Health assessment method of mechatronic transmission system based on core space gradient projection

Abstract

The present invention relates to the field of fault diagnosis technology, and specifically to a health assessment method for an electromechanical composite transmission system based on core space gradient projection, which constructs a multi-core group convolution module and stacks it to construct a deep assessment network to extract multi-scale degradation features without increasing the computational cost, and focuses the feature extraction of the deep assessment network on a preset feature extraction scale. Then, the full life data set of the monitored components of the electromechanical composite transmission system is gradually collected, and the deep assessment network is continuously learned based on gradient projection learning to enhance the learning of new knowledge while preventing the old knowledge from being interfered with. At the same time, knowledge compression and redundancy removal are performed during the continuous learning process to ensure the scalability of the network during the continuous learning process. The present invention has the advantages of higher health assessment accuracy, better stability, and stronger robustness, and can achieve better efficiency and effect in the continuous learning of the assessment network.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and more particularly to a health assessment method for an electromechanical hybrid transmission system based on core space gradient projection. Background Art

[0002] Electromechanical transmission systems are widely used in transportation equipment, industrial manufacturing equipment, special equipment and armored vehicles. During the operation of the equipment, due to overload, improper installation, poor lubrication, complex operating conditions and other reasons, the key components of the electromechanical transmission system are prone to various faults. If not handled in time, these faults will continue to escalate, eventually leading to the paralysis of the transmission system or even the entire equipment, causing greater property losses and even casualties. To prevent such situations from happening, the health assessment of the electromechanical transmission system has become a key technology for providing fault warnings and improving maintenance plans. Health assessment technology can avoid catastrophic failures, reduce maintenance costs, and effectively improve the safety, reliability and economy of the electromechanical transmission system. Thanks to the popularization of industrial Internet of Things and artificial intelligence, the health assessment of electromechanical transmission systems has entered the "big data" era, and many intelligent health assessment methods based on deep learning have emerged.

[0003] In recent years, the problem of how to adapt neural networks to dynamic and gradually collected data sets has become the key to the practical application of health assessment methods based on deep learning, and has received widespread attention. Continuous learning technology has also emerged. Continuous learning technology uses special strategies to enable the network to learn new knowledge while retaining existing knowledge. It enables the evaluation network to gradually improve its performance without laborious retraining. This technology expands the upper bound of the learning ability of the evaluation network and can better adapt to the changing distribution of data sets. However, there are still several important problems that need to be solved in the existing health assessment methods based on continuous learning. 1) The operating conditions of the electromechanical hybrid transmission system will inevitably change, resulting in the degradation characteristics of its key components presenting different time scales. Existing methods usually find it difficult to find an appropriate balance between feature extraction scale and computational cost. 2) Existing methods do not fully consider the forward knowledge transfer, knowledge compression and redundancy removal on the learning trajectory, resulting in poor continuous learning performance in variable operating conditions. Summary of the invention

[0004] In view of this, the present invention provides a health assessment method for an electromechanical hybrid transmission system based on core space gradient projection, which can solve at least one technical problem in the above-mentioned background technology.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A health assessment method for a mechatronic transmission system based on core space gradient projection comprises the following steps:

[0007] Construct a multi-core group convolution module and stack it to build a deep evaluation network; extract the degradation features of the monitored components at different time scales through the multi-core group convolution module, and focus the feature extraction of the deep evaluation network on the preset feature extraction scale;

[0008] Obtain the original vibration signal and speed signal of the monitored component during its entire life cycle under variable speed conditions as the initial data set;

[0009] Entering the initial learning phase, the initial data set is used to train the deep evaluation network. After the initial training, the core input space basis of the initial data set is identified and saved, and a small amount of samples are randomly extracted from the initial data set and saved.

[0010] Enter the continuous learning phase, obtain a new data set and integrate it with the trace samples saved in the previous phase as the data set for the current learning phase;

[0011] The deep evaluation network trained in the previous stage is trained using the data set of the current learning stage. During the training process, the gradient descent direction is guided based on the core input space basis of the previous stage.

[0012] After each continuous learning phase, the core input space basis of the current phase is identified and saved, and a small number of samples are randomly extracted from the current phase data set and saved;

[0013] The continuous learning phase is repeated multiple times until the performance of the deep evaluation network meets the preset accuracy.

