An experimental apparatus and method for predicting the remaining life of offshore wind turbine gearboxes.

Abstract

This invention discloses an experimental apparatus and method for predicting the remaining life of offshore wind turbine gearboxes. The experimental apparatus includes a gearbox vibration module, a vibration signal amplifier, and a base. The vibration signal amplifier includes an accelerometer, a transmission rod, and an output shaft coupling. The vibration signal amplifier is connected to the planetary gearbox via the output shaft coupling. This invention utilizes a convolutional neural network to fuse acceleration data from an accelerometer and temperature data from a temperature sensor at the same time to construct a life prediction dataset. It then uses a long short-term memory neural network to construct a deep learning remaining life prediction model, selecting the number of network layers, learning rate, batch size, and iteration count for training until the average relative error between the predicted remaining life and the actual remaining life is less than 10%. This invention utilizes a mechanical structure to amplify vibration signal variations, improving the accuracy of monitored minute vibration data and providing data support for gearbox remaining life prediction.

Description

Technical Field

[0001] This invention relates to vibration monitoring devices, and more particularly to a test device and method for predicting the remaining life of offshore wind turbine gearboxes. Background Technology

[0002] Currently, the rapid development of offshore wind power projects is bringing increasing maintenance pressure. For a long time, the maintenance of offshore wind power equipment has relied on planned maintenance or repairs only after a malfunction occurs. The former easily leads to untimely maintenance or premature replacement of parts, causing economic losses, while the latter results in prolonged turbine downtime, causing even greater economic losses.

[0003] Even minor faults in offshore wind turbine gearboxes will manifest in vibration parameters. Early vibration detection and appropriate measures can predict the gearbox's remaining service life, allowing for the development of maintenance plans in advance and reducing equipment wear and maintenance costs. However, revealing the relationship between gearbox vibration signals and remaining service life requires establishing a vibration testing platform for research.

[0004] Regarding vibration testing platforms, Chinese utility model patent "A Vibration Testing Platform Device" (patent number: CN207300523U) discloses an experimental device for monitoring vibration, belonging to the field of monitoring technology. However, this vibration testing platform lacks accuracy for minute vibration data, and how to amplify the vibration signal to improve the accuracy of the monitored minute vibration data has become an urgent technical problem to be solved. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to address the shortcomings of the prior art by providing a test device and method for predicting the remaining life of offshore wind turbine gearboxes. By using a test device that can amplify vibration signals in a simulated natural environment, the vibration data of the gearbox during operation is acquired, and the vibration signals are amplified by a mechanical structure, thereby improving the accuracy of the monitored minute vibration data and providing data support for predicting the remaining life of the gearbox.

[0006] Technical solution: The experimental device for predicting the remaining life of offshore wind turbine gearboxes according to the present invention includes a gearbox vibration module, a vibration signal amplifier, and a base;

[0007] The gearbox vibration module includes a planetary gearbox, an input shaft coupling, a counterweight, a drive motor, and a temperature sensor.

[0008] The vibration signal amplifier includes an accelerometer, a transmission rod, and an output shaft coupling;

[0009] The vibration signal amplifier is connected to the planetary gearbox via an output shaft coupling.

[0010] The method for predicting the remaining life of offshore wind turbine gearboxes according to the present invention includes the following steps:

[0011] (1) Install a temperature sensor in the planetary gearbox and add coolant, and connect the temperature sensor to the computer via Bluetooth.

[0012] (2) Connect the accelerometer to the computer for measurement;

[0013] (3) Load the load onto the input shaft coupling;

[0014] (4) Start the drive motor and adjust the speed to the calculated speed value. The speed formula of the drive motor is n = 9550P / T, where P is the output power of the drive motor and T is the torque of the drive motor. Here, the torque T of the drive motor is the same as the simulated actual load.

[0015] (5) Record and save the data measured by the accelerometer and the temperature sensor;

[0016] (6) Use convolutional neural networks to extract features, fuse acceleration data from the accelerometer and temperature data from the temperature sensor at the same time, and construct a lifetime prediction dataset.

[0017] (7) Construct a deep learning remaining life prediction model using a long short-term memory neural network, select the number of network layers, learning rate, batch and iteration number for training, until the average relative error between the remaining life prediction value output by the remaining life prediction model and the actual life is less than 10%.

[0018] In step (2), the measured value of the accelerometer

[0019] Where F represents the applied force, l represents the length of the transmission rod, EI is the bending stiffness of the transmission rod material, and t is the time it takes for the accelerometer to record one data point.

