CN118410279A - Deep learning-based turbine unit fault processing method and system - Google Patents

Abstract

The invention discloses a turbine unit fault processing method and system based on deep learning, and belongs to the technical field of mechanical fault diagnosis. The method comprises the following steps: acquiring operation state data of the pretreated turbine unit; inputting the running state data into a pre-trained running fault diagnosis model, and outputting to obtain a fault diagnosis result of the turbine unit; and determining an optimal treatment scheme of the turbine unit fault based on the fault diagnosis result. The invention improves the accuracy and the comprehensiveness of fault diagnosis by fusing the data of multiple sensors; the accuracy and the adaptability of the model are improved through the self-adaptive operation fault diagnosis model; the system can automatically recommend repair measures or maintenance plans according to the diagnosis results and support decision making.

Description

Deep learning-based turbine unit fault processing method and system

Technical Field

The invention relates to the technical field of mechanical fault diagnosis, in particular to a turbine unit fault processing method and system based on deep learning.

Background

Steam turbines are the main equipment of modern thermal power plants and are also used in metallurgical industry, chemical industry and ship power plants, and traditional turbine set vibration fault diagnosis mainly relies on experienced technicians to determine fault types and positions by observing and analyzing vibration data.

These conventional diagnostic methods tend to have high technician dependence and are difficult to diagnose complex faults, resulting in reduced diagnostic accuracy for the turbine unit faults.

Based on this, there is a need for a turbine unit fault handling method and device based on deep learning to solve the above-mentioned technical problems.

Disclosure of Invention

The invention provides a turbine unit fault processing method and device based on deep learning, which can effectively improve the fault diagnosis accuracy of a turbine unit. The technical proposal is as follows:

In a first aspect, a turbine unit fault handling method based on deep learning is provided, the method comprising:

acquiring operation state data of the pretreated turbine unit; wherein the operating state data includes acceleration, temperature, pressure, voltage, and current;

Inputting the running state data into a pre-trained running fault diagnosis model, and outputting to obtain a fault diagnosis result of the turbine unit;

And determining an optimal treatment scheme of the turbine unit fault based on the fault diagnosis result.

Preferably, the preprocessing includes a data cleaning process, a data synchronization process, a feature normalization process, and a default filling process; the data cleaning process is used for removing noise and irrelevant data in the real-time data of the turboset, the data synchronization process is used for ensuring consistency of time labels of the real-time data of the turboset, the characteristic normalization process is used for eliminating influences of different dimensions in the real-time data of the turboset, and the default filling process is used for filling missing data in the real-time data of the turboset.

Preferably, the operation fault diagnosis model is specifically obtained through training in the following manner:

determining a selection scheme of the deep learning network based on the fault type and the characteristic property of the turbine unit to be identified;

Training the selected deep learning network based on a known sample set to obtain an initial model; wherein the known sample set includes historical fault data and historical normal data;

and carrying out parameter and structure adjustment on the initial model based on a preset independent verification data set to obtain the operation fault diagnosis model.

Preferably, the fault diagnosis result is specifically obtained by:

Performing feature extraction processing on the running state data by using a convolutional neural network to obtain advanced features for representing the running mode and abnormal state of the turbine unit;

Performing recognition processing on the advanced features by using a cyclic neural network to obtain a pattern recognition result of the steam turbine unit;

classifying the advanced features and the pattern recognition result by using the operation fault diagnosis model to obtain a fault state of the steam turbine unit; wherein the fault status fault type and fault location;

and determining the fault diagnosis result based on the advanced feature, the pattern recognition result and the fault state.

Preferably, the determining the fault diagnosis result based on the advanced feature, the pattern recognition result, and the fault state includes:

summarizing, analyzing and processing the advanced features, the pattern recognition result and the fault state to obtain an analysis result;

Converting the analysis result based on a rule engine to obtain diagnosis information of the turbine unit fault; wherein the diagnostic information includes a cause of failure and maintenance measures;

and processing the fault state and the diagnosis information based on a natural language generation technology to obtain the fault diagnosis result.

Preferably, the determining an optimal treatment scheme for the turbine unit fault based on the fault diagnosis result includes:

Matching the fault diagnosis result with a preset database to obtain fault background information; the database comprises a history maintenance record, a fault database and an equipment technical manual;

performing evaluation processing on the fault background information based on a preset evaluation algorithm to obtain a fault evaluation result of the turbine unit; wherein the evaluation algorithm comprises a risk scoring algorithm, a rule inference algorithm and a historical data analysis algorithm;

Calculating the fault evaluation result based on a preset decision algorithm to obtain an optimal treatment scheme of the turbine unit fault; wherein the decision algorithm includes a decision tree algorithm, a priority algorithm, and a cost-benefit algorithm.

In a second aspect, there is provided a deep learning-based turbine unit fault handling apparatus, the apparatus comprising:

The acquisition unit is used for acquiring the operation state data of the pretreated turbine unit; wherein the operating state data includes acceleration, temperature, pressure, voltage, and current;

The diagnosis unit is used for inputting the running state data into a pre-trained running fault diagnosis model and outputting a fault diagnosis result of the steam turbine unit;

And the decision unit is used for determining an optimal processing scheme of the turbine unit fault based on the fault diagnosis result.

In a third aspect, a computer device is provided, where the computer device includes a memory for storing a computer program and a processor for executing the computer program stored on the memory to implement the steps of the deep learning-based turbine set fault handling method described above.

In a fourth aspect, a computer readable storage medium is provided, in which a computer program is stored, which when executed by a processor, implements the steps of the deep learning based turbine unit fault handling method described above.

In a fifth aspect, a computer program product is provided, comprising a computer program, which when executed by a processor, implements the steps of the deep learning based turbine set fault handling method described above.

The technical scheme provided by the invention has at least the following beneficial effects:

1. The accuracy and the efficiency of the fault diagnosis of the turbine unit are improved. Traditional fault diagnosis methods generally rely on manual experience and threshold judgment, and are prone to missed diagnosis and misdiagnosis. The invention can automatically learn fault characteristics from massive historical operation data by utilizing the deep learning model to form more accurate and comprehensive diagnosis capability and reduce human factor interference.

2. Potential fault symptoms can be found in time. The early symptoms of the faults can be sensitively captured by monitoring key parameters such as vibration, temperature and the like of the turbine unit in real time and inputting a trained fault diagnosis model, so that basis is provided for preventing further deterioration and loss of the faults for preventive maintenance.

