Edge layer data acquisition and analysis method and system for management and control of a combustion engine power plant

Abstract

The application discloses an edge layer data acquisition and analysis method and system for gas turbine power plant management and control, relates to the technical field of intelligent power plant management and control, and is characterized in that a gas turbine power plant management and control server analyzes key characteristic values of multi-source sensor data, generates a light version model based on the key characteristic values of the multi-source sensor data, and deploys the light version model in communication with an edge acquisition node; the edge acquisition node monitors a core equipment partition of the gas turbine power plant based on the light version model, and transmits monitoring results to the gas turbine power plant management and control server after the light version model processes the monitoring data. The application can provide an edge layer data acquisition and analysis method and system for gas turbine power plant management and control, the edge acquisition node is arranged close to the core equipment partition, can directly and closely acquire sensor data, and reduces the length of a data transmission path; meanwhile, data cleaning, feature extraction and light model processing are all completed on the edge side, so that centralized processing delay after original data is fully uploaded to the gas turbine power plant management and control server is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent power plant management and control, and in particular to an edge layer data acquisition and analysis method and system for gas turbine power plant management and control. BACKGROUND

[0002] Current gas turbine power plant management and control mainly relies on the mode of combining traditional industrial control systems with information systems, and the core technologies include:

[0003] Data acquisition technology: based on DCS (Distributed Control System) and SCADA (Supervisory Control and Data Acquisition System) to realize the acquisition of key parameters (such as turbine speed, temperature, pressure, vibration, etc.), supplemented by sensors and PLCs to complete equipment state monitoring.

[0004] A hierarchical structure using field bus plus industrial Ethernet is adopted, with the field layer connecting devices through Profibus, Modbus, etc. and the control layer aggregating data to the server in the control room through industrial Ethernet (such as EtherNet / IP).

[0005] The communication mode is mainly wired communication (optical fiber, twisted pair), and mobile terminal access is realized by WiFi or 4G in some scenarios. Data transmission is mostly one-way aggregation, lacking real-time two-way interaction.

[0006] The data analysis mode adopts centralized storage plus post-analysis. After the data is uploaded to the plant-level data center, performance evaluation and fault tracing are performed through offline modeling, and some power plants introduce a simple AI model to realize simple early warning (such as over-temperature alarm).

[0007] The main defect of the prior art is the lack of real-time performance. Data needs to be forwarded through multiple layers to the control room or cloud analysis, with latency usually in the range of hundreds of milliseconds to seconds, making it difficult to meet the real-time regulation and control requirements of fast dynamic processes of gas turbines (such as combustion instability and surge).

[0008] Gas turbine power plants generate GB-level data (such as high-frequency vibration signals and infrared thermal imaging data) per second. Centralized transmission leads to congestion of the backbone network and a sharp increase in data center storage costs.

[0009] Different manufacturers' equipment (such as GE gas turbines and Siemens generators) use private protocols, and the data formats are not unified, forming "information silos" and making cross-system collaborative analysis difficult.

[0010] Analysis lag: traditional analysis relies on manual triggering or scheduled tasks, and fault warning is mostly "after-the-fact response", which cannot actively intervene in the early stages of faults (such as early bearing wear).

[0011] Edge side capability missing: field devices only have data acquisition functions, without deploying local computing units, and cannot complete real-time processing (such as noise filtering and feature extraction) near the data source, resulting in invalid data occupying transmission resources. SUMMARY

[0012] In order to solve the defects of the above-mentioned medical instrument inventory management technology, the present application provides an edge layer data acquisition and analysis method and system for the management and control of a gas turbine power plant. The technical scheme adopted is as follows:

[0013] The edge layer data acquisition and analysis method for the management and control of a gas turbine power plant comprises the following steps:

[0014] Step 1: partitioning according to core equipment of the gas turbine power plant, and installing edge acquisition nodes in the partitions;

[0015] Step 2: real-time acquisition of multi-source sensor data by the edge acquisition nodes, and edge-end data fusion;

[0016] Step 3: data cleaning and feature extraction by the edge acquisition nodes, and uploading only the key feature values of the multi-source sensor data to the gas turbine power plant management and control server;

[0017] Step 4: analysis of the key feature values of the multi-source sensor data by the gas turbine power plant management and control server, adaptive generation of a lightweight model according to the key feature values of the multi-source sensor data, and communication and deployment of the lightweight model with the edge acquisition nodes;

[0018] Step 5: monitoring of the core equipment partitions of the gas turbine power plant by the edge acquisition nodes based on the lightweight model, and transmission of the monitoring results to the gas turbine power plant management and control server after processing by the lightweight model;

[0019] Step 6: global monitoring result integration and display by the gas turbine power plant management and control server based on the monitoring results.

[0020] Optionally, the edge acquisition node is provided with a protocol conversion engine, which automatically identifies the communication protocol of the connected device.

[0021] Optionally, the method for automatically identifying the communication protocol by the protocol conversion engine is as follows:

[0022] The edge acquisition node first identifies the baud rate and signal rate, signal level and coding method, transmission medium and interface type through signal feature recognition, and performs preliminary screening and matching based on a pre-set communication protocol database to obtain a plurality of preliminary matching communication protocols;

[0023] Further, the structure characteristics of the data frame are analyzed, including frame boundary and fixed identifier, check mode and length characteristics, data field encoding and semantic characteristics; and the structure characteristics are matched with multiple preliminary matching communication protocols based on the structure characteristics to obtain a precise matching communication protocol.

[0024] Finally, the protocol is confirmed based on the interactive logic through a communication handshake process.

