Abnormal mode detection method and system for generator sets based on deep learning

Abstract

The present application proposes a method and system for detecting abnormal patterns of generator sets based on deep learning. The method includes: obtaining historical operating data of the generator set, and performing steady-state discrimination and operating condition division on the historical operating data after removing abnormal data; determining the reference value of threshold-type signs under each operating condition through a Gaussian mixture model, and training a bidirectional long short-term memory network to predict data trends; comparing real-time data with the reference value of threshold-type signs under corresponding operating conditions, and extracting threshold-type signs; obtaining target data in an adjacent preset time period before the real-time data, predicting the data trend of the target data through the trained bidirectional long short-term memory network, determining the degree of deviation between the real-time data and the data trend, and extracting trend-type signs; determining abnormal patterns according to the extracted signs. The method can accurately identify data in the abnormal stage from a large amount of operating data, and improve the accuracy of abnormal pattern detection.

Description

Technical Field

[0001] The present application relates to the technical field of generator maintenance, and in particular to a method and system for detecting abnormal modes of generator sets based on deep learning. Background Art

[0002] At present, timely elimination of faults caused by generator sets and ensuring the safe operation of generator sets have attracted more and more attention. The identification of abnormal operation modes of generator sets is a necessary condition before fault diagnosis and reasoning. According to the abnormal data during the operation of the generator set, the specific abnormal mode can be identified, which can be used as the entry point for fault diagnosis and reasoning of the generator set. During the operation of the generator set, the operation data of the generator set becomes abnormal due to the occurrence of faults. How to correctly identify the abnormal mode is an important basis for subsequent fault diagnosis and reasoning.

[0003] The abnormality detection of generator set operation data is to measure whether the operating status of the generator set is abnormal under the current working conditions. It is the premise of carrying out health maintenance work. At the same time, the operating data of the components in a healthy operating state is the basis for the prediction of the power generation of the generator set. By monitoring and analyzing the real-time operating data of the generator set, the difference between the current data and the data during healthy and normal operation is compared to identify abnormal patterns. Therefore, the focus of abnormal pattern identification of generator set operation status is to establish a data baseline model under a healthy operating state.

[0004] In the related art, the benchmark model is usually established by traditional generator set performance tests and mathematical models. However, due to the mutual coupling between the operating parameters of the generator set and the influence of external boundary conditions, it presents high-dimensional, dynamic and nonlinear characteristics. In addition, as the generator set equipment ages and the external conditions change, the actual benchmark state of the component often deviates greatly from the design value. Therefore, the methods in the above-mentioned related art cannot accurately perform benchmark modeling, resulting in large errors in abnormal mode detection. Summary of the invention

[0005] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0006] To this end, the first purpose of this application is to propose a method for detecting abnormal patterns of generator sets based on deep learning. This method uses deep learning models to mine data information, solves the multivariable coupling relationship, and performs accurate benchmark modeling under different operating boundary conditions of the generator set. A data cleaning process is proposed, and the cleaned data is used for model training to ensure that the generator set operating data used for model training is in a healthy state. The accuracy of establishing the benchmark model is improved, and then the accuracy and applicability of abnormal pattern detection are improved, which is conducive to timely and accurate diagnosis of unit faults.

[0007] The second object of the present application is to propose an abnormal mode detection system for a generator set based on deep learning;

[0008] A third object of the present application is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first embodiment of the present application is to propose a method for detecting abnormal modes of a generator set based on deep learning, the method comprising the following steps:

[0010] Obtain the historical operating data of the generator set, perform steady-state judgment on the historical operating data after removing abnormal data, and divide the steady-state historical operating data into working conditions;

[0011] The Gaussian mixture model is used to determine the baseline value of the threshold-type signs under each working condition, and the bidirectional long short-term memory network is trained with the historical operating data after the working condition division as a prediction model for data trends;

[0012] Acquire real-time data of the generator set, compare the real-time data with a reference value of a threshold-type symptom under a corresponding working condition, and extract the threshold-type symptom;

[0013] Acquire target data in an adjacent preset time period before the real-time data, predict the data trend of the target data through a trained bidirectional long short-term memory network, determine the degree of deviation between the real-time data and the data trend, and extract trend signs;

[0014] An abnormal pattern is determined based on the extracted threshold-type symptoms and trend-type symptoms.

[0015] In addition, the abnormal mode detection method of a generator set based on deep learning proposed in the above embodiment of the present application may also have the following additional technical features:

[0016] Optionally, in one embodiment of the present application, after determining the abnormal pattern based on the extracted threshold-type signs and the trend-type signs, it also includes: obtaining a supplementary abnormal pattern obtained through manual detection or sensor detection; inputting the threshold-type abnormal pattern, the trend-type abnormal pattern and the supplementary abnormal pattern into the diagnostic knowledge base of the generator set for reasoning, and obtaining the returned fault mode, fault cause and fault maintenance measures.

[0017] Optionally, in one embodiment of the present application, a bidirectional long short-term memory network includes an input layer, a hidden layer and an output layer, the hidden layer includes a loop structure, and the trained bidirectional long short-term memory network predicts the data trend of the target data, including: inputting multiple feature parameters in the target data into the bidirectional long short-term memory network, and mining the connection between each feature parameter in the adjacent preset time periods and the current time period through the hidden layer.

[0018] Optionally, in one embodiment of the present application, determining the degree of deviation between the real-time data and the data trend includes: when the real-time data and the data trend begin to differ, judging the degree of deviation based on a confidence interval, wherein the expression of the confidence interval is as follows:

[0019] p{c1 <m<c2}=1-α

[0020] Among them, p represents probability, α represents significance level, c1 and c2 represent two statistics under significance level, and m represents the difference between real-time data and data trend.

