the sensor sampling frequencies differ, they need to be unified to the same time base. The target frequency is set to 1Hz (one point per second). Temperature and pressure data at 1Hz are directly retained. For flow rate data occurring once per minute, a linear interpolation method is used. Based on the values ​​of two consecutive minute points (e.g., 1000 t / h in the first minute, 1020 t / h in the second minute), the interpolation value for each second is calculated (e.g., 1000.33 t / h for the 61st second). For enthalpy data occurring every 5 minutes, linear interpolation is also used to generate a sequence of one point per second. Finally, a fused data table with four columns (T, P, F, H) and strictly aligned timestamps is formed.

[0023] The aligned data is initially screened using the "3σ principle". For example, the mean of the steam temperature series over the past 10 minutes is calculated to be 535℃, with a standard deviation of 2.5℃. If a data point is found to be 600℃, far exceeding the normal fluctuation range (535±3*2.5), it is marked as a suspected outlier.

[0024] The ADF test was used to examine data from a continuous, stable operating condition (e.g., running at full load for half an hour). If the test result showed a p-value greater than 0.05, the series was considered non-stationary and may contain trend or periodic components, which should be considered in subsequent analyses.

[0025] Z-score standardization based on a moving-time window is employed. Using the current moment as a baseline, data from the previous 10 minutes (600 data points) are taken as a window. The mean (μ) and standard deviation (σ) of each sensor data point within the window are calculated. For each original data value x within the window, the standardized value z is calculated as z = (x - μ) / σ. For example, if the original steam pressure at the current moment is 16.2 MPa, and the mean of the previous 10 minutes is 16.0 MPa with a standard deviation of 0.3 MPa, then the standardized value is (16.2 - 16.0) / 0.3 ≈ 0.67. This method eliminates dimensional differences between different sensors and transforms the data into a distribution with a mean of 0 and a standard deviation of 1, while also adapting to slow changes in operating conditions.

[0026] Through the above steps, the original multi-frequency, multi-dimensional sensor data is transformed into a clean data sequence that is time-synchronized, preliminarily cleaned, and conforms to a standard normal distribution.

[0027] 102. Causal Module: Used to build a weighted causal graph based on the correlation between the physical structure of the equipment and standardized data.

[0028] Specifically, based on the piping and instrumentation diagrams and functional connections of power plant equipment, the physical connections between sensor nodes are determined; the time-delay cross-correlation sequences of different sensor data pairs in the standardized data are calculated within a set time window; the peak values ​​and their corresponding time delay values ​​in the time-delay cross-correlation sequences are extracted; the dynamic association weights between sensor nodes are calculated based on the peak values ​​and time delay values; and the dynamic association weights are assigned to the connection edges between the corresponding nodes.

[0029] It should be noted that the physical connection topology of the four sensor nodes is determined based on the piping and instrumentation diagram (P&ID) and process principles of the power plant boiler system. Feedwater flow rate (F) is the input, indirectly affecting the main steam temperature (T) and pressure (P) by influencing the drum water level and the evaporation process; steam temperature and pressure together determine the steam enthalpy (H). Therefore, the physical connection is: F→{T,P}; simultaneously, T and P together determine H, i.e., {T,P}→H.

[0030] Data from two hours of continuous and stable operation, standardized in step 101, was collected. To investigate the dynamic impact of feedwater flow rate F on main steam temperature T, a sliding time window of 10 minutes (600 points) was set. The cross-correlation coefficient series of the F and T sequences under different time delays (-600 seconds to 600 seconds) was calculated. Based on the process mechanism and actual data analysis, it was found that the cross-correlation coefficient reached a peak of 0.75 when the F sequence led the T sequence by approximately 180 seconds (3 minutes). This aligns with the common knowledge that boilers have high thermal inertia: changes in feedwater require a relatively long heating and evaporation process to significantly alter the steam temperature.

[0031] Repeat the above calculations for all preset sensor pairs (e.g., (F,T), (F,P), (T,H), (P,H)). For example, the following key results are obtained: when F leads P by 120 seconds, the peak correlation coefficient is 0.82; the peak correlation coefficient between T and H (0-second delay) is 0.90; and the peak correlation coefficient between P and H (0-second delay) is 0.85.

[0032] The association weight is determined by both the peak value of the correlation coefficient and the latency. The basic rule is: the higher the peak value and the shorter the latency, the greater the weight. A simplified weight calculation logic is defined as follows: Assuming a time constant of 100 seconds, the dynamic weights for the node pair (F, T) are calculated as follows: Similarly, the weights of (F, P) are... The weight of (T,H) is approximately 0.90 (with a delay of 0), and the weight of (P,H) is approximately 0.85.

