Fault diagnosis method and apparatus for sensor of electromechanical actuation system

Abstract

Disclosed in the embodiments of the present description is a fault diagnosis method for a sensor of an electromechanical actuation system. The method comprises: on the basis of a reference sensor signal and an actual sensor signal of a target sensor, determining whether the target sensor has a fault, wherein the reference sensor signal is output by an electromechanical actuation system model; if the target sensor has a fault, inputting the difference between the reference sensor signal and the actual sensor signal into a first neural network, so as to determine whether the fault of the target sensor belongs to a first fault type; and if it is not determined that the fault of the target sensor belongs to the first fault type, inputting the actual sensor signal into a second neural network, so as to determine whether the fault of the target sensor belongs to a second fault type.

Description

Electromechanical actuation system sensor fault diagnosis method and device TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to an electromechanical actuation system sensor fault diagnosis method and device. BACKGROUND

[0002] Due to the many advantages of electromechanical actuation systems, they have played an important role in many aspects. For example, electromechanical actuators are widely used in control of rudders, landing gear actuation, and engine reversers. Electromechanical actuation systems reduce the number of hydraulic pumps and hydraulic circuits on the aircraft, thereby reducing the weight of the aircraft and the complexity of the maintenance procedures. In addition, electromechanical actuation systems also have the characteristics of significantly better start-stop performance than hydraulic actuation systems, and improved effects of closed-loop frequency.

[0003] A variety of sensors are involved in electromechanical actuation systems, and sensors are an important part of control and health management technology. However, in a large number of practical applications, sensors are relatively weak links because their reliability is usually not as good as that of mechanical components, which is also determined by the working principle of the sensors. Since sensor failures can have a significant impact on the working state of electromechanical actuation systems, how to diagnose sensor failures has become an important issue. In particular, external interference and system dynamics have brought challenges to sensor fault diagnosis.

[0004] Current electromechanical actuation system sensor fault diagnosis methods mainly focus on the application of mathematical methods combined with computer technology, and are mostly based on filter observers and wavelet analysis methods. These methods are based on signal processing or analytical redundancy relationship, consider the signal characteristics of the sensor, and design a fault detection scheme considering the actual sensor redundancy design. They are easily limited by modeling uncertainties and are difficult to ensure the fidelity of the fault scenario. For example, Wang J et al. studied an angular displacement sensor fault detection method based on the Z-test algorithm, which uses Kalman filter estimation method to adjust the mismatch fault in the solver to complete fault adjustment. This method can monitor the angular displacement sensor fault, but it does not consider noise interference.

[0005] With the progress of artificial intelligence technology, neural networks have gradually received attention. In 2009, Edward Balaban et al. established an artificial neural network and based on the NASA electromechanical actuation system test bench to detect sensor faults, but the detection time is relatively long and noise is not considered. NASA AMES Research Center used a neural network-based fault parameter extraction method to detect electromechanical actuation system faults, but this method has fewer types of sensor fault detection, does not consider intermittent and jamming faults, and requires a lot of information for network detection, which prolongs the time required for fault detection.

[0006] It can be seen that although researchers have done a lot of research on sensor fault diagnosis of electromechanical actuation system, there are still problems such as not comprehensively considering fault types, only selecting part of faults in detection, narrow coverage of fault diagnosis, and not considering noise of sensors themselves which will seriously affect fault detection results. In addition, the detection cost of existing sensor fault detection research results of electromechanical actuation system is relatively high, and the corresponding detection efficiency is relatively low. For example, the above research results use the same network for sensor fault detection, which increases the information required for sensor fault detection and prolongs the time required for sensor fault detection.

[0007] SUMMARY

[0008] The embodiments of the present specification provide a sensor fault diagnosis method and device of electromechanical actuation system to solve the technical problem of how to more effectively and efficiently diagnose sensor faults of electromechanical actuation system.

[0009] To solve the above technical problems, the embodiments of the present specification provide the following technical solutions:

[0010] The embodiments of the present specification provide a sensor fault diagnosis method of electromechanical actuation system, the method comprising:

[0011] According to the reference sensor signal and the actual sensor signal of the target sensor, it is judged whether the target sensor has a fault; wherein the reference sensor signal is output by the electromechanical actuation system model;

[0012] If yes, the difference between the reference sensor signal and the actual sensor signal is input into the first neural network to determine whether the fault of the target sensor belongs to the first fault type;

[0013] If it is not determined that the fault of the target sensor belongs to the first fault type, the actual sensor signal is input into the second neural network to determine whether the fault of the target sensor belongs to the second fault type.