[0021] Optionally, in one embodiment of the present application, a steady-state judgment is performed on the historical operating data after the abnormal data is eliminated, including: selecting the power generation of the generator set as a discriminant feature for steady-state judgment; calculating the difference in power generation between two adjacent moments; estimating the rate of change of the power generation by the mean of the sample statistic in the time window; based on a confidence interval estimation method, estimating whether the confidence interval of the rate of change of the power generation includes zero, and if so, determining that the generator set is in a steady-state operating condition during the time period corresponding to the time window.

[0022] Optionally, in one embodiment of the present application, the steady-state historical operation data is divided into operating conditions, including: manually determining multiple boundary conditions, and determining the size of the characteristic interval corresponding to each boundary condition according to the characteristics of the boundary conditions; combining each characteristic interval, and performing equal-interval operating condition division according to the following formula:

[0023] C i =(A i ,A i +ΔA)∩(B i ,B i +ΔB)...(D i ,D i +ΔD)

[0024] sA i <<A min ,A i +ΔA>>A max ,B i <<B min,B i +ΔB>>B max ,...D i <<D min ,D i +ΔD>>D max

[0025] Among them, A, B and D represent different boundary conditions, C i Represents the Cth interval determined by the intersection of multiple feature intervals i working condition, ΔA, ΔB and ΔD represent the working condition division intervals of different boundary conditions respectively.

[0026] To achieve the above objectives, the second aspect of the present application also proposes an abnormal mode detection system for a generator set based on deep learning, comprising the following modules:

[0027] A division module is used to obtain the historical operation data of the generator set, and to perform steady-state judgment on the historical operation data after removing abnormal data, and to divide the steady-state historical operation data into working conditions;

[0028] A training module is used to construct a Gaussian mixture model, determine the reference value of the threshold-type sign under each working condition through the Gaussian mixture model, and train a bidirectional long short-term memory network as a prediction model for data trends through the historical operation data after the working condition division;

[0029] A first extraction module is used to obtain real-time data of the generator set, compare the real-time data with a reference value of a threshold-type symptom under a corresponding working condition, and extract the threshold-type symptom;

[0030] The second extraction module is used to obtain target data in an adjacent preset time period before the real-time data, predict the data trend of the target data through the trained bidirectional long short-term memory network, determine the degree of deviation between the real-time data and the data trend, and extract trend signs;

[0031] A determination module is used to determine an abnormal pattern according to the extracted threshold-type symptoms and the trend-type symptoms.

[0032] Optionally, in one embodiment of the present application, the system also includes: a first acquisition module, used to obtain a supplementary abnormality pattern obtained through manual detection or sensor detection; a second acquisition module, used to input the threshold-type abnormality pattern, the trend-type abnormality pattern and the supplementary abnormality pattern into the diagnostic knowledge base of the generator set for reasoning, and obtain the returned fault mode, fault cause and fault maintenance measures.

[0033] Optionally, in one embodiment of the present application, the bidirectional long short-term memory network includes an input layer, a hidden layer and an output layer, the output layer includes a loop structure, and the second extraction module is specifically used to: input multiple feature parameters in the target data into the bidirectional long short-term memory network, and mine the connection between each feature parameter in the adjacent preset time periods and the current time period through the hidden layer.

[0034] In order to implement the above-mentioned embodiments, the third aspect embodiment of the present application also proposes a non-temporary computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the abnormal mode detection method of the generator set based on deep learning in the above-mentioned embodiment is implemented.

[0035] The technical solution provided by the embodiment of the present application brings at least the following beneficial effects: the present application uses a deep learning model to mine the complex mapping rules between the normal operating state and parameters of the generator set, solves the problem that the multivariable coupling relationship affects the benchmark modeling, and can perform accurate benchmark modeling under different operating boundary conditions of the generator set. The abnormal pattern recognition method of the threshold type abnormal pattern recognition method and the trend type abnormal pattern recognition method are used to identify the abnormal pattern of the real-time operating data of the generator set, which is applicable to different working conditions when the generator set is actually running. In addition, a data cleaning process is proposed, and the cleaned data is used for model training to ensure that the generator set operation data used for model training is in a healthy state of the generator set. As a result, the accuracy and pertinence of establishing the benchmark model are improved, and then the accuracy and applicability of abnormal pattern detection are improved, which is conducive to timely and accurate diagnosis of unit faults. Through early identification of fault signs, timely adjustment and maintenance are carried out to avoid abnormal parameters from developing backward to generate faults, achieve early fault warning, and improve the safety of the generator set.

[0036] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0038] Figure 1 A flowchart of a method for detecting abnormal modes of a generator set based on deep learning proposed in an embodiment of the present application;

[0039] Figure 2 A flow chart of a method for determining a steady-state operating condition of a generator set proposed in an embodiment of the present application;

[0040] Figure 3 A schematic diagram of a loop structure proposed in an embodiment of the present application;

[0041] Figure 4 A flowchart of a specific method for detecting abnormal modes of a generator set based on deep learning proposed in an embodiment of the present application;

[0042] Figure 5 A schematic diagram of the structure of an abnormal mode detection system for a generator set based on deep learning proposed in an embodiment of the present application. DETAILED DESCRIPTION

[0043] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0044] A method and system for detecting abnormal modes of a generator set based on deep learning according to an embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Figure 1 A flowchart of a method for detecting abnormal modes of a generator set based on deep learning proposed in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method comprises the following steps:

[0046] Step S101, obtaining historical operating data of the generator set, and performing steady-state determination on the historical operating data after removing abnormal data, and dividing the steady-state historical operating data into operating conditions.

[0047] Among them, the generator set targeted by the present application can be a generator set in a new energy system such as a wind turbine, or it can be other types of generator sets such as a gas turbine.

