Water turbine wear fault diagnosis method based on deep learning model

Abstract

The invention discloses a water turbine wear fault diagnosis method based on a deep learning model, and belongs to the technical field of water turbine fault signal processing. Comprising the following steps: acquiring a water turbine wear fault sound signal; a Tent chaotic mapping function is adopted to optimize an initial population of a white shark optimization algorithm WSO; introducing a long and short term memory (LSTM) network model to solve the limitation of a convolutional neural network (CNN) in processing a long sequence problem, and constructing a CNN-LSTM water turbine wear fault diagnosis model; an improved white shark optimization algorithm IWSO is introduced into the CNN-LSTM model, and hyper-parameters of the model are optimized, so that the diagnosis precision is improved, and the calculation amount is reduced; and importing the collected fault signal as an input into the model for training and testing to obtain a final diagnosis result. Diagnosis and identification of the water turbine wear fault signal are completed through the steps, the deep learning model has good robust performance, the diagnosis precision of the water turbine wear fault can be improved, and the deep learning model serves as a beneficial supplement of an existing water turbine state monitoring and fault diagnosis system.

Description

Technical Field

[0001] The present invention relates to the technical field of hydraulic turbine fault diagnosis, and in particular to a hydraulic turbine wear fault diagnosis method based on a deep learning model. Background Art

[0002] The function of a turbine is to convert the potential energy of water into mechanical energy. It is the core equipment for the utilization of water resources, and its reliable and safe operation is the foundation. Sediment erosion is an inevitable process in a hydropower station, and sediment wear and erosion are one of the important causes of turbine failure. Under the continuous impact and friction of sediment particles, the surface of the turbine runner will be worn and damaged, thereby changing the flow trajectory inside the turbine, causing the unit efficiency and stability to decrease, the plant noise to increase, the vibration to intensify, and the maintenance cost to increase. If the damage is serious, it will pose a serious threat to the safety of the hydropower station units and operators.

[0003] Traditional turbine fault diagnosis methods rely heavily on manual feature extraction and domain expert knowledge, which not only increases time and cost, but also may lead to subjectivity and incompleteness in feature selection. In addition, traditional methods have limited performance when processing complex, nonlinear and large-scale data, and often have difficulty in accurately identifying and classifying fault types, thus affecting the operational safety and reliability of turbines.

[0004] The turbine fault diagnosis method based on deep learning can automatically extract high-level features from raw data, reduce reliance on feature engineering, and perform well in processing complex and nonlinear data, significantly improving the accuracy and robustness of diagnosis. In addition, the deep learning method can also combine multiple data sources for comprehensive analysis to further improve the comprehensiveness and accuracy of fault diagnosis. Among them, convolutional neural network CNN is widely used in fault diagnosis of rotating machinery due to its powerful automatic feature extraction ability. The network shows excellent performance in identifying the health status of rotating machinery. However, CNN can only process static features in the data, which is a limitation for processing lengthy time series problems. Although increasing the number of nodes and network layers in the neural network can improve the learning and fitting effects, it will lead to a significant increase in the number of hyperparameters. In order to evaluate each hyperparameter combination, a large number of iterative calculations are required, which consumes a lot of time and energy. Therefore, it is very necessary to optimize the hyperparameters. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a turbine wear fault diagnosis method based on a deep learning model. Firstly, a long short-term memory network model LSTM is introduced to process the long sequence time data in the convolutional neural network model; at the same time, the Tent chaotic mapping function is used to optimize the initial population of the white shark optimization algorithm WSO to improve the uniformity and randomness of the initial population; and IWSO is used for hyperparameter optimization of the CNN-LSTM model; finally, the turbine wear fault is diagnosed and identified through the proposed deep learning fault diagnosis model.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for diagnosing wear faults of a hydraulic turbine based on a deep learning model comprises the following steps: Step 1, obtain the sound signal of turbine wear fault; Step 2, using the Tent chaotic mapping function to optimize the initial population of the white shark optimization algorithm WSO, and obtain the improved white shark optimization algorithm IWSO; Step 3, introduce the long short-term memory network model LSTM to solve the limitations of convolutional neural network CNN in processing long sequence problems, and construct a CNN-LSTM turbine wear fault diagnosis model; Step 4: Use the algorithm in Step 2 to optimize the hyperparameters of the CNN-LSTM model in Step 3. Step 5: Import the fault signal in Step 1 as input to the IWSO-CNN-LSTM model in Step 4, train and test the model, and obtain the turbine wear fault diagnosis result.