[0014] Furthermore, the deep evaluation network consists of N multi-core group convolution modules, N maximum pooling layers and an output layer; N multi-core group convolution modules and N maximum pooling layers are cross-stacked to form a feature extraction network; the feature surface output by the last layer of N maximum pooling layers passes through the flattening function to obtain a degradation feature with elastic scale ; The feature After inputting to the output layer, we get the same monitoring data as the input The corresponding health assessment values ​​of the monitored components of the mechatronic transmission system.

[0015] Furthermore, each multi-core group convolution module consists of P parallel and independent representation learning paths, and the P representation learning paths have different feature extraction scales;

[0016] For the p-th representation learning path of the l-th multi-core group convolution module, the input is divided into Group, apply convolution kernels of the same shape to each group of input feature maps independently, and concatenate the output feature maps along the channel dimension to obtain the output of the representation learning path ;

[0017] The P feature maps obtained by representing the learning path are padded to make them of the same dimension, and then spliced ​​along the channel dimension to obtain the output of the lth multi-core group convolution module.

[0018] Furthermore, the output of the lth multi-core group convolution module It is expressed as:

[0019] ;

[0020] ;

[0021] in, , represents the output of the p-th learning path of the l-th multi-core group convolution module; represents the linear rectification activation function; represents the attention weight, obtained by a two-layer fully connected network, and Represent the weights of the two-layer fully connected network respectively; , represents the i-th group of input feature maps; , indicating the The i-th convolution kernel of the group; , represents the output of the p-th learning path, Indicates the number of input channels, Represents the number of output channels, where and Can be Integer division; ksize p represents the size of the convolution kernel of the p-th representation learning path.

[0022] Furthermore, the size of the p-th convolution kernel representing the learning path is , the pth represents the number of groups corresponding to the learning path .

[0023] Further, the step of identifying the core input space basis of the initial data set includes:

[0024] Input the initial data set into the deep evaluation network after initial training and perform forward propagation to obtain the input representation of the deep evaluation network at layer l and construct a representation matrix ;

[0025] Representation matrix Perform singular value decomposition;

[0026] According to the given threshold, the low-rank matrix is ​​approximated to obtain the matrix The low-rank matrix approximation of ;

[0027] The core input space basis of the input space where the initial dataset is located is constructed based on the low-rank matrix approximation.

[0028] Furthermore, the representation matrix The expression for performing singular value decomposition is:

[0029] ;

[0030] in, is the left singular matrix of the lth layer of the depth evaluation network in the initial stage, is the right singular matrix of the lth layer of the depth evaluation network in the initial stage, and are all orthogonal matrices, is a rectangular diagonal matrix with decreasing non-negative singular values ​​on its diagonal; m represents the dimension of the left singular matrix, and n0 represents the dimension of the right singular matrix;

[0031] The expression for low-rank matrix approximation based on a given threshold is:

[0032] ;

[0033] ;

[0034] in, represents a given threshold; To represent the matrix The k-rank matrix approximation of , represents the Frobenius norm; represents the i-th value in the left singular matrix; represents the i-th singular value; represents the i-th value in the right singular matrix;

[0035] The core input space basis for the initial learning phase is constructed in the following way:

[0036]

[0037] in, Represents the core input space basis for the initial learning phase.

[0038] Furthermore, in the continuous learning stage, the training process of the deep evaluation network is expressed as:

[0039] ;

[0040] ;

[0041] in, Represents the gradient projection of the gradient vector of each layer of the depth evaluation network in the orthogonal direction of the previous core input space, Indicates the gradient of the lth layer generated by the back propagation of the current stage s deep evaluation network; represents the core input space basis of the previous stage s-1; represents the transpose operation, represents the projection of the gradient on the core input space basis of the previous stage, represents the updated weight, Represents the learning rate.

[0042] Furthermore, in the continuous learning stage, the steps of identifying the core spatial basis of the input space where the current stage data set is located include:

[0043] Input the current stage data set into the trained deep evaluation network and perform forward propagation to obtain the input representation of the network at layer l and construct a representation matrix ;

[0044] Calculate the s-1 core input space basis of the previous stage The squares of the singular values ​​corresponding to each basis in ;

[0045] Eliminate the core input space base that already exists in the previous stage The common direction basis of ;

[0046] Representation matrix Perform singular value decomposition, the mathematical expression is as follows:

[0047] ;

[0048] in, Represents the left singular matrix of the lth layer at the current stage, and its column vector is the corresponding new basis ; represents the right singular matrix of the lth layer at the current stage; is the rectangular diagonal matrix of the lth layer in the current stage;

[0049] Input the core of the previous stage into the spatial base The square of the singular value corresponding to each basis in and new base Stitch together to get the vector ;

[0050] The core space foundation of the previous stage and the left singular matrix of the current stage Connect together to get the vector ;

[0051] Sort descending , and reorder accordingly ;

[0052] right Perform low-rank matrix approximation from Select a new base , so that The sum of the first k elements in is greater than ;

[0053] connect and , get the core space base of the current stage .