[0014] Optionally, before judging whether the target sensor has a fault, the method further comprises:

[0015] Establishing an electromechanical actuation system model.

[0016] Optionally, the first neural network is obtained by training in a time sequence training mode.

[0017] Optionally, the second neural network is obtained by training in an eigenvalue training mode.

[0018] Optionally, the method further comprises:

[0019] The first neural network and / or the second neural network are trained based on a GA-BP neural network.

[0020] Optionally, the method further comprises:

[0021] During the training of the first neural network and / or the second neural network based on the GA-BP neural network, the LM algorithm is used for training to obtain the first neural network and / or the second neural network.

[0022] Optionally, the method further comprises:

[0023] The electromechanical actuation system model and sensor data of the electromechanical actuation system are obtained.

[0024] The sensor data of the electromechanical actuation system is denoised, and training data is obtained based on the sensor data of the electromechanical actuation system model and the denoised data.

[0025] Optionally, the reference sensor signal and the actual sensor signal of the target sensor are the reference sensor signal and the actual sensor signal of the target sensor within a certain time range from the actual actuation time.

[0026] Optionally, the first fault type is a liquid fault type, and / or the second fault type is a non-liquid fault type.

[0027] The embodiments of the present specification provide an electromechanical actuation system sensor fault diagnosis device, the device comprises:

[0028] An initial judgment module is configured to judge whether the target sensor has a fault based on a reference sensor signal and an actual sensor signal of the target sensor, wherein the reference sensor signal is output by an electromechanical actuation system model.

[0029] A fault classification module is configured to input a difference between the reference sensor signal and the actual sensor signal into a first neural network to judge whether the fault of the target sensor belongs to a first fault type if it is judged that the target sensor has a fault.

[0030] If it is not judged that the fault of the target sensor belongs to the first fault type, the actual sensor signal is input into a second neural network to judge whether the fault of the target sensor belongs to a second fault type.

[0031] The above at least one technical solution adopted by the embodiments of the present specification can achieve the following beneficial effects:

[0032] The reference sensor signal and the actual sensor signal are compared to determine whether a fault occurs, and then the first neural network or the second neural network is used to determine the fault type. In this way, the sensor fault diagnosis is performed through the multi-level model architecture (electromechanical actuation system model, first neural network, and second neural network), which can effectively expand the fault diagnosis range, improve the fault diagnosis accuracy, shorten the fault diagnosis time, and thus improve the sensor fault diagnosis effect and efficiency of the electromechanical actuation system. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present specification or the prior art, the drawings required in the description of the embodiments of the present specification or the prior art will be briefly described below. Obviously, only the drawings required by some embodiments of the present application are described below, and other drawings can be obtained by those skilled in the art without creative labor.

[0034] FIG. 1 is a flowchart of the electromechanical actuation system sensor fault diagnosis method in the first embodiment of the present specification.

[0035] FIG. 2 is a schematic diagram of the electromechanical actuation system sensor fault diagnosis process in the first embodiment of the present specification.

[0036] FIG. 3 is a schematic diagram of the GA-BP neural network construction process in the first embodiment of the present specification.

[0037] FIG. 4 is a schematic diagram of the overall construction process of the two-level network in the first embodiment of the present specification.

[0038] FIG. 5 is a schematic diagram of the wavelet packet denoising process in the first embodiment of the present specification.

[0039] FIG. 6 is a schematic diagram of the structure of the electromechanical actuation system sensor fault diagnosis device in the second embodiment of the present specification. DETAILED DESCRIPTION

[0040] In order for those skilled in the art to better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be clearly and completely described below with reference to the drawings. Obviously, the embodiments involved in the specific embodiments are only some of the embodiments of the present application, not all. All other embodiments obtained by those skilled in the art based on the embodiments in the specific embodiments should be within the protection scope of the present application without creative labor.

[0041] The first embodiment of the present specification (hereinafter referred to as "Embodiment One") provides a sensor fault diagnosis method of an electromechanical actuation system. The execution subject of Embodiment One can be a terminal such as a personal computer or an embedded computer. The execution subject can be set, used or transformed as needed. In addition, a third-party application program can also assist the execution subject to execute Embodiment One. For example, the server can execute the sensor fault diagnosis method of the electromechanical actuation system in Embodiment One, and the corresponding application program can be installed on the terminal (held by the user). The terminal or the application program can transmit data with the server, and the terminal or the application program can collect or input or output or process data, thereby assisting the server to execute the sensor fault diagnosis method of the electromechanical actuation system in Embodiment One.

