Multi-strategy fusion water-turbine generator set fault diagnosis method

Abstract

The invention provides a multi-strategy fusion water-turbine generator set fault diagnosis method. The method comprises the following steps: S1, constructing a time sequence model based on an industrial internet platform; s2, constructing a decision model based on industrial internet platform data; s3, using the trained GRU model and the generated optimal random forest to construct a comprehensive model; and S4, performing fault diagnosis on the working condition of the current unit by using the real-time data of the industrial internet platform. Historical data and real-time data in multi-source data of an industrial internet platform are adopted, and a GRU recurrent neural network and a random forest are adopted. A multi-strategy fusion mode is adopted, a GRU-based neural network time sequence model and a random forest decision model are combined to construct a comprehensive model, the advantages of different algorithms are fully utilized, the limitation of fault judgment by a single algorithm is avoided, and the accuracy of fault diagnosis is improved.

Description

Technical Field

[0001] The invention relates to the field of fault diagnosis of a hydro-generator set, and in particular to a multi-strategy fusion fault diagnosis method for a hydro-generator set. Background Art

[0002] The hydro-turbine generator set is the core equipment of the hydropower station, and its operating status is directly related to the power supply capacity and quality of the power system. However, during operation, the hydro-turbine generator set is easily affected by various factors, resulting in various faults and reducing its operating capacity. Therefore, fault diagnosis of the hydro-turbine generator set is of great significance in ensuring power supply, improving operational reliability, and optimizing maintenance management.

[0003] The data center of the hydropower station industrial Internet platform gathers panoramic data of the hydropower station's turbine generator units, including data from the computer monitoring system in Zone I and various online monitoring system data from Zone II, such as the online monitoring system for unit vibration swing and the online monitoring system for gas in oil of the main transformer. The complete data set facilitates unit fault diagnosis. The data storage of the industrial Internet of Things platform is divided into a time series library and a historical library. The time series library is responsible for storing real-time data and short-term historical data. Its data has continuous time attributes and is more suitable for time series data mining and analysis. The historical library stores long-term historical data and is more suitable for model construction and training.

[0004] Current fault diagnosis of hydro-turbine generator sets is mostly based on a single algorithm for fault judgment, or based on a single monitoring system data. At the same time, some advanced applications such as unit fault diagnosis are mainly based on simple models (threshold judgment method, expert experience method), lack the application of historical data, and require expert experience guidance. The accuracy and scientificity need to be considered. Summary of the invention

[0005] The main purpose of the present invention is to provide a multi-strategy integrated hydro-generator fault diagnosis method to solve the problem that the current hydro-generator fault diagnosis is mostly based on a single algorithm for fault judgment, and the accuracy and scientificity need to be considered.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a multi-strategy fusion fault diagnosis method for a hydro-generator set, comprising the following steps: S1. Build a time series model based on the industrial Internet platform; S2. Build a decision-making model based on industrial Internet platform data; S3, build a comprehensive model using the trained GRU model and the generated best random forest; S4. Use the real-time data from the industrial Internet platform to perform fault diagnosis on the current unit operating conditions.

[0007] In the preferred embodiment, the specific method in step S1 includes: S11. Select the historical data of telemetry parameters related to the hydro-turbine generator set in the industrial Internet platform to construct training data, construct the telemetry at the same time point and its corresponding unit operating condition as a sample data, and construct a continuous data set in the time dimension according to a certain time step. The data set format is ; Where X is the telemetry value; Y is the unit operating condition, the normal unit operating condition corresponds to 0, and the abnormal unit operating condition corresponds to 1; the data set contains m telemetry attributes, with a total of n sample data. The telemetry values ​​in the data set are normalized, and the formula is: ; in is the current telemetry value, is the normalized telemetry value, is the lower limit of the current telemetry measurement, is the upper limit of the telemetry measurement; S12. Divide the data set into continuous training sets and validation sets using a sliding window in a ratio of 8:2, build a GRU-based neural network, and use the training set, validation set, and GRU-based neural network to train the time series model.