[0048] Among them, abnormal data refers to abnormal points in the measured data that do not conform to the normal output characteristics of the generator set due to aging of the data acquisition device, poor contact of electronic components, and environmental factors in actual applications. Steady-state discrimination is to distinguish whether the unit is in a steady-state condition or an unsteady-state condition. During the actual operation of the generator set, due to the influence of unexpected factors such as weather factors, there may be scenes where unsteady-state conditions and steady-state conditions alternate during the operation of the generator set, so it is necessary to distinguish between steady-state and unsteady-state conditions.

[0049] Specifically, after obtaining the historical operating data of the generator set through different methods, such as reading the historical data stored in the system database, the abnormal data is first eliminated. As an example, historical experience and expert knowledge can be combined in advance, and big data analysis technology can be used to perform statistical analysis on massive amounts of generator set data, and a data threshold that meets the normal output characteristics of the generator set can be set. When the operating data is within the range of the threshold limit, it means that the data meets the basic requirements of the normal output characteristics of the generator set. Then, the historical operating data of the generator set obtained is screened through the threshold range, and abnormal data outside the threshold range is eliminated.

[0050] In one embodiment of the present application, after eliminating abnormal data, the screened blank values ​​can be filled by interpolation or averaging, so as to ensure the integrity of the acquired unit historical data at different time stages as much as possible while ensuring the accuracy of the data.

[0051] Furthermore, a steady-state judgment is performed on the historical operating data after the abnormal data is eliminated. Since a strong consistency cannot be guaranteed between the input and output parameters of the generator set under non-steady-state conditions, in order to establish an accurate benchmark model, this application first divides the operating conditions into steady-state and non-steady-state conditions, and eliminates the non-steady-state data generated under non-steady-state conditions.

[0052] In one embodiment of the present application, in order to more clearly illustrate the specific implementation process of the steady-state determination of the present application, the following is an exemplary description of a method for determining the steady-state condition of a generator set proposed in an embodiment of the present application. Figure 2 As shown, the method comprises the following steps:

[0053] Step S201, selecting the power generation of the generator set as a discriminant feature for steady-state discrimination.

[0054] Among them, the generator set is in a steady-state process, and the characteristic parameters remain unchanged or change slightly over a period of time. As a characteristic parameter, the power generation has the characteristics of small fluctuations in the steady-state process and large fluctuations in the unsteady-state process. Therefore, selecting the power generation as the discriminant feature can well reflect the steady-state and unsteady-state processes.

[0055] In specific implementation, the output power of the wind turbine generator set is first expressed as an expression including the rate of change of the generated power as shown in the following formula:

[0056] p t =p0+mt+δ

[0057] Among them, p tand p0 represent the measured value and true value of the generated power at time t respectively, t is any positive integer, m represents the rate of change of the generated power, δ represents the random error of the generated power, and δ obeys the normal distribution.

[0058] Step S202, calculating the difference between the power generated at two adjacent moments.

[0059] It should be noted that the difference between steady-state and non-steady-state processes lies in whether the power generation response rate is 0, so it is necessary to estimate the value of m.

[0060] Specifically, the value of m is estimated by using the difference in power generation at two adjacent moments, that is, subtracting the above expression of the rate of change at two adjacent moments to obtain the difference. The calculation process is shown in the following formula:

[0061] Δp=p t -p t-1 =m+(δ t -δ t-1 )

[0062] Among them, p t-1 It represents the measured value of the generated power at time t-1. The meanings of the other parameters can be referred to the explanations in the above example.

[0063] Step S203, estimating the rate of change of the generated power by using the mean of the sample statistics within the time window.

[0064] Specifically, the expected difference between the power generation at the above two moments is m, and due to ε t ~N(0,σ 2 ), that is, Δp~N(m,2σ 2 ), according to the nature of the time series, m can be estimated by the mean of the sample statistics in the time window. The specific calculation formula is as follows:

[0065]

[0066] Where h represents the number of samples in the sliding time window.

[0067] Step S204, based on the confidence interval estimation method, estimate whether the confidence interval of the rate of change of the generated power includes zero, and if so, determine that the generator set is in a steady-state condition during the period corresponding to the time window.

[0068] Specifically, to ensure the reliability of the estimate, this application adopts the confidence interval estimation method, and the specific formula is as follows:

[0069] p{m1 <m<m2}=1-α

[0070] Among them, m1 and m2 are two statistics under a given confidence level α. When there is a confidence level of 1-α, it is considered that the true value of m is within the confidence interval (m1, m2). Therefore, whether the target wind turbine is in a steady-state condition within the time window t-1 to t can be judged based on whether the confidence interval includes zero. In this embodiment, if the interval (m1, m2) does not include 0, it is determined that it is in a non-steady-state condition within the time period.

[0071] Therefore, the method for determining steady-state operating conditions in the embodiment of the present application can accurately determine whether the generator set is in a steady-state or non-steady-state operating condition within a certain time period, and further, when determining that the generator set is in a non-steady-state operating condition within the time period, the non-steady-state data generated under the non-steady-state operating condition can be eliminated to ensure the consistency between the input and output parameters of the generator set in the acquired historical data.

[0072] This application removes invalid and non-steady data mixed in the acquired historical operation data, realizes data cleaning, and uses the cleaned data for model training, thereby ensuring the accuracy of the benchmark model determined by subsequent deep learning based on the data.

[0073] Furthermore, after the steady-state and non-steady-state discrimination is performed on the generator set, the steady-state data is divided into working conditions. The working condition division in this application refers to the division of the working conditions of the generator set under different boundary conditions such as workload, ambient temperature and wind speed. The division of working conditions uses specific multiple boundary conditions. The boundary conditions are parameters that have a direct or indirect impact on the state characteristics of the unit. The specific parameters and quantity of the boundary conditions can be determined according to actual needs.