[0042] As shown in FIG. 1 and FIG. 2, the sensor fault diagnosis method of the electromechanical actuation system provided by Embodiment One comprises:

[0043] S101: judging whether the target sensor fails according to the reference sensor signal and the actual sensor signal of the target sensor; wherein the reference sensor signal is output by the electromechanical actuation system model;

[0044] In Embodiment One, any sensor (including but not limited to position sensor and / or speed sensor and / or current sensor) of the electromechanical actuation system (i.e. electromechanical actuator) can be used as the target sensor, and the reference sensor signal and the actual sensor signal of the target sensor can be obtained. The reference sensor signal is the signal of the target sensor output by the electromechanical actuation system model, and the actual sensor signal is the actual signal of the target sensor (for example, the actual signal of the sensor can be obtained by monitoring the electromechanical actuation system). Specifically, the position command can be input to the electromechanical actuation system model and the electromechanical actuation system (i.e. electromechanical actuation system in kind, the same below), and the reference sensor signal output by the electromechanical actuation system model according to the position command and the actual sensor signal of the target sensor of the electromechanical actuation system can be obtained (the electromechanical actuation system takes the position command as the input, and the target sensor monitors the system state to feedback, thereby obtaining the actual sensor signal).

[0045] Of course, the electromechanical actuation system model can be constructed in advance before judging whether the target sensor fails. Specifically, the electromechanical actuation system model to be established refers to the electromechanical actuation system, is based on the reference parameters of the electromechanical actuation system, and respectively establishes the models of each part (such as the driving motor, the reducer gear, the ball screw, and the rudder surface) of the electromechanical actuation system. Then, the preliminary electromechanical actuation system model is formed by combining the models of each part of the electromechanical actuation system, and a three-closed-loop control strategy is established, including the current loop, the position loop, and the speed loop. The current loop adopts PWM inverter hysteresis control, and the difference between the given current and the actual current is analyzed. If the positive edge and the negative edge of the hysteresis loop width are reached, the current is increased or decreased. The position loop and the speed loop adopt PID control, the deviation is calculated by the PID regulator, and the PID variable of the given value and the actual value is calculated, which is integrated into the control quantity to complete the control of the position and the speed. After the preliminary electromechanical actuation system model is established, the model and the parameters of the electromechanical actuation system are identified, the nonlinear modeling parameters are obtained, and the establishment of the electromechanical actuation system model is completed. Moreover, the electromechanical actuation system model can be verified, the output of the electromechanical actuation system model is compared with the output of the electromechanical actuation system, the electromechanical actuation system model is improved or optimized, and the required electromechanical actuation system model is obtained. The electromechanical actuation system model can simulate or simulate the electromechanical actuation system.

[0046] In embodiment one, whether the target sensor fails can be judged according to the reference sensor signal and the actual sensor signal of the target sensor. Specifically, the reference sensor signal and the actual sensor signal of the target sensor can be compared (including difference), and whether the target sensor fails can be judged according to the difference (including difference value) between the two. Generally, if the difference between the two is large or meets the preset condition, it can be judged that the target sensor fails.

[0047] S103: If it is judged that the target sensor fails, the difference between the reference sensor signal and the actual sensor signal is input into the first neural network to judge whether the failure of the target sensor belongs to the first failure type.

[0048] If it is judged that the target sensor fails, the specific failure type diagnosis is performed by the secondary neural network. The secondary neural network includes the first neural network and the second neural network.

[0049] First, the construction of the first neural network is described:

[0050] In the embodiment one, the BP network (i.e. BP neural network) can be taken as the basis, the BP network adopts the three-layer structure design, including the input layer, the implicit layer and the output layer. Firstly, the network topology structure, i.e. the input layer, the implicit layer, the output layer node quantity is determined, then the optimal learning method of the BP network is determined. After determining the optimal learning method, the genetic algorithm (Genetic Algorithm, GA) is used to optimize the BP network, the initial weight and threshold of the BP network are improved, thereby the construction of the GA-BP network is realized, and the first neural network can be obtained based on the training of the GA-BP network. In the process of obtaining the first neural network based on the training of the GA-BP network, the above optimal learning method (the LM algorithm is adopted in the embodiment one) can be used for training, thereby the first neural network is obtained.

