Gas turbine real-time operation risk early warning method and system based on sequence matching

Abstract

The invention belongs to the technical field of industrial process safety and intelligent monitoring, and particularly relates to a gas turbine real-time operation risk early warning method and system based on sequence matching, and the method comprises the steps: constructing a standard operation model and an abnormal operation model; selecting a standard operation model or an abnormal operation model according to the current operation background, and calculating the deviation degree between the current operation sequence and the standard operation model and the likelihood between the current operation sequence and the abnormal operation model; and when the likelihood / deviation degree exceeds a corresponding threshold value, activating a corresponding risk evolution model, performing dynamic prediction on the type, probability and time window of a fault possibly occurring in the future in combination with the current working condition parameter, and generating risk early warning. The method is mainly used for identifying and early warning potential risks caused by improper operation in advance, and the safety defense line is moved forward.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial process safety and intelligent monitoring, and particularly relates to a gas turbine real-time operation risk early warning method and system based on sequence matching. BACKGROUND

[0002] As the core equipment of modern power plants, the operation process of gas turbines is highly complex, and the accuracy and timeliness of operation are extremely strict. Industry data shows that a large number of equipment abnormalities, performance degradation and serious failures are directly or indirectly related to non-standard, delayed or incorrect manual operation sequences.

[0003] Although there are standard operating procedures (SOPs), operators may deviate from the procedures and perform high-risk operation combinations due to insufficient experience, judgment errors or environmental pressure (such as start-stop stages and rapid load adjustment) under emergency or complex working conditions. The current safety monitoring of gas turbine power plants relies on equipment state parameter threshold alarms (such as temperature / pressure overrun triggering alarms), which belong to post-response or in-process response. This approach only intervenes when the equipment state is already significantly abnormal, and cannot take advantage of the valuable time window between the start of the risky operation and the parameter abnormality.

[0004] Therefore, the prior art lacks effective monitoring and evaluation means for the potential risks of operation behavior, and cannot identify dangerous operation sequences before the equipment parameter alarm. When a risky operation sequence occurs or in the early stage when the equipment parameter has not yet appeared abnormal, it is impossible to intelligently identify risky operations and perform risk early warning, forming “pre-event” prevention or “in-process” early intervention. SUMMARY

[0005] The purpose of the present application is to provide a gas turbine real-time operation risk early warning method and system based on sequence matching, to overcome the problem that the prior art cannot intelligently identify risky operations and perform risk early warning, forming “pre-event” prevention or “in-process” early intervention.

[0006] To achieve the above-mentioned purpose, the technical solutions adopted by the present application are as follows: The present application discloses a gas turbine real-time operation risk early warning method based on sequence matching, comprising the following steps: Constructing a standard operation model corresponding to each operating background according to a plurality of preset operating backgrounds and normal operation sequences corresponding to each operating background; Determining a plurality of abnormal operation sequences according to alarm data and historical working condition change information, and constructing a corresponding abnormal operation model according to each abnormal operation sequence; Constructing an associated risk evolution model for each abnormal operation model according to each abnormal operation sequence, historical working condition change information and alarm data; According to a preset rule, at least one standard operation model and at least one abnormal operation model are selected as matching models respectively; A current operation sequence is obtained and input into each matching model respectively, and a deviation output by each standard operation model and a likelihood output by each abnormal operation model in the matching model are obtained; Each likelihood is compared with a first preset threshold respectively to obtain a plurality of first comparison results, and each deviation is compared with a second preset threshold respectively to obtain a plurality of second comparison results, a risk operation sequence is identified according to the plurality of first comparison results and the plurality of second comparison results, and a risk evolution model associated with a preset abnormal operation model is triggered to perform real-time operation risk warning.

[0007] Preferably, after the corresponding abnormal operation model is constructed according to each abnormal operation sequence, the method further comprises: At least one absolute abnormal model is determined from the plurality of abnormal operation models according to a preset prohibited operation standard, and at least one relative abnormal model is determined from the plurality of abnormal operation models for each preset running background; According to a preset rule, at least one standard operation model and at least one abnormal operation model are selected as matching models respectively, specifically: A current running background is determined, and at least one standard operation model and at least one relative abnormal model are selected according to the current running background; All absolute abnormal models, selected standard operation models and selected relative abnormal models are taken as matching models.