3. And giving an optimal fault handling scheme. The simple fault diagnosis may not direct the actual fault handling. The invention can intelligently generate corresponding optimal treatment measures such as adjusting operation parameters, replacing wearing parts, overhauling specific parts and the like on the basis of diagnosis, and provides decision support for on-line maintenance personnel.

4. Accumulating optimized diagnosis and treatment knowledge. The diagnosis and processing method can be continuously iterated, optimized and summarized to form the knowledge base culture and inheritance expert experience of the field through continuously accumulating various fault case data of the turbine unit and training the turbine unit for a deep learning model.

6. And the safety and the economy of the energy system are improved. The safe and stable operation of the turbine unit, which is core equipment of energy systems of thermal power plants and the like, is directly related to power supply and cost effectiveness. The invention can furthest reduce the unplanned downtime, prolong the service life of equipment and ensure the safe and economic operation of an energy system.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a method for handling turbine unit faults based on deep learning according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a deep learning-based turbine unit fault handling system according to an embodiment of the present invention;

fig. 3 is a hardware architecture diagram of a computer device according to a deep learning-based turbine unit fault handling method according to an embodiment of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments, and all other embodiments obtained by those skilled in the art without making any inventive effort based on the embodiments of the present invention are within the scope of protection of the present invention.

As described above, the embodiment of the invention adopts a multi-sensor data fusion technology and combines machine learning and an artificial intelligence algorithm to monitor and analyze the vibration data of the turbine set in real time. The system cleans and integrates data by collecting various sensor data such as vibration, temperature, pressure and the like of the turbine unit and using a data preprocessing technology. Then, the convolutional neural network (Convolutional Neural Network, CNN) and/or the cyclic neural network (Recurrent Neural Network, RNN) are used for carrying out feature extraction and pattern recognition on the processed data, so that the automatic recognition and positioning of the fault type of the turbine unit are realized.

Specific implementations of the above concepts are described below.

Referring to fig. 1, the method for processing a turbine unit fault based on deep learning according to the embodiment of the present invention includes:

step 100, acquiring operation state data of a pretreated turbine unit; wherein the operational status data includes vibration, temperature, pressure, voltage, and current;

102, inputting the running state data into a pre-trained running fault diagnosis model of the turbine unit, and outputting to obtain a fault diagnosis result of the turbine unit;

step 104, determining an optimal treatment scheme of the turbine unit faults based on the fault diagnosis result.

In the embodiment of the invention, the data information acquired by various sensors is firstly acquired, then the data information is cleaned according to a preset processing method, the cleaned data information is integrated, then the integrated data is subjected to feature extraction and pattern recognition according to a preset neural network model, and the neural network model is continuously updated according to a pattern recognition result, so that the accuracy of the fault type recognition of the turbine unit is improved. The multi-dimensional operation state of the turbine unit is comprehensively monitored by a multi-sensor data fusion technology, so that the accuracy and the comprehensiveness of fault diagnosis of the turbine unit are improved; the model parameters are automatically adjusted according to the characteristics of the operation state data through the self-adaptive operation turbine unit fault diagnosis model, so that the accuracy and the adaptability of the turbine unit operation fault diagnosis model are improved; the system can automatically recommend repair measures or maintenance plans according to fault diagnosis results and support decision making.

The manner in which the individual steps shown in fig. 1 are performed is described below.

First, for step 100, the operation state data of the turbine group after the pretreatment is acquired.

In the embodiment of the invention, in order to accurately detect the faults of the turbine unit, various types of sensors are required to be deployed at key positions of the turbine unit, the running state data of the turbine unit, including vibration, temperature, pressure, voltage and current, are acquired in real time, and the running state data are transmitted to a data preprocessing module. Current sensors, voltage sensors, among other things, monitor current and voltage parameters of motors and electrical systems in real time, which data is critical to identifying insulation faults, poor contacts, overloads, or other electrical problems.

Specifically, the mounting positions of the respective sensors are as follows:

Accelerometers, i.e., vibration sensors, are used to measure and monitor the vibration level of a turbine unit, and these vibration data can be used to detect imbalance, misalignment, bearing failure, or other mechanical problems, and are therefore mounted near the bearings, on rotating parts, or on the housing of the turbine unit to capture the vibration characteristics of the entire turbine unit.

Temperature sensors are used to monitor the temperature of critical components such as bearings, cabinets, steam pipes. The temperature change may indicate problems such as overheating, increased friction, or cooling system failure, and is thus installed at hot spot areas, such as bearings, in the cabinet, steam inlet and outlet pipes, etc., to monitor the temperature change in real time.

The pressure sensor is used for measuring the steam pressure in the turbine set. The pressure change may reflect a change in steam flow or a decrease in sealing performance, and is therefore mounted at strategic locations in the steam line, cylinder inlet and outlet, etc., for monitoring steam pressure and detecting leaks or decreased performance.

Current sensors are used to monitor the current of the generator and associated electrical equipment. Current anomalies may indicate motor overload, insulation damage, or electrical connection problems, and are therefore installed in the generator power supply lines, control systems, and circuits of critical electrical components to monitor current conditions in real time.

In particular, in a turbine group, the main relevant electrical equipment associated with the current sensor comprises:

The mechanical energy of the generator and the turbine is finally converted into electric energy, which is needed to be realized through the generator. The generator is one of the most critical electrical devices in the turbine unit, and the change condition of the winding current directly reflects the power generation state and the load level. The current sensors are arranged in the stator winding and the exciting winding of the generator, so that the operation condition of the generator can be monitored.

And the excitation system provides direct current for an excitation winding of the generator and controls reactive output and voltage of the generator. Current anomalies in the excitation system may cause faults such as unstable generator voltage or loss of excitation. It is therefore necessary to arrange current sensors at the positions of the excitation current transformer, the exciter output circuit, etc.

And the main transformer is used for boosting the electric energy generated by the generator and then transmitting the boosted electric energy to the power grid. The current of the primary and secondary windings of the transformer reflects the output of the generator and the operating state of the transformer itself. These key current information can be obtained by installing current transformers on the high and low voltage sides of the transformer.

And the station power system provides power for various auxiliary equipment of the steam turbine generator unit, including a circulating water pump, a condensate pump, an oil pump and the like. The current sensor is arranged on the power supply line of the station service system, so that the running state and the energy consumption level of the equipment can be mastered.

And an accident oil pool pump, which is used for collecting oil or other mediums which are possibly leaked when the steam turbine unit has an accident, and transferring the oil or other mediums to the accident oil pool pump. The current of the accident oil pool pump is monitored to judge whether the accident oil pool pump is in a starting state or not, and whether the leakage accident occurs or not is indirectly reflected.