[0051] The construction process of the first neural network can be shown in FIG. 3 and FIG. 4, including:

[0052] Initialize population: each individual represents the weight and bias of a BP neural network. The population size is self-determined, for example, the population size is 2.

[0053] Calculate fitness value: for each individual, use the BP neural network for training, and calculate its fitness, the fitness function can be the sum of squares of prediction errors.

[0054] Selection: use the selection operation to select the parent individual according to the fitness function.

[0055] Mutation: use the crossover operation to cross the parent individual to generate a new individual. The mutation rate is self-determined, for example, the mutation rate is 0.1.

[0056] Crossover: use the mutation operation to mutate the new individual to introduce new gene information. The crossover rate is self-determined, for example, the crossover rate is 0.9.

[0057] Add the new individual to the population and delete the individual with lower fitness.

[0058] Repeat the above calculation of fitness value, selection, mutation, crossover operation until the stopping condition (such as reaching the maximum genetic generation, for example, the maximum genetic generation is 100) is reached. Select the maximum fitness individual and decode to obtain the optimized weight and threshold, thereby the GA-BP network is established. Then the GA-BP network is trained, and the first neural network is obtained based on the iterative training of the accuracy requirement.

[0059] In actual application, there is a certain delay between the reference sensor signal of the electromechanical actuator system model and the electromechanical actuator system, and the delay value can be determined, for example, 0.1 seconds. Since the first neural network determines the fault type by using the data (specifically, the difference between the reference sensor signal and the actual sensor signal of the target sensor in a certain time range (for example, 0.5 seconds) from the actual actuation time (i.e., the time when the fault occurs, but it does not mean that the fault exists at this time, and the same below) and considering the above calculation delay, the time series difference signal (i.e., the reference sensor signal and the actual sensor signal used for actual fault diagnosis, used to determine the above difference) is selected as the data (including the reference sensor signal and the actual sensor signal of the target sensor) within 0.4 seconds from the actual actuation time, and the sampling interval can be 0.02s. It can be known that the input quantity for inputting the first neural network is 20 data points, so the input layer node can be determined as 20. In this way, in the case of sensor failure, the input layer node contains the difference between the sensor fault signal (i.e., the actual sensor signal) within a certain time range (i.e., 0.4 seconds) from the time of fault occurrence and the correct signal (the reference sensor signal output by the electromechanical actuator system model can be used as the ideal signal or correct signal).

[0060] During training, the time series difference signal (i.e., the actual signal of the sensor used for training and the sensor signal output by the electromechanical actuator system model, used to determine the difference for training) can be selected as the data (including the actual signal of various sensors and the sensor signal output by the electromechanical actuator system model) within 0.4 seconds from the actual actuation time, and the sampling interval is 0.02s, and is input into the input layer.

[0061] The number of hidden layer neurons is determined as 41 according to Kolmogorov's law.

[0062] The output layer is used to output the corresponding fault category, i.e., the first fault type. For example, the first fault type can include faults of two states of bias and noise (not limited), and the number of network nodes of the output layer can be determined as 2.

[0063] The first layer transfer function (the transfer function between the input layer and the hidden layer) is selected as tansig, and the output layer transfer function is selected as purelin after experimental error analysis.

[0064] In the training process of the GA-BP network, in order to obtain better learning effect, the upper limit of the training times can be selected as 100000, and the learning method is selected as LM algorithm. The fault states included in the first fault type can be numbered, such as bias state 1 and noise state 2, and an array is established to make the training data one-to-one mapped with the fault state number, so that after inputting a set of data (i.e., the difference), the first neural network feeds back the fault state number result after calculation.

[0065] From the above, the first neural network can be trained by a time sequence training manner, that is, the training data used in the above GA optimization process and the training GA-BP network process is sensor data with time information, specifically, the difference between the actual signals of various sensors and the various sensor signals output by the electromechanical actuation system model, and with time information or corresponding to time. How to obtain the training data is described below:

[0066] In embodiment one, the electromechanical actuation system model and the electromechanical actuation system input position command can be input, the sensor data of the electromechanical actuation system model and the electromechanical actuation system (the sensor data includes the sensor signals output by the electromechanical actuation system model and the actual signals of the sensors of the electromechanical actuation system, the same below) within a certain time range from the actual actuation time) can be obtained, and a training data set can be established according to the sensor data of the electromechanical actuation system model and the electromechanical actuation system, which is used to train the first neural network.