[0008] Preferably, the risk operation sequence is identified according to the plurality of first comparison results and the plurality of second comparison results, specifically: When any one of the likelihoods is greater than the first preset threshold or any one of the deviations is greater than the second preset threshold, the current operation sequence is identified as a risk operation sequence.

[0009] Preferably, the alarm data includes a fault occurrence time and a fault type, Accordingly, the risk evolution model associated with the preset abnormal operation model is triggered to perform real-time operation risk warning, specifically: When the first likelihood greater than the first preset threshold occurs, the risk evolution model associated with the abnormal operation model corresponding to the likelihood is triggered, which is recorded as a target evolution model; The current working condition parameters corresponding to the risk operation sequence are obtained, the risk operation sequence and the current working condition parameters are input into the target evolution model, and a risk prediction result is output by the target evolution model, the risk prediction result including a fault prediction time, a fault prediction type and a probability distribution corresponding to the fault prediction type caused by the risk operation sequence.

[0010] Preferably, after outputting the risk prediction result using the target evolution model, the method further includes: Obtain the actual results, which include the actual failure time and actual failure type caused by the risky operation sequence; The target evolution model is updated and corrected based on the error between the actual results and the risk prediction results.

[0011] Preferably, the target evolution model is updated and corrected based on the error between the actual result and the risk prediction result, specifically as follows: Under multiple real-time operational risk warnings, the error between the actual result and the risk prediction result for each instance is obtained; The parameters of the first preset threshold or the target evolution model are updated based on the error.

[0012] Preferably, the current running background is determined as follows: Obtain current operating parameters; The current operating background is determined based on the current operating condition parameters.

[0013] Preferably, the operating background includes at least one of the following: gas turbine start-up and shutdown phases, load adjustment rate, equipment maintenance status, and equipment warnings indicating activity.

[0014] Preferably, the baseline model used to construct the standard operation model and the abnormal operation model is a hidden Markov model or a long short-term memory network.

[0015] This invention also discloses a real-time operation risk early warning system for gas turbines based on sequence matching, comprising: The standard module is used to construct a standard operation model for each operating scenario based on multiple preset operating scenarios and the normal operation sequence corresponding to each operating scenario. The anomaly module is used to determine various abnormal operation sequences based on alarm data and historical operating condition change information, and to build a corresponding abnormal operation model based on each abnormal operation sequence. The correlation module is used to construct a correlation risk evolution model for each abnormal operation model based on each abnormal operation sequence, historical operating condition change information, and alarm data; The matching module is used to select at least one standard operation model and at least one abnormal operation model as matching models according to preset rules. The deviation likelihood module is used to obtain the current operation sequence and input the current operation sequence into each matching model to obtain the deviation degree output by each standard operation model and the likelihood degree output by each abnormal operation model in the matching model. The early warning module is used to compare each likelihood with a first preset threshold to obtain multiple first comparison results, and to compare each deviation with a second preset threshold to obtain multiple second comparison results. Based on the multiple first comparison results and the multiple second comparison results, the module identifies risk operation sequences and triggers a risk evolution model associated with a preset abnormal operation model to provide real-time operation risk warnings.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: (1) This invention establishes multiple standard operation models, multiple abnormal operation models, and risk evolution models associated with each abnormal operation model. It matches the current operation sequence with the standard operation model and the abnormal operation model respectively, identifies risky operations based on the matching results, and then activates the corresponding risk evolution model for risk warning. This can identify potential risks before obvious abnormalities occur in equipment parameters, and shifts the safety defense line from "status monitoring" after the fact or during the incident to "behavioral monitoring" before the incident. This realizes the transformation from passive response to active warning, which is conducive to more timely risk prevention and control. (2) By dynamically predicting the type, probability distribution and time of future failures, this invention provides operators with valuable predictive information and intervention time, effectively avoiding or mitigating the occurrence of potential failures and improving the timeliness and predictability of early warnings. (3) The present invention can perform dynamic analysis in combination with the current operating background of the gas turbine, accurately identify high-risk operation sequences in specific contexts, make up for the deficiencies of static operating procedures, and enhance the ability to identify operational risks under complex working conditions. (4) By comparing the current operation with the standard model and providing corrective suggestions, this invention helps guide operators to perform standard operations, reduce human error, improve the overall operation and maintenance level, and promote the standardization and normalization of operations. (5) Through closed-loop feedback and online learning of the early warning effect, the predictive ability of the risk evolution model can be continuously optimized, adapting to changes in equipment status and operating environment, and possessing self-learning and continuous evolution capabilities. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of one embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention 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 of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus 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 apparatus.

