An intelligent decoupling diagnosis method for composite faults of electrostatic hydraulic actuators

Abstract

The present invention discloses an intelligent decoupling diagnosis method for EHA composite faults, comprising the following steps: first, according to the fault type of EHA, a flow meter, a pressure sensor and a grating ruler are used to collect data at a specific position of a hydraulic system to obtain real multi-sensor operation data in various health states; then, the collected multi-sensor data in normal, single fault and composite fault states are sliced ​​and processed to divide the training data set and the test data set; then, the constructed maximum aggregation attention convolution capsule network is trained using the training data set, and the extended loss function is used as the loss function; finally, the trained maximum aggregation attention convolution capsule network is tested using the test data set to obtain the result of EHA composite fault diagnosis. In addition, the present invention is simple and easy to implement, and is suitable for the decoupling and diagnosis of EHA composite faults with high composite fault data collection costs, limited single sensor information perception and unclear fault characteristics.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis in mechanical equipment, and in particular to an intelligent decoupling diagnosis method for composite faults of an electrostatic hydraulic actuator (EHA). Background Art

[0002] The electrostatic hydraulic actuator is a device that integrates an electric actuator and a hydraulic transmission system. It generates hydraulic energy by driving the hydraulic pump with electricity, and controls the hydraulic cylinder or hydraulic motor to achieve high-precision movement. It has the characteristics of high power-to-weight ratio, high integration, and low pollution. It is widely used in aerospace, industrial automation, marine engineering and other fields. Due to the complex internal structure and harsh working environment of the electrostatic hydraulic actuator, the safety and reliability of the system are greatly challenged. The diagnosis of single faults in hydraulic systems has received widespread attention from scholars. However, practice has shown that electrostatic hydraulic actuators may have compound faults, that is, multiple faults occur at the same time or one fault triggers another fault, which puts higher requirements on the fault diagnosis of hydraulic systems.

[0003] At present, the research on composite fault diagnosis of electrostatic hydraulic actuators is still in its infancy. Although some progress has been made, it still faces many challenges. For the composite fault diagnosis of mechanical systems, there are mainly the following methods. Composite fault diagnosis method based on failure mechanism: This method predicts and identifies faults by understanding the failure mechanism of each component. It relies on professional knowledge and detailed system analysis, and emphasizes an in-depth understanding of the equipment structure and working principle. Composite fault diagnosis method based on signal processing: This method extracts or separates single fault signal features from composite fault signals by applying a variety of signal processing techniques, thereby identifying possible faults. Composite fault diagnosis method based on artificial intelligence: This method uses artificial intelligence technologies such as machine learning and deep learning to analyze and learn sensor data in different health states to establish a fault diagnosis model.

[0004] It should be noted that the capsule network type artificial intelligence algorithm has achieved remarkable results in the intelligent diagnosis of complex faults in mechanical systems due to its vector input and vector output characteristics. The routing algorithm between capsule layers is the core of the capsule network, which determines the transmission relationship between two adjacent capsule layers. Hinton proposed the first routing algorithm for capsule networks: the dynamic routing algorithm. At present, this algorithm is widely used in complex fault diagnosis based on capsule networks. However, the dynamic routing algorithm lacks a statistical theoretical basis. Later, Hinton proposed the expectation maximization routing algorithm. This algorithm overcomes the limitations of the dynamic routing algorithm through global perspective optimization based on the probability model. Through the iterative process of E-step and M-step, the EM routing algorithm can better capture the complex relationships in the data and improve the robustness of the model.

[0005] Considering the high cost of data collection for composite faults in hydraulic systems, limited information perception of single sensors, and unclear fault features, this paper proposes a maximum aggregation attention convolutional capsule network algorithm. The algorithm uses convolutional neural networks as primary feature extractors and capsule networks as deep feature extractors; then, through the maximum aggregation routing algorithm, which is an improved expectation maximization algorithm, the main capsules with similar features are integrated into the same digital capsule; finally, a decoupled classification layer is added to enable the ability to output multiple single fault types, and the network is optimized using an extended loss function. Summary of the invention

[0006] In order to realize intelligent decoupling diagnosis of composite faults of electrostatic-hydraulic actuators, the present invention designs an intelligent decoupling diagnosis method of composite faults of electrostatic-hydraulic actuators based on maximal aggregation attention convolutional capsule network, aiming to utilize the maximal aggregation algorithm and extended loss function to highlight the more critical features for fault diagnosis and realize more efficient and robust intelligent decoupling diagnosis of composite faults of electrostatic-hydraulic actuators.