[0008] In the preferred embodiment, the method for constructing a GRU-based neural network in step S12 is: S13, dividing the data set into continuous training sets and validation sets in a ratio of 8:2, and the division method is sliding window division, that is, from the first data to the 4n / 5th data, select random data as the breakpoint, take the data and the subsequent n / 5-1 data as the validation set, and the remaining data as the training set; S14. Use the training set, validation set and GRU-based neural network to train the time series model. The batch_size is set to 128, the hidden layer width hidden_size is set to 2, the sequence_length is set to 8, the input_size is set to m, the dropout is set to 0.02, the number of training epochs is set to n / 32, and the stochastic gradient descent method is used as the optimizer. The loss function is the cross entropy loss function. The model parameters with the lowest loss on the validation set are saved during the training process.

[0009] In the preferred embodiment, the specific method in step S2 includes: Based on the historical data of telemetering related to the parameters of the unit on the industrial Internet platform, discrete training data T is constructed. The random forest algorithm is used to build a decision model. The size of the random forest is set to K. A new sub-training set D is formed based on the training data T. The decision tree is trained using the sub-training set D to construct a random forest decision model.

[0010] The preferred solution specifically includes: S21: Construct discrete training data T based on the historical data of telemetering quantities related to unit parameters on the industrial Internet platform. According to the type of unit operating condition, select the same number of samples of the operating condition to ensure the balance of the samples.

[0011] S22: Use the random forest algorithm to build a decision model, set the random forest size to K, select m features from the training set T with replacement sampling, sample 2N / 3 samples, form a new sub-training set D, and use the sub-training set D to train the CD4.5 decision tree: First, calculate the information entropy of the training set D. Since the unit condition diagnosis is a binary classification task, the information entropy calculation formula is:

[0012] Where p_k is the probability of occurrence of the kth type of sample. For example, if there are 1 / 3 abnormal working condition data and 2 / 3 normal working condition data in the sub-training set D, then p_abnormal = 1 / 3, p_normal = 2 / 3; calculate the information entropy of the data attributes in the sub-training set D. For any attribute a, since there are only three cases (0, 1 and offline) of the telesignaling attribute, the sub-training set D is divided into three subsets {a1, a2, a3} through it, and the information gain is calculated as:

[0013] Calculate the information gain rate of any attribute a, formula:

[0014] Select the attribute with the highest information gain rate as the partition attribute, and repeat the above steps for the three subsets until no further training is possible, thus training a C4.5 tree. Repeat the training K times to obtain K CD4.5 decision trees, and set the weight of each decision tree to 1 / K to construct a random forest.

[0015] In the preferred embodiment, the specific steps of step S3 are: S3: Build a comprehensive model: Use the trained GRU model and the generated best random forest to build a comprehensive model, including: S31: Filter the latest 1-8 data in the time series library, select the same telemetry data as S11 to build the model input, use the trained GRU time series model to predict the unit operating conditions, and obtain the confidence level C1; S32: Filter the current real-time data in the time series library, select the same telesignal as S21 to build the model input, use the trained decision model to predict the working condition, and obtain the confidence C2, where the decision model is a random forest decision model; S33: Repeat S31 and S32 multiple times to obtain a large number of confidence samples and construct paired sets of C1, C2, and Y.

[0016] S34: Build a linear regression model and use the mean square error method to obtain the optimal weights W1 and W2 of C1 and C2, and obtain the bias, C3=W1*C1+W2*C2+bias.

[0017] In the preferred embodiment, the specific steps of step S4 are: Use the trained time series model to predict the unit operating conditions and obtain the confidence level C1; Use the trained random forest decision model to predict the working condition and obtain the confidence level C2; And calculate the comprehensive model confidence C3; Construct an important telesignaling quantity set; when a telesignaling quantity value in the important telesignaling quantity set is 1, it is considered that the unit operating condition is abnormal; When the telesignal values ​​in the important telesignal set are all 0, the unit operating condition is considered abnormal when C2≥0.5; when the telesignal values ​​in the important telesignal set are all 0, the unit operating condition is considered abnormal when C2<0.5 and C3≥0.5; When the telesignal values ​​in the important telesignal set are all 0, the unit is considered to be in normal operating condition when C1<0.5, C2<0.5, and C3<0.5.