The air cooling island fan is a key device for cooling steam exhausted from the tail of the steam turbine for the air cooling steam turbine unit. The current of the fan motor can reflect the running state and cooling effect of the fan, so that current monitoring is also needed.

Other electrical auxiliary machines, such as high pressure oil pumps, drain pumps, vacuum pumps, electric valves, on-site control panels, etc., can also provide certain status information on their supply current.

By reasonably configuring the current sensors in the power supply lines of the electrical equipment and analyzing and modeling the collected current data, the comprehensive monitoring and fault diagnosis of the electrical system of the steam turbine unit can be realized, and a reliable basis is provided for safe and efficient operation. Meanwhile, the current data can also be used for evaluating the energy efficiency level of equipment, and the potential of energy saving optimization is mined.

Voltage sensors are used to monitor the voltage of the motor and electrical system. Voltage anomalies may indicate supply instability, electrical faults or equipment grounding problems, so they are installed on the main power lines of the motor and electrical system, as well as at critical nodes, such as transformers, distribution plates or main control rooms, for monitoring voltage levels and stability.

By installing the sensors at the key positions, the running state of the turbine unit can be comprehensively monitored, and signals which can cause faults can be timely captured. Such comprehensive monitoring is critical for early identification of problems and preventive maintenance.

In the embodiment of the invention, in order to facilitate the analysis of the real-time operation data of the subsequent turbine unit, the acquired real-time operation data needs to be processed, and the processing process comprises data cleaning processing, data synchronization processing, feature standardization processing and value deficiency filling processing.

In particular, the data cleaning process may remove noise and extraneous data during collection, such as outliers resulting from sensor failure, or non-run-time data logging; the data synchronization processing can synchronize the data acquired by different sensors, so that the consistency of the time labels is ensured; the feature normalization process can normalize data in different ranges to a uniform scale, such as converting temperature data and pressure data to between 0 and 1, so as to eliminate the influence of different dimensions; the missing value filling process is to fill the missing data by adopting interpolation, regression or a neighbor-based method.

For example, assume that some data points are missing from the collected vibration data due to a short failure of a certain sensor. The data preprocessing module first fills in these missing data points by linear interpolation. Meanwhile, if some data value is detected to be abnormally high (e.g., beyond the possible maximum vibration range of the apparatus), it is regarded as noise and replaced with an average value of surrounding data. In addition, vibration data is recorded with a millisecond-level time stamp, temperature data can be recorded with a second-level time stamp, and the data preprocessing module can perform time synchronization processing on the data so as to ensure consistency of the data during analysis.

Then, for step 102, the operation state data is input into a pre-trained turbine unit operation fault diagnosis model, and a fault diagnosis result of the turbine unit is output.

In the embodiment of the invention, before the preprocessed running state data is input into the running fault diagnosis model of the turboset, the model is required to be trained, so that the fault diagnosis accuracy of the model is improved, and the training comprises the following steps: determining a selection scheme of the deep learning network based on the fault type and the characteristic property of the turbine unit to be identified; training the selected deep learning network based on a known sample set to obtain an initial model; wherein the known sample set comprises historical fault data and historical normal data of the turbine unit; and carrying out parameter and structure adjustment on the initial model based on a preset independent verification data set to obtain the turbine unit operation fault diagnosis model.

The fault type refers to various abnormal conditions possibly occurring in the steam turbine unit, such as bearing faults, blade faults, rotor imbalance, overspeed, overheat and the like. Each fault has its unique characteristic properties, and is manifested by abnormal patterns in vibration, temperature, pressure, etc. parameters. For example, bearing failure may be characterized by the occurrence of bearing characteristic frequencies and their multiples in the vibration spectrum, temperature rise; blade faults can cause abrupt change of vibration amplitude, and blade frequency and frequency multiplication occur in a frequency spectrum; rotor imbalance is mainly manifested by an abnormal increase in vibration amplitude near the rotor frequency.

The known sample set is a data set for training a fault diagnosis model, and comprises historical operation data of various known fault working conditions of the turbine unit and data of normal working conditions. For example, the historical fault data may be a record of sensors such as vibration, temperature, etc. for a period of time before and after a bearing fault occurs; the normal data may be an operational record of the turbine set during a certain period of health. These data are pre-processed by washing, alignment, labeling, etc. to form a structured sample set.

The independent verification data set is a part of sample data which does not participate in model training, is used for objectively evaluating the performance of the trained initial model, and provides basis for further optimization. For example, 20% may be randomly left out of the historical data as a validation set without participating in training. And testing the model by using a verification set, adjusting the layer number, the neuron number, the activation function, the regularization parameter and the like of the model according to indexes such as errors, accuracy and the like, and retraining on a training set for continuous iteration to obtain the model with optimal generalization performance.

Specifically, the deep learning network is first selected and configured, i.e., the appropriate deep learning network is selected according to the type of fault and the nature of the feature that needs to be identified. For example, convolutional Neural Networks (CNNs) are suitable for processing data having a spatial structure (e.g., images or spectrograms), while Recurrent Neural Networks (RNNs) or long-term memory networks (LSTM) are suitable for processing time-series data. Parameters that need to be configured include, among other things, the number of neural network layers, the number of neurons, and the activation function to accommodate a particular failure feature and data structure. And selecting a proper deep learning model according to the fault type and the data characteristics to be diagnosed. For example, for vibration signals, a one-dimensional convolutional neural network can be selected to extract time-frequency domain features; for multi-parameter time sequence data such as temperature, pressure and the like, a long-short-term memory network (LSTM) can be used for learning the time correlation; for end-to-end fault classification, deep feed forward neural networks (DNNs) may be employed to directly learn the mapping of raw data to fault classes.

Further, the selected deep learning network is trained using the fault cases and the normal operation data in the historical data set. In the training process, the neural network learns how to identify different fault types according to the input advanced features, so that repeated iterative training is performed, and model weights are adjusted through a back propagation algorithm to minimize prediction errors and improve the accuracy of fault classification. Specifically, back propagation is an algorithm for training a neural network, and the model output is made to continuously approximate to a true value by calculating the gradient of the loss function to weights of each layer and updating the weights along the gradient descent direction. For example, assuming that the mean square error is used as the loss function, the current batch of training samples produces a large error at the output layer, the gradient of the error for the neuron weight of the previous layer can be obtained by the chain rule, the gradient of the previous layer can be obtained, and the like until the input layer. And then updating the weight of each layer according to a gradient descent formula at a certain learning rate to reduce the loss function value. This process is repeated until the model converges.