[0025] Optionally, in step 4, the adaptive generation of the lightweight model includes the following steps:

[0026] In step 41, the key feature values uploaded by the edge collection node are set as a set X = {x1, x2, x3, …, xn}, where n is the feature dimension, and the feature contribution rate of each key feature value in X is calculated by the gas turbine power plant management and control server through principal component analysis, and the first m principal components with a cumulative contribution rate greater than a set contribution rate threshold are selected, and the feature dimension is compressed to m, and the value of m and the type of key feature value are used as core constraint conditions for model structure screening. n} are calculated by the gas turbine power plant management and control server through principal component analysis, and the first m principal components with a cumulative contribution rate greater than a set contribution rate threshold are selected, and the feature dimension is compressed to m, and the value of m and the type of key feature value are used as core constraint conditions for model structure screening.

[0027] In step 42, the basic model library is called based on the device type, and models whose input layer dimension cannot be adjusted to m are excluded, and then the matching degree of the type of key feature value and the remaining basic model input item is calculated, and the basic model with the largest matching degree is selected as the target model.

[0028] In step 43, the target model is compressed to obtain a lightweight model based on the resource parameters of the edge collection node.

[0029] In step 44, the lightweight model is packaged and transmitted to the edge collection node using the MQTT protocol, and after receiving the lightweight model, the edge collection node verifies the file integrity, executes the decompression instruction, loads the model and runs, and returns a deployment success signal.

[0030] Optionally, in step 42, the formula for calculating the matching degree of the type of key feature value and the remaining basic model input item is:

[0031]

[0032] Where Acc is the accuracy of the model on the historical feature set matched with the m-dimensional feature set, FLPs is the model calculation amount, d opt is the optimal input dimension of the model, a is the precision weight, and g is the dimension adaptation weight. The proximity between m and d opt is enhanced by an exponential function exp.

[0033] Optionally, in step 43, the resource constraints of the edge collection node are: the maximum memory M max , the maximum computing power C max , and the maximum real-time memory occupancy Mused and maximum real-time computing capability occupation C used , calculate the minimum idle memory M free , M free = M max -M used ; calculate the minimum idle computing capability C free , C free =C max -C used ; take the minimum idle memory M free and the minimum idle computing capability C free as resource constraints.

[0034] Optionally, in step 43, the target model is compressed to obtain a lightweight model by model pruning, model quantization and knowledge distillation, so that the volume of the lightweight model is less than M free and the computing capability requirement is less than C free .

[0035] Optionally, after the edge collection node loads the lightweight model, the edge collection node performs local inference on the real-time collected feature data to obtain a monitoring result, and transmits the local inference result to the gas turbine power plant management and control server, the gas turbine power plant management and control server performs cross-domain fusion on the monitoring results of the edge collection nodes in each partition, and displays the original monitoring result and the cross-domain fusion result in digital twinning.

[0036] Optionally, the gas turbine power plant management and control server inputs the original monitoring result and the cross-domain fusion result into the gas turbine power plant operation AI early warning model, the gas turbine power plant operation AI early warning model outputs an early warning type and an early warning result, and the gas turbine power plant management and control server controls the early warning unit to perform an early warning action based on the early warning type and the early warning result.

[0037] The edge layer data collection and analysis system for gas turbine power plant management and control is used to realize the edge layer data collection and analysis method for gas turbine power plant management and control, and the edge layer data collection and analysis system comprises a plurality of edge collection nodes, a gas turbine power plant management and control server and an early warning unit, the plurality of edge collection nodes are respectively installed in a plurality of partition physical centers of core equipment of the gas turbine power plant, and are respectively in communication connection with the gas turbine power plant management and control server, the plurality of edge collection nodes respectively perform inference on monitoring data based on a lightweight model at an edge to obtain a monitoring result, and communicate and transmit the monitoring result to the gas turbine power plant management and control server, the gas turbine power plant management and control server performs cross-domain fusion on the monitoring result, inputs the original monitoring result and the cross-domain fusion result into a gas turbine power plant operation AI early warning model, and controls an early warning unit to perform an early warning action based on an early warning result.

[0038] In summary, the present application has at least one of the following beneficial technical effects:

[0039] The application can provide an edge layer data acquisition and analysis method and system for gas turbine power plant management and control, edge acquisition nodes are deployed close to core equipment partitions, can directly and closely collect sensor data, and reduce data transmission path length; at the same time, data cleaning, feature extraction and lightweight model processing are completed on the edge side, avoiding centralized processing delay after uploading all original data to the gas turbine power plant management and control server.

[0040] The gas turbine power plant management and control server generates a lightweight model based on the actual collected key feature values, which means that the model can dynamically match the real-time running state of the equipment, avoiding the limitations of fixed models and improving the adaptability of the model to the actual scene.

[0041] Through the hierarchical architecture of edge side localization processing and server global cooperation, the real-time performance, resource efficiency, monitoring accuracy and system adaptability are improved, and the intelligent level and operation reliability of the gas turbine power plant management and control are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flowchart of the edge layer data acquisition and analysis method for gas turbine power plant management and control. DETAILED DESCRIPTION

[0043] The application will be further described in detail below with reference to the accompanying drawings.

[0044] The application embodiment discloses an edge layer data acquisition and analysis method and system for gas turbine power plant management and control.

[0045] Referring to Figure 1 , embodiment 1, an edge layer data acquisition and analysis method for gas turbine power plant management and control, comprising the following steps:

[0046] Step 1, partitioning according to core equipment of the gas turbine power plant, and installing edge acquisition nodes in the partitions;

[0047] Step 2, the edge acquisition nodes collect multi-source sensor data in real time, and perform edge-end data fusion;

[0048] Step 3, the edge acquisition nodes complete data cleaning and feature extraction, and only upload key feature values of the multi-source sensor data to the gas turbine power plant management and control server;

[0049] Step 4, the gas turbine power plant management and control server analyzes the key feature values of the multi-source sensor data, adaptively generates a lightweight model according to the key feature values of the multi-source sensor data, and communicates and deploys the lightweight model with the edge acquisition nodes;

[0050] Step 5, the edge acquisition nodes monitor the core equipment partitions of the gas turbine power plant based on the lightweight model, and the lightweight model transmits the monitoring results to the gas turbine power plant management and control server after processing the monitoring data.