[0033] Using sensors as nodes, a directed weighted graph is constructed based on physical connections and calculated dynamic weights. For example: There is a directed edge from node F to node T with a weight of 0.124.

[0034] There is a directed edge from node F to node P with a weight of 0.247.

[0035] There is a directed edge from node T to node H with a weight of 0.90.

[0036] There is a directed edge from node P to node H with a weight of 0.85.

[0037] This diagram not only illustrates the physical structure, but also quantifies the strength and latency of dynamic data-level associations through weights (especially the smaller F->T weights).

[0038] 103. Processing module, used to calculate the information entropy value of each sensor based on the distribution characteristics of standardized data.

[0039] Specifically, for each sensor's standardized data sequence, an adaptive binning scheme is determined based on the dispersion of its numerical distribution and the number of samples; the number of data points falling into each data interval is counted to obtain the statistical frequency of each data interval; the probability distribution of data points falling into each data interval is calculated based on the statistical frequency; and the information entropy value representing the uncertainty of the state is calculated based on the probability distribution.

[0040] Furthermore, the continuity of data points in the standardized data sequence is analyzed, and abnormal data discontinuities are identified and marked. Based on the data sequence after removing abnormal data discontinuities, a statistic characterizing its overall dispersion is calculated. According to the statistic and the number of valid data points, the number of bins and the boundary values ​​of each bin are determined. Based on the determined bin configuration, the complete standardized data sequence is binned and the frequency is counted.

[0041] It should be noted that we examine the standardized data sequence of sensor T. We assume its values ​​are primarily distributed between -2.5 and 2.5 (i.e., Z-score values). We use the "effective data point statistics method" to determine the binning: the standard deviation of the sequence is calculated to be 1.02, and the number of effective data points (after removing the very few outliers marked in the previous step) is 3598. Based on common rules of thumb, we set the number of bins to 7. We use equal-width binning, with a bin width of (2.5 - (-2.5)) / 7 ≈ 0.714. Therefore, the boundaries of the 7 data intervals are determined as: [-2.5, -1.786), [-1.786, -1.072), [-1.072, -0.358), [-0.358, 0.356), [0.356, 1.070), [1.070, 1.784), [1.784, 2.5]. To cover all data points, in actual statistics, points less than -2.5 are assigned to the first interval, and points greater than 2.5 are assigned to the last interval.

[0042] The 3598 valid data points were categorized into the seven intervals mentioned above. The statistical results might be as follows: Interval 1: 185 points; Interval 2: 800 points; Interval 3: 1200 points; Interval 4: 900 points; Interval 5: 450 points; Interval 6: 60 points; and Interval 7: 3 points. Then, the probability of each interval was obtained by dividing the number of data points in each interval by the total number of points, 3598. For example, the probability of the most concentrated interval, the third interval, is approximately 1200 / 3598 ≈ 0.333, while the probability of the seventh interval at both ends is extremely low, approximately 0.001.

[0043] According to the principles of information theory, information entropy is calculated from the probability of each interval, and it measures the degree of disorder in the data distribution. The more uniform the probability distribution (i.e., the more dispersed the data), the higher the entropy value; the more concentrated the distribution (the more stable the data), the lower the entropy value. Based on the above probability distribution, the information entropy value of sensor T in the current hour is approximately 1.82 (the unit is usually "bits").

[0044] Repeating the above process: The feedwater flow rate (F) data fluctuates significantly due to adjustments, resulting in a more dispersed distribution and a higher entropy value of approximately 2.45 bits. The main steam pressure (P) data is very stable, with its distribution highly concentrated around the mean, and its entropy value is lower, approximately 1.35 bits. The steam enthalpy (H), calculated from T and P, is also typically stable, with an entropy value of approximately 1.20 bits.

[0045] The entropy values ​​of the four sensors were obtained within the current time window. These entropy values ​​serve as a quantitative indicator of the "uncertainty" of their operating state; an abnormal increase in entropy values ​​may indicate that the data at that measurement point is beginning to become disordered.

[0046] 104. Anomaly Module: This module is used to calculate anomaly scores by combining information entropy values, node centrality in the causal graph, and entropy change trends.

[0047] Specifically, a reference baseline for the information entropy value of each sensor is determined based on historical operating conditions, and the deviation of the current information entropy value from the reference baseline is calculated; a node centrality metric is calculated based on the weight of the connecting edges of each node in the weighted causal graph; the entropy change rate is calculated based on the difference between the current information entropy value and the historical information entropy values ​​in adjacent time windows; and an anomaly score is calculated by weighted summation based on the deviation, node centrality metric, and entropy change rate.