[0020] This invention discloses a real-time operational risk early warning method for gas turbines based on sequence matching, comprising the following steps: A standard operation model corresponding to each operation scenario is constructed based on multiple preset operation scenarios and the normal operation sequence corresponding to each operation scenario. Preferably, the operating background includes at least one of the following: the start-up and shutdown phase of the gas turbine, the load adjustment rate, the equipment maintenance status, and the presence of equipment warnings for activity, used to describe the macroscopic state of the gas turbine.

[0021] In this invention, different operating backgrounds of the gas turbine (such as "normal start-up", "full load frequency regulation", "emergency shutdown" etc.) are set and classified by those skilled in the art based on existing data. Taking the normal start-up operating background as an example: when constructing the standard operation model, all "normal operation sequences" belonging to the task background of "normal start-up" within a preset historical time period will be collected and trained together to form a standard operation model representing "standard start-up operation". Similarly, standard operation models will be established for other different operating backgrounds such as "normal shutdown" and "load adjustment".

[0022] This invention establishes one or more standard operation models for each different operating background because the standard operation procedures and timing requirements are completely different under different backgrounds, and such a setting is more in line with actual operating conditions.

[0023] The standard operating procedure model takes a real-time sequence of operations as input and outputs a metric (deviation) that measures how well the input sequence matches the standard operating procedure sequence represented by the model. This metric can be a probability value or another form of deviation score.

[0024] Based on alarm data and historical operating condition change information, various abnormal operation sequences are determined, and a corresponding abnormal operation model is constructed based on each abnormal operation sequence. In this invention, historical operation sequences, alarm data, and historical operating condition change information are the key bases for constructing and labeling historical abnormal operation sequences.

[0025] In one embodiment, a person skilled in the art acquires multiple alarm data (such as "high vibration" alarm), and selects abnormal operation sequences from the gas turbine's historical operating condition change information (such as changes in vibration value curves) corresponding to each alarm data, and manually labels the categories.

[0026] In another embodiment, both standard operation sequences and abnormal operation sequences are obtained by performing causal inference or association rule mining on massive historical operation logs and corresponding equipment operation data (such as historical operating parameters, alarm data, etc.) and operational background, and then automatically identifying and labeling standard and abnormal operation sequences using causal inference or association algorithms. Furthermore, this invention stores multiple standard operation models and multiple abnormal operation models in the same operation sequence knowledge base so that the invention can be used multiple times in different scenarios.

[0027] In this invention, the abnormal operation model is a model established for combinations of erroneous operations that may occur under any operating context and have historically been proven to lead to failure. For example, whether at startup or full load (different operating contexts), incorrectly and rapidly closing a critical valve (abnormal operation sequence) may cause problems. Therefore, the abnormal operation model is classified based on the type of abnormal operation sequence itself, rather than on the operating context.

[0028] Based on each abnormal operation sequence, historical operating condition change information, and alarm data, an associated risk evolution model is constructed for each abnormal operation model; During the construction of the abnormal operation model, the fault type and fault occurrence time caused by each abnormal operation can be analyzed based on the alarm type and time contained in the alarm data. Furthermore, the historical operating condition parameter change information from the occurrence of the abnormal operation to the occurrence of the fault can be found from the historical operating condition data.

[0029] Using a certain abnormal operation sequence and corresponding historical operating condition change information as input, and alarm data as labels, a risk evolution model associated with this abnormal operation sequence (which is also the abnormal operation model) can be trained through causal inference or association algorithms. The baseline model for the risk evolution model can be a time series model (such as LSTM), a probabilistic model [such as survival analysis (Cox model) or Bayesian network], or a joint prediction time + type multi-task model (such as multi-output neural network, Transformer multi-task encoder), etc.

[0030] Risk evolution models are used to describe the temporal evolution, probability distribution, and fault type from the start of a corresponding abnormal operation sequence to the occurrence of a potential fault. For example, when a certain abnormal operation sequence is matched, the associated risk evolution model will combine the current operating conditions to give a comprehensive prediction result, such as: "The probability of a 'high vibration' fault occurring within the next 5-15 minutes is 85%".