[0007] In the above-mentioned Step 1, acoustic signals of the normal operating condition and the wear condition of the turbine are collected.

[0008] The above-mentioned Step 2 to obtain the White Shark Optimization Algorithm IWSO includes the following steps: Step 2.1, the speed at which the white shark moves towards its prey; the swimming speed is: ; in, i It's the white shark population, ; is the shrinkage factor, used to enhance the robustness of the algorithm, ; is the speed of the i-th white shark at the k-th iteration; and are scale parameters, respectively controlling and , , ; is the optimal position of the white shark after the kth iteration; is the position of the i-th white shark at the k-th iteration; and is a random number, the value range is ; is the i-th best position vector of the population; is the i-th index vector of the best-positioned white shark, ; Step 2.2, position update to move towards the best prey; the position update strategy of the white shark moving towards the prey is expressed as: ; in, It is The new position of the i-th white shark at the iteration; is the negation operator; is a logical vector; u and l represent the upper and lower limits of the search space respectively; a and b are one-dimensional binary vectors; f represents the frequency of the white shark's wave-like motion; Indicated in Random numbers generated within the range of uniform distribution; mv Indicates the sensory intensity of the white shark. mv When the value is smaller, the WSO algorithm Conduct a smaller local search nearby; when mv When the value is larger, the WSO algorithm is far away from Conduct a wider search in the area; Step 2.3, Cluster behavior; the behavior of white sharks moving toward the best white shark position: ; in, Iterate to the The new position of the ith white shark relative to the prey; , and They are A random number in a range; Control the direction and randomness of local search, and the probability of taking 1 and -1 is the same; is the distance between the prey and the white shark; Represents the olfactory and visual parameters of a white shark group targeting a white shark in the best position.

[0009] In the above Step 2, the Tent chaotic mapping function is used to optimize the initial population of the WSO algorithm and improve the convergence ability and local search ability of the algorithm. The expression is as follows: ; in, is the position of the current population individual; It is the position of the next population individual generated by tent mapping.

[0010] The above Step 3 specifically includes the following steps: Step 3.1, take the turbine wear fault signal as input, import it into the convolutional neural network for convolution CONV calculation, and use the convolution kernel to adaptively extract fault features. The convolution operation is as follows: ; in, yes l Layer i Features, f is the activation function, It is l Layer i convolution kernel weight matrix, operator represents the convolution operation, It is The output of the layer, is the bias term; Step 3.2, import the convolved fault feature data as input into the pooling layer to reduce model parameters and suppress overfitting while retaining the salient features of the output data; the expression is as follows: ; in, Represents the maximum pooling after l +1 level i Elements in a feature; Indicates l Tier i The elements of features in the pooling kernel, It is j Pooling area; Step 3.3, after dimensionality reduction, the processed data is imported into the LSTM layer, and the LSTM model is trained and learned through three gate structures: input gate, output gate and forget gate; The input gate is used to update the unit state. It determines whether the new information can be remembered by the unit. The input gate expression is as follows: ; in, is the input gate output; is the weight matrix between the input and the input gate; is the weight matrix between the historical output and input gates; is the bias term of the input gate; The forget gate filters the memory content and decides whether the stored data should be retained or discarded. Its expression is as follows: ; in, is the forget gate output; Represents the activation function, usually the Sigmoid function, the output range , used to indicate the degree of opening of the Gate; is the weight matrix between the current input and the forget gate; is the current input state; is the weight matrix between the historical output and the forget gate; is the output state at the previous moment; is the bias term of the forget gate; The output gate is the final output of the cell state. After the output of the output gate is multiplied by the current cell state information, the cell output is obtained through the tanh function. The output of the output gate is expressed as: ; in, is the weight matrix between the input and output gates; is the weight matrix between the historical output and the output gate; is the bias term of the output gate.

[0011] Step 3.4, use the Softmax activation function to classify faults and finally complete the model training phase. The Softmax function expression is as follows: ; in, The output layer j The output of a neuron; m is the number of categories of turbine failure severity; It is the probability output of the neuron after the Softmax function; Step 3.5: Test the effectiveness of the deep learning training model in the test set, output the results and end the diagnosis.