[0074] In one embodiment of the present application, the steady-state historical operating data is divided into operating conditions, including the following steps: first, manually determine multiple boundary conditions, and determine the size of the characteristic interval corresponding to each boundary condition according to the characteristics of the boundary conditions. Specifically, this embodiment can adopt the method of manually determining the boundary conditions, and determine the operating condition grid in an equal interval division manner, and then put the sample data into different operating conditions according to different boundary conditions. This division method first determines the boundary conditions, and determines the interval size according to the characteristics of the boundary conditions. When it is specifically implemented, the cleaned historical operating data can be traversed, and the maximum and minimum values ​​of each boundary condition can be determined by comparison, so as to determine the interval size, and then obtain the preset operating condition division interval of each operating condition characteristic parameter.

[0075] Then, each characteristic interval is combined and divided into equally spaced working conditions using the following formula:

[0076] C i =(A i ,A i +ΔA)∩(B i ,B i+ΔB)...(D i ,D i +ΔD)

[0077] sA i <<A min ,A i +ΔA>>A max ,B i <<B min ,B i +ΔB>>B max ,...D i <<D min ,D i +ΔD>>D max

[0078] Among them, A, B and D represent different boundary conditions, for example, A, B and D can represent temperature, load and torque, etc. This formula indicates that the selection of a specific boundary condition can include multiple boundary conditions, that is, there can be other boundary conditions between boundary condition B and boundary condition D. min represents the minimum value of the boundary condition, A max represents the maximum value of the boundary condition, and so on. i Represents the Cth interval determined by the intersection of multiple feature intervals i working condition, ΔA, ΔB and ΔD represent the working condition division intervals of different boundary conditions respectively.

[0079] Therefore, the present application divides the operating conditions of the generator set to facilitate the subsequent establishment of benchmark models for different operating conditions and improve the pertinence of benchmark value determination.

[0080] Step S102, determining the baseline value of the threshold-type symptom under each operating condition through a Gaussian mixture model, and training a bidirectional long short-term memory network as a prediction model for data trends through the historical operating data after the operating condition division.

[0081] It should be noted that the present application detects abnormal patterns, that is, determines corresponding abnormal signs. Abnormal signs are evidence for identifying the fault mode and cause. For example, the sign may be "the output voltage of the generator set drops". The abnormal signs targeted by the present application include threshold signs and trend signs.

[0082] Among them, the extraction of threshold-type symptoms is based on the mining of historical data, dividing the historical data into working conditions, determining the benchmark value under each working condition, and then determining the boundary of the threshold-type symptoms, and then comparing it with the real-time data to achieve the purpose of judging the threshold-type symptoms. Different benchmark value determination methods can be used according to the different changes in characteristic parameters. Trend-type symptoms are mainly manifested in the "increase", "decrease" and "unstable fluctuation" trends of related characteristic indicators. The characteristic parameters of trend-type symptoms are time-series and gradually begin to change with the change of time.

[0083] Among them, the Gaussian mixture model is a semi-parametric probability density estimation method that combines the advantages of parameter estimation and non-parametric estimation methods. It is not limited to a specific probability density function form. If there are enough members in the model, it can approximate any continuous distribution with arbitrary accuracy. The bidirectional long short-term memory network (Bidirectional Long Short Term Memory Network, referred to as BiLSTM) is developed based on the long short-term memory artificial neural network (LSTM). It is a time recurrent neural network that can perform forward and backward propagation and has strong capabilities in mining and predicting time series data.

[0084] Specifically, after the operating conditions of the generator set are divided, it is necessary to perceive the operating status of the unit under the operating condition. For the characteristic parameters of the threshold-type signs, the baseline value can be obtained, and then the real-time value can be compared with the baseline value to make an abnormal judgment. Therefore, the extraction of threshold-type signs first requires the establishment of a baseline value. This application uses a Gaussian mixture model to obtain the baseline value.

[0085] In the specific implementation, as a possible implementation method, first determine the probability distribution expression of the multivariate Gaussian mixture model as shown in the following formula:

[0086]

[0087] Where X is the L-dimensional parameter data column vector, X = [x1x2,...,x L ] T , K is the number of mixtures of sub-models, ω k is the weight coefficient of the mixed model, and φ(X|θ k ) is the multivariate single Gaussian probability density function of the k-th submodel, which can be calculated by the following formula:

[0088]

[0089] Then, the parameters of the Gaussian mixture model are estimated by the Expectation Maximum (EM) algorithm. EM is an iterative algorithm that iteratively obtains the maximum likelihood parameter value and is used for parameter estimation of probability models containing implicit variables. Its objective function is shown in the following formula:

[0090]

[0091] Among them, X is the parameter variable, K is the number of sub-models, φ k (X|θ k ) is the probability density function of the Kth sub-model, ω k is the weight coefficient of the Kth sub-model. In this embodiment, when estimating, the rough value of the GMM parameter can be estimated first, and then the obtained rough value is used to maximize the likelihood function.

[0092] Then, the number of sub-models of the Gaussian mixture model is determined by the Akaike Information Criterion AIC. The AIC evaluation criterion is based on the concept of entropy and provides an evaluation method that balances model complexity and good fitting data. Its definition is shown in the following formula:

[0093] AIC=2K-2ln(L)

[0094] Among them, K is the number of sub-models, and ln(L) represents the log-likelihood function of the model.

[0095] It should be noted that as the number of sub-models increases, the model complexity also increases, and the value of AIC will decrease and then increase. Therefore, the embodiment of the present application selects the number of sub-models corresponding to the minimum AIC as the optimal number of sub-models for GMM. Therefore, the embodiment of the present application determines the number of sub-models by estimating the parameters of GMM and combining the AIC criterion to cluster the data samples under a single working condition based on similarity. Then, the data samples under each divided working condition can be clustered in turn through GMM.