[0031] According to preset rules, at least one standard operation model and at least one abnormal operation model are selected as matching models respectively. The preset rules can be used to select between the standard operation model and the abnormal operation model based on the current operating background and current operating parameters. Those skilled in the art can set these rules themselves.

[0032] Preferably, after constructing the corresponding exception operation model based on each exception operation sequence, the method further includes: Based on preset prohibited operation criteria, at least one absolute abnormal model is determined from multiple abnormal operation models, and at least one relative abnormal model is determined from multiple abnormal operation models for each preset operating background. The preset operating standards include operations that are prohibited by law as recorded in relevant industry operating standards or operations that are common knowledge in the field and are strictly prohibited from being violated.

[0033] Accordingly, based on preset rules, at least one standard operation model and at least one abnormal operation model are selected as matching models, specifically: Determine the current operational context and select at least one standard operation model and at least one relative anomaly model based on the current operational context; Preferably, the current running background is determined as follows: Obtain current operating parameters; The current operating background is determined based on the current operating parameters.

[0034] This invention, through an edge computing gateway or PLC host computer set in the area where the gas turbine is located, can automatically identify the current operating background from the operation sequence knowledge base based on the current operating condition parameters (for example, judging that the unit is in the "load increase" stage based on parameters such as load and speed). Then, it automatically loads the standard operation sequence model and all relatively abnormal operation sequence models corresponding to the "load increase" operating background from the knowledge base for subsequent matching calculations.

[0035] All absolute anomaly models, the selected standard operating model, and the selected relative anomaly models are used as matching models.

[0036] In this invention, the selection logic for abnormal operation models differs from the one-to-one matching of standard models: while selecting a unique standard model based on the current operational context (e.g., "increased load and / or increased load"), the system simultaneously selects at least one relevant subset from all abnormal operation models for parallel matching. This subset is selected based on two criteria: first, it includes abnormal operation models corresponding to all universally applicable, severely abnormal operations that should be prohibited under any operational context, denoted as absolute abnormal models; second, it includes abnormal operation models that have been proven in historical data or mechanistically marked as highly relevant to the current specific operational context, denoted as relative abnormal models. Therefore, based on the two conditions of "general danger" and "specific danger under the current specific operational context," this invention intelligently filters out multiple most likely abnormal operation models for real-time similarity calculation, thereby achieving both comprehensive and focused risk monitoring.

[0037] Preferably, the baseline model used to construct the standard operation model and the abnormal operation model is a hidden Markov model or a long short-term memory network. Both the standard operation model and the abnormal operation model are used to characterize the temporal dependencies and probability transitions between operations contained in each operation sequence.

[0038] Obtain the current operation sequence and input it into each matching model to obtain the deviation of each standard operation model and the likelihood of each abnormal operation model in the matching model. The method for obtaining the current operation sequence and current operating parameters in this invention is mainly as follows: the current operation sequence of the gas turbine operator is collected in real time through an industrial data platform, and the real-time operating parameters accompanying the current operation sequence are obtained simultaneously.

[0039] This invention employs a matching algorithm based on sequence alignment or sequence embedding to calculate the deviation, first likelihood, and second likelihood of the current operation sequence in real time.

[0040] Each likelihood is compared with a first preset threshold to obtain multiple first comparison results, and each deviation is compared with a second preset threshold to obtain multiple second comparison results. Based on the multiple first comparison results and multiple second comparison results, risk operation sequences are identified, and a risk evolution model associated with a preset abnormal operation model is triggered to provide real-time operation risk warnings.

[0041] Specifically, determining the likelihood / deviation includes: quantitatively calculating and evaluating the differences between the timing, order, and magnitude of operations in the current operation sequence and the corresponding standard / abnormal timing, order, and magnitude of operations in the standard operation sequence model / abnormal operation model.

[0042] Preferably, the risk operation sequence is identified based on multiple first comparison results and multiple second comparison results, specifically as follows: When any likelihood exceeds a first preset threshold or any deviation exceeds a second preset threshold, the current operation sequence is identified as a risky operation sequence.