[0051] Step 6, the gas turbine power plant management server integrates and displays the global monitoring results based on the monitoring results.

[0052] By adopting the above technical solution, the edge collection node is deployed close to the core equipment partition, which can directly and closely collect sensor data, reducing the data transmission path length. At the same time, data cleaning, feature extraction and lightweight model processing are all completed on the edge side, avoiding the centralized processing delay after uploading the full amount of raw data to the gas turbine power plant management server.

[0053] The lightweight model runs locally on the edge, which can analyze the equipment state changes in real time and quickly output the results. Compared with the mode of edge collection and server full processing, the response time can be greatly shortened, providing a more timely basis for equipment abnormal early warning and real-time regulation.

[0054] The edge collection node first completes data cleaning and feature extraction, and only uploads key feature values instead of raw data, which can significantly reduce data transmission volume, reduce network bandwidth occupation, and avoid transmission congestion.

[0055] The gas turbine power plant management server does not need to process massive raw data, and only analyzes key feature values, reducing the computational load of the server. At the same time, the lightweight model adapts to the computational power limit of the edge collection node, avoiding the processing bottleneck caused by insufficient computational power on the edge side.

[0056] The gas turbine power plant management server generates a lightweight model based on the actual collected key feature values, which means that the model can dynamically match the real-time running state of the equipment (such as feature differences under different loads and working conditions), avoiding the limitations of fixed models and improving the adaptability of the model to actual scenarios.

[0057] The management server integrates the monitoring results of each edge node globally, which can associate the states of each partition from the overall perspective of the power plant, avoiding the one-sidedness of local monitoring, providing complete and global equipment running portraits for management personnel, and assisting in formulating systematic control strategies.

[0058] The method forms a collaborative advantage in real-time performance, resource efficiency, monitoring accuracy and system adaptability through the hierarchical architecture of local processing on the edge side and global coordination on the server, which can significantly improve the intelligent level and operation reliability of gas turbine power plant management.

[0059] Embodiment 2, the edge collection node is built-in protocol conversion engine, and the protocol conversion engine automatically identifies the communication protocol of the access equipment.

[0060] Embodiment 3, the method for automatically identifying the communication protocol by the protocol conversion engine is as follows:

[0061] The edge collection node first identifies the baud rate and signal rate, signal level and encoding method, transmission medium and interface type through signal feature recognition, and preliminarily filters and matches multiple preliminary matching communication protocols based on a preset communication protocol database;

[0062] Then, through analysis of the structural features of the data frame, including frame boundary and fixed identifier, check method and length feature, data field encoding and semantic feature, the precise matching communication protocol is obtained through structural feature matching of the structural features and the multiple preliminary matching communication protocols respectively.

[0063] Finally, the protocol is confirmed based on the interactive logic through the communication handshake process.

[0064] Through the above technical solution, the edge collection node judges the device communication protocol type by identifying signal features, data formats and communication handshake process information, and completes the specific method of protocol conversion based on the protocol type, which can be divided into the following three core identification links and subsequent protocol conversion processes, which are described in detail as follows:

[0065] Signal feature recognition: preliminary classification based on physical layer parameters

[0066] Signal feature recognition focuses on the physical layer attributes of the communication signal, and through the collection of signal electrical characteristics and transmission parameters, the protocol is preliminarily filtered, and the core identification parameters and methods are as follows:

[0067] Baud rate and signal rate: the edge collection node has a built-in signal sampling module to monitor the transmission rate of the signal in real time (such as baud rate common values are 9600, 19200, 38400, etc., and Ethernet protocol is 10Mbps / 100Mbps / 1Gbps). For example, ModbusRTU protocol mostly uses 9600 baud rate, while Profinet protocol is based on 100Mbps Ethernet by default, and through rate matching, protocols that do not match can be preliminarily excluded (such as signals with baud rate 38400 are excluded as Profinet protocol).

[0068] Signal level and encoding method: monitor the signal level standard (such as RS485 is a differential signal, the difference between high and low level is usually greater than 200mV; TTL level is a single-ended signal, high level 35V, low level 00.8V) and encoding method (such as Manchester encoding, NRZ encoding). For example, Ethernet protocol uses Manchester encoding (signal transition represents logic 1 / 0), while ModbusRTU mostly uses NRZ encoding (fixed level represents logic state), and through encoding features, the protocol range can be further narrowed.

[0069] Transmission medium and interface type: Determine the protocol transmission medium characteristics in combination with the access interface (such as RS232 / RS485 interface, Ethernet RJ45 interface, wireless LoRa module). For example, the RS485 interface usually corresponds to the bus protocols such as Modbus RTU and Profibus DP; the RJ45 interface corresponds to the Ethernet protocols such as TCP / IP, Modbus TCP and Profinet, and the interface type and signal characteristics can be quickly classified through the association.

[0070] Data format identification: Accurate matching based on frame structure characteristics;

[0071] Data format identification accurately distinguishes protocols by analyzing the structure characteristics of data frames (such as frame header, frame tail, check bit, data field format), and the core identification logic is as follows:

[0072] Frame boundary and fixed identifier:

[0073] Detect whether the monitoring data frame contains a fixed starting symbol / ending symbol (such as Modbus RTU frame without a fixed frame header, but separated by idle time; Modbus TCP MBAP header is fixed at 7 bytes, with the first byte being 0x00; OPC UA protocol frame uses a specific GUID as an identifier). By detecting the starting / ending identifier of the frame, the complete range of the frame can be located and the characteristics can be extracted.