After model training is completed, the performance of the model needs to be tested by using an independent verification data set, the accuracy and reliability of the model in terms of fault classification and positioning are evaluated, and model parameters or structures (such as increasing the layer number and adjusting the learning rate) are adjusted according to verification results so as to improve the performance and accuracy of the model, and thus, an operation fault diagnosis model is obtained.

Specifically, when validating and tuning the initial model, some evaluation metrics are used to measure the performance of the model on the independent validation dataset, such as accuracy, precision, recall, F1 score, ROC (Receiver Operating Characteristic, subject work feature) area under the curve, and the like. If these criteria are not as expected, adjustments to the parameters and structure of the model are required to improve its fault diagnosis capability. The following illustrates how this can be done.

Assuming that the embodiment trains a Convolutional Neural Network (CNN) model for diagnosing the faults of the turbine unit, the input is a time-frequency diagram of vibration signals, and the output is a fault type. The initial model achieves 95% accuracy on the training set, but only 80% on the validation set, and some fault types have lower recall, e.g., bearing faults of only 60%. This indicates that the model may have overfitting and insufficient generalization capability, and tuning is required.

First, the present embodiment can adjust parameters of a model, which mainly includes:

Learning rate: the step size of updating the weight for each iteration is controlled. If the accuracy rate oscillates or is slowly improved, the learning rate can be properly reduced; if the convergence speed is too slow, the learning rate can be appropriately increased.

Batch size: the number of samples per training input. Increasing the batch may increase the computational efficiency, but may affect convergence accuracy; reducing the batch may make the weight update more frequent, speeding up convergence.

Regularization coefficient: the weights of the L1 or L2 regularization terms are used for controlling the complexity of the model and reducing the overfitting. If the overfitting is severe, the regularization coefficient may be increased.

Dropout ratio: the probability of randomly discarding neurons is also a regularization means, which can be adjusted according to the degree of overfitting.

Secondly, the embodiment can also adjust the structure of the model, mainly including:

number and size of convolution layers: the number of convolution layers determines the depth of feature extraction and the size of the convolution kernel determines the receptive field size. If the fitting is not performed, the number of layers can be increased or the convolution kernel can be enlarged; if overfitting occurs, the number of layers can be reduced or the convolution kernel can be scaled down.

Pooling layer type and size: the maximum pooling can extract significant features, the average pooling can retain more information, and the pooling size influences the compression degree of the feature map.

Full connection layer number and width: the full connection layer is used for feature combination and classification decision. Increasing the number of full connection layers and width can increase the model capacity, but can also lead to overfitting.

Activation function: reLU, sigmoid, tanh, etc., can affect the expressive power and convergence properties of the model.

For example, in response to the problem of low bearing failure recall, the present embodiment may attempt to add a convolution layer, expand the convolution kernel size, and extract more texture features associated with bearing failures; reducing the width of the full connection layer and increasing the dropout proportion so as to reduce the risk of overfitting; while the learning rate is appropriately reduced to achieve finer weight adjustment.

After making a series of parameter and structure adjustments, the present embodiment again evaluates the model performance on the validation set. If the indexes such as the accuracy rate, the recall rate and the like are improved, and the diagnosis of each fault class is more balanced, the final fault diagnosis model can be determined; if the boost is not significant, further iterative optimization is possible.

It is worth noting that through the detailed process, the fault classification and positioning module can effectively identify and position various faults in the turbine unit by using a deep learning technology, and provides important technical support for subsequent maintenance decisions.

In the embodiment of the invention, after the operation fault diagnosis model is obtained, a fault diagnosis result can be obtained and a corresponding fault diagnosis report can be generated according to the processed operation state data, and the method specifically comprises the following steps: performing feature extraction processing on the running state data by using a convolutional neural network to obtain advanced features for representing the running mode and abnormal state of the turbine unit; the method comprises the steps of performing identification processing on advanced features by using a cyclic neural network to obtain a pattern identification result of a turbine unit; classifying the advanced features and the pattern recognition result by using an operation fault diagnosis model to obtain a fault state of the turbine unit; wherein, the fault status fault type and fault location; based on the advanced features, the pattern recognition result, and the fault state, a fault diagnosis result is determined.

Specifically, the Convolutional Neural Network (CNN) is used for extracting the characteristics of the data, identifying important characteristics and modes in the data, and is suitable for processing the data with spatial structure characteristics (such as time sequence data, spectrograms and the like); and (3) analyzing the time series data by using a cyclic neural network (RNN), and capturing time dependence and long-term dependence, so as to accurately identify fault characteristics in the vibration data.

Thus, in feature extraction of operating state data, a Convolutional Neural Network (CNN) or other feature extraction algorithm is used to automatically identify key or advanced features from the data that can reveal important patterns and anomalies in the data, such as frequency peaks, periodic variations in time series, and so forth. For time series data (e.g., vibration signals), the CNN may extract important time or frequency features that affect system performance. For example, in processing vibration data, CNN can identify abnormal peaks or patterns of specific frequencies by analyzing the time-series of spectrograms, which may be indicative of bearing damage or unbalanced rotating components.

It is worth noting that features extracted from vibration data may include energy peaks of specific frequencies, which are associated with mechanical faults such as bearing damage; alternatively, in the temperature data, feature extraction may reveal a trend of temperature over time, indicating a change in cooling efficiency, all of which are advanced features.

Further, the data obtained in the feature extraction stage (advanced features obtained in the feature extraction stage) is analyzed using a Recurrent Neural Network (RNN) or a Long Short-Term Memory (LSTM). These networks are capable of processing and analyzing long-term dependencies and periodic variations in time series data to identify behaviors and patterns in the data, i.e., pattern recognition results. For example, using LSTM processing of vibration data, a tendency for the amplitude of vibration to increase gradually due to mechanical fatigue or loss, or periodic vibration patterns that occur under certain operating conditions, can be identified. In the temperature sequence, the LSTM may find periodic temperature fluctuations, indicating periodic failures of the heat dissipating system. For pressure sequences, pattern recognition can find pressure fluctuations due to unstable operation of the valve.

After the advanced features and the pattern recognition result are obtained, the advanced features and the pattern recognition result are input into a trained turbine unit operation fault diagnosis model, so that different faults are classified, and the faults are positioned to specific parts of the turbine unit.