[0018] The present invention provides a multi-strategy fusion method for diagnosing faults of hydro-generator sets, which adopts historical data and real-time data from multi-source data of an industrial Internet platform, and a method for diagnosing abnormal operating conditions of hydro-generator sets using a GRU recurrent neural network and a random forest.

[0019] A multi-strategy fusion approach is adopted to combine the GRU-based neural network time series model and the random forest decision model to build a comprehensive model, which fully utilizes the advantages of different algorithms, avoids the limitations of a single algorithm for fault judgment, and improves the accuracy of fault diagnosis.

[0020] Reduced reliance on expert experience: Traditional fault diagnosis methods for hydropower generators are mostly based on a single algorithm or a single monitoring system data, and some advanced applications such as unit fault diagnosis are based on simple models and require expert experience guidance. This method reduces reliance on expert experience and improves the scientificity and objectivity of diagnosis by integrating multiple algorithms and multi-source data.

[0021] Using the real-time data of the industrial Internet platform to diagnose the current unit operating conditions can timely discover the abnormal conditions of the unit, provide timely decision-making basis for maintenance management, help shorten the fault handling time, and reduce fault losses. This multi-strategy fusion method for fault diagnosis of hydropower generator sets has many beneficial effects, such as improving diagnostic accuracy, enhancing scientificity and reliability, and optimizing maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a neural network diagram of the GRU of the present invention; Figure 2 It is a GRU unit structure diagram of the present invention; DETAILED DESCRIPTION Embodiment 1 like Figure 1-2 As shown, a multi-strategy fusion fault diagnosis method for a hydro-generator unit includes the following steps: Step 1: Build a time series model based on the Industrial Internet platform, which includes: select the historical data of telemetry parameters related to the hydropower generator set in the Industrial Internet platform to build training data, build the telemetry at the same time point and its corresponding unit operating condition into a sample data, and build a continuous data set in the time dimension according to a certain time step. The data set format is ; Where X is the telemetry value; Y is the unit operating condition, the normal unit operating condition corresponds to 0, and the abnormal unit operating condition corresponds to 1; the data set contains m telemetry attributes, with a total of n sample data. The telemetry values ​​in the data set are normalized, and the formula is: ; in is the current telemetry value, is the normalized telemetry value, is the lower limit of the current telemetry measurement, is the upper limit of telemetry; the data set is divided into continuous training set and validation set by sliding window in the ratio of 8:2, a GRU-based neural network is constructed, and the time series model is trained using the training set, validation set and GRU-based neural network; The data set is divided into continuous training sets and validation sets using a sliding window in a ratio of 8:2. This division method can better simulate the distribution of actual data and improve the generalization ability and reliability of the model.

[0023] Step 2: Construct a decision model based on the data of the industrial Internet platform; specifically, it includes: constructing discrete training data T based on the historical data of the telemetering quantity related to the parameters of the industrial Internet platform unit, using the random forest algorithm to construct a decision model, setting the size of the random forest to K, forming a new sub-training set D based on the training data T, and using the sub-training set D to train the decision tree, thereby constructing a random forest decision model; Step 3: Use the trained GRU model and the generated best random forest to build a comprehensive model; Step 4: Use the real-time data of the Industrial Internet platform to perform fault diagnosis on the current unit operating condition, including: using the trained time series model to predict the unit operating condition and obtain the confidence C1; using the trained random forest decision model to predict the condition and obtain the confidence C2; ​​and calculating the comprehensive model confidence C3; constructing a set of important telesignal quantities; when the telesignal value in the important telesignal quantity set is 1, the unit operating condition is considered abnormal; when the telesignal values ​​in the important telesignal quantity set are all 0, the unit operating condition is considered abnormal when C2≥0.5; when the telesignal values ​​in the important telesignal quantity set are all 0, the unit operating condition is considered abnormal when C2<0.5 and C3≥0.5; when the telesignal values ​​in the important telesignal quantity set are all 0, the unit operating condition is considered normal when C1<0.5, C2<0.5, and C3<0.5.