[0074] Check method and length characteristics:

[0075] Analyze the check field type of the frame (such as Modbus RTU using CRC16 check, Modbus ASCII using LRC check, and TCP / IP using checksum) and the data field length rule (such as Profibus DP data frame length variable, but limited by the maximum transmission unit MTU; BACnet protocol frame contains an explicit length field). For example, if the frame tail contains a 2-byte CRC check and the data field length is ≤256 bytes, it can be initially determined as a Modbus RTU protocol.

[0076] Data field encoding and semantic characteristics:

[0077] Analyze the encoding format (binary, ASCII, JSON, etc.) and semantic identifier (such as function code, register address) of the data field. For example, the Modbus protocol data field contains a function code (0x03 represents reading a holding register, 0x06 represents writing a single register), and the register address is a 16-bit integer; the Profinet protocol data field contains IO data and alarm information, and has a device unique identifier (DeviceID), and through these semantic characteristics, the protocol type can be accurately matched.

[0078] The communication handshake process recognition identifies the protocol type by monitoring the message interaction sequence, handshake rules and state machine when the device establishes a connection. The core recognition dimensions are as follows:

[0079] Connection establishment process:

[0080] Record the initial interaction steps of device communication:

[0081] If there is a "request response" interaction, it may be a master-slave bus protocol such as Modbus RTU / ASCII;

[0082] If there is a "three-way handshake" (SYN, SYN+ACK, ACK) and "four-way handshake" process, it is a TCP protocol;

[0083] If there is a device discovery (Discovery), parameter configuration (Parameterization), and cyclic communication step (such as Profinet DCP protocol discovery device, IRT real-time channel establishment), it is an industrial Ethernet real-time protocol.

[0084] Timeout and retransmission mechanism: Monitor the message retransmission rules (such as TCP retransmission timeout time dynamic adjustment, Modbus RTU retransmission interval fixed at 3.5 character times, LoRa protocol based on ACK confirmation retransmission). For example, if no response is received after sending a message, retransmit at an interval of 100ms ± 20ms, and the number of retransmissions is ≤3 times, it can be determined to be consistent with the characteristics of industrial wireless protocols (such as WirelessHART).

[0085] Session layer identification interaction: For high-level protocols (such as OPCUA, MQTT), the handshake process includes session establishment information (such as MQTT's CONNECT message containing client ID, protocol version; OPCUA's Hello message containing endpoint URL, security policy). By analyzing these session layer identifiers, the protocol type can be directly matched.

[0086] When the protocol type is confirmed through the above three links, the protocol conversion engine performs the following operations:

[0087] Protocol mapping: Call the built-in protocol mapping library (contains the mapping relationship between mainstream industrial protocols and edge node internal standard protocols).

[0088] Format unification: unify the timing, unit, and data type (such as floating point, integer) of different protocols into an internal format that the edge node can handle (such as timestamp accurate to milliseconds, temperature unit unified to ℃, pressure unit unified to MPa), laying the foundation for subsequent edge data fusion (step 2).

[0089] Dynamic adaptation: if a new protocol is accessed (not in the built-in library), the protocol conversion engine records its signal characteristics, data format and handshake rules, generates a temporary mapping table and feeds back to the management server, which updates the protocol library and synchronizes to the edge node, realizing self-iteration of protocol identification.

[0090] The edge collection node can automatically identify the communication protocol of the access device (such as Modbus, Profinet, OPCUA, etc.), and complete protocol conversion, ensuring that multi-source sensor data can be uniformly collected, fused and processed, providing standardized data input for subsequent data cleaning, feature extraction and lightweight model deployment.

[0091] In step 4 of embodiment 4, the adaptive generation of the lightweight model includes the following steps:

[0092] Step 41, let the key feature values uploaded by the edge collection node be a set X = {x1, x2, x3, …, xn}, where n is the feature dimension, and the gas turbine power plant management server calculates the feature contribution rate of each key feature value in X through principal component analysis, selects the first m principal components with cumulative contribution rate greater than the set contribution rate threshold, and compresses the feature dimension to m, m value and key feature value type as the core constraint condition of model structure selection; n}, wherein n is the feature dimension, and the gas turbine power plant management server calculates the feature contribution rate of each key feature value in X through principal component analysis, selects the first m principal components with cumulative contribution rate greater than the set contribution rate threshold, and compresses the feature dimension to m, m value and key feature value type as the core constraint condition of model structure selection;

[0093] Step 42, based on the device type, call the basic model library, first exclude the models whose input layer dimension cannot be adjusted to m, then calculate the matching degree of the key feature value type and the remaining basic model input item, and select the basic model with the largest matching degree as the target model;

[0094] Step 43, based on the resource parameters of the edge collection node, compress the target model to obtain a lightweight model;

[0095] Step 44, package the lightweight model and transmit it to the edge collection node by using the MQTT protocol, after receiving, the edge collection node first verifies the file integrity, then executes the decompression instruction, finally loads the model and runs, and returns a deployment success signal.

[0096] In step 42 of embodiment 5, the formula for calculating the matching degree of the key feature value type and the remaining basic model input item is:

[0097]

[0098] Where Acc is the accuracy of the model on the historical feature set matched with m dimension, FLPs is the model calculation amount, d opt is the optimal input dimension of the model, a is the precision weight, and g is the dimension adaptation weight, which is enhanced by the exponential function exp m and d opt proximity.

[0099] In step 43 of Embodiment 6, the resource constraints of the edge collection node are set as: maximum memory M max , maximum computing capacity C max , maximum real-time memory occupancy M used , and maximum real-time computing capacity occupancy C used , the minimum idle memory M free , M free = M max -M used , the minimum idle computing capacity C free , C free = C max -C used , and the minimum idle memory M free and the minimum idle computing capacity C free are taken as the resource constraints.