[0020] The specific process of step (6) is as follows:

[0021] (6.1) The acceleration data generated by the accelerometer and the temperature data generated by the temperature sensor are numbered in row order (1, 2, 3, ..., i) and column order (1, 2, 3, ..., j), respectively, and the acceleration data and temperature data are constructed into a matrix M. ij Then, normalization is performed.

[0022] (6.2) Construct the normalized data into a dataset P = [x ij ];

[0023] (6.3) Construct two convolutional layers and two max pooling layers, arranged as convolutional layer 1, max pooling layer 1, convolutional layer 2, max pooling layer 2;

[0024] (6.4) Using the dataset P constructed in step (6.2) as input, the feature matrix H obtained by fusing the accelerometer data and the temperature sensor data is obtained. t =[h t1 ,h t2 ,...,h ti ].

[0025] In step (6.1), the normalization process uses zero-mean normalization and is performed according to the following formula: Where x ij This represents the normalized data in the i-th row and j-th column, m. ij For the sample data collected in the i-th row and j-th column, m j均 Let σ represent the mean of the data corresponding to the j-th column. j标准差 This represents the standard deviation of the data corresponding to column j.

[0026] In step (7), the input layer of this deep learning remaining lifespan prediction model is the feature matrix H from step (6.4). t =[h t1 ,h t2 ,...,h ti ].

[0027] In step (7), the remaining lifetime prediction model includes an input layer, a hidden layer, and an output layer; the hidden layer includes a forget gate, an input gate, and an output gate.

[0028] In step (7), the output layer of the deep learning remaining lifetime prediction model is the average relative error between the remaining lifetime prediction value and the lifetime of the training set and the test set.

[0029] In step (7), the forget gate is: f t =σ(W f ·[h t-1 ,x t ]+b f ), where f t It is the output of the forget gate, σ is the activation function sigmoid function, and h t-1 It is the output value of the previous cell, x t This is the current input value, W. f It is a learnable weight, b f It is a learnable bias.

[0030] In step (7), the output gate is: i t =σ(W i ·[h t-1,x t ]+b i ), Where i t These are the weighting coefficients for updating information, and σ is the sigmoid activation function. It is the candidate state vector, tanh is the activation function tanh function, W i W c It is the weight of learning, b i and b c It's a deviation.

[0031] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0032] (1) The present invention utilizes a drive motor to drive the input shaft of the planetary gearbox to rotate, simulating the force of wind in the natural environment that drives the wind turbine blades and hub to rotate. This force drives the low-speed shaft of the gearbox to rotate, and then the gearbox vibrates due to the acceleration of the gearbox.

[0033] (2) The test apparatus of the present invention provides an experimental platform for the correlation between vibration and gearbox remaining life, which helps to reveal the correspondence between vibration and gearbox remaining life and provides theoretical and data support for the remaining life of offshore wind turbine gearboxes.

[0034] (3) This test device utilizes the increased deflection at the end of the transmission shaft to increase the vibration signal, thereby improving the accuracy of the monitored minute vibration data. Attached Figure Description

[0035] Figure 1 This is an exploded view of the experimental apparatus for predicting the remaining life of offshore wind turbine gearboxes according to the present invention.

[0036] Figure 2 This is a three-dimensional structural schematic diagram of the experimental device of the present invention;

[0037] Figure 3 This is a two-dimensional cross-sectional view of the vibration signal amplifier of the present invention;

[0038] Figure 4 This is a structural diagram of the planetary gearbox of the present invention;

[0039] in, Figure 4 (a) is a side view of the planetary gearbox; Figure 4 (b) is a sectional view of the planetary gearbox;

[0040] Figure 5 This is a flowchart of the test method of the present invention;

[0041] Figure 6 This is a schematic diagram of the Long Short-Term Memory (LSTM) neural network unit used in this invention;

[0042] Figure 7 This is a schematic diagram of the forgetting gate of the Long Short-Term Memory (LSTM) neural network unit used in this invention;

[0043] Figure 8 This is a schematic diagram of the input gate of the Long Short-Term Memory (LSTM) neural network unit used in this invention;

[0044] Figure 9 This is a schematic diagram of the state update of the Long Short-Term Memory (LSTM) neural network unit used in this invention;

[0045] Figure 10 This is a schematic diagram of the output gate of the Long Short-Term Memory (LSTM) neural network unit used in this invention. Detailed Implementation

[0046] like Figure 1 As shown, the experimental device for predicting the remaining life of offshore wind turbine gearboxes according to the present invention includes a gearbox vibration module 1, a vibration signal amplifier 2, and a base 3. The gearbox vibration module 1 includes a planetary gearbox 1a, an input shaft coupling 1b, a counterweight 1c, a drive motor 1d, and a temperature sensor.