[0054] Furthermore, in each learning stage, after obtaining the data set, it is preprocessed, including:

[0055] The vibration monitoring signal in the data set Perform normalization processing to obtain the processed vibration monitoring signal , get the sample ,in, is the original speed signal obtained at the tth sampling moment;

[0056] Set a time window of size S to integrate the data And the monitoring signal samples obtained at the previous S-1 sampling moments, the sample is obtained ,in, , ;

[0057] For sample Set the health status value label.

[0058] It can be seen from the above technical solution that compared with the prior art, the present invention has the following beneficial effects:

[0059] 1) The multi-core group convolution module constructed in the present invention enables the evaluation network to extract features from different scales while maintaining a lightweight computing scale, and calibrates high-dimensional features by combining operating condition information to reduce the impact of variable operating conditions on network evaluation performance, thereby improving the accuracy of health assessment of key components of mechatronic transmission systems in continuous learning scenarios.

[0060] 2) In the continuous learning process, the present invention guides the direction of gradient descent based on the core space basis of the previous stage, so as to transfer old knowledge to the new learning stage, thereby enhancing the learning of new knowledge while preventing the old knowledge from being disturbed.

[0061] 3) The present invention can perform knowledge compression and redundancy removal in the process of continuous learning, and ensure the scalability of the network in the process of continuous learning by identifying the core space of previous learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0063] Figure 1 A flow chart of a health assessment method for an electromechanical hybrid transmission system based on core space gradient projection provided by the present invention.

[0064] Figure 2 A schematic diagram of the structure of the multi-core group convolution module provided by the present invention.

[0065] Figure 3 A schematic diagram of the gradient projection learning process provided by the present invention.

[0066] Figure 4 This is an example diagram of the rotational speed value and vibration signal of the monitored components of the electromechanical composite transmission system provided by the present invention.

[0067] Figure 5 This is a diagram of the health assessment results of the monitored components.

[0068] Figure 6 The figure shows the comparison of health assessment performance of various methods in the ablation experiment. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] like Figure 1 As shown, the embodiment of the present invention discloses a health assessment method for a mechatronic transmission system based on core space gradient projection, comprising the following steps:

[0071] Construct a multi-core group convolution module and stack it to build a deep evaluation network; extract the degradation features of the monitored components at different time scales through the multi-core group convolution module, and focus the feature extraction of the deep evaluation network on the preset feature extraction scale;

[0072] Obtain the original vibration signal and speed signal of the monitored component during its entire life cycle under variable speed conditions as the initial data set;

[0073] Entering the initial learning phase, the initial data set is used to train the deep evaluation network. After the initial training, the core input space basis of the initial data set is identified and saved, and a small amount of samples are randomly extracted from the initial data set and saved.

[0074] Enter the continuous learning phase, obtain a new data set and integrate it with the trace samples saved in the previous phase as the data set for the current learning phase;

[0075] The deep evaluation network trained in the previous stage is trained using the data set of the current learning stage. During the training process, the gradient descent direction is guided based on the core input space basis of the previous stage.

[0076] After each continuous learning phase, the core input space basis of the current phase is identified and saved, and a small number of samples are randomly extracted from the current phase data set and saved;

[0077] The continuous learning phase is repeated multiple times until the performance of the deep evaluation network meets the preset accuracy.

[0078] The above steps are further explained below.

[0079] (1) Construction of deep evaluation network.

[0080] The input data of the deep evaluation network is Indicates that is the vibration signal of the monitored component of the electromechanical hybrid transmission system. The value is recorded for the rotational speed of the component being monitored.

[0081] The deep evaluation network consists of N multi-core group convolution modules, N maximum pooling layers and an output layer. N multi-core group convolution modules and N maximum pooling layers are cross-stacked to form a feature extraction network. The feature surface output by the last layer of N maximum pooling layers is passed through the flattening function to obtain a degradation feature with elastic scale. ; The feature After inputting to the output layer composed of the fully connected layer, the input monitoring data is obtained The corresponding health assessment values ​​of the monitored components of the mechatronic transmission system.

[0082] Specifically, if Figure 2 As shown in Fig. 1, each multi-core group convolution module consists of P parallel and independent representation learning paths, and the P representation learning paths have different feature extraction scales. The input vibration signal of the l-th multi-core group convolution module can be expressed as .