[0048] Furthermore, based on the device type and the importance of sensor functions, basic weight coefficients are assigned to the degree of deviation, node centrality measure, and entropy change rate, respectively. According to the subgraph structure where the sensor node is located in the weighted causal graph, the weight coefficients corresponding to the node centrality measure are adjusted. The adjusted weight coefficients are multiplied by the corresponding degree of deviation, node centrality measure, and entropy change rate, and then summed to obtain a preliminary anomaly score. Based on the preliminary anomaly scores of other nodes with strong connections to the sensor node, the preliminary anomaly score is corrected.

[0049] It should be noted that, based on the historical knowledge base, the information entropy reference baseline of sensor T under the current "full load stable operation" condition is determined to be 1.50 bits. From step 103, we know that the current entropy value of T (E) curr The value is 1.82 bits. Its deviation (D) is calculated as: (current value - baseline value) / baseline value. Therefore, D = (1.82 - 1.50) / 1.50 ≈ 0.213, meaning a deviation of 21.3%.

[0050] Based on the weighted causal graph constructed in step 102 (using the current weights), sensor T has two edges: one as the target, receiving the influence from the feedwater flow rate F (weight 0.124); the other as the source, influencing the steam enthalpy H (weight 0.90). Node centrality (C) can be simplified to the sum of the weights of all its connected edges. Therefore, C = 0.124 + 0.90 = 1.024. For normalization, we assume the maximum node centrality in the graph is 2.0, then the normalized centrality metric is approximately 1.024 / 2.0 = 0.512.

[0051] The entropy value of 1.82 bits for the current time window (the most recent hour) is compared with the entropy value of 1.60 bits for the previous time window (one hour ago). The rate of change (R) is calculated as: (current value - previous value) / previous value. Therefore, R = (1.82 - 1.60) / 1.60 = 0.1375, meaning the entropy value has increased by 13.75% recently.

[0052] Assign basic weights to the three indicators. Given that temperature is a key parameter, its deviation from the baseline has the highest weight; node centrality reflects its ability to propagate influence; and the rate of change reflects the speed of deterioration. Set: Deviation weight (W) d ) = 0.5, node centrality weight (W) c =0.3, rate of change weight (W) r The initial anomaly score is 0.2. initial )for: S initial =D*W d +C*W c +R*W r =0.213*0.5+0.512*0.3+0.1375*0.2 =0.1065+0.1536+0.0275 =0.2876 In the causal graph, the node with a strong connection (weight greater than 0.5) to sensor T is H (affected by T, weight 0.90). The initial anomaly score of node H is 0.28. The correction rule is to combine the initial score of this node with the weighted average of the scores of its strongly connected neighbor nodes. Assuming the correction factors are 0.3 (itself) and 0.7 (strongly connected neighbors), the final anomaly score of T (S) is... final The revised value is: 0.3*0.2876 + 0.7*0.28 ≈ 0.282. This score takes into account the user's own status, network location, and associated influences.

[0053] Through the above steps, we obtain a quantitative comprehensive anomaly score for sensor T, which simultaneously reflects its own state deviation, its importance in the system, its deterioration trend, and its associated impacts.

[0054] 105. Early warning module, used to determine the early warning level and abnormal correlation based on the anomaly score and the connection relationship in the cause-effect graph.

[0055] Specifically, the current operating conditions of the equipment are matched with a variety of pre-set standard operating conditions to determine the corresponding dynamic early warning thresholds; the anomaly score of each sensor is compared with the dynamic early warning thresholds to determine the state level; in a weighted causal graph, path search is performed starting from sensor nodes whose state levels exceed the thresholds to identify the critical paths for anomaly propagation; based on the state level and the critical paths, comprehensive early warning information including the warning level and the associated impact range is generated.

[0056] It should be noted that this step aims to transform the quantified anomaly score into specific and actionable early warning information. Assuming the system has been matched to the current "full load stable operation" condition, its dynamic early warning threshold is set as follows: anomaly score ≥0.4 triggers the "warning" level, ≥0.3 but <0.4 triggers the "attention" level, and <0.3 is "normal".

[0057] Based on the final anomaly score calculated in step 104, and compared with the aforementioned dynamic threshold, the results are shown in Table 1 below: Table 1 In the weighted causal graph established in step 102, the influence propagation path is searched starting from node F, which has a state level of "warning". From node F, the connecting edges are F→T (weight 0.124) and F→P (weight 0.247). Both T and P currently have normal scores, but T (score 0.282) is close to the "attention" threshold and is a direct downstream of F. From node T, the strong connecting edge is T→H (weight 0.90). H's score (0.28) is also close to the "attention" threshold and is a direct downstream of T. In summary, a key potential influence path is identified: F (warning) → T (normal, but requires observation) → H (normal, but requires observation). This indicates that the anomaly may have originated from the water supply system, its impact has initially manifested in temperature, and there is a clear path for propagation to enthalpy.