[0007] The technical solution to realize the present invention is: an intelligent decoupling diagnosis method for composite faults of electrostatic hydraulic actuators based on a maximally aggregated attention convolutional capsule network, comprising the following steps:

[0008] Step 1: According to the fault type of the electrostatic hydraulic actuator, the relevant signals of the electrostatic hydraulic actuator are collected through different sensors, and the multi-sensor related signals of the normal state, single fault state and compound fault state collected are sliced ​​and processed to divide the training data set and the test data set.

[0009] Step 2: Construct a maximum aggregation attention convolutional capsule network. The maximum aggregation attention convolutional capsule network uses a one-dimensional convolutional neural network as a primary feature extractor and a capsule network as a deep feature extractor. Then, through the maximum aggregation routing algorithm, the main capsules with similar features are integrated into the same digital capsule. Finally, a decoupled classification layer is added to enable the network to output multiple single fault types.

[0010] Step 3: Input the training data set into the maximum aggregated attention convolutional capsule network to train the intelligent decoupling diagnosis model of electrostatic hydraulic actuator composite faults, and use the extended loss function to optimize the network to obtain the trained fault diagnosis model.

[0011] Step 4: Input the test data set into the trained fault diagnosis model to perform intelligent decoupling diagnosis of electrostatic-hydraulic actuator compound faults.

[0012] Compared with the prior art, the present invention has the following significant advantages:

[0013] (1) The present invention proposes for the first time a maximum aggregation attention convolutional capsule network algorithm and applies it to the intelligent decoupling diagnosis of complex faults of electrostatic hydraulic actuators. First, the relevant signals of the electrostatic hydraulic actuator are collected through different sensors. Then, the multi-sensor related signals of the normal state, single fault state and complex fault state collected are sliced ​​and processed to divide the training data set and the test data set. Then, the maximum aggregation attention convolutional capsule network algorithm is used to realize the fusion and extraction of multi-channel features. The extended loss is used as the loss function to enhance the separation between categories and improve the generalization ability of the model.

[0014] (2) The designed maximum aggregated attention convolutional capsule network algorithm only needs to be trained with the single fault data of the electrostatic hydraulic actuator to realize the intelligent fault diagnosis of complex faults. Compared with the capsule network based on dynamic routing, the maximum aggregated attention convolutional capsule network algorithm has better stability and achieves higher fault diagnosis accuracy in the electrostatic hydraulic actuator complex fault diagnosis task. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a composite fault diagnosis method for an electrostatic hydraulic actuator based on a maximally aggregated attention convolutional capsule network according to the present invention.

[0016] Figure 2 is the output probability of the maximized aggregated attention convolutional capsule network.

[0017] Figure 3 It is the t-SNE clustering effect of the maximum aggregated attention convolutional capsule network. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] At present, capsule networks have achieved remarkable results in the intelligent diagnosis of complex faults in mechanical systems due to their vector input and vector output characteristics. In order to solve the problems of high cost of complex fault data collection, limited information perception of a single sensor, and unclear fault characteristics, we collected multi-source data through multiple sensors, introduced an efficient channel attention mechanism, designed a maximum aggregation routing algorithm, and combined it with an extended loss function, so that we can complete the efficient intelligent decoupling diagnosis of complex faults by only training on the single fault data of the electrostatic hydraulic actuator.