[0047] The vibration signal amplifier 2 includes an acceleration sensor 2a, a transmission rod 2b, an acceleration sensor mounting plate 2c, an output shaft coupling 2d, fastening screws 2e, and screws 2f. The planetary gearbox 1a is fixed to the base 3 using bolts, and the drive motor 1d is fixed to the base 3 using bolts.

[0048] Figure 3 This is a two-dimensional cross-sectional view of vibration signal amplifier 2. (See image below.) Figure 2 , Figure 3 As shown, the vibration signal amplifier 2 is connected to the gearbox output shaft 1a.2 using an output shaft coupling 2d, wherein the output shaft coupling 2d is secured to the gearbox output shaft 1a.2 and the transmission rod 2b using a fastening screw 2e. The acceleration sensor 2a is fixed to the center of the sensor mounting plate 2c at the end of the transmission rod 2b using a screw 2f.

[0049] According to the approximate differential equation of the torsion curve, The measured values ​​of accelerometer 2a were derived. Where F is applied at the end of the drive shaft 2b and perpendicular to it, l is the length of the drive rod 2b, EI is the bending stiffness of the material used for the drive rod 2b, and t is the time it takes for the accelerometer 2a to record one data point. F, EI, and t do not change in a single measurement. According to formula ②, the measured acceleration a of the accelerometer 2a is proportional to the cube of the length l of the drive rod 2b, thus amplifying the vibration signal of the planetary gearbox 1a and enabling the accelerometer 2a to monitor even finer vibration data.

[0050] like Figure 4 As shown, the planetary gearbox 1a includes an input shaft 1a.1, an output shaft 1a.2, a sun gear 1a.3, planet gears 1a.4, and a housing 1a.5. The rotation of the input shaft 1a.1 drives the sun gear 1a.3 to rotate, which in turn drives the planet gears 1a.4 to rotate through gear meshing, and then drives the output shaft 1a.2 to rotate.

[0051] The method for predicting the remaining life of offshore wind turbine gearboxes according to the present invention includes the following steps:

[0052] (1) Install a temperature sensor inside the housing 1a.5 of the planetary gearbox 1a and add coolant, and connect the temperature sensor to the computer via Bluetooth.

[0053] (2) Connect the accelerometer 2a to the computer using a USB cable to perform measurements;

[0054] (3) Select loads 1c of different masses according to experimental requirements and mount them on the load. Figure 2 The mass of load 1c at position 1c is shown. Where M is the rated torque of the gearbox, L is the axial distance from the mounting position to the center of gravity of the gearbox, and g is the acceleration due to gravity.

[0055] (4) Start the drive motor 1d and adjust it to the corresponding speed. The formula for the drive motor speed is n = 9550P / T, where P is the output power of the drive motor and T is the torque of the drive motor. Here, the torque T of the drive motor is the same as the simulated actual load.

[0056] (5) Record and save the data m from the accelerometer 2a and the data n from the temperature sensor;

[0057] (6) Use convolutional neural networks to extract features, fuse the acceleration data m generated by the accelerometer 2a and the temperature data n generated by the temperature sensor at the same time, and construct a lifetime prediction dataset.

[0058] The specific process of step (6) is as follows:

[0059] (6.1) The acceleration data generated by the accelerometer 2a and the temperature data generated by the temperature sensor are numbered in row order (1, 2, 3, ..., i) and in column order (1, 2, 3, ..., j), respectively. The acceleration data and temperature data are then constructed into a matrix M. ij Then, normalization is performed. Normalization uses zero-mean normalization and is performed according to the following formula: Where x ij This represents the normalized data in the i-th row and j-th column, m.ij For the sample data m directly collected in the i-th row and j-th column j均 Let σ represent the mean of the data corresponding to the j-th column. j标准差 It represents the standard deviation of the data corresponding to the j-th column.

[0060] (6.2) Construct the normalized data into a dataset P = [x ij ].