[0100] In step 43 of Embodiment 7, the target model is compressed to a lightweight model by model pruning, model quantization, and knowledge distillation, so that the volume of the lightweight model is less than M free and the computing capacity requirement is less than C free .

[0101] By adopting the above technical solution, the feature dimension compression (from n to m) based on principal component analysis (PCA) has a clear quantification target: only the first m principal components with cumulative contribution rate exceeding the threshold are retained.

[0102] Principal component analysis extracts the most representative features (principal components with high contribution rate) in the data, compresses the dimension while avoiding loss of key information, and ensures that the feature set of the input model can effectively reflect the device state (such as the correlation of core indicators such as temperature and vibration).

[0103] After the dimension is compressed from n to m, the computational complexity of subsequent model training and inference decreases linearly or exponentially with the reduction of dimension, which not only reduces the burden of server-generated models, but also lays a foundation for lightweight operation of edge nodes.

[0104] Step 42 filters the target model by dimension matching and matching quantification calculation, and combines specific matching formulas (considering accuracy, computational complexity, optimal input dimension, and weight) to achieve precise adaptation of the model to the scene:

[0105] First, exclude models whose input layer dimension cannot be adjusted to m to avoid performance waste caused by forcing large models into small dimensions or precision loss caused by "strong expansion of small model dimensions", and reduce adaptation errors from the source.

[0106] Quantitative balance of precision and efficiency: in the matching formula, α·Acc ensures the basic monitoring accuracy of the model, Inhibit high-computational model (avoid edge node insufficient computing power), γ·exp(-|m-d opt |) through exponential function to strengthen the optimal input dimension d of the model opt The proximity to the actual dimension m. This quantization mechanism makes the selected target model meet the current feature dimension requirement and maintain high accuracy under edge computing power constraints, avoiding extreme cases of sacrificing accuracy for lightweight or ignoring computing power for accuracy.

[0107] Step 43 realizes the tailor-made of the model and the edge node hardware through resource constraint calculation (minimum free memory, minimum free computing power) and targeted compression techniques (pruning, quantization, knowledge distillation):

[0108] Model pruning removes redundant neurons (such as removing connections that have little effect on output), reducing model parameters; model quantization converts 32-bit floating-point parameters to 8-bit integers, reducing memory usage and computational load (usually 4 times the volume can be compressed); knowledge distillation transfers knowledge from the teacher model (server complex model) to the student model (edge lightweight model), preserving core reasoning ability while compressing.

[0109] The combination of the three makes the lightweight model reduce the volume and computing power demand while keeping the key monitoring performance loss within a small range.

[0110] Step 44 ensures the reliability of the entire model deployment process through MQTT protocol transmission, integrity check, and feedback mechanism: low-consumption and efficient transmission suitable for long-distance or weak network environment communication between edge nodes and servers, avoiding transmission delay affecting model update timeliness.

[0111] The edge node receives the model and first checks the file integrity to avoid model running errors caused by data damage during transmission; after decompression and loading, it returns a deployment success signal, allowing the server to real-time perceive the deployment status and trigger a retry mechanism for failed nodes to ensure model consistency across all partition edge nodes.

[0112] Example 8, after the edge collection node loads the lightweight model, it performs local inference on the real-time collected feature data to obtain monitoring results, and transmits the local inference results to the gas turbine power plant management server. The gas turbine power plant management server performs cross-domain fusion on the monitoring results of each partition edge collection node, and displays the original monitoring results and cross-domain fusion results through digital twinning.

[0113] By adopting the above technical solution, the core equipment of the gas turbine power plant, such as the gas turbine, the waste heat boiler and the steam turbine, has a strong coupling relationship. For example, the exhaust gas temperature of the gas turbine affects the efficiency of the waste heat boiler, and the pressure of the boiler affects the operation of the steam turbine. The single partition edge node can only monitor the local equipment state, and it is difficult to identify the cross-partition associated fault. Cross-domain fusion can construct a global equipment state association model by integrating the monitoring results of each partition (such as temperature, pressure, vibration and other associated features), and identify the systemic risk that cannot be perceived by a single partition.

[0114] Improve the accuracy of monitoring results: Cross-domain fusion can reduce local misjudgment through multi-source result verification. For example, a partition edge node mistakenly reports a temperature anomaly due to temporary sensor drift. The server can determine that the anomaly is an isolated error rather than a real fault by fusing the temperature, flow and other associated data of adjacent partitions, reducing the false positive rate. Conversely, if the multi-partition monitoring results point to the same potential problem (such as abnormal vibration detected by multiple partitions near the gas turbine), cross-domain fusion can strengthen the confidence of the fault and avoid false negatives.

[0115] The cross-domain fusion result can reflect the overall operation state of the whole plant area equipment (such as the load distribution and energy efficiency level of each partition), providing data support for the server to develop a global optimization strategy (such as dynamic load distribution and maintenance plan coordination), avoiding the problem of local optimization but global inefficiency caused by developing a strategy based on single partition data.

[0116] In embodiment 9, the gas turbine power plant control server inputs the original monitoring result and the cross-domain fusion result into the gas turbine power plant operation AI early warning model, and the gas turbine power plant operation AI early warning model outputs the early warning type and the early warning result. The gas turbine power plant control server controls the early warning unit to perform the early warning action based on the early warning type and the early warning result.

[0117] By adopting the above technical solution, the AI early warning combined with the cross-domain fusion result can break through the monitoring limitations of a single device or partition, identify multi-partition collaborative anomalies (such as the associated fault of the gas supply system and the turbine set), and achieve global risk perception of the overall operation state of the power plant. At the same time, the combination of the early warning result and the digital twin display can intuitively present the abnormal position, the influence range and the development trend, providing panoramic decision support for management personnel and improving the overall safety control level of the power plant.