[0058] Based on the above analysis, a structured early warning message is generated. Warning Level: Level 2 (Local anomaly, requires attention). This is because a critical parameter triggered the "warning," and a clear impact path points to two other important parameters at critical points. Core anomaly: Feedwater flow rate (F). Related impact scope: The anomaly may originate from the feedwater system. The scores for main steam temperature (T) and steam enthalpy (H) are approaching the "attention" threshold and are on the affected critical path, requiring close monitoring of their trends. Main steam pressure (P) is currently normal and relatively stable. Recommended measures: Prioritize checking the feedwater automatic control loop, feedwater pump status, and flow measurement unit; closely monitor the trends of steam temperature and enthalpy changes to confirm whether combustion conditions are compatible. Through this step, the system not only identifies the specific anomaly but also infers "potential impact directions and risk points" through a cause-and-effect diagram, providing operators with relevant and predictive decision support.

[0059] Please see Figure 2 Another embodiment of the intelligent AI data analysis system for monitoring the status of power plant equipment in this invention further includes: 106. The monitoring module is used to continuously store standardized data sequences, dynamic cause-effect graphs, information entropy values ​​and anomaly scores generated during the monitoring process, forming a historical monitoring knowledge base. Based on the historical monitoring knowledge base, the reference baseline of information entropy values ​​and dynamic early warning thresholds under various operating conditions are updated regularly through statistical analysis and manual verification.

[0060] Furthermore, abnormal score sequences and early warning records under various operating conditions within the historical monitoring period are extracted; the statistical distribution of abnormal score sequences is analyzed to determine the typical normal fluctuation range of abnormal scores under different operating conditions, and the dynamic early warning threshold is updated; abnormal patterns that frequently trigger early warnings in historical monitoring are identified, and their corresponding sensor node sets and abnormal propagation paths are associated; based on the abnormal patterns, the connection weights between related sensor nodes in the dynamic cause-effect graph are adjusted.

[0061] It should be noted that the system has been running for a quarter based on the aforementioned example (using the current model parameters), accumulating a large amount of historical data records, and has begun its first periodic knowledge base update and model optimization.

[0062] During continuous monitoring, the results of each cycle (hourly), including the standardized data sequences from the four sensors, the dynamic causal graph (current F→T weight 0.124), the information entropy value (e.g., T baseline 1.50 bits), and the anomaly score, are all stored in the historical monitoring knowledge base. For example, the knowledge base already contains records of over 2000 hours under the "full load stable operation" condition, forming the statistical basis for analysis.

[0063] Analysis of all anomaly score records under the "full load stable operation" condition in the previous quarter revealed that sensor T's anomaly score was below 0.35 in 99% of cases. Furthermore, manually verified scores between 0.30 and 0.35 were considered normal fluctuations and did not correspond to actual faults. The original "attention" level threshold was 0.3. To reduce redundant alerts and improve warning specificity, the system increased the "attention" level threshold for T under this condition from 0.3 to 0.35. The main parameter updates are compared in Table 2 below. Table 2 Analysis of historical data revealed that in multiple cases of "feedwater fluctuations," the correlation between the anomaly of sensor F and the subsequent response of sensor T was statistically weaker than that reflected by the initial model (weight 0.124). Experts judged that this might be because the boiler control system has a strong ability to compensate for and adjust to feedwater disturbances. Therefore, combining statistical analysis and manual verification, the weight of the "F→T" edge in the causal graph was reduced from 0.124 to 0.100 to better reflect the dynamic response relationship of the actual system.

[0064] The system calculates the moving average of the information entropy of each sensor under all "full load stable operation" conditions in the previous quarter. It was found that the baseline entropy value of the feedwater flow rate (F) slowly increased from the previously set 2.45 bits to 2.50 bits. This was confirmed to be due to slight changes in the characteristics of the feedwater regulating valve caused by seasonal variations, and is considered a normal, slow drift. The system updates this new statistical mean (2.50 bits) as the new reference baseline for the F entropy value for future calculations of deviation.