[0043] Preferably, the alarm data includes the time of fault occurrence and the type of fault. Correspondingly, the risk evolution model associated with the preset abnormal operation model is triggered to provide real-time operational risk warnings, specifically as follows: When the first likelihood value greater than the first preset threshold is generated, the risk evolution model associated with the abnormal operation model corresponding to that likelihood value is triggered, which is denoted as the target evolution model. Specifically, when multiple likelihood values ​​exceed the first preset threshold, the risk evolution model is selected according to the following rules: Although this invention calculates the similarity with multiple abnormal operation models in parallel during the real-time matching phase, in most cases, the likelihood of a specific abnormal operation model (e.g., "abnormal operation model A") first exceeds the first preset threshold. Once this occurs, the system will uniquely activate the "risk evolution model A" associated with "abnormal operation model A" to dynamically predict future risks.

[0044] In other embodiments, when the basis for identifying risky operations is high deviation and the likelihood is less than or equal to a first preset threshold, the risk evolution model corresponding to the abnormal operation model with the highest likelihood is selected for risk warning; when multiple likelihoods are greater than the first preset threshold, the risk evolution model corresponding to the abnormal operation model with the highest likelihood is also selected; when there are multiple abnormal operation models with the highest likelihood, the risk evolution model corresponding to any one of the abnormal operation models is selected.

[0045] Obtain the current operating condition parameters corresponding to the risk operation sequence, input the risk operation sequence and the current operating condition parameters into the target evolution model, and use the target evolution model to output the risk prediction result. The risk prediction result includes the fault prediction time, fault prediction type and the probability distribution corresponding to the fault prediction type caused by the risk operation sequence.

[0046] Specifically, the fault prediction time, fault prediction type, and the probability distribution corresponding to the fault prediction type are all visualized and output through the user monitoring interface.

[0047] In other embodiments, the risk prediction results also include a visual comparison of the differences between the current operation sequence and the matched standard operation sequence model, as well as recommended operation steps for correcting the current operation sequence to the standard operation sequence.

[0048] Preferably, after outputting the risk prediction results using the target evolution model, the method further includes: Obtain real results, which include the actual failure time, actual failure type, and probability distribution of the actual failure type caused by the risk operation sequence. The target evolution model is updated and corrected based on the error between the actual results and the risk prediction results.

[0049] As mentioned earlier, since the system uniquely activates the "Risk Evolution Model A" associated with the "Abnormal Operation Model A" to dynamically predict future risks, the subsequent feedback and learning process is also highly targeted. That is, the system will compare the final actual result (whether the failure occurs as predicted) with the prediction of the "Risk Evolution Model A" and use this result (success or failure) as feedback, which is only used to update and optimize the pair of models "Abnormal Model A" and "Risk Model A".

[0050] Preferably, the risk evolution model is updated and corrected based on the error between the actual results and the risk prediction results, specifically as follows: Under multiple real-time operational risk warnings, the error between the actual result and the risk prediction result for each time is obtained, and the parameters of the first preset threshold or the target evolution model are updated based on the error.

[0051] The parameters of the target evolution model are updated based on the error, specifically as follows: Obtain the number of errors that are less than the third preset threshold; The first preset threshold is updated based on the number of errors that are less than the third preset threshold. If the quantity is greater than or equal to the fourth preset threshold, then the first preset threshold is reduced. If the quantity is less than the fourth preset threshold, then the first preset threshold is increased.

[0052] The parameters of the target evolution model are updated based on the error, specifically as follows: If the error between the actual result and the risk prediction result for the corresponding number of times is less than the third preset threshold, the risk prediction result is taken as a positive sample; if it is greater than the third preset threshold, the risk prediction result is taken as a negative sample, and the risk evolution model is optimized and updated using the positive and negative samples.

[0053] After each risk warning occurs, this invention optimizes the sensitivity of the target risk model by adjusting the trigger threshold (i.e., the first preset threshold), thereby calibrating the abnormal operation model at the application level. At the same time, the successful or failed risk prediction results are used as learning samples to update the internal parameters of the corresponding target risk model and the abnormal operation model that triggers it (e.g., adjusting the weights of the neural network or the probability matrix of the HMM), thereby making the prediction ability of the target evolution model more accurate.