[0118] The embodiment 10 is an edge layer data acquisition and analysis system for gas turbine power plant management and control, which is used to realize an edge layer data acquisition and analysis method for gas turbine power plant management and control. The edge layer data acquisition and analysis system comprises a plurality of edge acquisition nodes, a gas turbine power plant management and control server and an early warning unit. The plurality of edge acquisition nodes are respectively installed in a plurality of partition physical centers of core equipment of the gas turbine power plant, and are respectively in communication connection with the gas turbine power plant management and control server. The plurality of edge acquisition nodes respectively perform reasoning on monitoring data based on a lightweight model at an edge end to obtain monitoring results, and communicate and transmit the monitoring results to the gas turbine power plant management and control server. The gas turbine power plant management and control server performs cross-domain fusion on the monitoring results, and inputs the original monitoring results and the cross-domain fusion results into an AI early warning model for gas turbine power plant operation, and controls the early warning unit to perform early warning actions based on a warning result.

[0119] The implementation principle of the present application is described below by using specific embodiments:

[0120] A certain gas turbine power plant comprises a gas turbine (GT25000 type), a waste heat boiler (double-pressure horizontal type), a steam turbine (single-shaft combined cycle), a generator and auxiliary systems (fuel supply, cooling system, etc.), and the core equipment is distributed in five functional partitions (gas turbine area, waste heat boiler area, steam turbine area, generator area and auxiliary system area). A plurality of types of sensors (temperature, vibration, pressure, flow, current, etc.) are arranged in each partition, and the communication protocols of the sensors include ModbusRTU, Profinet, OPCUA, WirelessHART, etc. Real-time monitoring of the running state of the equipment, abnormal early warning and global management and control need to be realized.

[0121] I. Partition and edge node deployment:

[0122] 1. Partition division: divided into five partitions (numbered Z1 to Z5) according to the functions of the core equipment. One edge acquisition node (model ECU800, computing power 8TOPS, memory 8GB, supporting multi-protocol access) is arranged at the physical center of each partition. There are five edge nodes (E1 to E5) corresponding to Z1 to Z5 respectively.

[0123] 2. Node deployment: the edge nodes are connected with the sensors / intelligent instruments in the partition through an industrial switch, adopt distributed power supply (DC 24V), have dustproof and anti-vibration design (adapt to the workshop environment of the power plant), and establish communication with the gas turbine power plant management and control server (deployed in the central control room, computing power 200TOPS, memory 128GB) through a 5G industrial gateway.

[0124] II. Protocol conversion and multi-source data acquisition:

[0125] The edge node E1 (gas turbine area) needs to access the following equipment data:

[0126] Gas turbine shaft vibration sensor (2 units, Modbus RTU protocol, RS485 interface);

[0127] Combustion chamber temperature sensor (4 units, Profinet protocol, RJ45 interface);

[0128] Fuel flow intelligent instrument (1 unit, OPCUA protocol, Ethernet interface);

[0129] Cooling system pressure sensor (2 units, WirelessHART protocol, wireless transmission).

[0130] Protocol conversion process:

[0131] 1. Signal feature recognition:

[0132] E1 built-in protocol conversion engine detects through signal sampling module: Modbus RTU sensor baud rate 9600, RS485 differential signal (level difference 250mV), NRZ encoding; Profinet device signal rate 100Mbps, Manchester encoding, RJ45 interface; Based on the preset protocol database (containing 12 types of industrial protocol features), 4 candidate protocols (Modbus RTU, Profinet, OPCUA, WirelessHART) are matched.

[0133] 2. Data frame structure matching:

[0134] Parse Modbus RTU frame: no fixed frame header (separated by 3.5 character idle time), frame tail contains 2 bytes of CRC check, data field contains function code (0x03) and register address (0x0001 0x0008), completely matches the preset Modbus RTU structure feature;

[0135] Parse Profinet frame: contains DCP device discovery message (DeviceID=0x1A3B), data field is divided into IO data (16 bytes) and alarm information (8 bytes), matches Profinet structure feature.

[0136] 3. Communication handshake confirmation:

[0137] When establishing a connection with the OPCUA instrument, the "Hello message (endpoint URL=opc.tcp: / / 192.168.1.10:4840) + security policy negotiation" interaction logic is detected, confirming that the protocol is OPCUA;

[0138] When interacting with the WirelessHART sensor, the "ACK retransmission (timeout 100ms±10ms, retransmission 3 times)" mechanism is monitored, matching the WirelessHART protocol.

[0139] 4. Protocol conversion result: the engine converts the above 4 types of protocol data into the internal standard format of the edge node (timestamp accurate to milliseconds, temperature unit ℃, pressure unit MPa, flow rate unit kg / s), laying the foundation for data fusion.

[0140] III. Data acquisition and feature extraction:

[0141] 1. Multi-source data acquisition:

[0142] The edge nodes E1E5 collect data in real time at a frequency of 10 Hz, for example:

[0143] Gas turbine area (E1): shaft vibration (X / Y direction, 050 mm / s), combustion chamber outlet temperature (650850 ℃), natural gas flow rate (2040 kg / s), cooling water pressure (0.81.2 MPa);

[0144] Waste heat boiler area (E2): economizer outlet water temperature (280320 ℃), steam pressure (4.55.5 MPa), flue gas oxygen content (35%).

[0145] 2. Edge data fusion:

[0146] E1 performs spatio-temporal alignment on vibration, temperature, and flow rate data with the same timestamp (based on GPS synchronous clock, error ≤1 ms), and eliminates isolated values caused by sensor drift (such as a temperature sensor instantaneously jumping to 1000 ℃, which is determined as abnormal and replaced with a moving average value).