[0067] Specifically, the sensor data of the electromechanical actuation system model and the electromechanical actuation system can be used as raw data. Because the actual physical or system or platform is affected by various nonlinear factors, and the actual sensor itself also has measurement noise, which may cause noise oscillation in the measurement signal of the sensor. For example, in actual scenarios, the sensor data of the electromechanical actuation system can be measured through an experimental platform, and the speed sensor signal and the current sensor signal often have noise oscillation. Therefore, in embodiment one, the sensor data of the electromechanical actuation system is also subjected to denoising processing, for example, wavelet packet denoising is used.

[0068] As shown in FIG. 5, the denoising processing can include: determining a suitable wavelet basis function (such as a dmey function) according to the sensor signal, and determining the corresponding decomposition level N (which can be determined by the denoising result), and then performing N-level wavelet decomposition. The best wavelet packet decomposition tree is calculated by the standard Shanoon entropy. An appropriate threshold is selected to quantize the high-frequency coefficients after decomposition of different scales, for example, the balance sparsity-norm threshold method can be used to determine the threshold, which is to keep the percentage of the total energy the same as the percentage of 0 after processing, and the threshold processing considers the energy distribution of different frequency bands, making the denoising effect more smooth. Finally, the wavelet is reconstructed according to the wavelet packet decomposition coefficients of different layers to complete the denoising. The data obtained by the denoising processing is referred to as denoising data, and the sensor data of the electromechanical actuation system model is referred to as model data.

[0069] In the embodiment one, the training data can be obtained according to the denoised data and the model data, and specifically can include: subtracting the denoised data from the model data to obtain a series of difference data. Since the sensor data corresponds to time, the denoised data and the model data corresponding to the same time can be subtracted respectively to obtain a series of difference values corresponding to time (which can also be referred to as time series signal data or time series signal difference values). These difference values can be used as training data for training the first neural network, used in the GA optimization process and the process of training the GA-BP network to obtain the first neural network.

[0070] Continue to describe the fault diagnosis process:

[0071] In actual application, if it is judged that the target sensor has a fault, the difference between the reference sensor signal and the actual sensor signal of the target sensor within a certain time range from the actual actuation time of the electromechanical actuation system can be input into the first neural network to determine whether the fault of the target sensor belongs to the first fault type. For example, the first neural network outputs a fault state number, and according to the fault state number, it is determined whether the fault of the target sensor belongs to the first fault type.

[0072] Of course, in the previous judgment of whether the target sensor has a fault, the reference sensor signal and the actual sensor signal of the target sensor within a certain time range from the actual actuation time can also be used.

[0073] S105: If it is not judged that the fault of the target sensor belongs to the first fault type, the actual sensor signal is input into the second neural network to determine whether the fault of the target sensor belongs to the second fault type.

[0074] First, describe the construction of the second neural network:

[0075] Referring to the first neural network, the second neural network can also be based on a BP network to establish the topological relationship of the network input layer, the hidden layer and the output layer, and determine the optimal learning method of the BP network. After determining the optimal learning method, the genetic algorithm (Genetic Algorithm, GA) is used to optimize the BP network to improve the initial weight and threshold of the BP network, thereby realizing the construction of the GA-BP network, and the second neural network can be obtained based on the training of the GA-BP network. In the process of training the second neural network based on the GA-BP network, the above optimal learning method (LM algorithm in the embodiment one) can be used for training to obtain the second neural network, which can be specifically shown in FIG. 3 and FIG. 4.

[0076] The difference between the construction process of the second neural network and the first neural network mainly lies in the training process based on the GA-BP network, that is, the parameters or elements of the GA-BP network based on which the second neural network is established, and the training method or training data are different from those of the GA-BP network based on which the first neural network is established, so that the topological structure of the second neural network is also different from that of the first neural network.

[0077] In one aspect, during the training process of the GA-BP network, the GA-BP network can also adopt a three-layer structure design, including an input layer, a hidden layer and an output layer.

[0078] In actual application, the second neural network needs to use the data (specifically, the actual sensor signals of the target sensor within a certain time range, for example, 0.5 seconds) within a certain time range to calculate the characteristic values, so as to diagnose whether it belongs to the second fault type, and the second neural network does not need to be subtracted, so it can fully use the data within the certain time range.