[0054] This invention also discloses a real-time operation risk early warning system for gas turbines based on sequence matching, comprising: The standard module is used to construct a standard operation model for each operating scenario based on multiple preset operating scenarios and the normal operation sequence corresponding to each operating scenario. The anomaly module is used to determine various abnormal operation sequences based on alarm data and historical operating condition change information, and to build a corresponding abnormal operation model based on each abnormal operation sequence. The correlation module is used to construct a correlation risk evolution model for each abnormal operation model based on each abnormal operation sequence, historical operating condition change information, and alarm data; The matching module is used to select at least one standard operation model and at least one abnormal operation model as matching models according to preset rules. The deviation likelihood module is used to obtain the current operation sequence and input the current operation sequence into each matching model to obtain the deviation degree output by each standard operation model and the likelihood degree output by each abnormal operation model in the matching model. The early warning module is used to compare each likelihood with a first preset threshold to obtain multiple first comparison results, and to compare each deviation with a second preset threshold to obtain multiple second comparison results. Based on the multiple first comparison results and multiple second comparison results, the module identifies risk operation sequences and triggers a risk evolution model associated with a preset abnormal operation model to provide real-time operation risk warnings.

[0055] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for real-time operation risk warning of gas turbines based on sequence matching.

[0056] likeFigure 2 As shown, the following is an embodiment of the present invention: Step a: Construction of the operation sequence knowledge base The goal of this step is to build a model library that can describe the "good operation" and "bad operation" of a gas turbine.

[0057] Step a1: Data source: Massive historical data from the industrial data platform, mainly including: Operation logs are all operations recorded by the operator station (HMI) of the DCS or SCADA system, such as button clicks, setpoint modifications, valve opening and closing, start and stop commands, etc. They must include the operation object, operation type, operation value, operation timestamp, and operator employee number.

[0058] Operating data, including gas turbine operating parameters (temperature, pressure, vibration, load, etc.) synchronized with the operating timestamp.

[0059] Event and alarm data, DCS alarm records, and SIS action records.

[0060] Text data, Standard Operating Procedures (SOP) documents, incident analysis reports, and expert knowledge bases.

[0061] Step a2: Construction of the standard operating model: Step a21: Sequence Extraction: From historical data, identify the "golden operation sequence" that was executed by an experienced operator in a specific operational context (such as normal startup, shutdown, load adjustment) and did not cause any abnormalities; record this as the normal operation sequence. Alternatively, standard operating procedures can be parsed from SOP documents as normal operation sequences.

[0062] Step a22: Model Training: Using the extracted normal operation sequence samples, train sequence models under different operating conditions. For example, a Hidden Markov Model (HMM) or an autoencoder based on a Recurrent Neural Network (RNN) such as a Long Short-Term Memory (LSTM) network can be trained for "normal start-stop".

[0063] The trained model (such as the transition and emission probabilities of an HMM, or the weights of an LSTM model) constitutes a mathematical description of the standard operation.

[0064] Step a3: Construction of the exception operation model: Step a31: Sequence Extraction: Analyze alarm data and historical operating condition change data using causal inference or association rule mining algorithms to identify specific operational sequences that frequently occur before equipment anomalies (i.e., abnormal operating condition changes) or failures. For example, analyzing all operation logs prior to "high turbine vibration" failures may reveal a common, inappropriate load increase operation pattern.

[0065] Step a4: Risk evolution modeling: For each extracted abnormal operation sequence and historical operating condition change information, further analyze its relationship with subsequent failures. Construct a risk evolution model, which can be a probability function or a time series prediction model. Its input is the abnormal operation sequence and the operating conditions at that time, and its output is the probability distribution of a specific failure (such as "high vibration") or the predicted value of the failure occurrence time in the future.

[0066] Store the abnormal operation model (which can also be an HMM or RNN model) and its associated risk evolution model together.

[0067] Step b: Real-time data acquisition The system listens for and collects all operation commands from the industrial data platform's data bus or database in real time, forming a real-time operation sequence. Simultaneously, it collects key operating parameters corresponding to the operation timestamps for subsequent determination of the operational context.

[0068] Step c: Real-time matching and deviation calculation of operation sequences Step c1: Operational Background Identification and Model Selection: The system first determines the real-time operational background based on real-time operating parameters. For example, it determines whether the unit is "starting up," "at full load," or "shutting down" based on load values ​​and speed. Based on this operational background, the system loads the corresponding standard operation model and relative abnormal operation model from the operation model knowledge base.