[0147] 3. Feature extraction and upload:

[0148] Feature extraction is performed on the original data (about 720,000 per hour), and only key feature values are retained:

[0149] Time domain features: vibration peak value (daily maximum / minimum value), temperature mean value (5-minute moving average), flow rate fluctuation variance;

[0150] Frequency domain features: gas turbine rotor 1X / 2X vibration frequency (extracted through FFT conversion);

[0151] Correlation features: Pearson correlation coefficient of combustion chamber temperature and fuel flow rate.

[0152] Finally, only 120 key feature values per hour are uploaded (data volume reduced by 99.98%), and sent to the management server through the MQTT protocol.

[0153] IV. Lightweight model generation and deployment:

[0154] Taking the gas turbine area (E1) as an example, the server generates a lightweight "device anomaly monitoring model":

[0155] 1. Feature dimension compression:

[0156] The server receives the 15-dimensional key features (vibration, temperature, flow, etc.) uploaded by E1, calculates the feature contribution rate through principal component analysis (PCA): the cumulative contribution rate of the top 5 principal components (vibration peak, temperature mean, 1X frequency, flow variance, and temperature-flow correlation coefficient) reaches 92% (threshold set at 85%), so the dimension is compressed to m = 5.

[0157] 2. Target model screening:

[0158] The server calls the base model library (including random forest, LSTM, CNN), excludes CNN with input layer dimension that cannot be adjusted to 5 (fixed input dimension 16), and calculates the matching degree of the remaining models (random forest, LSTM):

[0159] Random forest: historical accuracy rate 91%, calculation amount 2.3 TOPS, optimal input dimension 5 (consistent with m = 5);

[0160] LSTM: historical accuracy rate 89%, calculation amount 3.8 TOPS, optimal input dimension 8;

[0161] Substitute the matching degree formula (precision weight α = 0.6, dimension adaptation weight β = 0.4), random forest matching degree (0.87) is higher than LSTM (0.72), and random forest is selected as the target model.

[0162] 3. Model lightweight compression:

[0163] E1 resource constraints: maximum memory 4GB, maximum calculation capacity 5TOPS; minimum free memory = 4GB, current occupation 1.2GB = 2.8GB, minimum free calculation capacity = 5TOPS, current load 1.5TOPS = 3.5TOPS.

[0164] Lightweight random forest (original volume 256MB, calculation amount 2.3TOPS):

[0165] Pruning: remove 15% redundant decision trees (impact on accuracy <2%);

[0166] Quantization: convert 32-bit floating-point parameters to 8-bit integers (volume compression 4 times);

[0167] Knowledge distillation: use the server complex model (XGBoost) as the "teacher" to optimize the "student model" output layer weight.

[0168] The final lightweight model volume is 32MB (<2.8GB), and the calculation amount is 1.8TOPS (<3.5TOPS), which meets the E1 resource constraints.

[0169] 4. Deployment process:

[0170] The server sends the model package (with CRC check) to E1 through the MQTT protocol; E1 checks the file integrity (CRC match) and decompresses, loads the model, and returns a "deployment success" signal, with a total time of 12 seconds.

[0171] Five, edge monitoring and cross-domain fusion:

[0172] 1. Edge local monitoring:

[0173] E1 performs inference on real-time feature data (updated every 100ms) based on a lightweight model: when "peak vibration > 45mm / s and 1X frequency fluctuation > 5Hz" are detected, it is determined as "rotor imbalance warning", and the result (including timestamp, confidence 96%) is uploaded to the server.

[0174] 2. Server cross-domain fusion:

[0175] The server receives the "rotor imbalance warning" from E1 (gas turbine area), synchronously retrieves the "steam pressure fluctuation" data from E2 (waste heat boiler area) and the "bearing temperature rise" data from E3 (steam turbine area), and through correlation analysis finds that: gas turbine rotor imbalance → waste heat boiler steam flow pulsation → steam turbine bearing load imbalance, forming a causal chain, and determining as "systematic mechanical vibration anomaly" (excluding single sensor false alarm).

[0176] Six, AI warning and execution:

[0177] 1. Warning model output:

[0178] The server inputs the cross-domain fusion result into the "gas turbine power plant operation AI warning model" and outputs the warning type "secondary mechanical failure", with the warning result: the fault location is the No. 3 bearing of the gas turbine, the impact scope involves the waste heat boiler superheater, and it is recommended to shut down for maintenance within 1 hour.

[0179] 2. Warning execution:

[0180] The server controls the warning unit (audible alarm, large screen in the control room, and management personnel's mobile phone APP):

[0181] The audible alarm in the plant area starts (red flashing + buzzing);

[0182] The large screen digital twin interface highlights the fault area and associated equipment;

[0183] Send a message to the operations director + APP push (including fault details and disposal suggestions).

[0184] The edge side inference response time is shortened from 2.3 seconds of traditional server centralized processing to 80 ms, and the abnormal early warning is 15 minutes in advance;

[0185] The data transmission volume is reduced by 92% (from 1.8 TB per day to 144 GB), and the server computing load is reduced by 65%;

[0186] Through cross-domain fusion, the false alarm rate is reduced from 8% to 1.2%, and two times of unit trip caused by gas turbine rotor imbalance are successfully avoided;

[0187] Adaptability: The light model is dynamically updated according to the working condition (such as automatically adjusting the temperature threshold in summer high load), and the model adaptation accuracy reaches 94%.

[0188] Through the above implementation, the gas turbine power plant realizes intelligent management and control of edge localization processing and server global cooperation, and the operation reliability is greatly improved, and the annual reduction and downtime loss are greatly reduced.