[0012] After the IWSO algorithm is optimized in Step 4 above, the optimal combination of hyperparameters of CNN-LSTM is output as feature_num=[27,93], LayerSizes=47, and Stride=2.

[0013] Before importing the fault signal in the above Step 5, the signal data is sorted and divided into fixed segments. The invention provides a method for diagnosing turbine wear faults based on a deep learning model, which has the following beneficial effects: 1. Using sound signal method to diagnose and identify turbine wear faults is applicable to the complex structure inside the hydropower unit, and can refine the acoustic signal characteristics of key components and detect and analyze them.

[0014] 2. The turbine fault diagnosis method based on deep learning can automatically extract high-level features from raw data, reduce dependence on feature engineering, and perform well in processing complex and nonlinear data, significantly improving the accuracy and robustness of diagnosis.

[0015] 3. Introducing the long short-term memory network model LSTM to process the long sequence time data in the convolutional neural network model CNN can simultaneously extract spatial features and time features, thereby improving the processing effect of complex time series data and improving the accuracy and robustness of fault diagnosis. Using the Tent chaos mapping function to optimize the initial population of the White Shark Optimization Algorithm WSO can increase the randomness and uniformity of the population, and further improve the convergence speed and accuracy of the algorithm. Through IWSO to optimize the hyperparameters of the CNN-LSTM model, it can reduce the time and effort of manual parameter adjustment, improve efficiency, and effectively avoid falling into the local optimum, thereby improving the effect of hyperparameter optimization and the final performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 are the acoustic signals of the turbine under different working conditions in Example 1; Figure 3 It is the iterative process of optimizing the algorithm; Figure 4 The comparison curve of training results of different deep learning fault diagnosis models for turbines; Figure 5 This is a comparison curve of the test results of different deep learning fault diagnosis models for turbines. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0018] like Figure 1 As shown in , a turbine wear fault diagnosis method based on deep learning model includes the following steps: Step 1, obtain the sound signal of turbine wear failure; in this step, the sound signal of sand-containing water flow impacting the turbine runner blades is collected; because sand wear and erosion is one of the important causes of turbine failure; Step 2, using the Tent chaotic mapping function to optimize the initial population of the white shark optimization algorithm WSO, and obtain the improved white shark optimization algorithm IWSO; Step 3, introduce the long short-term memory network model LSTM to solve the limitations of convolutional neural network CNN in processing long sequence problems, and construct a CNN-LSTM turbine wear fault diagnosis model; Step 4: Use the algorithm in Step 2 to optimize the hyperparameters of the CNN-LSTM model in Step 3. Step 5: Import the fault signal in Step 1 as input to the IWSO-CNN-LSTM model in Step 4, train and test the model, and obtain the turbine wear fault diagnosis result.

[0019] In the above-mentioned Step 1, acoustic signals of the normal operating condition and the wear condition of the turbine are collected.

[0020] The above-mentioned Step 2 to obtain the White Shark Optimization Algorithm IWSO includes the following steps: Step 2.1. The speed at which the white shark moves towards its prey: The great white shark uses its sensitive senses to track its prey. When the prey moves, it creates ripples in the water. The white shark senses the location of the prey based on the sound of the ripples and swims towards the prey in a wave-like manner. The swimming speed is: ; in, i It's the white shark population, ; is the shrinkage factor, used to enhance the robustness of the algorithm, ; is the speed of the i-th white shark at the k-th iteration; and are scale parameters, respectively controlling and , , ; is the optimal position of the white shark after the kth iteration; is the position of the i-th white shark at the k-th iteration; and is a random number, the value range is ; is the i-th best position vector of the population; is the i-th index vector of the best-positioned white shark, ; Step 2.2, position update to move towards the best prey; white sharks usually spend a lot of time looking for potential prey, so its position is constantly changing; when the white shark hears the waves generated by the prey or smells the scent of the prey during the search process, it will move towards the location of the prey; in some cases, the prey will leave the original position; the white shark will track the prey again based on the smell left by the prey or the waves generated by the swimming; in this case, the position update strategy of the white shark moving towards the prey is expressed as: ; in, It is The new position of the i-th white shark at the iteration; is the negation operator; is a logical vector; u and l represent the upper and lower limits of the search space respectively; a and b are one-dimensional binary vectors; f represents the frequency of the white shark's wave-like motion; Indicated in Random numbers generated within the range of uniform distribution; mv Indicates the sensory intensity of the white shark. mv When the value is smaller, the WSO algorithm Conduct a smaller local search nearby; when mv When the value is larger, the WSO algorithm is far away from Conduct a wider search in the area; Step 2.3, Cluster behavior; white sharks are highly social animals, and they will cooperate highly when hunting. The white shark group will move towards the white shark closest to the prey; the following formula simulates the behavior of the white shark group moving towards the best white shark position: ; in, Iterate to the The new position of the ith white shark relative to the prey; , and They are A random number in a range; Control the direction and randomness of local search, and the probability of taking 1 and -1 is the same; is the distance between the prey and the white shark; Represents the olfactory and visual parameters of a white shark group targeting a white shark in the best position.