[0096] Finally, after clustering is completed, the average value of the class with the highest power generation efficiency is selected as the benchmark sample class. In the embodiment of the present application, the benchmark is defined as the parameter value corresponding to the highest unit operating efficiency that can actually be achieved under the current operating boundary conditions. Therefore, the data with the highest power generation efficiency is selected as the benchmark sample.

[0097] Furthermore, after determining the baseline value of the threshold-type signs through the Gaussian mixture model, a bidirectional long short-term memory neural network is trained to predict the trend of characteristic parameters for trend-type signs.

[0098] It should be noted that the process from the initiation of a fault to its impact is a process. In the early stage of a fault, the fault-related fault symptoms can be reflected by the relevant characteristic parameters. The abnormal symptoms reflecting the abnormal mode of the generator set involve changes in the operating parameters, where the parameters also have different effects over time. In view of the characteristics of the time series of the generator set parameters, this application uses the advantages of bidirectional long short-term memory network to learn and memorize the long-term information of the sequence to predict the trend of the characteristic parameters.

[0099] Among them, the data trend can be approximately regarded as the changing trend of the feature parameters from the previous moment to the next moment. It is understandable that since the bidirectional long short-term memory network can accurately obtain the context information of each node in the input sequence and mine the connection between the future and past moments of data, the trained bidirectional long short-term memory network can predict the trend of the input data.

[0100] In one embodiment of the present application, a bidirectional long short-term memory network is first constructed, including an input layer, a hidden layer and an output layer. The hidden layer includes Figure 3 The loop structure shown in the figure, each x represents the input feature parameter, y represents the predicted value of the corresponding feature parameter, S and A can be regarded as intermediate calculation variables, and each input sequence is set to pass through the recurrent neural network once in the forward and reverse directions, so as to obtain the complete past and future context information of each node. As an example, the formal representation of the LSTM constructed in this application is as follows:

[0101] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0102] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0103]

[0104]

[0105] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0106] h t =o t ⊙tanh(c t )

[0107] where x t 、h t-1 、c t-1 They represent the input at time t, the output at time t-1, and the unit state at time t-1. ⊙ represents the dot multiplication operation, σ is the sigmoid function, tanh represents the hyperbolic tangent function, and W * Respectively represent the weights of the corresponding states, b * Represents a bias term.

[0108] Furthermore, after constructing the bidirectional long short-term memory network, the data after the working condition division is used as training data to train the bidirectional long short-term memory network. In specific implementation, in one embodiment of the present application, the bidirectional long short-term memory network is trained for each working condition to predict the data trend under the working condition, and the historical operating data under each working condition is divided into a training set, a validation set, and a test set according to a preset ratio. The model is trained with the data in the training set. The specific training method can refer to the training method in the relevant technology, for example, including forward calculation of the output value of each neuron, reverse calculation of the error term value of each neuron, calculation of the gradient of each weight, and updating the weight through the stochastic gradient descent algorithm, etc. The specific training process will not be repeated here. After the training is completed, the accuracy of the prediction to the long short-term memory network is tested with the data in the test set to see if it meets the requirements.

[0109] Step S103, acquiring real-time data of the generator set, comparing the real-time data with a reference value of a threshold-type symptom under a corresponding working condition, and extracting the threshold-type symptom.

[0110] Specifically, the real-time data of the generator set is obtained, the operating condition corresponding to the real-time data is determined, and then the reference value of the threshold-type symptom under the predetermined operating condition is obtained. The reference value of the threshold-type symptom is then compared with the real-time data to determine whether the real-time data is outside the reference value of the threshold-type symptom. If so, the threshold-type symptom is extracted, that is, it is recognized that the current real-time data has a threshold-type symptom.

[0111] It should be noted that the threshold-type symptom extracted in the present application can be zero, that is, when the real-time data is within the reference value range of the threshold-type symptom, it is determined that there is no threshold-type abnormal data in the real-time data.

[0112] Step S104, obtaining target data in an adjacent preset time period before the real-time data, predicting the data trend of the target data through the trained bidirectional long short-term memory network, determining the degree of deviation between the real-time data and the data trend, and extracting trend signs.

[0113] Among them, the target data is the historical data in the time period before the real-time data. After the threshold-type signs are extracted in step S103, even if the data at the current moment is in the normal range corresponding to the benchmark value in the specified time period, the data in this time period may have an abnormal trend as a whole. For example, the data gradually increases. Therefore, the present application also obtains the target data in the adjacent preset time period before the real-time data, and inputs it into the trained bidirectional long short-term memory network to predict the data trend of the target data. When the actual data deviates from the trend predicted by the prediction model to an unacceptable degree, it can be considered that a trend-type sign has occurred.

[0114] In specific implementation, the data trend of the target data is predicted through the trained bidirectional long short-term memory network, including inputting multiple characteristic parameters in the target data into the bidirectional long short-term memory network, and mining the connection between each characteristic parameter in the adjacent preset time period and the current time period through the hidden layer, that is, predicting the trend of the target data, including the values ​​of each parameter at different times in the current time period. Then, the degree of deviation between the real-time data and the predicted data trend is determined, and when the degree of deviation is greater than the corresponding value, the trend-type fault sign is identified.

[0115] In one embodiment of the present application, the degree of deviation between the real-time data and the data trend is determined, including judging the degree of deviation based on a confidence interval when the real-time data and the data trend begin to differ. The expression of the confidence interval is as follows:

[0116] p{c1 <m<c2}=1-α

[0117] Among them, p represents probability, α represents significance level, c1 and c2 represent two statistics at significance level, and m represents the difference between real-time data and data trend. That is, in the embodiment of the present application, the degree of deviation is judged by confidence interval method. When the trend of characteristic parameters is different from the effect predicted by the model and continues to exceed the confidence interval of the prediction model, it is determined that an abnormality occurs.