[0079] During training, the specific content of the characteristic value can be determined according to the content of the second fault type. For example, the second fault type can include two faults of signal disconnection and blockage, which can be determined according to the number of signal 0 points and the number of signal 1 points, and the characteristic value can include the number of signal 0 points and the number of signal 1 points. For another example, the second fault type can also include two faults of intermittent and scaling, so that the characteristic value can include mean value, root mean square value, root amplitude value, absolute average value, skewness, kurtosis, variance, number of signal 0 points, number of signal 1 points, a total of 9 parameters (but not limited). The input layer node is determined according to the characteristic value parameter, for example, the characteristic value includes 9 parameters, then the input layer node is 9, and then the number of hidden layer nodes is determined to be 19 according to the Kolmogorov law.

[0080] The output layer is used to output the corresponding fault category, that is, the second fault type. For example, the second fault type includes the above four faults, and the number of output layer nodes can be determined as 4.

[0081] In order to obtain better learning effect, the upper limit of the training times can be selected as 100000, and the learning method is selected as LM algorithm. The fault states contained in the first fault type can be numbered, such as disconnection as state 1, blockage as state 2, intermittent as state 3, and scaling as state 4, and an array is established to make the training data one-to-one mapping with the fault state number, so that after inputting a group of data (i.e. the actual signal of the sensor), the second neural network feeds back the fault state number result after calculation (including characteristic value calculation).

[0082] From the above, the second neural network can be trained by eigenvalue training, that is, the training data used in the above GA optimization process and the training of the GA-BP network process are obtained according to the characteristic values of the sensor data.

[0083] On the other hand, how to obtain the training data is described: the position command can be input to the electromechanical actuator system, the sensor data of the electromechanical actuator system within a certain time range from the actual actuation time is obtained, and the training data set is established according to the sensor data of the electromechanical actuator system, which is used to train the second neural network.

[0084] Specifically, as described above, the sensor data of the electromechanical actuator system can also be denoised, for example, using wavelet packet denoising, which will not be repeated here. The training data used to train the second neural network can be obtained from the denoised data, for example, calculating the above various characteristic values of the denoised data, and using the various characteristic values as training data for the GA optimization process and the process of training the GA-BP network to obtain the second neural network.

[0085] The parts of the second neural network that are not described in detail, such as the maximum genetic generation, population size, crossover rate, and mutation rate in the genetic algorithm optimization process, the actual signal of the sensor with time signal, or corresponding to time, etc. can be referred to the training part of the first neural network. The actual signal of the sensor in the training data of the first neural network and the second neural network needs to include normal signals and fault signals to correspond to different fault type flags.

[0086] It should be noted that the first neural network and the second neural network can be established in advance, and there is no absolute order of establishment. After the first neural network and the second neural network are established, it means that the network building elements have been confirmed, including the number of nodes in different layers and the weight threshold between nodes. From this, the network can be coded into a function, and the model deployment can be completed after compiling in the form of a two-level function.

[0087] Continue to describe the fault diagnosis process:

[0088] In actual application, if the fault of the target sensor does not belong to the first fault type according to the output of the first neural network (such as the fault state signal output by the first neural network does not belong to the first fault type or no fault state signal is output), the actual sensor signal of the target sensor within a certain time range from the actual actuation time of the electromechanical actuator system can be input to the second neural network to determine whether the fault of the target sensor belongs to the second fault type. For example, the second neural network outputs a fault state number, and according to the fault state number, it is determined whether the fault of the target sensor belongs to the second fault type.

[0089] In actual applications, the specific content of the first fault type and the second fault type is not limited. For example, the first fault type can be a liquid or superimposed fault type, for example, including bias or noise faults; and / or, the second fault type can be a non-liquid or non-superimposed fault type, for example, a disconnection, jamming, intermittent or scaling fault. Among them, the fault performance of the liquid fault is based on the original signal or the ideal signal, such as bias and noise, which are superimposed on the original signal, so that the fault feature can be obtained by subtracting the actual sensor signal from the reference sensor signal. The difference signal is still a time sequence signal. For non-liquid faults, a second neural network of the feature value training mode is established, and the fault feature is obtained by calculating the feature quantity of the signal, so as to detect the non-liquid fault.