[0061] (6.3) Construct two convolutional layers and two max pooling layers, arranged as Convolutional Layer 1, Max Pooling Layer 1, Convolutional Layer 2, and Max Pooling Layer 2, respectively. Convolutional Layer 1 and Max Pooling Layer 1 have 32 kernels, a stride of 1, and are padded with "same"; Convolutional Layer 2 and Max Pooling Layer 2 have 64 kernels, a stride of 1, and are padded with "same"; the activation function for all layers is "relu".

[0062] (6.4) Using the dataset P constructed in step (6.2) as input, the feature matrix H obtained by fusing the accelerometer data and the temperature sensor data is obtained. t =[h t1 ,h t2 ,...,h ti ].

[0063] (7) Using long short-term memory neural networks as a regression prediction method, a deep learning remaining life prediction model is constructed. The number of network layers, learning rate, batch and number of iterations are selected for training until the average relative error between the remaining life prediction value output by the remaining life prediction model and the actual life is less than 10%.

[0064] Step (7) includes the following process:

[0065] (7.1) Construct a remaining lifetime prediction model based on a long short-term memory neural network (LSTM). The remaining lifetime prediction model includes an input layer, a hidden layer, and an output layer. The input layer of the model contains the feature matrix in step (6.4); the output layer is the average relative error between the remaining lifetime prediction value described in 7.2 and the actual lifetime of the training set and the test set; the hidden layer contains three gate states, namely the forget gate, the input gate, and the output gate.

[0066] like Figure 6 As shown, a Long Short-Term Memory (LSTM) neural network unit contains three gates to control the cell state. These three gates are called the forget gate, the input gate, and the output gate.

[0067] like Figure 7The diagram shows a forgetting gate. Before transmitting the acceleration and temperature sensor signals from the previous moment to the next moment from the offshore wind turbine gearbox, a selective discarding process is often required. This embodiment uses a forgetting gate to accomplish this operation. (The last sentence, "h," appears to be a fragment and doesn't translate directly. It's left as is.) t-1 and x t Substituting the values ​​into the forgetting gate formula yields a vector belonging to the range [0, 1], where the value of the vector represents the previous cell state C. t-1 This indicates how much information is retained or discarded. 0 represents no retention, 1 represents all retention, and values ​​from 0 to 1 represent some retention. The forget gate formula is f. t =σ(W f ·[h t-1 ,x t ]+b f ), where f t This is the output of the forget gate, σ is the activation function sigmoid, whose output value is in [0, 1], h t-1 It is the output value of the previous cell, x t This is the current input value, W. f It is a learnable weight and b f It is a learnable bias.

[0068] like Figure 8 The diagram shows an input gate. This step determines which operational data of the offshore wind turbine gearbox needs to be retained at the current moment and which new information needs to be added to the unit status. This operation is completed through the input gate. t It is the weighting coefficient for updating information, h t-1 and x t Substituting into the first formula below, we obtain the result, and then using the activation function tanh, we apply h. t-1 and x t Generate new candidate state vectors The corresponding formulas for the output gate are: i t =σ(W i ·[h t-1 ,x t ]+b i ), Where i t These are the weighting coefficients for updating information, and σ is the sigmoid activation function. It is the candidate state vector, tanh is the activation function tanh function, W i W c It is the weight of learning, b i and b c It's a deviation.

[0069] like Figure 9The diagram illustrates the cell update process. The operating data of the offshore wind turbine gearbox that needed to be retained from the previous moment and the operating data that needs to be retained from the current moment are updated using the following cell update formula. According to the formula, the update covers both a portion of the state information from the previous moment and a portion of the state information from the current moment. The cell update calculation formula is: Where C t This is a new unit state.

[0070] like Figure 10 The diagram shows the output gate. Finally, the output gate needs to perform the corresponding operation steps to output the next moment's offshore wind turbine gearbox operating data, and then output h... t-1 and x t Substitute o t =σ(W o [h t-1 ,x t ]+b o Obtain the judgment condition, and then C t Substituting the tanh activation function, we calculate a vector belonging to the range [-1, 1], and then multiply it by the condition o. t The final output is then obtained; the formula for this step is h. t =o t *tanh(C t ), where o t It is a condition for judgment, h t It is the final output in this cellular state, W O It is a learnable weight, b o It is a learnable bias.