[0118] Step S105, determining an abnormal pattern according to the extracted threshold-type signs and trend-type signs.

[0119] Specifically, if a threshold-type sign is extracted, it is determined that a threshold-type abnormal pattern is detected; if a threshold-type sign is extracted, it is determined that a threshold-type abnormal pattern is detected.

[0120] In one embodiment of the present invention, after determining the abnormal pattern based on the extracted threshold-type signs and trend-type signs, it also includes obtaining a supplementary abnormal pattern obtained through manual detection or sensor detection, inputting the threshold-type abnormal pattern, trend-type abnormal pattern and supplementary abnormal pattern into the diagnostic knowledge base of the generator set for reasoning, and obtaining the returned fault mode, fault cause and fault maintenance measures.

[0121] Specifically, supplementary abnormal patterns are other types of abnormal patterns corresponding to signs that cannot be obtained through data, and can be roughly regarded as abnormal conditions detected by naked eye observation or instruments, such as artificially seeing white smoke, hearing roaring sounds, or detecting ice on the surface of the generator set through sensors. In addition to detecting abnormal patterns by analyzing data, this application also detects abnormal patterns through actual detection methods such as manual detection or sensor detection, thereby improving the comprehensiveness of abnormal pattern recognition.

[0122] Furthermore, the various abnormal modes obtained in the above embodiments will be used as input signs of the unit fault diagnosis inference engine and input into the mechanism diagnosis knowledge base corresponding to each generator set for reasoning to determine the corresponding fault mode, fault cause and fault measures. Through early identification of fault signs, timely adjustments and maintenance can be carried out to avoid abnormalities caused by the development of abnormal parameters, thereby avoiding the occurrence of faults and achieving early fault warning.

[0123] Therefore, the present application analyzes the threshold-type symptoms and trend-type symptoms in the fault mode. First, the historical data of the generator set is divided into operating conditions to distinguish between steady-state conditions and non-steady-state conditions, providing support for the training of the subsequent feature extraction model. Then, different extraction methods are used for different symptom types. For threshold-type symptom extraction, a method of obtaining a reference value using a Gaussian mixture model is used. The characteristic parameters are compared with the reference values. When the characteristic parameters leave the reference value interval, a threshold-type abnormal mode is determined to have occurred. For trend-type symptom extraction, a method of predicting the trend of characteristic parameters using a bidirectional long short-term memory neural network is used. The characteristic parameters are input into the prediction model. When the trend of the characteristic parameters is different from the effect predicted by the model and continues to exceed the confidence interval of the prediction model, a trend-type abnormal mode is determined to have occurred.

[0124] In summary, the abnormal mode detection method of the generator set based on deep learning in the embodiment of the present application uses a deep learning model to mine the complex mapping rules between the normal operating state of the generator set and the parameters, solves the problem that the multivariable coupling relationship affects the benchmark modeling, and can perform accurate benchmark modeling under different operating boundary conditions of the generator set. The abnormal mode recognition method and the trend abnormal mode recognition method are used to identify the abnormal mode of the real-time operating data of the generator set, which is suitable for different working conditions when the generator set is actually running. In addition, a data cleaning process is proposed, and the cleaned data is used for model training to ensure that the generator set operation data used for model training is in a healthy state of the generator set. As a result, the method improves the accuracy and pertinence of establishing the benchmark model, and then improves the accuracy and applicability of abnormal mode detection, which is conducive to timely and accurate diagnosis of unit faults. Through early identification of fault signs, timely adjustment and maintenance are carried out to avoid abnormal parameters from developing backward to generate faults, achieve early fault warning, and improve the safety of the generator set.

[0125] In order to more clearly illustrate the process of the abnormal mode detection method of a generator set based on deep learning in an embodiment of the present application, a specific embodiment of detecting the abnormal mode of a generator set is described in detail below. Figure 4 The following is a flowchart of a specific method for detecting abnormal modes of a generator set based on deep learning proposed in an embodiment of the present application. Figure 4 As shown, the method comprises the following steps:

[0126] Step S401, obtain historical data.

[0127] Step S402, dividing the working conditions.

[0128] Specifically, this step includes the following sub-steps:

[0129] Step S4021, remove the start and stop data. This step is to remove the abnormal data points when the generator set is started and stopped.

[0130] Step S4022, dividing the steady-state working condition into the unsteady-state working condition. In this step, the data under the unsteady-state working condition is also eliminated.

[0131] Step S4023, divide the working conditions based on boundary conditions.

[0132] Step S403: perform model training.

[0133] Specifically, this step includes the following sub-steps:

[0134] Step S4031, training the GMM model clustering benchmark value.

[0135] Step S4032, training the BiLSTM model based on deep learning to predict parameter trends.

[0136] Step S404: extract fault features.

[0137] Specifically, this step includes the following sub-steps:

[0138] Step S4041, extracting threshold-type signs.

[0139] Step S4042, extracting trend signs.

[0140] It should be noted that, in this embodiment, the specific implementation method of each step can refer to the description of the above embodiment, and the specific implementation structure is similar, which will not be repeated here.

[0141] In order to implement the above embodiment, the present application also proposes an abnormal mode detection system for a generator set based on deep learning. Figure 5This is a structural diagram of an abnormal mode detection system for a generator set based on deep learning proposed in an embodiment of the present application, such as Figure 5 As shown, the system includes a division module 100 , a training module 200 , a first extraction module 300 , a second extraction module 400 and a determination module 500 .

[0142] The division module 100 is used to obtain the historical operating data of the generator set, perform steady-state judgment on the historical operating data after removing abnormal data, and divide the steady-state historical operating data into operating conditions.

[0143] The training module 200 is used to construct a Gaussian mixture model, determine the baseline value of the threshold-type sign under each operating condition through the Gaussian mixture model, and train a bidirectional long short-term memory network as a prediction model for data trends through the historical operating data after the operating condition division.