[0090] In example one, fault diagnosis can be continuous. Specifically, the reference sensor signal and the actual sensor signal of the target sensor after each actual actuation can be continuously monitored. Once the reference sensor signal and the actual sensor signal of the target sensor at a certain time meet the preset conditions, it can be judged that the target sensor has a fault. Then, the reference sensor signal and the actual sensor signal of the target sensor within a certain time range from the time are used to determine the fault type according to the above content, and then corresponding measures are taken.

[0091] Example one can achieve the following beneficial effects:

[0092] By comparing the reference sensor signal and the actual sensor signal, it is determined whether a fault occurs, and then the first neural network or the second neural network is used to determine the fault type. In this way, the sensor fault diagnosis is performed through a multi-level model architecture (electromechanical actuation system model, first neural network, second neural network), which can effectively expand the fault diagnosis range, improve the fault diagnosis accuracy, and shorten the fault diagnosis time, thereby improving the sensor fault diagnosis effect and efficiency of the electromechanical actuation system.

[0093] If a single network is used to judge the fault, either only a few faults can be diagnosed, or the network structure will be complex, the training will easily enter a complex decision surface, including in the face of many fault types, the network feature value characteristics are complex, the results need to be generated are many, the network topology structure is complex, the training information required is much, the training difficulty is high, the overfitting phenomenon is easy to occur, it is easy to fall into a local minimum value, and the fault diagnosis time is long.

[0094] In the first embodiment, different secondary neural networks are used for nonlinear mapping of different types and characteristics of faults to diagnose sensor faults, effectively distinguishing fault types, and classifying based on fault classification characteristics. The first neural network is trained in a time sequence, taking into account the feature that the fault signal is based on the original signal, and can establish a relationship between the fault signal and the normal signal. The second neural network is trained by feature value, which can train the fault signal of the sensor and establish a mapping between the fault signal and the corresponding fault. In this way, different secondary neural networks can be used to diagnose more types of sensor faults, and can be applied to fault diagnosis of various sensors (including but not limited to displacement sensors, speed sensors, and current sensors), improving the comprehensiveness and range of fault diagnosis. Different secondary neural networks use different training methods, which on the one hand ensures that different secondary neural networks can more accurately and more specifically diagnose different types of faults, such as the first neural network that can effectively target liquid faults and the second neural network that can effectively diagnose non-liquid faults. On the other hand, it helps to reduce the amount of information, training, and difficulty of training of a single neural network, reduce the complexity of the decision surface of a single neural network, and improve the network convergence speed, fault diagnosis speed, and diagnosis accuracy of a single neural network. Through these contents, the sensor fault diagnosis effect and efficiency can be further improved.

[0095] During the training of the second neural network, multiple feature values can be selected, so that the trained second neural network can effectively distinguish different types of non-liquid faults, further improving the sensor fault diagnosis effect and efficiency.

[0096] The neural network is applied to sensor fault diagnosis. Because sensor fault diagnosis requires real-time, in the first embodiment, the reference sensor signal and the actual sensor signal within a certain time range are used for fault diagnosis, and the certain time range can be 0.5 seconds or less, thereby improving the fault diagnosis efficiency and speed.

[0097] During the training of the neural network, based on the BP network, the optimal learning method can be determined by comparing the results of different learning methods, and the genetic algorithm and the optimal learning method are used for optimization, so as to obtain the secondary neural network, overcoming the shortcomings of the pure BP network (such as slow learning convergence speed and inability to guarantee convergence to the global minimum point), and ensuring the excellent use effect of the secondary neural network.

[0098] When determining the training data, the sensor data is denoised, solving the influence of sensor noise on fault diagnosis and further improving the fault diagnosis effect. And using wavelet packet denoising can consider low and high frequency signals at the same time, with more accurate analysis ability and better denoising effect.

[0099] Through actual verification on the embodiment one, the fault type diagnosis of various sensors of the electromechanical actuation system is realized with 100% accuracy, which proves the excellent effect and efficiency of the embodiment one.

[0100] The embodiment one can be used for sensor fault diagnosis in various scenes, especially for fault diagnosis of sensors of the electromechanical actuation system of the flight control system of a civil aircraft.

[0101] As shown in FIG. 6, the second embodiment of the present specification provides a sensor fault diagnosis device of an electromechanical actuation system corresponding to the embodiment one, which comprises:

[0102] An initial judgment module 202 is configured to judge whether a target sensor is faulty according to a reference sensor signal and an actual sensor signal of the target sensor, wherein the reference sensor signal is output by a model of the electromechanical actuation system.