[0083] 1.1) For the p-th representation learning path of the l-th multi-core group convolution module, the input Divided into channels group, get ; Then, convolution kernels of the same shape are applied to each set of input feature maps independently, and the output feature maps are concatenated along the channel dimension to obtain the output of the representation learning path , the mathematical expression is as follows:

[0084] ;

[0085] in, , , indicating the The i-th convolution kernel of the group; , represents the output of the pth learning path; is the number of input channels, is the number of output channels. Note that in order to ensure that the input has the appropriate number of channels, it is necessary to first transform the input channel dimension using point-by-point convolution so that its number of channels can be Divisibility, that is, guarantee and Can be Divisible.

[0086] The P feature maps obtained by representing the learning paths are padded to make them of the same dimension, for example, the low-dimensional feature maps are padded with 0 to make the feature map dimensions of each path consistent, and then spliced ​​along the channel dimension to obtain the output of the lth multi-core group convolution module. It should be noted that in the first multi-core group convolution module of the network, the feature map is recalibrated based on the working condition information to adjust the importance of different scales, aiming to dynamically focus the network's attention on the appropriate scale according to the working condition information of the electromechanical hybrid transmission. The recalibration process is as follows:

[0087] The output of the lth multi-core group convolution module It is expressed as:

[0088] ;

[0089] ;

[0090] in, , represents the output of the p-th learning path of the l-th multi-core group convolution module; Represents the Rectified linear unit (ReLU) activation function; represents the attention weight, obtained by a two-layer fully connected network, and Represent the weights of the two-layer fully connected network respectively; , represents the i-th group of input feature maps; ksize p represents the size of the convolution kernel of the p-th representation learning path.

[0091] It should be noted that the convolution kernels representing the learning paths (also called convolution paths) have different sizes and different numbers of groups. The larger the convolution kernel size, the more corresponding groups there are. The size of the convolution kernel representing the learning path is , the pth represents the number of groups corresponding to the learning path . It is the number of groups of the first path with the smallest convolution kernel size.

[0092] 1.2) The output of the multi-core group convolution module Input to the maximum pooling layer to obtain the maximum value of the elements in the non-overlapping pooling area and obtain the pooled features ; The calculation expression is as follows:

[0093] ;

[0094] in, represents the maximum pooling function; p is the pooling size; s is the pooling moving step.

[0095] 1.3) Since N multi-core group convolution modules and N maximum pooling layers are cross-stacked, the tiling function is applied to the feature surface output by the last layer to obtain degraded features with elastic scale. , the calculation expression is as follows:

[0096]

[0097] in, Represents the tiling operation function; is the output feature of the last pooling layer.

[0098] 1.4) Features Input to the fully connected layer to get the same monitoring data as the input Corresponding health assessment values ​​of monitored components of electromechanical hybrid transmission system , the calculation expression is as follows:

[0099] ;

[0100] in, represents the weight matrix; represents the bias vector; Represents the ReLU activation function.

[0101] (2) Acquisition of the initial data set.

[0102] Obtain the original vibration and speed signals of the monitored components of the electromechanical hybrid transmission system under variable speed conditions throughout their lifespan ,in, represents the original vibration signal obtained at the tth sampling moment, represents the original speed signal obtained at the tth sampling time, Q is the number of signal samples, H1 is the number of data points contained in each signal sample, and C1 is the number of vibration acceleration sensors.

[0103] (3) Preprocessing of the initial dataset.

[0104] Preprocess the original vibration signal and speed signal to form a data set: First, the vibration monitoring signal Perform Z-score normalization processing, and then embed the valid time information data before the sampling time into the current data through the time window embedding strategy. The specific steps are as follows:

[0105] 3.1) Vibration monitoring signals in the dataset using Z-score Perform normalization processing to obtain the processed vibration monitoring signal , the Z-score standardization operation expression is as follows:

[0106] ;

[0107] in, is the original vibration monitoring signal; is the mean value of the original vibration signal; is the standard deviation of the original signal.

[0108] 3.2) The pre-processed vibration monitoring signal With the original speed signal Combine and get samples .

[0109] Set a time window of size S to integrate the data And the monitoring signal samples obtained at the previous S-1 sampling moments, the sample is obtained ,in, , ;

[0110] 3.3) For the sample Set the health status value label, calculated as follows:

[0111] ;

[0112] in, For sample The corresponding health status value, is the remaining service life corresponding to the sample, that is, the time span from the operation of the monitored component to complete failure, and Y is the actual full life value of the monitored component corresponding to the sample.