[0021] In the above Step 2, the Tent chaotic mapping function is used to optimize the initial population of the WSO algorithm and improve the convergence ability and local search ability of the algorithm. The expression is as follows: ; in, is the position of the current population individual; It is the position of the next population individual generated by tent mapping.

[0022] Tent mapping is a typical chaotic mapping function, which iterates an initial value according to a certain rule to obtain a series of pseudo-random numbers; these values ​​show chaotic characteristics, with good distribution and uncertainty; the Tent mapping function is used to optimize the initial population position of the WSO algorithm; by introducing the Tent mapping, the position of the population will be iterated according to the chaotic characteristics, thereby effectively improving the overall performance of the algorithm.

[0023] The above Step 3 specifically includes the following steps: Step 3.1, take the turbine wear fault signal as input, import it into the convolutional neural network for convolution CONV calculation, and use the convolution kernel to adaptively extract fault features. The convolution operation is as follows: ; in, yes l Layer i Features, f is the activation function, It is l Layer i convolution kernel weight matrix, operator represents the convolution operation, It is The output of the layer, is the bias term; Step 3.2, import the convolved fault feature data as input into the pooling layer, reduce model parameters and suppress overfitting, while retaining the salient features of the output data; commonly used pooling layers include average pooling and maximum pooling. This article uses the maximum pooling layer, and the expression is as follows: ; in, Represents the maximum pooling after l +1 level i Elements in a feature; Indicates l Tier i The elements of features in the pooling kernel, It is j Pooling area; The maximum pooling layer aggregates the extracted data features, reduces the data dimension and retains the main features; Step 3.3, after dimensionality reduction, the processed data is imported into the LSTM layer. The LSTM model is mainly trained and learned through three gate structures: input gate, output gate and forget gate; The input gate is used to update the unit state. It determines whether the new information can be remembered by the unit. The input gate expression is as follows: ; in, is the input gate output; is the weight matrix between the input and the input gate; is the weight matrix between the historical output and input gates; is the bias term of the input gate; The forget gate filters the memory content and decides whether the stored data should be retained or discarded. Its expression is as follows: ; in, is the forget gate output; Represents the activation function, usually the Sigmoid function, the output range , used to indicate the degree of opening of the Gate; is the weight matrix between the current input and the forget gate; is the current input state; is the weight matrix between the historical output and the forget gate; is the output state at the previous moment; is the bias term of the forget gate; The output gate is the final output of the cell state. After the output of the output gate is multiplied by the current cell state information, the cell output is obtained through the tanh function. The output of the output gate is expressed as: ; in, is the weight matrix between the input and output gates; is the weight matrix between the historical output and the output gate; is the bias term of the output gate.

[0024] The neural network is further trained in this step, during which the network learns to autonomously identify features associated with faults; Step 3.4, use the Softmax activation function to classify faults and finally complete the model training phase. The Softmax function expression is as follows: ; in, The output layer j The output of a neuron; m is the number of categories of turbine failure severity; It is the probability output of the neuron after the Softmax function; Step 3.5: Test the effectiveness of the deep learning training model in the test set, output the results and end the diagnosis.

[0025] In the above Step 4, the IWSO algorithm optimizes the hyperparameters of the CNN-LSTM model. After 50 iterative calculations, as the speed and position are continuously updated, the population continues to evolve, and the fitness function value of the optimal individual continues to decrease. The IWSO algorithm can complete convergence in less than 10 iterations and will not fall into the local optimum. After the IWSO algorithm is optimized, the optimal combination of hyperparameters of CNN-LSTM is output as feature_num=[27,93], LayerSizes=47, and Stride=2.