[0103] A fault classification module 204 is configured to input a difference between the reference sensor signal and the actual sensor signal into a first neural network to judge whether a fault of the target sensor belongs to a first fault type if it is judged that the target sensor is faulty.

[0104] If it is not judged that the fault of the target sensor belongs to the first fault type, the actual sensor signal is input into a second neural network to judge whether the fault of the target sensor belongs to a second fault type.

[0105] Optionally, the device further comprises:

[0106] A model building module is configured to build a model of the electromechanical actuation system before judging whether the target sensor is faulty.

[0107] Optionally, the first neural network is obtained through time sequence training.

[0108] Optionally, the second neural network is obtained through eigenvalue training.

[0109] Optionally, the device further comprises:

[0110] A network training module is configured to obtain the first neural network and / or the second neural network based on GA-BP neural network training.

[0111] Optionally, the network training module is configured to use the LM algorithm for training to obtain the first neural network and / or the second neural network during the process of obtaining the first neural network and / or the second neural network based on GA-BP neural network training.

[0112] Optionally, the device further comprises:

[0113] a network training module, configured to acquire sensor data of the electromechanical actuation system and a model of the electromechanical actuation system; perform denoising processing on the sensor data of the electromechanical actuation system, and obtain training data according to the sensor data of the electromechanical actuation system model and the data obtained by the denoising processing.

[0114] Optionally, the reference sensor signal and the actual sensor signal of the target sensor are the reference sensor signal and the actual sensor signal of the target sensor within a certain time range from the actual actuation time.

[0115] Optionally, the first fault type is a liquid fault type, and / or the second fault type is a non-liquid fault type.

[0116] The above only describes the embodiments of the present specification and is not intended to limit the present application. Various modifications and changes can be made by those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. A method for diagnosing sensor faults in an electromechanical actuation system, characterized in that: The method comprises: Determining whether a target sensor fails based on a reference sensor signal and an actual sensor signal of the target sensor, wherein the reference sensor signal is output by an electromechanical actuation system model; If so, inputting the difference between the reference sensor signal and the actual sensor signal into a first neural network to determine whether the fault of the target sensor belongs to a first fault type; If it is not determined that the fault of the target sensor belongs to the first fault type, the actual sensor signal is input into the second neural network to determine whether the fault of the target sensor belongs to the second fault type.

2. The method according to claim 1, wherein Before determining whether the target sensor fails, the method further includes: Establish a model of the electromechanical actuation system.

3. The method according to claim 1, wherein The first neural network is trained by a time series training method.

4. The method according to claim 1, wherein The second neural network is trained by an eigenvalue training method.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises: The first neural network and / or the second neural network are obtained based on GA-BP neural network training.

6. The method according to claim 5, wherein The method further comprises: In the process of obtaining the first neural network and / or the second neural network based on GA-BP neural network training, LM algorithm is used for training to obtain the first neural network and / or the second neural network.

7. The method according to any one of claims 1 to 4, characterized in that The method further comprises: obtaining a mechatronic actuation system model and sensor data of the mechatronic actuation system; The sensor data of the electromechanical actuator system is subjected to denoising processing, and the electromechanical actuator system model is used to The sensor data and the denoised data are used to obtain the training data.

8. The method according to any one of claims 1 to 4, characterized in that The reference sensor signal and the actual sensor signal of the target sensor are the reference sensor signal and the actual sensor signal of the target sensor within a certain time range from the actual actuation time.

9. The method according to any one of claims 1 to 4, characterized in that The first fault type is a liquid fault type, and / or the second fault type is a non-liquid fault type.

10. A sensor fault diagnosis device for an electromechanical actuation system, characterized in that: The device comprises: an initial judgment module, configured to judge whether a target sensor has failed based on a reference sensor signal and an actual sensor signal of the target sensor; wherein the reference sensor signal is output by the electromechanical actuation system model; a fault classification module, configured to, if it is determined that the target sensor has failed, input a difference between the reference sensor signal and the actual sensor signal into a first neural network to determine whether the failure of the target sensor belongs to a first failure type; If it is not determined that the fault of the target sensor belongs to the first fault type, the actual sensor signal is input into the second neural network to determine whether the fault of the target sensor belongs to the second fault type.

Images