[0113] (4) Initial learning stage, that is .

[0114] 4.1) Repeat steps (2) and (3) until an initial dataset of a certain size is obtained that can be used to train the network. ,in, is the number of samples in the initial data set, The monitoring signals of the monitored components of the electromechanical hybrid transmission system obtained in the initial stage (i.e., vibration signals and operating status records) are is the corresponding health status value.

[0115] 4.2) Using the Initial Dataset A deep evaluation network built based on the Stochastic Gradient Descent (SGD) optimization algorithm training , the mathematical expression is as follows:

[0116] ;

[0117] in, is the weight set of all layers of the network, L is the number of network layers, is the weight of the lth layer; is the loss value generated by the network forward propagation, is the loss function, The gradients for each layer generated by backpropagating through the network.

[0118] 4.3) Set the number of iterations E, repeat step 4.2), iteratively update the parameters of the deep evaluation network to obtain the optimal health evaluation network .

[0119] 4.4) After the initial stage of training, identify the core space of the input space where the initial data set is located. The specific steps are as follows:

[0120] 4.4.1) Each sample in the initial data set Input to the deep evaluation network after initial training and perform forward propagation to obtain the input representation of the deep evaluation network at layer l and construct a representation matrix , the mathematical expression is as follows:

[0121] ;

[0122] 4.4.2) Representation Matrix Perform singular value decomposition (SVD), the mathematical expression is as follows:

[0123] ;

[0124] in, is the left singular matrix of the lth layer of the depth evaluation network in the initial stage, is the right singular matrix of the lth layer of the depth evaluation network in the initial stage, and are all orthogonal matrices, It is a rectangular diagonal matrix with decreasing non-negative singular values ​​on the diagonal; m represents the dimension of the left singular matrix, and n0 represents the dimension of the right singular matrix.

[0125] 4.4.3) Perform low-rank matrix approximation (LRA) according to the given threshold to obtain the matrix The low-rank matrix approximation of is expressed as:

[0126] ;

[0127] ;

[0128] in, represents a given threshold; To represent the matrix The k-rank matrix approximation of , represents the Frobenius norm; represents the i-th value in the left singular matrix; represents the i-th singular value; Represents the i-th value in the right singular matrix.

[0129] The core input space basis of the input space where the initial dataset is located is constructed based on the low-rank matrix approximation, which is constructed in the following way:

[0130] ;

[0131] in, Represents the core input space basis for the initial learning phase.

[0132] 4.4.4) Save , and from the initial dataset Randomly sample micro samples in the process for continuous learning in subsequent stages.

[0133] (5) Continuous learning stage, that is ,like Figure 3 As shown, the specific steps include:

[0134] 5.1) Repeat 2) and step 3) until a new dataset of a certain size is obtained that can be used for continuous learning of the network. ,in, is the number of samples in the newly collected dataset, The monitoring signals (vibration signals and operating status records) of the monitored components of the electromechanical hybrid transmission system obtained during the continuous learning phase are: For the corresponding real assessment.

[0135] 5.2) Integrating new datasets and the trace samples preserved in the previous stage , and based on the gradient projection learning training the optimized depth evaluation network obtained in the previous stage , the mathematical expression is as follows:

[0136] ;

[0137] ;

[0138] in, Represents the gradient projection of the gradient vector of each layer of the depth evaluation network in the orthogonal direction of the previous core input space, Indicates the gradient of the lth layer generated by the back propagation of the current stage s deep evaluation network; represents the core input space basis of the previous stage s-1; represents the transpose operation, represents the projection of the gradient on the core input space basis of the previous stage, represents the updated weight, Represents the learning rate. Through the above-mentioned gradient projection learning, the knowledge of the previous learning stage is explicitly transferred to the new learning stage through the modified gradient, thus avoiding the interference of old knowledge and strengthening the learning of new knowledge.

[0139] 5.3) Set the number of iterations E, repeat the training process of step 4.2), iteratively update the parameters of the deep evaluation network to obtain the optimal health evaluation network .

[0140] 5.4) After each continuous learning phase, identify the core input space basis of the data set for that phase. The specific steps include:

[0141] 5.4.1) Each sample in the current stage of the data set Input to the trained deep evaluation network And perform forward propagation to obtain the input representation of the network at layer l and construct a representation matrix , the expression is:

[0142] ;

[0143] 5.4.2) Calculate the s-1 core input space basis of the previous stage The square of the singular value corresponding to each basis in is calculated as follows:

[0144] ;

[0145] in, yes The base of is the square of its corresponding singular value, Reflects the base and the correlation between the learning stage s.