[0020] Combination Figure 1, the present invention discloses a method for intelligent decoupling diagnosis of electrostatic hydraulic actuator composite faults of a maximally aggregated attention convolutional capsule network. First, according to the fault type of the electrostatic hydraulic actuator, a flow meter, a pressure sensor and a grating ruler are used to collect data at a specific position of the hydraulic system to obtain the multi-sensor real operation data of the system in various health states under different working conditions; then, the collected multi-sensor data of normal state, single fault state and composite fault state are sliced ​​and processed to divide the training data set and the test data set; then, the constructed maximally aggregated attention convolutional capsule network is trained using the training data set, and the extended loss function is used as the loss function to increase the gap between the correct category output probability and the wrong category output probability; finally, the trained maximally aggregated attention convolutional capsule network is tested using the test data set to obtain the result of composite fault diagnosis of the electrostatic hydraulic actuator. In addition, the present invention is simple and easy to implement, and is suitable for the decoupling and diagnosis of composite faults of electrostatic hydraulic actuators with high composite fault data acquisition cost, limited single sensor information perception and unclear fault characteristics. The specific steps are as follows:

[0021] Step 1: According to the fault type of the electrostatic hydraulic actuator, the relevant signals of the electrostatic hydraulic actuator are collected through different sensors, and the multi-sensor related signals of the normal state, single fault state and compound fault state collected are sliced ​​and processed to divide the training data set and the test data set.

[0022] The fault types of electrostatic hydraulic actuators include hydraulic cylinder leakage fault, pipeline blockage fault, and hydraulic cylinder leakage and pipeline blockage fault. The relevant signals of electrostatic hydraulic actuators include hydraulic cylinder high-pressure chamber flow, hydraulic cylinder leakage flow, hydraulic cylinder high-pressure chamber pressure, hydraulic cylinder low-pressure chamber pressure, control signal size, hydraulic cylinder piston displacement, and control tracking error.

[0023] Step 2: Construct a maximum aggregation attention convolutional capsule network. The maximum aggregation attention convolutional capsule network uses a one-dimensional convolutional neural network as a primary feature extractor and a capsule network as a deep feature extractor. Then, through the maximum aggregation routing algorithm, the main capsules with similar features are integrated into the same digital capsule. Finally, a decoupled classification layer is added to enable the network to output multiple single fault types.

[0024] Furthermore, a maximum aggregation attention convolutional capsule network is constructed as follows:

[0025] Step 2.1: The network structure of the maximally aggregated attention convolutional capsule network includes a one-dimensional convolutional neural network, an efficient channel attention (ECA) mechanism, a capsule network, and a decoupled classification layer; the one-dimensional convolutional neural network mainly consists of four parts: a one-dimensional convolutional layer, a batch normalization layer, an activation function, and an average pooling layer.

[0026] LeakyReLU is a variant of ReLU that allows small gradient values ​​to pass through, thereby effectively solving the problem of neuron "death" in the traditional ReLU activation function; the process is expressed as follows:

[0027] T l =LReLU(BN(W l *x+b l ))

[0028] Among them, W l and b l They represent the convolution kernel and bias term of the lth layer respectively, * is the convolution operation, x is the input signal, BN means applying batch normalization to the convolution result, LReLU means applying the LeakyReLU activation function, T l Represents the convolution output of layer l.

[0029] As a common downsampling method, the average pooling layer can effectively reduce the amount of computation of the subsequent network. The pooling process is expressed as follows:

[0030] y l =avgpooling(T l )

[0031] Among them, avgpooling(·) represents the average pooling operation, y l Represents the pooled output of the lth layer. The one-dimensional convolutional neural network is composed of the above core structural units.

[0032] Step 2.2: After the one-dimensional convolutional neural network feature extraction, an efficient channel attention mechanism is used to achieve feature-level data fusion, and the efficiency and accuracy of the neural network in processing information is improved by learning the importance weights between different channels.

[0033] Perform global average pooling on each channel to extract the global feature information of each channel; use a one-dimensional convolution kernel to capture the local dependencies between channels and generate a vector representing the channel weights; in the efficient channel attention mechanism, the size of the convolution kernel is adaptively determined and is usually calculated based on the number of channels; the size of the convolution kernel can be determined by the following formula:

[0034]

[0035] In the above formula, (·) odd Make sure k is an odd number; b and γ are hyperparameters, take b=1, γ=2; c is the number of channels, the larger the number of channels, the larger the k value, and vice versa.