[0026] Before importing the fault signal in the above Step 5, the signal data needs to be sorted and divided into fixed segments. Each sample contains 1024 data points; each working condition contains 70 groups of samples; samples are randomly selected to divide the training set and the test set to form the final data set, as shown in the following table: Table 1

[0027] Embodiment 1: The method of the present invention is experimentally illustrated in combination with a specific use example. The parameters of the hydro-generator set in this embodiment are: rated power of 5kW, rated frequency of 50Hz, rated speed of 1500r / min, rated voltage of 230V, rated current of 21.8A, power factor of 1.0, and the structural form is steel structure. The parameters of the acoustic sensor are: sampling rate of 48kHz, standard measurement range of 25-130dBA, measurement frequency range of 10-20000Hz, and communication interface of USB Audio + USB HID.

[0028] The experiment content is as follows: This experiment uses acoustic sensors to collect acoustic signals of turbines under normal working conditions and sand-containing working conditions. The fault signals are classified by using CNN model, LSTM model, CNN-LSTM model and IWSO-CNN-LSTM model described in Example 1 of the present invention to diagnose faults, and compare the diagnostic effects of different deep learning methods on turbine wear faults: The present invention diagnoses turbine wear faults according to the following steps: Step 1: Collect the wear fault signal of the turbine impacted by the sand-containing water flow, such as Figure 2 As shown in the figure, it can be seen that the waveforms of the three working conditions are chaotic; Step 2: Use Tent chaotic mapping function to optimize the initial population of the white shark optimization algorithm WSO to obtain the improved white shark optimization algorithm IWSO; The WSO algorithm has good optimization performance in low-dimensional optimization problems, but has low convergence accuracy in high-dimensional optimization problems. It is easy to fall into local extreme values ​​and converge prematurely. This paper proposes an improved WSO algorithm to improve the optimization speed and accuracy of the original algorithm. The chaos mapping algorithm has the characteristics of sensitivity to initial conditions and chaotic ergodicity. The chaos optimization algorithm is evolved based on these characteristics. During optimization, the chaos optimization algorithm can filter out local optimal solutions and converge quickly. The initial population of the WSO algorithm is generated randomly, which will cause the initial solutions to be unevenly distributed and lack diversity in the search domain. To this end, the tent chaos optimization algorithm is introduced to initialize the population, replace the random value selection process, and improve the convergence ability and local search ability of the WSO algorithm.

[0029] Step 3: Introduce the long short-term memory network model LSTM to solve the limitations of the convolutional neural network CNN in processing long sequence problems, and construct a CNN-LSTM turbine wear fault diagnosis model; In the CNN-LSTM fault diagnosis model, CNN is responsible for extracting the spatial features of the input data and reducing the data dimension. LSTM is responsible for mining the hidden temporal features in the data, and its long-term memory characteristics make it more suitable for data classification processing. The entire model includes the acoustic vibration signal input layer, CNN convolution layer, pooling layer, LSTM layer, classification layer and output layer. By combining CNN and LSTM in a synergistic manner, the model can give full play to their respective advantages and realize the diagnosis of turbine wear faults. The main steps of the diagnosis process are as follows: (1) Collect the acoustic vibration signals of turbine faults, classify and segment the signals, and form a fixed segmented data sample set; (2) Input the data sample set into the CNN convolution layer and use the convolution kernel to adaptively extract fault features; (3) After the convolution layer, the extracted features will be passed through the pooling layer for maximum pooling operation. This operation can reduce the data dimension of the features while retaining the main feature information; (4) The feature data after dimensionality reduction will be used as the input of the LSTM layer to train the neural network and automatically learn the fault characteristics; (5) Finally, the Softmax activation function is used to classify the turbine wear fault features.

[0030] The above steps outline the main process of the CNN-LSTM model in turbine fault diagnosis. By giving full play to the advantages of CNN and LSTM, combined with the ability of spatial and temporal feature extraction, features are extracted from fault signals and fault classification is performed.