[0189] The above are preferred embodiments of the present application, not to limit the protection scope of the present application, therefore: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for edge-layer data acquisition and analysis for the management and control of gas turbine power plants, characterized in that, Includes the following steps: Step 1: Divide the gas turbine power plant into zones according to the core equipment, and install edge acquisition nodes in each zone; Step 2: The edge acquisition node collects multi-source sensor data in real time and performs edge-end data fusion; Step 3: The edge acquisition nodes complete data cleaning and feature extraction, and only upload the key feature values ​​of the multi-source sensor data to the gas turbine power plant control server; Step 4: The gas turbine power plant control server analyzes the key feature values ​​of multi-source sensor data, adaptively generates a lightweight model based on the key feature values ​​of multi-source sensor data, and communicates with the edge acquisition nodes to deploy the lightweight model. Step 5: The edge acquisition nodes monitor the core equipment zones of the gas turbine power plant based on the lightweight model. After processing the monitoring data, the lightweight model transmits the monitoring results to the gas turbine power plant management and control server. Step 6: The gas turbine power plant management server integrates and displays the global monitoring results based on the monitoring results; In step 4, the adaptive generation of the lightweight model includes the following steps: Step 41, let the key feature values ​​uploaded by the edge acquisition nodes be set. Where n is the feature dimension, the gas turbine power plant control server calculates them using principal component analysis. The feature contribution rate of multiple key feature values ​​is used to select the top m principal components whose cumulative contribution rate is greater than the set contribution rate threshold. The feature dimension is compressed to m, and the value of m and the type of key feature value are used as the core constraints for model structure selection. Step 42: Based on the device type, call the basic model library. First, exclude models whose input layer dimension cannot be adjusted to m. Then, calculate the matching degree between the key feature value type and the input items of the remaining basic models. Select the basic model with the highest matching degree as the target model. Step 43: Based on the resource parameters of the edge acquisition nodes, the target model is lightly compressed to obtain a lightweight model; Step 44: The lightweight model is packaged and transmitted to the edge acquisition node via the MQTT protocol. After receiving the model, the edge acquisition node first verifies the file integrity, then executes the decompression command, and finally loads and runs the model, returning a deployment success signal. In step 42, the formula for calculating the matching degree between the key feature value type and the remaining base model inputs is: ； in The accuracy of the model on the historical feature set matched with the m-dimensional dimension. For model computational cost, The optimal input dimension for the model. For precision weights, To adapt weights to dimensions, an exponential function is used. Strengthen m and Proximity has an impact.

2. The edge layer data acquisition and analysis method for gas turbine power plant management and control according to claim 1, characterized in that, The edge acquisition node has a built-in protocol conversion engine that automatically identifies the communication protocol of the access device.

3. The edge layer data acquisition and analysis method for gas turbine power plant management and control according to claim 1, characterized in that, The protocol conversion engine automatically identifies the communication protocol using the following method: The edge acquisition node first identifies the baud rate and signal speed, signal level and encoding method, transmission medium and interface type by recognizing signal characteristics, and then performs preliminary screening and matching based on a pre-set communication protocol database to obtain multiple preliminary matching communication protocols; Then, by analyzing the structural features of the data frame, including frame boundaries and fixed identifiers, verification methods and length features, data field encoding and semantic features, the exact matching communication protocol is obtained by performing structural feature matching with multiple preliminary matching communication protocols based on the structural features. Finally, protocol confirmation based on interaction logic is achieved through a communication handshake process.

4. The edge layer data acquisition and analysis method for gas turbine power plant management and control according to claim 3, characterized in that, In step 43, the resource constraints of the edge acquisition node are set as follows: maximum memory Maximum computing power Maximum real-time memory usage and maximum real-time computing power usage , Calculate minimum free memory , ; Calculate minimum idle computing power , ; with minimal free memory and minimum idle computing power Due to resource constraints.

5. The edge layer data acquisition and analysis method for gas turbine power plant management and control according to claim 4, characterized in that, In step 43, model pruning, model quantization, and knowledge distillation are used to compress the target model into a lightweight model, reducing its size to less than [a certain value]. The computing power requirement is less than .

6. The edge layer data acquisition and analysis method for gas turbine power plant management and control according to claim 5, characterized in that, After the edge acquisition node loads the lightweight model, it performs localized inference on the real-time acquired feature data to obtain monitoring results, and transmits the localized inference results to the gas turbine power plant control server. The gas turbine power plant control server performs cross-domain fusion on the monitoring results of the edge acquisition nodes in each zone, and displays the original monitoring results and cross-domain fusion results as a digital twin.

7. The edge layer data acquisition and analysis method for gas turbine power plant management and control according to claim 6, characterized in that, The gas turbine power plant management and control server inputs the original monitoring results and cross-domain fusion results into the gas turbine power plant operation AI early warning model. The gas turbine power plant operation AI early warning model outputs the early warning type and early warning result. Based on the early warning type and early warning result, the gas turbine power plant management and control server controls the early warning unit to execute early warning actions.

8. An edge-layer data acquisition and analysis system for the management and control of gas turbine power plants, characterized in that: To implement the edge layer data acquisition and analysis method for gas turbine power plant management and control as described in claim 7, the edge layer data acquisition and analysis system includes multiple edge acquisition nodes, a gas turbine power plant management and control server, and an early warning unit. The multiple edge acquisition nodes are respectively installed in multiple partition physical centers of the core equipment of the gas turbine power plant and are respectively connected to the gas turbine power plant management and control server. The multiple edge acquisition nodes respectively perform inference on the monitoring data at the edge based on a lightweight model to obtain monitoring results, and transmit the monitoring results to the gas turbine power plant management and control server. The gas turbine power plant management and control server performs cross-domain fusion of the monitoring results, inputs the original monitoring results and cross-domain fusion results into the gas turbine power plant to run the AI ​​early warning model, and controls the early warning unit to execute early warning actions based on the early warning results.

Images