[0036] The output of the one-dimensional convolution is normalized to a value between 0 and 1 through the Sigmoid activation function σ(·), indicating the importance of each channel. The calculated channel weight is applied to each channel of the original one-dimensional input to achieve channel recalibration.

[0037] Step 2.3: The capsule network includes a main capsule layer, a digital capsule layer, and a maximum aggregation routing algorithm. The main capsule layer is used to further extract features from the fusion results from the ECA mechanism, and transforms the basic elements of the network from neuron scalars into capsule vectors through dimensionality transformation; the digital capsule layer is the output layer of the capsule network, and each digital capsule represents a specific category; through the output of the main capsule layer and the maximum aggregation routing algorithm, the digital capsule layer can capture more complex and high-level features in the input data.

[0038] The main capsule layer is composed of multiple parallel one-dimensional convolutional layers stacked together, each of which captures different types of features. The output features of the efficient channel attention layer are used as input to obtain the feature map of the main capsule layer. After dimension transformation, multiple primary capsules are obtained, each of which is a vector. The main capsule layer uses the nonlinear function Squash as the activation function, which shrinks the length of the input vector to [0, 1] without changing its direction. The modulus of the vector represents the probability of the feature existing. The Squash function is defined as follows:

[0039]

[0040] Among them, s jj is the input vector of the jjth capsule, T jj is the output vector of the jjth capsule after applying the Squash function.

[0041] A maximum aggregation routing algorithm is designed between the main capsule layer and the digital capsule layer. The capsule vector of the main capsule layer is used as the input vector of the maximum aggregation routing algorithm, and the capsule vector of the digital capsule layer is used as the output vector of the maximum aggregation routing algorithm. The details are as follows:

[0042] S2.3.1. Calculate activation value based on capsule vector of main capsule layer

[0043]

[0044] in, It is the element in the input vector of the maximizing aggregation routing algorithm, the number of input vectors is e=1,2,...n (inp) , n (inp) is the total number of input vectors, the number of elements in each input vector f=1,2,...d (inp) , d (inp) is the total number of elements in each input vector, is the corresponding learnable weight value, is the bias vector.

[0045] Initialize routing probability R em , R em =1 / n (out) , the number of output vectors of the maximum aggregation routing algorithm m = 1, 2, ... n (out) , n (out) Represents the total number of output vectors of the maximized aggregation routing algorithm.

[0046] S2.3.2. Calculation of input vectors used and ignore the input vector The formula is as follows:

[0047]

[0048] Here, σ(·) is the sigmoid function, which is used to limit the output value to [0,1].

[0049] S2.3.3. Through the intermediate variable φ em It represents the difference between the "use" and "ignore" weights of the input vector on the output vector. The calculation formula is as follows:

[0050]

[0051] in, represents the net benefit of using the input vector, represents the net cost of ignoring the input vector.

[0052] By performing linear transformation and bias adjustment on the intermediate variables, the contribution of the input vector is integrated to generate the final output vector The calculation formula is as follows:

[0053]

[0054] in, is a learnable parameter, the number of elements in each output vector n = 1, 2, ... d (out) , d (out) is the total number of elements in each output vector.

[0055] S2.3.4. Reversely calculate the input vector through the generated output vector to obtain the predicted input vector The calculation process is as follows:

[0056]

[0057] in, is a learnable parameter, LN(·) is a layer normalization operation; calculate the consistency S between the actual input vector and the predicted input vector em , the calculation formula is as follows:

[0058]

[0059] in, is a learnable parameter, logf(·) is the logarithmic sigmoid activation operation; update the routing probability R em :

[0060]

[0061] S2.3.5. Determine the relationship between the total number of iterations T' and the required number of iterations T. If T' <= T, continue to execute S2.3.2-S2.3.4. Otherwise, output

[0062] Step 2.4: The decoupled classification layer needs to calculate the bi-norm of the output vector of each digital capsule layer. The calculation formula can be expressed as:

[0063] p j =||ν j ||

[0064] Among them, ν j is the vector output by the jth capsule unit in the capsule network, p j is the probability of the jth fault category, To limit the number of output labels, if The classifier output at this label is 1, otherwise the output is 0.