For example, if a turbine unit operational failure diagnostic model identifies a particular frequency anomaly enhancement and a trend in vibration amplitude over time, it may be diagnosed as an early warning that the bearing will reach its useful life. The system can further locate the specific location of the failed bearing in combination with the design and operating parameters of the turbine unit. Based on the abnormal pattern of pressure data, the system can diagnose a pipe leak or valve failure and locate the approximate location of the leak or failure point. If the current or voltage data shows abnormal fluctuations, problems in the motor or electrical system, such as insulation failure or poor contact, may be indicated. Fault classification example: if the operational fault diagnostic model identifies successive high frequency peaks and increases in vibration amplitude in the vibration data, it will be classified as a signal that the bearing is about to fail. Similarly, by analyzing the abnormal patterns in the pressure and temperature data, the model may identify pipe leaks or cooling system faults and locate the faults to a particular system or component.

In the embodiment of the invention, based on the advanced features, the pattern recognition result, the fault type and the positioning information obtained in the process, a detailed fault diagnosis result can be obtained, and a corresponding fault diagnosis report is generated, and the process comprises the following steps: summarizing, analyzing and processing the advanced features, the pattern recognition result and the fault state to obtain an analysis result; converting the analysis result based on the rule engine to obtain diagnosis information of the turbine unit fault; wherein the diagnostic information includes a cause of failure and maintenance measures; and processing the fault state and the diagnosis information based on a natural language generation technology to obtain a fault diagnosis result.

Specifically, all relevant data and analysis results obtained in the process are summarized, including advanced features, pattern recognition results, fault types and fault positions; the technical analysis results are then converted into specific diagnostic information using a rules engine. These rules are formed based on expert knowledge and historical maintenance data to interpret the analysis results and suggest possible failure causes and suggested maintenance measures; the analysis results are then structurally presented in the report using predefined report templates. The report template comprises a fault type, a fault position, fault cause analysis, suggested maintenance measures and the like; finally, in some advanced systems, natural language generation techniques or algorithms may be used to convert the data and analysis results into readable textual descriptions, making the fault diagnosis report easier to understand.

The generation of the corresponding fault diagnosis report is described in more detail below:

The summary analysis processing is to comprehensively evaluate the information about the state of the turbine set obtained from different sources to obtain the judgment of the overall operation condition of the equipment. These information include: advanced features, such as characteristic frequencies in the vibration spectrum and their magnitudes, trends in temperature and pressure, outliers, etc., reflect the health level of the various components of the device; the pattern recognition result, such as the confidence level of various faults given by the fault classification model, shows the most probable problems of the equipment; the fault condition, such as the fault component and severity as determined by the fault localization module, indicates the extent of damage to the device.

In summary analysis, the correlation and importance of the information need to be considered, and the actual state of the device needs to be comprehensively judged. For example, if both vibration characteristics and temperature trends are directed to a bearing failure and failure classification and localization results are consistent, it may be determined that the bearing is indeed problematic.

The analysis results may be a structured data set containing equipment health index, primary failure mode, critical component status, etc., and corresponding confidence levels. These data will be input to the rules engine.

The rule engine is an reasoning mechanism based on expert knowledge and is used for explaining the analysis result and giving fault reasons and maintenance suggestions. It is a series of IF-THEN rules shaped as: "IF (failure mode= = bearing failure) AND (bearing temperature >80 ℃) AND (bearing vibration amplitude >1 mm/s), THEN (failure cause= bearing wear) AND (maintenance measure= replacement of bearing)".

The rules relate analysis results with background knowledge such as failure mechanism, equipment structure, operation condition and the like to form a diagnosis logic chain. The rule engine automatically matches and triggers the corresponding rules to infer the root cause and disposition scheme of the fault.

For example, if the analysis results show that the #2 bearing fails, the temperature and vibration are over-rated, the rules engine deduces that the bearing has worn severely and needs replacement; if the analysis results show blade failure, vibration and efficiency degradation, the rules engine deduces that the blade foreign body impacts cause deformation, need maintenance, etc.

The natural language generation technology is an artificial intelligence application, and can convert structured data and logic relations into smooth and easily understood text descriptions, so that the diagnosis report is more humanized. The basic principle is as follows: according to the predefined grammar template and vocabulary, fault information is filled into proper positions, and the word sequence, the connective words and the like are adjusted to generate sentences and paragraphs conforming to language habits.

For example, diagnostic information "(failure mode= bearing failure) AND (failure cause= bearing wear) AND (maintenance measure= replacement bearing)" may be converted into: the "diagnosis result indicates that the turbine unit #2 bearing fails because of bearing wear caused by long-term operation, and replacement processing is recommended.

The final fault diagnosis result can be a report with complete structure, prominent key points and unobstructed language, which comprises the following steps: device basic information including model number, run time, etc.; fault summary, which includes time of failure, failure mode, severity, etc.; fault analysis, which includes relevant feature changes, fault locations, cause analysis, etc.; maintenance advice including maintenance, replacement, adjustment, etc. action advice to take; additional information, including related charts, data, notes, etc.

For example, in addition to the bearing overheating examples mentioned in the above process, the following faults may occur:

electrical system failure:

If the diagnosis result shows that the electrical system has insulation fault or poor contact, the fault diagnosis report indicates abnormal current or voltage mode, and inspection and insulation test of the electrical circuit are suggested to be carried out so as to avoid equipment shutdown or safety accidents caused by the electrical fault.

Cooling system leakage:

For cooling systems, if a leak or a drop in cooling efficiency is found by diagnostics, reports detailing abnormal patterns in temperature data and pressure data and suggesting inspection of cooling lines, pumps and fittings, as well as leak testing and cleaning or maintenance of the cooling system.

Compressor performance decreases:

When there is a problem with the compressor experiencing reduced performance, the diagnostic report indicates abnormal patterns in the pressure and flow data and suggests checking the compressor's valves, seals and filters, and performing performance tests and adjustments.

By the method, a detailed document is provided for fault diagnosis report, the specific type and position of the fault are clearly pointed out, the reason analysis and the repair and maintenance suggestions are provided, and important decision support information is provided for maintenance teams.

It should be noted that, in addition to the above steps of generating the fault diagnosis result, the actual maintenance result and the operation feedback may be used for the optimization model, and the accuracy and efficiency of the fault diagnosis may be continuously improved through the machine learning algorithm. That is, based on the repair team feedback, if the actual fault is found to be not exactly consistent with the system diagnostics, the data will be used to adjust and train the model, thereby improving future diagnostic accuracy. Through the step, the deep learning analysis module can realize accurate diagnosis and effective positioning of the vibration faults of the turbine unit, and provides powerful decision support for maintenance teams

Finally, for step 104, an optimal treatment scheme for the turbine unit fault is determined based on the fault diagnosis result.