[0024] Embodiment 2 like Figure 1-2 As shown, further described in combination with Example 1: S1: Building a time series model based on the industrial Internet platform, specifically including: S11: Select the historical data of telemetry parameters related to the hydro-turbine generator set in the industrial Internet platform to construct training data (such as upper frame vibration, upper guide, lower guide and water guide swing, excitation voltage, oil pressure of the oil pressure system, oil level of the oil pressure tank, etc.), and construct the telemetry at the same time point and its corresponding unit operating condition into a sample data. Construct a continuous data set in the time dimension with a time step of 0.1s. The data set format is ; Where X is the telemetry value, Y is the unit operating condition (normal unit operating condition corresponds to 0, abnormal operating condition corresponds to 1), the data set contains m telemetry attributes, and there are n sample data in total. The telemetry values ​​in the data set are normalized, and the formula is:

[0025] in is the current telemetry value, is the normalized telemetry value, is the lower limit of the current telemetry measurement, It is the upper limit value of telemetry.

[0026] S12: Divide the data set into continuous training sets and validation sets in a ratio of 8:2. The division method is sliding window division, that is, from the first data to the 4n / 5th data, select random data as the breakpoint, take the data and the subsequent n / 5-1 data as the validation set, and the rest of the data as the training set.

[0027] S13: Construct a GRU-based neural network such as Figure 1 As shown: In the figure, (i∈[1,2...,n]) is a piece of data in the training set at a certain time i, Refers to the data of the next moment, and the time interval is 0.1s, which is the time step when the data set is constructed; FC layer is the fully connected layer. GRU is the GRU module; For feature stitching; The attention module uses a fully connected layer to add attention in the embodiment; ' is the output result, which is the confidence level between 0 and 1. GRU is the GRU unit, and its structure is as follows Figure 2 As shown; in is the input of the unit, , is the input gate neuron, for function, is the input gate neuron weight, For feature stitching; , To reset the gate neuron, for its weight; is the memory after input, for function, is the weight; * is the memory at time t.

[0028] S14: Use the training set, validation set, and GRU-based neural network to train the time series model, with batch_size set to 128, hidden layer width hidden_size set to 2, sequence_length set to 8 (i.e., 8 consecutive sets of data are input each time), input_size (input feature dimension) set to m, dropout set to 0.02, training epochs set to n / 32, and stochastic gradient descent as the optimizer, with the cross entropy loss function as the loss function. During the training process, the model parameters with the lowest loss in the validation set are saved.

[0029] S2: Build a decision model based on the data of the Industrial Internet platform, including: S21: Based on the historical data of telemetering related to unit parameters on the industrial Internet platform (such as the opening and closing status of the CCB outlet switch on the high-voltage side of the main transformer, the opening and closing status of the GCB outlet switch of the unit, etc.), discrete training data T is constructed (its format is similar to that of the S11 data set, but there is no time series relationship. It is assumed that there are N samples in total and the number of features is M). According to the type of unit operating condition, the same number of samples of the operating condition are selected to ensure the balance of the samples.

[0030] S22: Use the random forest algorithm to build a decision model, set the random forest size to K, select m features from the training set T with replacement sampling, sample 2N / 3 samples, form a new sub-training set D, and use the sub-training set D to train the CD4.5 decision tree: First, calculate the information entropy of the training set D. Since the unit condition diagnosis is a binary classification task, the information entropy calculation formula is:

[0031] in is the probability of occurrence of the k-th type of sample. For example, in the training set D, there are 1 / 3 abnormal working condition data and 2 / 3 normal working condition data. Then =1 / 3, =2 / 3; Calculate the information entropy of the data attributes in the sub-training set D. For any attribute a, since the telemetering attribute has only three cases (0, 1 and offline), the sub-training set D is divided into three subsets {a1, a2, a3}, and the information gain is calculated as:

[0032] Calculate the information gain rate of any attribute a, the formula is: ; Select the attribute with the highest information gain rate as the partition attribute, and repeat the above steps for the three subsets until no further training is possible, thus training a C4.5 tree. Repeat the training K times to obtain K CD4.5 decision trees, and set the weight of each decision tree to 1 / K to construct a random forest.