[0146] 5.4.3) Eliminate the core input space basis that already exists in the previous stage The common direction basis of , the specific expression is as follows:

[0147] ;

[0148] in, To eliminate the representation matrix after the common direction, the above formula ensures that the newly added basis is unique and orthogonal to the existing basis in memory.

[0149] 5.4.4) Representation Matrix Perform singular value decomposition, the mathematical expression is as follows:

[0150] ;

[0151] in, Represents the left singular matrix of the lth layer at the current stage, and its column vector is the corresponding new basis ; represents the right singular matrix of the lth layer at the current stage; is the rectangular diagonal matrix of the lth layer in the current stage;

[0152] 5.4.5) Select the most important basis from the previous basis and the new basis, including:

[0153] Input the core of the previous stage into the spatial base The square of the singular value corresponding to each basis in and new base Stitch together to get the vector ;

[0154] The core space foundation of the previous stage and the left singular matrix of the current stage Connect together to get the vector ;

[0155] Sort descending , and reorder accordingly ;

[0156] right Perform low-rank matrix approximation from Select a new base , so that The sum of the first k elements in is greater than ;

[0157] connect and , get the core space base of the current stage .

[0158] 5.4.6) Save , and from the current dataset Random sampling of a very small number of samples , for continuous learning in subsequent stages.

[0159] (6) Repeat step (5) until the deep evaluation network achieves ideal health assessment performance.

[0160] (7) The pre-processed vibration and speed monitoring signals of the monitored components of the electromechanical hybrid transmission system are Input into the optimal deep evaluation network to assess the health of the component.

[0161] The effectiveness of the method of the present invention is verified below based on an accelerated degradation experimental data set of an electromechanical hybrid transmission system.

[0162] The electromechanical composite transmission system used in this embodiment is mainly composed of a motor, a rotating shaft, a bearing, and a bearing support. The bearing is the monitored component, and its full life data from operation to failure is collected. The model of the tested bearing is LDKUER204, which is driven by an AC motor, and its speed is controlled by a motor speed controller. During the experiment, the radial load of the tested bearing is applied through a hydraulic loading device. The IMI608A11 accelerometer is used to collect the vertical and horizontal vibration signals of the tested bearing, with a sampling frequency of 25.6kHz, and 2560 data points are recorded every 12 seconds (that is, 0.1 second of data is collected every 12 seconds).

[0163] As shown in Table 1, a total of 14 tested bearings were subjected to accelerated degradation experiments under three different radial forces, and the speed was randomly changed under each radial force.

[0164] Table 1

[0165]

[0166] The speed and vibration signal values ​​of bearing 3-5 are as follows: Figure 4 shown.

[0167] To build a continuous learning scenario, these bearings are arranged to be learned by the deep evaluation network in three stages, where the test dataset is fixed and the remaining labeled datasets are assigned to different stages in sequence, as shown in Table 2.

[0168] Table 2

[0169]

[0170] The hyperparameter settings are shown in Table 3.

[0171] Table 3

[0172]

[0173] The health assessment method of the present invention is used to perform the assessment results of bearings 3-5 at each learning stage. Figure 5 As shown. Figure 5 It can be seen that as the learning continues, the evaluation results of the network gradually become more accurate, which shows that the method of the present invention can use the newly acquired data to gradually improve the evaluation performance, rather than retraining the entire network from scratch using all the accumulated data.

[0174] In order to verify the effectiveness of the multi-core group convolution module in the method of the present invention, the method of the present invention was subjected to an ablation test. Method A is composed of a stack of ordinary convolution layers. Method B introduces multiple parallel convolution paths during the network construction process, and different paths contain convolution kernels of different sizes. Method C is similar to Method B and also includes multiple convolution paths with different kernel sizes. The difference is that Method C also introduces a self-attention mechanism in each convolution path to enhance the network feature description capability. In addition, consistent with the method proposed in the present invention, the feature maps output by each convolution path in Method B and Method C will be spliced ​​together as the input of the next layer. Method D uses multi-core group convolution, but the feature map is not recalibrated based on the working condition information during the network construction process.