[0065] Step 3: Input the training data set into the maximum aggregated attention convolutional capsule network to train the intelligent decoupling diagnosis model of electrostatic hydraulic actuator composite faults, and use the extended loss function to optimize the network to obtain the trained fault diagnosis model.

[0066] The extended loss function is selected as the loss function of the maximum aggregated attention convolution capsule network. The extended loss function can help the model distinguish multiple faults that occur simultaneously; by enlarging the activation value a of the correct category ti and the incorrect category activation value a ki The gap between them makes each fault type reach a sufficient activation level, thus reducing the confusion between categories; the total expansion loss is defined as:

[0067]

[0068] Where N is the total number of samples in the batch, t i is the correct category index of sample i, ki is the wrong category index of sample i, C is the number of categories, m1 is a constant threshold, and m1=0.9 is taken.

[0069] Step 4: Input the test data set into the trained fault diagnosis model to perform intelligent decoupling diagnosis of electrostatic hydraulic actuator compound faults.

[0070] Example 1: Data verification of electrostatic hydraulic actuator fault simulation test bench

[0071] The data used in this experiment comes from the electrostatic hydraulic actuator fault simulation test bench. The collected signals mainly include seven types: hydraulic cylinder high-pressure chamber flow, hydraulic cylinder leakage flow, hydraulic cylinder high-pressure chamber pressure, hydraulic cylinder low-pressure chamber pressure, control signal size, hydraulic cylinder piston displacement and control tracking error. This experiment set up four health states: normal state, hydraulic cylinder leakage fault, pipeline blockage fault and hydraulic cylinder leakage and pipeline blockage combined fault. The data collection time for each health state is about 200s.

[0072] The collected continuous data is divided into samples, each sample contains 7×1024 data points, the overlap rate of normal state and single fault samples is set to 0.5, and the compound fault samples do not overlap. The training data set of the network is composed of data of some normal state and single fault state (hydraulic cylinder leakage, pipeline blockage); the test data set of the network is constructed using data of compound fault state, another part of normal state and single fault state. It should be noted that the data of compound fault only exists in the test data set. The specific description of the data set is shown in Table 3.

[0073] Table 3 Description of data sets under various working conditions

[0074]

[0075]

[0076] Figure 2 The output probabilities of the Convolutional Capsule Network based on Maximized Aggregate Attention evaluated on the test dataset are shown. Figure 3 It can be seen that the features learned by the maximum aggregated attention convolution capsule network method not only have obvious differences between classes, but also have more compact features within classes. Only using the single fault data of the electrostatic hydraulic actuator for training can achieve intelligent fault diagnosis of compound faults with a fault diagnosis accuracy of up to 97.4%.

[0077] In summary, in order to solve the problems of high cost of data collection for composite faults in hydraulic systems, limited information perception of single sensors, and unclear fault characteristics, the present invention designs a fault diagnosis method for electrostatic hydraulic actuators based on the maximum aggregation attention convolutional capsule network algorithm. The designed fault diagnosis method uses convolutional neural networks as primary feature extractors and capsule networks as deep feature extractors; then, through the maximum aggregation routing algorithm, which is an improved expectation maximization algorithm, the main capsules with similar features are integrated into the same digital capsule; finally, a decoupled classification layer is added to enable the ability to output multiple single fault types, and the network is optimized using the extended loss function.