In the embodiment of the invention, the optimal fault treatment scheme is determined by the following steps: matching the fault diagnosis result with a preset database to obtain fault background information; the database comprises a history maintenance record, a fault database and an equipment technical manual; performing evaluation processing on fault background information based on a preset evaluation algorithm to obtain a fault evaluation result of the turbine unit; the evaluation algorithm comprises a risk scoring algorithm, a rule inference algorithm and a historical data analysis algorithm; calculating a fault evaluation result based on a preset decision algorithm to obtain an optimal treatment scheme of the turbine unit fault; the decision algorithm includes a decision tree algorithm, a priority algorithm and a cost-benefit algorithm.

Specifically, after obtaining the fault diagnosis result given by the operation fault diagnosis model, the information can be compared and integrated with the historical maintenance record, the fault database and the equipment technical manual to obtain comprehensive fault background information. For example, if the deep learning analysis indicates that a particular bearing is faulty, the fault diagnostic module may query the historical maintenance record and manufacturer specifications for that bearing to determine if a repeated fault or a particular fault pattern exists.

The comparison and integration processes are as follows: firstly, receiving fault diagnosis results generated by an operation fault diagnosis model, wherein the results comprise detailed fault type and fault position information, and the real-time performance and accuracy are required to be ensured; then verifying the received fault data, checking the integrity and consistency of the fault data, removing any abnormal or irrelevant data, and ensuring the accuracy and reliability of subsequent analysis; the historical fault records matching the received fault type and location are then queried through an interface with a fault database. This process involves query and data matching algorithms to ensure that the most relevant information is found, in addition to fault records, historical maintenance data for the device can also be retrieved. This includes previous repair reports, records of replacement parts, and any associated maintenance activities to gain insight into failure modes and maintenance strategies; in addition, faults can be further analyzed according to technical manuals and operation specifications of the equipment, and the documents provide equipment structure, function descriptions, typical fault cases and maintenance guidelines, so that problems can be accurately positioned and solutions can be recommended; finally, comprehensively analyzing all collected information (fault diagnosis results, historical maintenance records, fault database information and technical manual guidance); and comparing and contrasting to establish a complete scene of fault occurrence, and identifying fault reasons and influence factors.

For example, taking the cooling system as an example, assume that the operational failure detection model identifies an abnormal rise in temperature of the cooling system, which may be due to a failure of the cooling pump or leakage of cooling fluid, at which point the operational failure detection model locates the failure to the cooling system; then, the historical maintenance record of the cooling system is searched, and if the leakage point of the supercooling pump which is replaced before or the supercooling system is repaired is found, the information is taken as an important reference of the diagnosis process; simultaneously comparing the current anomalies with similar cases stored in the fault database, which would further support fault diagnosis if similar temperature anomalies were present in the database and associated with cooling pump faults or coolant leaks; it is furthermore necessary to access technical manuals provided by the equipment manufacturer to determine the temperature range that a properly operating cooling system should have, the operating specifications of the cooling pump and the standard flow rate of the cooling fluid. By comparing the actual monitoring data to the specifications, the nature and severity of the fault is facilitated to be determined.

By the method, real-time monitoring data, historical maintenance records, similar fault cases and technical specifications can be comprehensively considered, comprehensive fault background information is provided for technical teams, and therefore more accurate diagnosis and repair strategies are formulated. The integration method is not limited to the cooling system, but can be applied to fault diagnosis of various equipment and systems such as an electric system, a lubrication system, a control system and the like.

In the embodiment of the invention, after the fault background information is obtained, the potential influence of the current fault on production and safety needs to be automatically evaluated, and the severity and urgency of the fault and the possible consequences if actions are not immediately taken are considered. For example, the impact of bearing failure on the operational stability of the entire turbine unit is evaluated based on a preset algorithm, and if the failure may result in damage to the main shaft or unexpected shutdown, the failure is marked as a high priority process.

Specifically, the evaluation of the potential impact of the fault is achieved by the following algorithm:

(1) A risk scoring algorithm that calculates a risk score based on fault severity, fault frequency, equipment criticality and possible consequences (e.g., downtime, safety incidents, maintenance costs), as shown by the following equation:

M＝(W_s×S)+(W_f×F)+(W_c×C)+(W_o×O)

wherein M is a risk score; s is a fault severity score, graded according to the potential impact of the fault, such as a range of 1 to 10, 10 representing extremely severe; f is a fault frequency score, graded according to the frequency of occurrence of the fault, such as a range of 1 to 10, 10 representing very frequent; c is a device criticality score, graded according to the importance of the device, such as a range of 1 to 10, 10 representing extremely critical; o is a score of possible consequences, comprehensively considering factors such as downtime, safety accidents, maintenance cost and the like, and is also scored by 1 to 10, wherein 10 indicates that the consequences are extremely serious; w _s、W_f、W_c、W_o is in turn the weight of the corresponding factors, reflecting the relative importance of each factor in the overall risk score, the choice of weight depending on the specific needs of the organization and the risk management policy, the total weight being typically 1.

Through the formula, each fault can be quantitatively evaluated to obtain a comprehensive risk score so as to guide maintenance and operation decisions; this scoring helps determine which faults need to be prioritized and which devices need more frequent monitoring or maintenance.

(2) And a rule inference algorithm, which is combined with the expert system and the decision tree, and automatically infers the influence level of the fault according to a specific rule. For example, if a failure of a certain critical device may result in a full line outage, the failure is immediately marked as high priority. For example, in a turbine set, if a failure of a certain sensor may cause the entire turbine set to stop, the failure may be marked as high priority and require immediate maintenance.

(3) Historical data analysis algorithms, which analyze historical fault data and repair records using machine learning algorithms, such as random forests or neural networks, predict the specific impact of faults on production. This approach can discover potential associations between faults and production effects, improving the accuracy of the assessment.

For example, an electrical system short circuit fault:

If the electrical system is shorted, the algorithm evaluates the effect of the short on the power supplied to the production equipment and the possible resulting safety risks. And automatically calculating the priority and urgency of the fault according to the importance of the equipment in the production line and the occurrence frequency of the short circuit.

Cooling system leakage:

For leakage of the cooling system, the system will take into account the leakage rate and the remaining amount of cooling liquid, evaluating the risk of overheating the machine and possible downtime. The priority of fault handling is automatically determined based on the extent to which leakage affects the performance of the device and the urgency of repair.

Control system failure:

When a control system fails, such as inaccurate sensor readings or control logic failure, the system may analyze operational errors or process control problems that such a failure may cause. The nature of the fault, the role of the control system in the overall operation, and historical fault data are used to evaluate the impact that may have on production quality and efficiency.

The algorithm and the automation tool can objectively and accurately evaluate the influence degree of each fault, and accordingly establish the priority of maintenance and treatment, so that resources are effectively distributed to the most critical problems.