[0069] Step c2: Real-time matching and deviation calculation: Step c21: Compare with the standard model (calculate deviation): Input the real-time operating sequence into the selected standard operating model (such as HMM) and calculate the probability (likelihood) of the output under the standard operating model. A very low likelihood score indicates that the real-time operating sequence deviates significantly from the normal operating sequence.

[0070] Furthermore, the edit distance between the real-time operation sequence and the standard sequence template in the standard operation model (considering the addition, deletion, and modification of operations) can be calculated, and the timing difference, sequence difference, and magnitude difference of the operations can be quantified to obtain the deviation.

[0071] Step c22: Compare with anomaly models (calculate similarity): Similarly, input the real-time operation sequence into the relevant anomalous operation model and calculate its similarity. A high similarity score indicates that the real-time operation is very close to a historically known dangerous operation pattern.

[0072] Step d: Dynamic early warning of future risks Step d1: Risk Identification: Set dynamic thresholds. For example, if the deviation from any selected standard model is greater than threshold A, or the likelihood of the selected abnormal model is greater than threshold B, the system determines that the real-time operation sequence has potential risks. In other embodiments, thresholds A and B can be dynamically adjusted by those skilled in the art based on the operational context; for example, during critical start-up and shutdown phases, the two thresholds can be set to be more sensitive.

[0073] Step d2: Risk prediction: When the first likelihood exceeds the threshold B, the system will activate the "risk evolution model" associated with the matched abnormal operation sequence model.

[0074] The identified risk operation sequence is used as input, and the risk evolution model is used to output dynamic prediction results of future risks: possible failure types (e.g., "high vibration"), probability of occurrence (e.g., "85%)", and time window of occurrence (e.g., "within the next 5 to 15 minutes").

[0075] Furthermore, real-time operating parameters (such as real-time load and ambient temperature) can be used as input to make real-time corrections to the risk evolution model. For example, if the real-time load is high, the model may predict a shorter time window for failure.

[0076] Step d3: Early Warning Information Generation and Output: The system generates a detailed risk early warning message. This message not only includes the above prediction results, but should also: (a) Clearly indicate which / which operations constitute the anomalous sequence.

[0077] (b) Visually demonstrate the differences between real-time operation sequences and standard operation sequences.

[0078] (c) Provide corrective recommendations, such as “It is recommended to immediately stop the load increase and check the status of valve XX” or directly link to the relevant standard operating procedures.

[0079] The risk warning message pops up in a dedicated window of the monitoring system, alerting the operator in a prominent manner.

[0080] Step e: System closed loop and self-learning The system of this invention should also include a feedback and learning module. Once a risk warning is issued, the system continuously tracks the subsequent unit status. If the predicted fault actually occurs, this successful prediction will serve as a positive sample, used to strengthen the corresponding abnormal operation sequence model and risk evolution model. If the fault does not occur (possibly due to timely operator intervention), the system will also record this information to adjust the model's trigger threshold. In this way, the system can continuously learn and evolve during operation, and its warning accuracy will become increasingly higher.

[0081] Through the above specific implementation methods, the present invention fundamentally changes the monitoring paradigm of industrial processes, and constructs a proactive and forward-looking safety defense line by deeply understanding operator behavior and intelligently predicting future risks.

Claims

1. A method for real-time operational risk early warning of gas turbines based on sequence matching, characterized in that, Includes the following steps: A standard operation model is constructed for each operating scenario based on multiple preset operating scenarios and the normal operation sequence corresponding to each operating scenario. Based on alarm data and historical operating condition change information, various abnormal operation sequences are determined, and a corresponding abnormal operation model is constructed based on each abnormal operation sequence. Based on each abnormal operation sequence, historical operating condition change information, and alarm data, an associated risk evolution model is constructed for each abnormal operation model; According to preset rules, at least one standard operation model and at least one abnormal operation model are selected as matching models respectively. Obtain the current operation sequence and input it into each matching model to obtain the deviation of each standard operation model and the likelihood of each abnormal operation model in the matching model. Each likelihood is compared with a first preset threshold to obtain multiple first comparison results, and each deviation is compared with a second preset threshold to obtain multiple second comparison results. Based on the multiple first comparison results and the multiple second comparison results, a risk operation sequence is identified, and a risk evolution model associated with a preset abnormal operation model is triggered to provide real-time operation risk warning.