[0078] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. An intelligent decoupling diagnosis method for composite faults of electrostatic hydraulic actuators, characterized in that: The following steps are involved: Step 1: According to the fault type of the electrostatic hydraulic actuator, the relevant signals of the electrostatic hydraulic actuator are collected through different sensors, and the multi-sensor related signals of the normal state, single fault state and compound fault state collected are sliced ​​and processed to divide the training data set and the test data set; Among them, the relevant signals of the electrostatic hydraulic actuator include the flow rate of the hydraulic cylinder high-pressure chamber, the leakage flow rate of the hydraulic cylinder, the pressure of the hydraulic cylinder high-pressure chamber, the pressure of the hydraulic cylinder low-pressure chamber, the size of the control signal, the displacement of the hydraulic cylinder piston and the control tracking error; Step 2: Construct a maximum aggregation attention convolutional capsule network. The maximum aggregation attention convolutional capsule network uses a one-dimensional convolutional neural network as a primary feature extractor and a capsule network as a deep feature extractor. Then, through the maximum aggregation routing algorithm, the main capsules with similar features are integrated into the same digital capsule. Finally, a decoupled classification layer is added to enable the network to output multiple single fault types. Among them, the maximum aggregation attention convolution capsule network is constructed as follows: Step 2.1: The network structure of the maximum aggregated attention convolution capsule network includes a one-dimensional convolutional neural network, an ECA mechanism, a capsule network, and a decoupled classification layer; the one-dimensional convolutional neural network consists of four parts: a one-dimensional convolutional layer, a batch normalization layer, a leaky ReLU, and an average pooling layer; Leaky ReLU is a variant of ReLU. As an activation function, the process is expressed as follows: T l =LReLU(BN ( W l *x+b l )) Among them, W l represents the convolution kernel of layer l, b l represents the bias term of the lth layer, * is the convolution operation, x is the input signal, BN represents the batch normalization operation applied to the convolution result, LReLU represents the application of the Leaky ReLU activation function, T l Represents the convolution output of the lth layer; The pooling process of the average pooling layer is expressed as follows: y l =avgpooling(T l ) Among them, avgpooling(·) represents the average pooling operation, y l represents the pooled output of layer l; Step 2.2: After extracting features through the one-dimensional convolutional neural network, the ECA mechanism is used to achieve feature-level data fusion, and the efficiency and accuracy of the neural network in processing information is improved by learning the importance weights between different channels; Step 2.3: The capsule network includes a main capsule layer, a digital capsule layer, and a maximum aggregation routing algorithm. The main capsule layer is used to further extract features from the fusion results from the ECA mechanism, and transforms the basic elements of the network from neuron scalars into capsule vectors through dimensional transformation; the digital capsule layer is the output layer of the capsule network, and each digital capsule represents a category; through the output of the main capsule layer and the maximum aggregation routing algorithm, the digital capsule layer can capture more complex and high-level features in the input data; Among them, a maximum aggregation routing algorithm is designed between the main capsule layer and the digital capsule layer. The capsule vector of the main capsule layer is used as the input vector of the maximum aggregation routing algorithm, and the capsule vector of the digital capsule layer is used as the output vector of the maximum aggregation routing algorithm. The details are as follows: S2.3.

1. Calculate activation value based on capsule vector of main capsule layer in, It is the element in the input vector of the maximizing aggregation routing algorithm, the number of input vectors is e=1,2,...n (inp) , n (inp) is the total number of input vectors, the number of elements in each input vector f=1,2,...d (inp) , d (inp) is the total number of elements in each input vector, is the corresponding learnable weight value, is the bias vector; Initialize routing probability R em , R em =1 / n (out) , the number of output vectors of the maximum aggregation routing algorithm m = 1, 2, ... n (out) , n (out) represents the total number of output vectors of the maximization aggregation routing algorithm; S2.3.

2. Calculation of input vectors used and ignore the input vector The formula is as follows: Where σ(·) is the sigmoid function, which is used to limit the output value to [0,1]; S2.3.

3. Through the intermediate variable φ em It represents the difference between the "use" and "ignore" weights of the input vector on the output vector. The calculation formula is as follows: in, represents the net benefit of using the input vector, represents the net cost of ignoring the input vector; By performing linear transformation and bias adjustment on the intermediate variables, the contribution of the input vector is integrated to generate the final output vector The calculation formula is as follows: in, are all learnable parameters, and the number of elements in each output vector is n = 1, 2, ... d (out) , d (out) is the total number of elements in each output vector; S2.3.

4. Reversely calculate the input vector through the generated output vector to obtain the predicted input vector The calculation process is as follows: in, are all learnable parameters, LN(·) is the layer normalization operation; Calculate the consistency S between the actual input vector and the predicted input vector em , the calculation formula is as follows: in, are all learnable parameters, logf(·) is the logarithmic sigmoid activation operation; update routing probability R em : S2.3.