In the embodiment of the invention, after the fault evaluation result is obtained by using an evaluation algorithm, suggestions for maintenance or adjustment are formulated according to the fault type, the influence on the evaluation and the criticality of the operation of the equipment, and the suggestions comprise suggested maintenance measures, replacement parts, adjustment of operation parameters or further detailed inspection. For example, for an identified bearing failure, replacement of the bearing or lubrication may be recommended, and if the failure is not severe, only increased monitoring frequency and maintenance checks may be recommended.

Specifically, the optimal fault handling scheme is determined by the following decision algorithm:

(1) Decision tree or expert system:

logic for systematically processing service recommendations using decision trees. Nodes represent decision points (e.g., severity of failure, criticality of equipment), edges represent decision results (e.g., emergency maintenance, planned maintenance, status monitoring, etc.).

Expert systems utilize rule sets that are based on long-term accumulated experience and knowledge, such as: "if the fault is high priority and affects critical equipment, immediate repair is recommended.

(2) Priority algorithm:

And calculating maintenance priority according to the emergency degree of the fault and the importance of the equipment. Faults with high urgency and high criticality will get higher priority.

The priority may be quantified by a scoring system that sums the weights of different factors, such as the weight of the severity of the fault, the weight of the criticality of the equipment, and the weight of the impact of the fault on production.

(3) Cost-benefit analysis:

The relationship between the cost of the maintenance action and the loss that may be incurred by deferring the maintenance is analyzed. The least costly and most cost effective maintenance scheme is selected.

A mathematical model, such as linear programming or multi-objective optimization, may be used to determine an optimal repair strategy.

For example, an electrical control system fault:

If a software problem with the electrical control system is diagnosed, it may be advisable to reprogram or update the software. If hardware fails, such as a control board is damaged, replacement of the component is recommended. Decision trees make recommendations based on the nature of the problem (software or hardware), the urgency, and the role of the control system in the production process.

Cooling system leakage:

Based on the severity of the leak and the impact on production, if the leak is small and does not affect production, periodic monitoring and repair in the next scheduled maintenance may be recommended. If the leak is severe and there is a risk of immediate failure, the system will recommend immediate action to repair the leak.

Compressor vibration problem:

Vibration problems for compressors are analyzed for severity and frequency of vibration. If the vibration is slight and the analysis does not consider immediate impact on the performance of the device, it may be advisable to increase the vibration monitoring frequency and check it the next time a shutdown is planned. If the vibration is severe, possibly resulting in equipment damage, immediate shut down and maintenance is recommended.

Through the algorithmic decision process, maintenance or adjustment suggestions can be ensured to be based on objective data and systematic analysis results, so that artificial subjective judgment is reduced, and maintenance decision efficiency and accuracy are improved.

It is worth to be noted that, on the basis of the above decision process, the embodiment of the present invention can also automatically make a maintenance plan, including the required personnel, tools and parts; at the same time, the most appropriate maintenance time and personnel are arranged in view of production planning and resource availability. For example, the system may schedule bearing replacement work based on maintenance team's work schedule and component inventory, and ensure that all necessary tools and materials are ready.

Further, the automation provides detailed fault diagnosis reports and maintenance suggestions for technical teams and management layers, supporting the decision-making process. In the maintenance process, the execution condition is supervised, and the scheduled execution is ensured. For example, the generated repair advice reports will be submitted to maintenance managers and engineers for approval and the task status and progress updated in real-time during the repair process to ensure that the problem is effectively resolved.

Further, after the repair is completed, feedback information and repair results are collected for evaluating the repair effect and optimizing future fault diagnosis and decision processes. For example, feedback provided by a maintenance team may be used to assess the accuracy of fault diagnosis and the effectiveness of maintenance measures, which information may be used to continuously optimize the diagnostic algorithms and decision logic of the model.

In summary, the above process ensures efficient and stable operation of the turbine assembly while reducing unexpected downtime and maintenance costs.

Referring to fig. 2, an embodiment of the present invention provides a turbine unit fault handling apparatus based on deep learning, which includes:

an obtaining unit 200, configured to obtain operation state data of the pretreated turbine unit; wherein the operating state data includes acceleration, temperature, pressure, voltage, and current;

The diagnosis unit 202 is configured to input the operation state data into a pre-trained operation fault diagnosis model, and output a fault diagnosis result of the turbine unit;

and the decision unit 204 is used for determining an optimal treatment scheme of the turbine unit fault based on the fault diagnosis result.

In the embodiment of the invention, the preprocessing comprises data cleaning processing, data synchronization processing, feature normalization processing and default filling processing.

In the embodiment of the invention, the operation fault diagnosis model is specifically obtained by training in the following manner: determining a selection scheme of the deep learning network based on the fault type and the characteristic property of the turbine unit to be identified; training the selected deep learning network based on a known sample set to obtain an initial model; wherein the known sample set includes historical fault data and historical normal data; and carrying out parameter and structure adjustment on the initial model based on a preset independent verification data set to obtain the operation fault diagnosis model.

In the embodiment of the present invention, when the diagnosis unit 202 performs the input of the operation state data into the pre-trained operation fault diagnosis model and outputs the fault diagnosis result of the turbine unit, the diagnosis unit is specifically configured to perform the following operations: performing feature extraction processing on the running state data by using a convolutional neural network to obtain advanced features for representing the running mode and abnormal state of the turbine unit; the method comprises the steps of performing identification processing on advanced features by using a cyclic neural network to obtain a pattern identification result of a turbine unit; classifying the advanced features and the pattern recognition result by using an operation fault diagnosis model to obtain a fault state of the turbine unit; wherein, the fault status fault type and fault location; based on the advanced features, the pattern recognition result, and the fault state, a fault diagnosis result is determined.

In the embodiment of the invention, determining the fault diagnosis result based on the advanced features, the pattern recognition result and the fault state comprises the following steps: summarizing, analyzing and processing the advanced features, the pattern recognition result and the fault state to obtain an analysis result; converting the analysis result based on the rule engine to obtain diagnosis information of the turbine unit fault; wherein the diagnostic information includes a cause of failure and maintenance measures; and processing the fault state and the diagnosis information based on a natural language generation technology to obtain a fault diagnosis result.