2. The method for real-time operational risk early warning of gas turbines based on sequence matching according to claim 1, characterized in that, After constructing the corresponding exception operation model based on each exception operation sequence, the following is also included: Based on preset prohibited operation criteria, at least one absolute abnormal model is determined from multiple abnormal operation models, and at least one relative abnormal model is determined from multiple abnormal operation models for each preset operating background. Accordingly, based on preset rules, at least one standard operation model and at least one abnormal operation model are selected as matching models, specifically: Determine the current operational context and select at least one standard operation model and at least one relative anomaly model based on the current operational context; All absolute anomaly models, the selected standard operating model, and the selected relative anomaly models are used as matching models.

3. The method for real-time operation risk early warning of gas turbines based on sequence matching according to claim 2, characterized in that, The risk operation sequence is identified based on the plurality of first comparison results and the plurality of second comparison results, specifically as follows: When any likelihood exceeds a first preset threshold or any deviation exceeds a second preset threshold, the current operation sequence is identified as a risky operation sequence.

4. The method for real-time operational risk early warning of gas turbines based on sequence matching according to claim 3, characterized in that, The alarm data includes the time of fault occurrence and the type of fault. Correspondingly, the risk evolution model associated with the preset abnormal operation model is triggered to provide real-time operational risk warnings, specifically as follows: When the first likelihood value greater than the first preset threshold is generated, the risk evolution model associated with the abnormal operation model corresponding to that likelihood value is triggered, which is denoted as the target evolution model. Obtain the current operating condition parameters corresponding to the risk operation sequence, input the risk operation sequence and the current operating condition parameters into the target evolution model, and use the target evolution model to output the risk prediction result. The risk prediction result includes the fault prediction time, fault prediction type and probability distribution corresponding to the fault prediction type caused by the risk operation sequence.

5. The method for real-time operational risk early warning of gas turbines based on sequence matching according to claim 4, characterized in that, After outputting risk prediction results using the aforementioned target evolution model, the method further includes: Obtain the actual results, which include the actual failure time and actual failure type caused by the risky operation sequence; The target evolution model is updated and corrected based on the error between the actual results and the risk prediction results.

6. The method for real-time operational risk early warning of gas turbines based on sequence matching according to claim 5, characterized in that, The target evolution model is updated and corrected based on the error between the actual results and the risk prediction results, specifically as follows: Under multiple real-time operational risk warnings, the error between the actual result and the risk prediction result for each instance is obtained; The parameters of the first preset threshold or the target evolution model are updated based on the error.

7. The method for real-time operational risk early warning of gas turbines based on sequence matching according to claim 1, characterized in that, Determine the current running context, specifically: Obtain current operating parameters; The current operating background is determined based on the current operating condition parameters.

8. The method for real-time operation risk early warning of gas turbines based on sequence matching according to claim 1, characterized in that, The operational context includes at least one of the following: gas turbine start-up and shutdown phases, load adjustment rate, equipment maintenance status, and equipment warnings regarding activity.

9. The method for real-time operation risk early warning of gas turbines based on sequence matching according to claim 1, characterized in that, The baseline models used to construct the standard operation model and the abnormal operation model are Hidden Markov Models or Long Short-Term Memory Networks.

10. A real-time operational risk early warning system for gas turbines based on sequence matching, characterized in that, include: The standard module is used to construct a standard operation model for each operating scenario based on multiple preset operating scenarios and the normal operation sequence corresponding to each operating scenario. The anomaly module is used to determine various abnormal operation sequences based on alarm data and historical operating condition change information, and to build a corresponding abnormal operation model based on each abnormal operation sequence. The correlation module is used to construct a correlation risk evolution model for each abnormal operation model based on each abnormal operation sequence, historical operating condition change information, and alarm data; The matching module is used to select at least one standard operation model and at least one abnormal operation model as matching models according to preset rules. The deviation likelihood module is used to obtain the current operation sequence and input the current operation sequence into each matching model to obtain the deviation degree output by each standard operation model and the likelihood degree output by each abnormal operation model in the matching model. The early warning module is used to compare each likelihood with a first preset threshold to obtain multiple first comparison results, and to compare each deviation with a second preset threshold to obtain multiple second comparison results. Based on the multiple first comparison results and the multiple second comparison results, the module identifies risk operation sequences and triggers a risk evolution model associated with a preset abnormal operation model to provide real-time operation risk warnings.

Images