[0033] S3: Build a comprehensive model: Use the trained GRU model and the generated best random forest to build a comprehensive model, including: S31: Filter the latest 1-8 data in the time series library, select the same telemetry data as S11 to build the model input, use the trained GRU time series model to predict the unit operating conditions, and obtain the confidence level C1; S32: Filter the current real-time data in the time series library, select the same telesignal quantity as S21 to build the model input, use the trained decision model (random forest) to predict the working condition, and obtain the confidence level C2.

[0034] S33: Repeat S31 and S32 multiple times to obtain a large number of confidence samples and construct paired sets of C1, C2, and Y (true value).

[0035] S34: Build a linear regression model and use the mean square error method to obtain the optimal weights W1 and W2 of C1 and C2, and obtain the bias, C3=W1*C1+W2*C2+bias (C3 is the confidence of the comprehensive model).

[0036] S4: Use the real-time data from the Industrial Internet platform to diagnose the current unit operating conditions, including: S41: Filter the latest 1-8 data in the time series library, select the same telemetry data as S11 to build the model input, use the trained GRU time series model to predict the unit operating conditions, and obtain the confidence level C1; S42: Filter the current real-time data in the time series library, select the same telesignal quantity as S21 to build the model input, use the trained decision model (random forest) to predict the working condition, and obtain the confidence level C2.

[0037] S43: Calculate the comprehensive model confidence C3.

[0038] S44: Filter important telesignal quantities (mechanical overspeed action, shear pin shortening action, water inlet gate sliding down to the accident position, protection trip signal, main pressure regulating valve failure, hydraulic system oil level too low, etc.) to build a collection.

[0039] S44: (1) When the value of a telesignal in the important telesignal set is 1, the unit is considered to be in an abnormal condition; (2) When the values ​​of telesignal in the important telesignal set are all 0, the unit is considered to be in an abnormal condition when C2 ≥ 0.5; (3) When the values ​​of telesignal in the important telesignal set are all 0, the unit is considered to be in an abnormal condition when C2 < 0.5 and C3 ≥ 0.5; (4) When the telesignal values ​​in the important telesignal set are all 0, and when C1, C2, and C3 are all <0.5, the unit is considered to be in normal operating condition.

[0040] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A multi-strategy fusion fault diagnosis method for a hydro-generator set, characterized by: The steps include: S1. Build a time series model based on the industrial Internet platform; S2. Build a decision-making model based on industrial Internet platform data; S3, build a comprehensive model using the trained GRU model and the generated best random forest; S4. Use the real-time data from the industrial Internet platform to perform fault diagnosis on the current unit operating conditions.

2. The multi-strategy fusion fault diagnosis method for a hydro-generator set according to claim 1 is characterized by: The specific method in step S1 includes: S11. Select the historical data of telemetry parameters related to the hydro-turbine generator set in the industrial Internet platform to construct training data, construct the telemetry at the same time point and its corresponding unit operating condition as a sample data, and construct a continuous data set in the time dimension according to a certain time step. The data set format is ； Where X is the telemetry value; Y is the unit operating condition, the normal unit operating condition corresponds to 0, and the abnormal unit operating condition corresponds to 1; the data set contains m telemetry attributes, with a total of n sample data. The telemetry values ​​in the data set are normalized, and the formula is: ； in is the current telemetry value, is the normalized telemetry value, is the lower limit of the current telemetry measurement, is the upper limit of the telemetry measurement; S12. Divide the data set into continuous training sets and validation sets using a sliding window in a ratio of 8:2, build a GRU-based neural network, and use the training set, validation set, and GRU-based neural network to train the time series model.

3. According to the multi-strategy fusion method for fault diagnosis of a hydro-generator set as described in claim 2, it is characterized by: The method for constructing a GRU-based neural network in step S12 is: S13, dividing the data set into continuous training sets and validation sets in a ratio of 8:2, and the division method is sliding window division, that is, from the first data to the 4n / 5th data, select random data as the breakpoint, take the data and the subsequent n / 5-1 data as the validation set, and the remaining data as the training set; S14. Use the training set, validation set and GRU-based neural network to train the time series model. The batch_size is set to 128, the hidden layer width hidden_size is set to 2, the sequence_length is set to 8, the input_size is set to m, the dropout is set to 0.02, the number of training epochs is set to n / 32, and the stochastic gradient descent method is used as the optimizer. The loss function is the cross entropy loss function. The model parameters with the lowest loss on the validation set are saved during the training process.