In the embodiment of the present invention, when executing the optimal processing scheme for determining the turbine unit fault based on the fault diagnosis result, the decision unit 204 is specifically configured to execute the following operations: matching the fault diagnosis result with a preset database to obtain fault background information; the database comprises a history maintenance record, a fault database and an equipment technical manual; performing evaluation processing on fault background information based on a preset evaluation algorithm to obtain a fault evaluation result of the turbine unit; the evaluation algorithm comprises a risk scoring algorithm, a rule inference algorithm and a historical data analysis algorithm; calculating a fault evaluation result based on a preset decision algorithm to obtain an optimal treatment scheme of the turbine unit fault; the decision algorithm includes a decision tree algorithm, a priority algorithm and a cost-benefit algorithm.

It should be noted that: the deep learning-based turboset fault processing device provided in the above embodiment is only exemplified by the division of the above functional modules, and in practical application, the above functional allocation may be performed by different functional modules according to needs, i.e., the internal structure of the device is divided into different functional modules, so as to perform all or part of the functions described above. In addition, the deep learning-based turboset fault processing device and the deep learning-based turboset fault processing method provided in the foregoing embodiments belong to the same concept, and specific implementation processes thereof are detailed in the method embodiments and are not described herein.

The embodiment of the present application further provides a computer device, please refer to fig. 3, where the computer device includes a processor and a memory, where at least one instruction, at least one section of program, a code set or an instruction set is stored in the memory, and the at least one instruction, the at least one section of program, the code set or the instruction set is loaded and executed by the processor to implement a deep learning-based turbine set fault handling method provided in the foregoing method embodiments.

The embodiment of the application also provides a computer readable storage medium, wherein at least one instruction, at least one section of program, code set or instruction set is stored on the computer readable storage medium, and the at least one instruction, the at least one section of program, the code set or the instruction set is loaded and executed by a processor, so as to realize the deep learning-based turbine unit fault processing method provided by the embodiment of the method.

Embodiments of the present application also provide a computer program product comprising a computer program, which is read from a computer-readable storage medium by a processor of a computer device, the computer program being executed by the processor such that the computer device performs a deep learning based turbine set fault handling method as described in any of the above embodiments.

For convenience of description, the above system or apparatus is described as being functionally divided into various modules or units, respectively. Of course, the functions of each element may be implemented in the same piece or pieces of software and/or hardware when implementing the present application.

From the above description of embodiments, it will be apparent to those skilled in the art that the present application may be implemented in software plus a necessary general hardware platform. Based on such understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a storage medium, such as a ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the embodiments or some parts of the embodiments of the present application.

Finally, it is further noted that relational terms such as first, second, third, fourth, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The foregoing is merely a preferred embodiment of the present application and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present application, which are intended to be comprehended within the scope of the present application.

Claims (10)

1. The turbine unit fault processing method based on deep learning is characterized by comprising the following steps of:

acquiring operation state data of the pretreated turbine unit; wherein the operating state data includes vibration, temperature, pressure, voltage, and current;

Inputting the running state data into a pre-trained turbine unit running fault diagnosis model, and outputting a fault diagnosis result of the turbine unit;

And determining an optimal treatment scheme of the turbine unit fault based on the fault diagnosis result.

2. The method of claim 1, wherein the preprocessing comprises a data cleansing process, a data synchronization process, a feature normalization process, and a default population process; the data cleaning process is used for removing noise and irrelevant data in the real-time data of the turboset, the data synchronization process is used for ensuring consistency of time labels of the real-time data of the turboset, the characteristic normalization process is used for eliminating influences of different dimensions in the real-time data of the turboset, and the default filling process is used for filling missing data in the real-time data of the turboset.

3. The method according to claim 1, characterized in that the turbine unit operational failure diagnosis model is specifically trained by:

determining a selection scheme of the deep learning network based on the fault type and the characteristic property of the turbine unit to be identified;

training the selected deep learning network based on a known sample set to obtain an initial model; wherein the known sample set comprises historical fault data and historical normal data of the turbine unit;

and carrying out parameter and structure adjustment on the initial model based on a preset independent verification data set to obtain the turbine unit operation fault diagnosis model.

4. The method according to claim 1, wherein the fault diagnosis result of the turbine unit is specifically obtained by:

Performing feature extraction processing on the running state data by using a convolutional neural network to obtain advanced features for representing the running mode and abnormal state of the turbine unit;

Performing recognition processing on the advanced features by using a cyclic neural network to obtain a pattern recognition result of the steam turbine unit;

Classifying the advanced features and the pattern recognition result by using the turbine unit operation fault diagnosis model to obtain a fault state of the turbine unit; wherein the fault condition includes a fault type and a fault location;

And determining a fault diagnosis result of the turbine unit based on the advanced features, the pattern recognition result and the fault state.

5. The method of claim 4, wherein the determining a fault diagnosis of the turbine group based on the advanced features, the pattern recognition result, and the fault condition comprises:

summarizing, analyzing and processing the advanced features, the pattern recognition result and the fault state to obtain an analysis result;

Converting the analysis result based on a rule engine to obtain diagnosis information of the turbine unit fault; wherein the diagnostic information includes a cause of failure and maintenance measures;

and processing the fault state and the diagnosis information based on a natural language generation technology to obtain the fault diagnosis result.

6. The method of claim 5, wherein determining an optimal treatment scheme for the turbine unit fault based on the fault diagnosis results comprises:

Matching the fault diagnosis result with a preset database to obtain fault background information; the database comprises a history maintenance record, a fault database and an equipment technical manual;

performing evaluation processing on the fault background information based on a preset evaluation algorithm to obtain a fault evaluation result of the turbine unit; wherein the evaluation algorithm comprises a risk scoring algorithm, a rule inference algorithm and a historical data analysis algorithm;

Calculating the fault evaluation result based on a preset decision algorithm to obtain an optimal treatment scheme of the turbine unit fault; wherein the decision algorithm comprises one or more of a decision tree algorithm, a priority algorithm, and a cost-benefit algorithm.

7. A deep learning-based turbine unit fault handling system, the system comprising:

the acquisition unit is used for acquiring the operation state data of the pretreated turbine unit; wherein the operating state data includes vibration, temperature, pressure, voltage, and current;

The diagnosis unit is used for inputting the running state data into a pre-trained running fault diagnosis model and outputting a fault diagnosis result of the steam turbine unit;

And the decision unit is used for determining an optimal processing scheme of the turbine unit fault based on the fault diagnosis result.

8. A computer device, characterized in that it comprises a memory for storing a computer program and a processor for executing the computer program stored on the memory for carrying out the steps of the method according to any of the preceding claims 1-6.

9. A computer-readable storage medium, characterized in that the storage medium has stored therein a computer program which, when executed by a processor, implements the steps of the method of any of claims 1-6.

10. A computer program product comprising a computer program which, when executed by a processor, implements the steps of the method of any of claims 1-6.