[0065] Through the above process, the system completes knowledge iteration based on historical data. The "attention" threshold becomes more accurate, reducing interference alarms caused by normal fluctuations; the cause-effect graph better reflects the actual coupling strength of the current equipment; and the entropy baseline also slowly and adaptively updates along with the equipment status. This enables the entire monitoring model to adapt to the slow changes in equipment and the evolution of its operating characteristics, continuously maintaining the accuracy and reliability of early warnings.

[0066] The present invention also provides an intelligent AI data analysis device for monitoring the status of power plant equipment. The intelligent AI data analysis device for monitoring the status of power plant equipment includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the intelligent AI data analysis system for monitoring the status of power plant equipment in the above embodiments.

[0067] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the intelligent AI data analysis system for monitoring the status of power plant equipment.

[0068] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0069] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent AI data analysis system for monitoring the condition of power plant equipment, characterized in that, include: The data acquisition module is used to collect and process operating data from multiple sensors in power plant equipment. The causal module is used to build a weighted causal graph based on the correlation between the physical structure of the equipment and standardized data. The processing module is used to calculate the information entropy value of each sensor based on the distribution characteristics of standardized data; The anomaly module is used to calculate anomaly scores by combining information entropy values, node centrality in the causal graph, and entropy change trends. The early warning module is used to determine the early warning level and the correlation between anomalies based on the anomaly score and the connection relationship in the cause-effect graph.

2. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 1, characterized in that, include: Based on the piping and instrumentation diagrams and functional connections of power plant equipment, determine the physical connections between sensor nodes; Calculate the time-delay cross-correlation sequence of different sensor data pairs within a set time window in standardized data; Extract the peak value and its corresponding time delay value from the time-delay cross-correlation sequence; Calculate the dynamic association weights between sensor nodes based on the peak value and the time delay value; The dynamic association weights are assigned to the connection edges between the corresponding nodes.

3. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 1, characterized in that, include: For each sensor's standardized data sequence, an adaptive binning scheme is determined based on the dispersion of its numerical distribution and the number of samples. The number of data points falling into each data interval is used to obtain the statistical frequency of each data interval; The probability distribution of data points falling into each data interval is calculated based on the statistical frequency. The information entropy value representing the uncertainty of the state is calculated based on the probability distribution.

4. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 1, characterized in that, include: Based on historical operating conditions, a reference baseline for the information entropy value of each sensor is determined, and the deviation of the current information entropy value from the reference baseline is calculated. Calculate the node centrality metric based on the weight of the connecting edges of each node in the weighted causal graph. Calculate the rate of change of entropy value based on the difference between the current entropy value and the historical entropy value in adjacent time windows; Anomaly scores are calculated by weighted summation based on the degree of deviation, the node centrality metric, and the rate of change of entropy.

5. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 4, characterized in that, Based on the device type and the importance of sensor functions, basic weight coefficients are assigned to the deviation degree, node centrality measure, and entropy change rate, respectively. Based on the subgraph structure where the sensor nodes are located in the weighted causal graph, adjust the weight coefficients corresponding to the node centrality metric. The adjusted weight coefficients are multiplied by the corresponding deviation degree, node centrality measure and entropy change rate, and then summed to obtain the preliminary anomaly score; The preliminary anomaly score is corrected based on the preliminary anomaly scores of other nodes that have strong connections with the sensor node.

6. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 1, characterized in that, include: The current operating conditions of the equipment are matched with a variety of pre-set standard operating conditions to determine the corresponding dynamic early warning thresholds; The anomaly score of each sensor is compared with the dynamic early warning threshold to determine the status level; In the weighted causal graph, path search is performed starting from sensor nodes whose state level exceeds the threshold to identify critical paths for anomaly propagation. Based on the status level and the critical path, a comprehensive early warning information is generated, which includes the warning level and the associated impact range.

7. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 1, characterized in that, It also includes a monitoring module: Continuously store standardized data sequences, dynamic cause-effect graphs, information entropy values, and anomaly scores generated during the monitoring process to form a historical monitoring knowledge base; Based on the historical monitoring knowledge base, the reference baseline of information entropy value and the dynamic early warning threshold under various operating conditions are updated regularly through statistical analysis and manual verification.

8. The intelligent AI data analysis system for power plant equipment status monitoring as described in claim 7, characterized in that, Extract abnormal scoring sequences and early warning records under each operating condition within the historical monitoring period; Analyze the statistical distribution of the abnormal score sequence to determine the typical normal fluctuation range of abnormal scores under different operating conditions, and update the dynamic early warning threshold. Identify abnormal patterns that frequently trigger early warnings in historical monitoring, and associate them with the corresponding set of sensor nodes and the abnormal propagation path; Based on the abnormal pattern, adjust the connection weights between associated sensor nodes in the dynamic causal graph.

Images