[0071] (7.2) Based on the remaining life prediction model constructed in step (7.1) using the Long Short-Term Memory Neural Network (LSTM), select the appropriate number of network layers, learning rate, batch size, and number of iterations according to the actual data type and data size, so that the average relative error between the remaining life prediction value output by the LSTM neural network and the actual life of the training set and test set is less than 10%. Otherwise, repeat this step, reselect the number of network layers, learning rate, batch size, and number of iterations to train the network model until the average relative error between the remaining life prediction value output by the LSTM neural network and the actual life is less than 10%.

Claims

1. A method for predicting the remaining life of a marine wind turbine gearbox, characterized by: The application relates to a test device for predicting the residual life of a marine wind power gear box, which comprises a gear box vibration module (1), a vibration signal amplifier (2) and a base (3). The gear box vibration module (1) comprises a planetary gear box (1a), an input shaft coupling (1b), a counterweight (1c), a driving motor (1d) and a temperature sensor. The vibration signal amplifier (2) comprises an acceleration sensor (2a), a transmission rod (2b) and an output shaft coupling (2d). The vibration signal amplifier (2) is connected with the planetary gear box (1a) through the output shaft coupling (2d). The method comprises the following steps: (1) installing the temperature sensor in the planetary gear box (1a) and filling the cooling liquid, and connecting the temperature sensor with the computer through Bluetooth; (2) connecting the acceleration sensor (2a) with the computer for measurement; The measured value of the acceleration sensor in step (2) Wherein, F represents the acting force, l represents the length of the transmission rod, EI is the bending stiffness of the material of the transmission rod, and t is the time for recording the first data of the acceleration sensor; (3) hanging the load on the input shaft coupling (1b); (4) starting the driving motor (1d) and adjusting the rotating speed to the calculated rotating speed value, wherein the rotating speed formula of the driving motor is n=9550P / T, P is the output power of the driving motor, and T is the driving motor torque, wherein the driving motor torque T is the same as the simulated actual load; (5) recording and saving the data measured by the acceleration sensor (2a) and the data measured by the temperature sensor; (6) performing feature extraction by using a convolutional neural network, performing data fusion on the acceleration data generated by the acceleration sensor (2a) and the temperature data generated by the temperature sensor at the same time, and constructing a life prediction data set; the process is: (6.1) The acceleration data generated by the acceleration sensor and the temperature data generated by the temperature sensor are numbered in row order (1, 2, 3,..., i) and column order (1, 2, 3,..., j), and the acceleration data and the temperature data are constructed into a matrix M ij and then normalized; In step (6.1), the normalization processing is performed using 0-mean normalization, according to the following equation: where x ij represents the normalized data of the i-th row and j-th column, m ij is the sample data collected for the i-th row and j-th column, m j均 represents the mean value of the data corresponding to the j-th column, σ j标准差 represents the standard deviation of the data corresponding to the j-th column. (6.2) Construct the normalized data into a dataset P = [x ij ]; (6.3) constructing two layers of convolutional layers and two layers of maximum pooling layers, and arranging the convolutional layer 1, the maximum pooling layer 1, the convolutional layer 2 and the maximum pooling layer 2; (6.4) Taking the data set P constructed in step (6.2) as input, obtaining a feature matrix H after fusion of the acceleration sensor data and the temperature sensor data t = [h t1 ,h t2 ,...,h ti ] ; (7) constructing a deep learning residual life prediction model by using a long short-term memory neural network, selecting the number of network layers, the learning rate, the batch and the iteration number for training, and stopping until the average relative error between the residual life prediction value output by the residual life prediction model and the true life is less than 10%; The input layer of the deep learning residual life prediction model is the feature matrix H in step (6.4) t = [h t1 ,h t2 ,...,h ti ] ; In step (7), the residual life prediction model comprises an input layer, a hidden layer and an output layer; the hidden layer comprises a forgetting gate, an input gate and an output gate; The output layer of the deep learning residual life prediction model is the average relative error between the residual life prediction value and the life of the training set and the test set. The forget gate is f t = σ(W f ·[h t-1 , x t ]+b f ), where f t is the output of the forget gate, σ is the activation function sigmoid function, h t-1 is the output value of the previous cell, x t is the current input value, W f is the learnable weight, and b f is the learnable bias; The output gate is: i t = σ(W i · [h t-1 , x t ]+b i ), where i t is a weight coefficient of updating information, σ is an activation function sigmoid function, is a candidate state vector, tanh is an activation function tanh function, W i , W c is a weight of learning, b i and b c are biases.

Images