[0175] Methods A, B, C and D are used as comparison methods. Figure 6 The root mean square error (RMSE), number of parameters and floating point operations (FLOPs) of the comparison method and the method of the present invention on the test data set are shown. Figure 6It can be seen that the proposed method does not introduce too many parameters compared with the comparison method, and the amount of calculation is much lower than the comparison method, but it obtains performance similar to the comparison method, which proves the effectiveness of the multi-core group convolution module. Furthermore, a comparison method is constructed to verify the effectiveness of gradient projection learning in the method of the present invention in evaluating the continuous learning of the network. Method E does not add gradient projection learning during the network training process, while method F introduces gradient projection learning. Method E and method F have the same network structure, and both perform continuous learning based on the same initial stage. In addition, the embodiment designs an average forgetting value (Average forgetting, AF) to measure the forgetting of the evaluation network, and its calculation formula is:

[0176] ;

[0177] in, , represents the kth amount of forgetting after the network completes the sth stage of learning, is the mean absolute percentage error on the kth training set after the network completes the sth stage. The comparison results of the two methods are shown in Table 4.

[0178] Table 4

[0179]

[0180] As can be seen from Table 4, the RMSE value and AF value of method E on the test data set are always higher than those of method F. This shows that gradient projection learning can further improve the health assessment performance while avoiding catastrophic forgetting. This shows that the health assessment accuracy of the method of the present invention is higher, the stability is better, and the robustness is stronger in the continuous learning process.

[0181] Through the above health assessment results of the monitored components of the mechatronic transmission system and the comparison of the assessment performance with multiple comparative methods, it can be found that the multi-core group convolution module in the method of the present invention can enable the evaluation network to extract features from different time scales while maintaining a lightweight computing scale, and calibrate high-dimensional features by combining working condition information to reduce the impact of variable working conditions on the network evaluation performance, thereby improving the accuracy of mechanical health assessment in continuous learning scenarios. In addition, the core space gradient projection in the inventive method can prevent old knowledge from being interfered with while enhancing the learning of new knowledge, and perform knowledge compression and de-redundancy in the continuous learning process to ensure the scalability of the network in the continuous learning process. By combining the above-mentioned features, the method of the present invention can achieve better efficiency and effect in the continuous learning of the health assessment network of the mechatronic transmission system.

[0182] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0183] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A health assessment method for electromechanical hybrid transmission system based on core space gradient projection, characterized in that: The following steps are involved: Construct a multi-core group convolution module and stack it to build a deep evaluation network; extract the degradation features of the monitored components at different time scales through the multi-core group convolution module, and focus the feature extraction of the deep evaluation network on the preset feature extraction scale; Obtain the original vibration signal and speed signal of the monitored component during its entire life cycle under variable speed conditions as the initial data set; Entering the initial learning phase, the initial data set is used to train the deep evaluation network. After the initial training, the core input space basis of the initial data set is identified and saved, and a small amount of samples are randomly extracted from the initial data set and saved. Enter the continuous learning phase, obtain a new data set and integrate it with the trace samples saved in the previous phase as the data set for the current learning phase; The deep evaluation network trained in the previous stage is trained using the data set of the current learning stage. During the training process, the gradient descent direction is guided based on the core input space basis of the previous stage. After each continuous learning phase, the core input space basis of the current phase is identified and saved, and a small number of samples are randomly extracted from the current phase data set and saved; Repeat the continuous learning phase multiple times until the performance of the deep evaluation network meets the preset accuracy; Each multi-core group convolution module consists of P parallel and independent representation learning paths, and the P representation learning paths have different feature extraction scales; For the p-th representation learning path of the l-th multi-core group convolution module, the input is divided into G according to the channel dimension p The convolution kernels with the same shape are applied to each group of input feature maps independently, and the output feature maps are concatenated along the channel dimension to obtain the output h of the representation learning path. p l ; The P feature maps obtained by representing the learning path are padded to make them have the same dimension, and then spliced ​​along the channel dimension to obtain the output of the l-th multi-core group convolution module; The output O of the lth multi-core group convolution module l It is expressed as: in, represents the output of the p-th learning path of the l-th multi-core group convolution module; τ(·) represents the linear rectification activation function; α=[a1,a2,…a P ] represents the attention weight, obtained by a two-layer fully connected network. and Represent the weights of the two-layer fully connected network respectively; represents the i-th group of input feature maps, and L represents the number of network layers; Indicates G p The i-th convolution kernel of the group; represents the output of the p-th learning path, C i Indicates the number of input channels, C o Indicates the number of output channels, where C i and C o Can be G p Integer division; ksize p represents the size of the convolution kernel of the pth representation learning path; The steps to identify the core input spatial basis of the initial dataset include: Input the initial data set into the deep evaluation network after initial training and perform forward propagation to obtain the input representation of the deep evaluation network at layer l and construct a representation matrix Representation matrix Perform singular value decomposition; According to the given threshold, the low-rank matrix is ​​approximated to obtain the matrix A low-rank matrix approximation of ; Construct the core input space basis of the input space where the initial data set is located based on the low-rank matrix approximation; In the continuous learning phase, the steps to identify the core spatial basis of the input space where the current data set is located include: Input the current stage data set into the trained deep evaluation network and perform forward propagation to obtain the input representation of the network at layer l and construct a representation matrix Calculate the s-1 core input space basis of the previous stage The squares of the singular values ​​corresponding to each basis in ; Eliminate the core input space base that already exists in the previous stage The common direction basis of Representation matrix Perform singular value decomposition, the mathematical expression is as follows: in, Represents the left singular matrix of the lth layer at the current stage, and its column vector is the corresponding new basis m is the pair representation matrix The dimension of the left singular matrix when performing singular value decomposition, n s The number of newly collected dataset samples for stage s; represents the right singular matrix of the lth layer at the current stage; is the rectangular diagonal matrix of the lth layer in the current stage; Input the core of the previous stage into the spatial base The square of the singular value corresponding to each basis in and new base Stitch together to get the vector The core space foundation of the previous stage and the left singular matrix of the current stage Connect together to get the vector Sort descending and reorder accordingly right Perform low-rank matrix approximation from Select a new base Make The sum of the first k elements in is greater than connect and Get the core space base of the current stage 2. The method for health assessment of electromechanical transmission system based on core space gradient projection according to claim 1 is characterized in that: The deep assessment network consists of N multi-core group convolution modules, N maximum pooling layers and an output layer; N multi-core group convolution modules and N maximum pooling layers are cross-stacked to form a feature extraction network; the feature surface output by the last layer of N maximum pooling layers is passed through the flattening function to obtain the degradation feature H with elastic scale; after the feature H is input into the output layer, the health assessment value of the monitored component of the mechatronic transmission system corresponding to the input monitoring data {x, v} is obtained.