4. According to the multi-strategy fusion fault diagnosis method of a hydro-generator set as described in claim 1, it is characterized by: The specific method in step S2 includes: Based on the historical data of telemetering related to the parameters of the unit on the industrial Internet platform, discrete training data T is constructed. The random forest algorithm is used to build a decision model. The size of the random forest is set to K. A new sub-training set D is formed based on the training data T. The decision tree is trained using the sub-training set D to construct a random forest decision model.

5. According to the multi-strategy fusion fault diagnosis method of a hydro-generator set according to claim 4, it is characterized in that: include: S21: Based on the historical data of telemetering related to the parameters of the unit on the industrial Internet platform, discrete training data T is constructed, wherein the same number of samples of the operating condition are selected according to the type of unit operating condition to ensure the balance of the samples; S22: Use the random forest algorithm to build a decision model, set the random forest size to K, select m features from the training set T with replacement sampling, sample 2N / 3 samples, form a new sub-training set D, and use the sub-training set D to train the CD4.5 decision tree: First, calculate the information entropy of the training set D. Since the unit condition diagnosis is a binary classification task, the information entropy calculation formula is: Where p_k is the probability of occurrence of the kth type of sample. For example, if there are 1 / 3 abnormal working condition data and 2 / 3 normal working condition data in the sub-training set D, then p_abnormal = 1 / 3, p_normal = 2 / 3; calculate the information entropy of the data attributes in the sub-training set D. For any attribute a, since there are only three cases (0, 1 and offline) of the telesignaling attribute, the sub-training set D is divided into three subsets {a1, a2, a3} through it, and the information gain is calculated as: Calculate the information gain rate of any attribute a, formula: The attribute with the highest information gain rate is selected as the partition attribute, and the above steps are repeated for the three subsets until it cannot be carried out. That is, a C4.5 tree is trained. Repeat the training K times to obtain K CD4.5 decision trees. Set the weight of each decision tree to 1 / K to construct a random forest.

6. According to the multi-strategy fusion fault diagnosis method of a hydro-generator set as described in claim 1, it is characterized by: The specific steps of step S3 are: S3: Build a comprehensive model: Use the trained GRU model and the generated best random forest to build a comprehensive model, including: S31: Filter the latest 1-8 data in the time series library, select the same telemetry data as S11 to build the model input, use the trained GRU time series model to predict the unit operating conditions, and obtain the confidence level C1; S32: Filter the current real-time data in the time series library, select the same telesignal as S21 to build the model input, use the trained decision model to predict the working condition, and obtain the confidence C2, where the decision model is a random forest decision model; S33: Repeat S31 and S32 multiple times to obtain a large number of confidence samples and construct paired sets of C1, C2, and Y; S34: Build a linear regression model and use the mean square error method to obtain the optimal weights W1 and W2 of C1 and C2, and obtain the bias, C3=W1*C1+W2*C2+bias.

7. According to the multi-strategy fusion fault diagnosis method of a hydro-generator set as described in claim 1, it is characterized by: The specific steps of step S4 are: Use the trained time series model to predict the unit operating conditions and obtain the confidence level C1; Use the trained random forest decision model to predict the working condition and obtain the confidence level C2; And calculate the comprehensive model confidence C3; Construct an important telesignaling quantity set; when a telesignaling quantity value in the important telesignaling quantity set is 1, it is considered that the unit operating condition is abnormal; When the telesignal values ​​in the important telesignal set are all 0, the unit operating condition is considered abnormal when C2≥0.5; when the telesignal values ​​in the important telesignal set are all 0, the unit operating condition is considered abnormal when C2<0.5 and C3≥0.5; When the telesignal values ​​in the important telesignal set are all 0, the unit is considered to be in normal operating condition when C1<0.5, C2<0.5, and C3<0.5.

Images