[0144] The first extraction module 300 is used to obtain real-time data of the generator set, compare the real-time data with a reference value of a threshold-type symptom under a corresponding working condition, and extract the threshold-type symptom.

[0145] The second extraction module 400 is used to obtain target data within an adjacent preset time period before the real-time data, predict the data trend of the target data through the trained bidirectional long short-term memory network, determine the degree of deviation between the real-time data and the data trend, and extract trend signs.

[0146] The determination module 500 is used to determine an abnormal pattern according to the extracted threshold-type symptoms and trend-type symptoms.

[0147] Optionally, in one embodiment of the present application, the system also includes: a first acquisition module, used to obtain a supplementary abnormality pattern obtained through manual detection or sensor detection; a second acquisition module, used to input the threshold-type abnormality pattern, the trend-type abnormality pattern and the supplementary abnormality pattern into the diagnostic knowledge base of the generator set for reasoning, and obtain the returned fault mode, fault cause and fault maintenance measures.

[0148] Optionally, in one embodiment of the present application, the bidirectional long short-term memory network includes an input layer, a hidden layer and an output layer, the hidden layer includes a loop structure, and the second extraction module is specifically used to: input multiple feature parameters in the target data into the bidirectional long short-term memory network, and mine the connection between each feature parameter in adjacent preset time periods and the current time period through the hidden layer.

[0149] Optionally, in one embodiment of the present application, the second extraction module 400 is further used to: when the real-time data and the data trend begin to differ, determine the degree of deviation based on a confidence interval, and the expression of the confidence interval is as follows:

[0150] p{c1 <m<c2}=1-α

[0151] Among them, p represents probability, α represents significance level, c1 and c2 represent two statistics under significance level, and m represents the difference between real-time data and data trend.

[0152] Optionally, in one embodiment of the present application, the division module 100 is specifically used to: select the power generation of the generator set as a discriminant feature for steady-state discrimination; calculate the difference in power generation between two adjacent moments; estimate the rate of change of the power generation by the mean of the sample statistic in the time window; based on the confidence interval estimation method, estimate whether the confidence interval of the rate of change of the power generation includes zero, and if so, determine that the generator set is in a steady-state operating condition in the time period corresponding to the time window.

[0153] Optionally, in one embodiment of the present application, the division module 100 is further used to: manually determine a plurality of boundary conditions, and determine the size of a characteristic interval corresponding to each boundary condition according to the characteristics of the boundary conditions;

[0154] Combine each characteristic interval and divide the working conditions into equal intervals using the following formula:

[0155] C i =(A i ,A i +ΔA)∩(B i ,B i +ΔB)...(D i ,D i +ΔD)

[0156] sA i <<A min ,A i +ΔA>>A max ,B i <<B min ,B i +ΔB>>B max ,...D i <<D min ,D i +ΔD>>D max

[0157] Among them, A, B and D represent different boundary conditions, C i Represents the Cth interval determined by the intersection of multiple feature intervals i working condition, ΔA, ΔB and ΔD represent the working condition division intervals of different boundary conditions respectively.

[0158] It should be noted that the above explanation of the embodiment of the abnormal mode detection method of the generator set based on deep learning is also applicable to the system of this embodiment, and will not be repeated here.

[0159] In summary, the abnormal mode detection system of the generator set based on deep learning in the embodiment of the present application uses a deep learning model to mine the complex mapping rules between the normal operating state and parameters of the generator set, solves the problem that the multivariable coupling relationship affects the benchmark modeling, and can perform accurate benchmark modeling under different operating boundary conditions of the generator set. The threshold-type abnormal mode recognition method and the trend-type abnormal mode recognition method are used to identify the abnormal mode of the real-time operating data of the generator set, which is suitable for different working conditions during the actual operation of the generator set. In addition, the cleaned data is used for model training to ensure that the generator set operation data used for model training is in a healthy state. As a result, the system improves the accuracy and applicability of abnormal mode detection, is conducive to timely and accurate diagnosis of unit faults, and timely adjusts and maintains through early identification of fault signs to avoid abnormal parameters from developing into faults, achieves early fault warning, and improves the safety of the generator set.

[0160] In order to implement the above embodiments, the present application also proposes a non-temporary computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the abnormal mode detection method of a generator set based on deep learning as described in any of the above embodiments is implemented.