3. The method for health assessment of electromechanical transmission system based on core space gradient projection according to claim 1 is characterized in that: The size of the pth convolution kernel representing the learning path is ksize p =2p+1, the number of groups corresponding to the p-th learning path is G p =2 p G0.

4. The method for health assessment of electromechanical transmission system based on core space gradient projection according to claim 1 is characterized in that: Representation matrix The expression for performing singular value decomposition is: in, is the left singular matrix of the lth layer of the depth evaluation network in the initial stage, is the right singular matrix of the lth layer of the depth evaluation network in the initial stage, and are all orthogonal matrices, is a rectangular diagonal matrix with decreasing non-negative singular values ​​on its diagonal; m represents the dimension of the left singular matrix, and n0 represents the dimension of the right singular matrix; The expression for low-rank matrix approximation based on a given threshold is: Among them, ξ represents the given threshold; To represent the matrix The k-rank matrix approximation of , whose rank k≤min{m,n0}, ||·|| F represents the Frobenius norm; represents the i-th value in the left singular matrix; represents the i-th singular value; represents the i-th value in the right singular matrix; The core input space basis for the initial learning phase is constructed in the following way: in, Represents the core input space basis for the initial learning phase.

5. The method for health assessment of electromechanical transmission system based on core space gradient projection according to claim 1, characterized in that: In the continuous learning stage, the training process of the deep evaluation network is expressed as: in, Represents the gradient projection of the gradient vector of each layer of the depth evaluation network in the orthogonal direction of the previous core input space, Indicates the gradient of the lth layer generated by the back propagation of the current stage s deep evaluation network; represents the core input space basis of the previous stage s-1; (·) T represents the transpose operation, represents the projection of the gradient on the core input space basis of the previous stage, θ l represents the updated weight, and α represents the learning rate.

6. The method for health assessment of electromechanical transmission system based on core space gradient projection according to claim 1, characterized in that: In each learning stage, after obtaining the data set, it is preprocessed, including: For the vibration monitoring signal x in the data set t Perform normalization processing to obtain the processed vibration monitoring signal x Z t , we get the sample {x Z t , v t }, where v t is the original speed signal obtained at the tth sampling moment; Set the time window size to S and integrate the data {x Z t , v t } and the monitoring signal samples obtained at the previous S-1 sampling moments to obtain samples {x, v}, where v={v t-S+1 , v t-S+2 , …, v t }; Set the health status value label for the sample {x, v}.

Images