[0031] Step 4: Use the algorithm in step 2 to optimize the hyperparameters of the CNN-LSTM model in step 3; the optimization iteration process curve is as follows: Figure 3As shown, the WSO algorithm that has not been optimized by Teng's chaotic mapping function is selected as a comparison. It can be seen that with the continuous update of speed and position, the population continues to evolve, and the fitness function value of the optimal individual continues to decrease. After 50 iterative calculations, the IWSO algorithm can complete convergence in less than 10 iterations. The unoptimized WSO algorithm requires about 13 iterations to complete convergence, and the number of convergences is greater than that of the IWSO algorithm. And the WSO algorithm falls into a local optimum. This figure shows that the use of the IWSO algorithm can have higher work efficiency and more accurate calculation results. At this time, the optimal combination of hyperparameters is output as feature_num=[27,93], LayerSizes=47, Stride=2; Step 5: Import the fault signal in step 1 as input into the IWSO-CNN-LSTM model in step 4, train and test the model to obtain the turbine wear fault diagnosis result. In order to highlight the effectiveness of the method proposed in Example 1, CNN, LSTM and CNN-LSTM models are selected as comparison models.

[0032] The training results are as follows Figure 4 As shown in the figure, it can be seen that after 50 iterations, the IWSO-CNN-LSTM model performed outstandingly in terms of accuracy, reaching 97.2%, which is higher than the CNN, LSTM and CNN-LSTM models. In contrast, the LSTM model has the lowest accuracy, only 68.7%. Although the accuracy of the CNN model is higher than that of the LSTM model, the curve fluctuates greatly and the smoothness is poor. However, after the introduction of the LSTM model, the accuracy of the CNN-LSTM model reached 83.7%, which shows that the introduction of LSTM effectively overcomes the limitations of the CNN model in dealing with long sequence problems and improves the stability and accuracy of the model. The deep learning model proposed in Example 1 stands out for its excellent fault diagnosis performance. The model effectively combines the characteristics of CNN and LSTM, and combines the IWSO algorithm to search for the best hyperparameter combination, thereby improving the accuracy of the model and enhancing its stability.

[0033] Test results such as Figure 5 As shown in the figure, it can be seen that the accuracy of the LSTM model is only 65.3%, and it performs poorly in the diagnosis of working condition 2. In contrast, the accuracy of the CNN model reaches 77.8%, which is significantly higher than the LSTM model. The accuracy of the CNN-LSTM model is 84.2%. By applying the White Shark optimization algorithm to optimize the CNN-LSTM model, the accuracy of the IWSO-CNN-LSTM model reaches 96.2%, and it achieves 100% accuracy in the diagnosis of working conditions 1 and 3. This result verifies the excellent performance of the proposed method.

[0034] In summary, the method described in the embodiment of the present invention combines the two deep learning methods of CNN and LSTM, and introduces the improved WSO algorithm to optimize its hyperparameters, which can not only solve the limitations of CNN in processing long time series, but also maximize the robustness of the model, reduce the amount of calculation, and improve the diagnostic accuracy of turbine wear faults.

[0035] Although the invention has been described herein with reference to a number of illustrative embodiments thereof, it should be understood that numerous other modifications and embodiments can be devised by those skilled in the art that will fall within the scope of this disclosure.

Claims

1. A turbine wear fault diagnosis method based on deep learning model, characterized in that: The following steps are involved: Step 1, obtain the sound signal of turbine wear fault; Step 2, using the Tent chaotic mapping function to optimize the initial population of the white shark optimization algorithm WSO, and obtain the improved white shark optimization algorithm IWSO; Step 3, introduce the long short-term memory network model LSTM to solve the limitations of convolutional neural network CNN in processing long sequence problems, and construct a CNN-LSTM turbine wear fault diagnosis model; Step 4: Use the algorithm in Step 2 to optimize the hyperparameters of the CNN-LSTM model in Step 3. Step 5: Import the fault signal in Step 1 as input to the IWSO-CNN-LSTM model in Step 4, train and test the model, and obtain the turbine wear fault diagnosis result.

2. A method for diagnosing turbine wear faults based on a deep learning model according to claim 1, characterized in that: In the above-mentioned Step 1, acoustic signals of the normal operating condition and the wear condition of the turbine are collected.