5. Determine the relationship between the total number of iterations T' and the required number of iterations T. If T' <= T, continue to execute S2.3.2-S2.3.

4. Otherwise, output Step 2.4: The decoupled classification layer calculates the bi-norm of the output vector of each digital capsule layer to obtain the probability of the corresponding fault category: usually p j =||n j || Among them, ν j is the vector output by the jth capsule unit in the capsule network, p j is the probability of the fault category corresponding to the jth capsule unit, and a threshold is selected To limit the number of output labels, if The classifier output at this label is 1, otherwise the output is 0; Step 3: Input the training data set into the maximum aggregated attention convolutional capsule network to train the intelligent decoupling diagnosis model of electrostatic hydraulic actuator composite faults, and use the extended loss function to optimize the network to obtain the trained fault diagnosis model; Step 4: Input the test data set into the trained fault diagnosis model to perform intelligent decoupling diagnosis of electrostatic-hydraulic actuator compound faults.

2. The intelligent decoupling diagnosis method for composite faults of electrostatic hydraulic actuator according to claim 1 is characterized by: In step 1, the fault type of the electrostatic hydraulic actuator includes at least one of a hydraulic cylinder leakage fault and a pipeline blockage fault.

3. The intelligent decoupling diagnosis method for composite faults of electrostatic hydraulic actuator according to claim 2 is characterized by: In step 3, the training data set is input into the maximum aggregated attention convolutional capsule network to train the intelligent decoupling diagnosis model of the electrostatic hydraulic actuator composite fault, and the network is optimized using the extended loss function to obtain the trained fault diagnosis model, as follows: The extended loss function is selected as the loss function of the maximum aggregated attention convolution capsule network. The extended loss function can help the model distinguish multiple faults that occur simultaneously; by enlarging the activation value of the correct category and incorrect class activation values The gap between them; the total expansion loss L is defined as: Where N is the total number of samples in the batch, t i is the correct category index of sample i, k i is the wrong category index of sample i, C is the number of categories, m1 is a constant threshold, and m1=0.9 is taken.

4. The intelligent decoupling fault diagnosis method for an electrostatic hydraulic actuator according to claim 3, characterized in that: In step 2.2, after the one-dimensional convolutional neural network feature extraction, the ECA mechanism is used to achieve feature-level data fusion, and the efficiency and accuracy of the neural network in processing information is improved by learning the importance weights between different channels, as follows: First, global average pooling is performed on each channel to extract the global feature information of each channel; then, a one-dimensional convolution kernel is used to capture the local dependencies between channels and generate a vector representing the channel weights; in the ECA mechanism, the size of the convolution kernel is adaptively determined and calculated based on the number of channels; the size of the convolution kernel k is determined by the following formula: In the above formula, (·) odd represents a function that ensures that k is an odd number. b and γ are both hyperparameters, b=1, γ=2; c is the number of channels. The larger the number of channels, the larger the k value, and vice versa. Finally, the output of the one-dimensional convolution is normalized to between 0 and 1 through the Sigmoid activation function σ(·); the calculated channel weight is applied to each channel of the original one-dimensional input to achieve channel recalibration and use it as the input of the capsule network.

5. The intelligent decoupling fault diagnosis method for an electrostatic hydraulic actuator according to claim 1, characterized in that: In step 2.3, the main capsule layer is composed of multiple parallel one-dimensional convolutional layers stacked together, each convolutional layer captures different types of features; the output features of the ECA mechanism are used as input to obtain the feature map of the main capsule layer, and after dimension transformation, multiple primary capsules are obtained, each of which is a vector; the main capsule layer uses the nonlinear function Squash as the activation function, which shrinks the length of the input vector to [0, 1] without changing the direction of the input vector. The modulus of the input vector represents the probability of the existence of the feature; the Squash function is defined as follows: Among them, s jj is the input vector of the jjth capsule, T jj is the output vector of the jjth capsule after applying the Squash function.

Images