[0161] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0162] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0163] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0164] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0165] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0166] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0167] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0168] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for detecting abnormal modes of a generator set based on deep learning, characterized in that: The following steps are involved: The historical operation data of the generator set is obtained, and the steady-state judgment is performed on the historical operation data after the abnormal data is removed. The size of the characteristic interval corresponding to each boundary condition is determined according to the characteristics of multiple boundary conditions, each characteristic interval is combined, and the steady-state historical operation data is divided into equally spaced working conditions by the following formula: C i =(A i ,A i +ΔA)∩(B i ,B i +ΔB)...(D i ,D i +ΔD) s.t.Ai<<A min ,A i +ΔA>>A max ,B j <<B min ,B i +ΔB>>B max ,...D i <<D min ,D i +ΔD>>D max Among them, A, B and D represent different boundary conditions, C i Represents the Cth interval determined by the intersection of multiple feature intervals i working condition, ΔA, ΔB and ΔD represent the working condition division intervals of different boundary conditions respectively; The reference value of the threshold sign under each working condition is determined by the Gaussian mixture model, and the bidirectional long short-term memory network is trained as a prediction model of the data trend by the historical operation data after the working condition division, wherein the expression of the constructed bidirectional long short-term memory network is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ⊙tanh(c t ) Among them, x t 、h t-1 、c t-1 They represent the input at time t, the output at time t-1, and the unit state at t-1, ⊙ represents the dot multiplication operation, σ is the sigmoid function, tanh represents the hyperbolic tangent function, and W * Respectively represent the weights of the corresponding states, b * represents the bias term; Acquire real-time data of the generator set, compare the real-time data with a reference value of a threshold-type symptom under a corresponding working condition, and extract the threshold-type symptom; Obtain target data in an adjacent preset time period before the real-time data, predict the data trend of the target data through the trained bidirectional long short-term memory network, and when the real-time data and the data trend begin to differ, determine the degree of deviation between the real-time data and the data trend based on the confidence interval method, and extract trend signs, wherein the expression of the confidence interval is as follows: p{c1 <m<c2}=1-α Among them, p represents probability, α represents significance level, c1 and c2 represent two statistics under significance level, and m represents the difference between real-time data and data trend; Determining an abnormal pattern based on the extracted threshold-type symptoms and trend-type symptoms; Acquire supplementary anomaly patterns obtained through manual or sensor detection; Inputting the threshold-type abnormal pattern, the trend-type abnormal pattern and the supplementary abnormal pattern into the diagnostic knowledge base of the generator set for reasoning, and obtaining the returned fault mode, fault cause and fault maintenance measures; The steady-state judgment of the historical operating data after eliminating abnormal data includes: selecting the power generation of the generator set as a discriminant feature for steady-state judgment; calculating the difference in power generation between two adjacent moments; estimating the rate of change of the power generation by the mean of the sample statistic in the time window; and estimating whether the confidence interval of the rate of change of the power generation includes zero based on the confidence interval estimation method, and if so, determining that the generator set is in a steady-state operating condition in the time period corresponding to the time window.

2. The method according to claim 1, characterized in that The bidirectional long short-term memory network includes an input layer, a hidden layer and an output layer, the hidden layer includes a loop structure, and the bidirectional long short-term memory network completed by training predicts the data trend of the target data, including: A plurality of characteristic parameters in the target data are input into a bidirectional long short-term memory network, and the connection between each characteristic parameter in the adjacent preset time periods and the current time period is mined through the hidden layer.

3. The method according to claim 1, characterized in that The plurality of boundary conditions are determined manually.

4. A deep learning-based abnormal mode detection system for a generator set, characterized in that: include: The division module is used to obtain the historical operation data of the generator set, and perform steady-state judgment on the historical operation data after removing abnormal data, determine the size of the characteristic interval corresponding to each boundary condition according to the characteristics of multiple boundary conditions, combine each characteristic interval, and divide the steady-state historical operation data into equally spaced working conditions by the following formula: C i =(A i ,A i +ΔA)∩(B i ,B i +ΔB)...(D i ,D i +ΔD) s.t.Ai<<A min ,A i +ΔA>>A max ,B i <<B min ,B i +ΔB>>B max ,...D i <<D min ,D i +ΔD>>D max Among them, A, B and D represent different boundary conditions, C i Represents the Cth interval determined by the intersection of multiple feature intervals i working condition, ΔA, ΔB and ΔD represent the working condition division intervals of different boundary conditions respectively; The training module is used to construct a Gaussian mixture model, determine the reference value of the threshold-type sign under each working condition through the Gaussian mixture model, and train a bidirectional long short-term memory network as a prediction model for data trends through the historical operation data after the working condition division, wherein the expression of the constructed bidirectional long short-term memory network is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ⊙tanh(c t ) Among them, x t 、h t-1 、c t-1 They represent the input at time t, the output at time t-1, and the unit state at t-1, ⊙ represents the dot multiplication operation, σ is the sigmoid function, tanh represents the hyperbolic tangent function, and W * Respectively represent the weights of the corresponding states, b * represents the bias term; A first extraction module is used to obtain real-time data of the generator set, compare the real-time data with a reference value of a threshold-type symptom under a corresponding working condition, and extract the threshold-type symptom; The second extraction module is used to obtain target data in an adjacent preset time period before the real-time data, predict the data trend of the target data through the trained bidirectional long short-term memory network, and when the real-time data and the data trend begin to differ, determine the degree of deviation between the real-time data and the data trend based on the confidence interval method, and extract trend signs, wherein the expression of the confidence interval is as follows: p{c1 <m<c2}=1-α Among them, p represents probability, α represents significance level, c1 and c2 represent two statistics under significance level, and m represents the difference between real-time data and data trend; A determination module, configured to determine an abnormal pattern according to the extracted threshold-type symptoms and trend-type symptoms; A first acquisition module is used to acquire a supplementary abnormal pattern obtained by manual detection or sensor detection after determining the abnormal pattern according to the extracted threshold-type symptoms and the trend-type symptoms; A second acquisition module is used to input the threshold type abnormal mode, the trend type abnormal mode and the supplementary abnormal mode into the diagnostic knowledge base of the generator set for reasoning, and obtain the returned fault mode, fault cause and fault maintenance measure; The steady-state judgment of the historical operating data after eliminating abnormal data includes: selecting the power generation of the generator set as a discriminant feature for steady-state judgment; calculating the difference in power generation between two adjacent moments; estimating the rate of change of the power generation by the mean of the sample statistic in the time window; and estimating whether the confidence interval of the rate of change of the power generation includes zero based on the confidence interval estimation method, and if so, determining that the generator set is in a steady-state operating condition in the time period corresponding to the time window.

5. The system according to claim 4, characterized in that The bidirectional long short-term memory network includes an input layer, a hidden layer and an output layer, the hidden layer includes a loop structure, and the second extraction module is specifically used for: A plurality of characteristic parameters in the target data are input into a bidirectional long short-term memory network, and the connection between each characteristic parameter in the adjacent preset time periods and the current time period is mined through the hidden layer.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the abnormal mode detection method of a generator set based on deep learning as described in any one of claims 1 to 3 is implemented.

Images