3. The method for diagnosing turbine wear faults based on a deep learning model according to claim 1, characterized in that: The step 2 to obtain the White Shark Optimization Algorithm IWSO includes the following steps: Step 2.1, the speed at which the white shark moves towards its prey; the swimming speed is: ； in, i It's the white shark population. ; is the shrinkage factor, used to enhance the robustness of the algorithm, ; is the speed of the i-th white shark at the k-th iteration; and are scale parameters, respectively controlling and , , ; is the optimal position of the white shark after the kth iteration; is the position of the i-th white shark at the k-th iteration; and is a random number, the value range is ; is the i-th best position vector of the population; is the i-th index vector of the best-positioned white shark, ; Step 2.2, position update to move towards the best prey; the position update strategy of the white shark moving towards the prey is expressed as: ； in, It is The new position of the i-th white shark at the iteration; is the negation operator; is a logical vector; u and l represent the upper and lower limits of the search space respectively; a and b are one-dimensional binary vectors; f represents the frequency of the white shark's wave-like motion; Indicated in Random numbers generated within the range of uniform distribution; mv Indicates the sensory intensity of the white shark. mv When the value is smaller, the WSO algorithm Conduct a smaller local search nearby; when mv When the value is larger, the WSO algorithm is far away from Conduct a wider search in the area; Step 2.3, Cluster behavior; the behavior of white sharks moving toward the best white shark position: ； in, Iterate to the The new position of the ith white shark relative to the prey; , and They are A random number in a range; Control the direction and randomness of local search, and the probability of taking 1 and -1 is the same; is the distance between the prey and the white shark; Represents the olfactory and visual parameters of a white shark group targeting a white shark in the best position.

4. A method for diagnosing turbine wear faults based on a deep learning model according to claim 3, characterized in that: In the Step 2, the Tent chaotic mapping function is used to optimize the initial population of the WSO algorithm and improve the convergence ability and local search ability of the algorithm. The expression is as follows: ； in, is the position of the current population individual; It is the position of the next population individual generated by tent mapping.

5. The method for diagnosing turbine wear faults based on a deep learning model according to claim 3, characterized in that: The Step 3 specifically includes the following steps: Step 3.1, take the turbine wear fault signal as input, import it into the convolutional neural network for convolution CONV calculation, and use the convolution kernel to adaptively extract fault features. The convolution operation is as follows: ； in, yes l Layer i Features, f is the activation function, It is l Layer i convolution kernel weight matrix, operator represents the convolution operation, It is The output of the layer, is the bias term; Step 3.2, import the convolved fault feature data as input into the pooling layer to reduce model parameters and suppress overfitting while retaining the salient features of the output data; the expression is as follows: ； in, Represents the maximum pooling after l +1 level i Elements in a feature; Indicates l Tier i The elements of features in the pooling kernel, It is j Pooling area; Step 3.3, after dimensionality reduction, the processed data is imported into the LSTM layer, and the LSTM model is trained and learned through three gate structures: input gate, output gate and forget gate; The input gate is used to update the unit state. It determines whether the new information can be remembered by the unit. The input gate expression is as follows: ； in, is the input gate output; is the weight matrix between the input and the input gate; is the weight matrix between the historical output and input gates; is the bias term of the input gate; The forget gate filters the memory content and decides whether the stored data should be retained or discarded. Its expression is as follows: ； in, is the forget gate output; Represents the activation function, usually the Sigmoid function, the output range , used to indicate the degree of opening of the Gate; is the weight matrix between the current input and the forget gate; is the current input state; is the weight matrix between the historical output and the forget gate; is the output state at the previous moment; is the bias term of the forget gate; The output gate is the final output of the cell state. After the output of the output gate is multiplied by the current cell state information, the cell output is obtained through the tanh function. The output of the output gate is expressed as: ； in, is the weight matrix between the input and output gates; is the weight matrix between the historical output and the output gate; is the bias term of the output gate; Step 3.4, use the Softmax activation function to classify faults and finally complete the model training phase. The Softmax function expression is as follows: ； in, The output layer j The output of a neuron; m is the number of categories of turbine failure severity; It is the probability output of the neuron after the Softmax function; Step 3.5: Test the effectiveness of the deep learning training model in the test set, output the results and end the diagnosis.

6. The method for diagnosing wear faults of turbines based on a deep learning model according to claim 1, characterized in that: After the IWSO algorithm is optimized in Step 4, the optimal combination of hyperparameters of CNN-LSTM is output as feature_num=[27,93], LayerSizes=47, and Stride=2.

7. The method for diagnosing wear faults of turbines based on a deep learning model according to claim 1, characterized in that: Before importing the fault signal in the Step 5, the signal data is sorted and divided into fixed segments.

Images