A Dynamic Balancing Method for Modal Clearance in the Identification of the Health Category of Electromechanical Equipment

Abstract

The present invention discloses a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment, belonging to the technical field of electromechanical equipment, and comprising the following steps: S1: Input sensor data into a feature extraction network to obtain feature vectors; S2: Calculate the multi-modal classification loss and the multi-modal matching loss; S3: Based on weight parameters, perform weighted summation on the multi-modal classification loss and the multi-modal matching loss to obtain the multi-modal total loss; Based on the gradient descent method, the selected weight parameter update strategy, the multi-modal total loss, and the multi-modal classification loss, update the weight parameters and network parameters; S4: Iterate steps S1-S3 to obtain the final network parameters of the feature extraction network. The present invention can effectively solve the modal gap problem existing in the identification of the health category of electromechanical equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of electromechanical equipment, and more specifically, to a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment. Background Art

[0002] At present, with the continuous improvement of the complexity of electromechanical equipment and the increasingly harsh operating environment, the traditional health category identification method based on a single mode has shown obvious limitations. The multi-modal health category identification method can capture health category features from different physical dimensions by integrating multi-modal sensor data such as vibration, sound, and current, significantly improving the identification accuracy. However, there are significant distribution differences (i.e., modal gaps) in the feature space after the data of different modal sensors pass through the feature extraction network. This heterogeneity makes it difficult to synchronously optimize the feature extraction networks of each mode during joint training, resulting in gradient conflicts and asynchronous convergence problems between modes, thereby reducing the accuracy and robustness of the multi-modal health category identification model;

[0003] In response to the above problems, existing scholars mainly focus on alleviating the difference in convergence speed during the training process and have not fundamentally solved the modal gap problem existing in the identification of the health category of electromechanical equipment.

[0004] Therefore, how to provide a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment,

[0005] which can effectively solve the modal gap problem existing in the identification of the health category of electromechanical equipment, thereby improving the identification accuracy and robustness of the multi-modal health category identification model is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment.

[0007] To achieve the above purpose, the present invention adopts the following technical solutions:

[0008] A dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment includes the following steps:

[0009] S1: Input the sensor data of the electromechanical equipment into the feature extraction network to obtain the feature vector ; where represents the sensor data of the j-th mode in the i-th training sample at the t-th iteration; represents the feature extraction network corresponding to the j-th mode; represents the corresponding feature vector; t = 1, 2,... ; where \(i = 1, 2, \cdots, n\) and \(j = 1, 2, \cdots, m\); represents the maximum number of iterations; \(n\) represents the number of training samples used in one iteration; \(m\) represents the number of modalities included in one training sample;

[0010] S2: Input the feature vectors after feature fusion into the classification layer to obtain the predicted probability of the healthy category ; where represents the predicted probability of the healthy category of the \(i\)-th training sample at the \(t\)-th iteration;

[0011] Based on the predicted probability of the healthy category and the true healthy category label calculate the multi-modal classification loss ; where represents the true healthy category label of the \(i\)-th training sample at the \(t\)-th iteration; represents the multi-modal classification loss at the \(t\)-th iteration;

[0012] Based on the feature vectors calculate the multi-modal matching loss ; where represents the multi-modal matching loss at the \(t\)-th iteration;

[0013] S3: Based on the weight parameter perform weighted summation on the multi-modal classification loss and the multi-modal matching loss to obtain the multi-modal total loss ; where represents the weight parameter before update at the \(t\)-th iteration, is the initial value of the weight parameter; represents the multi-modal total loss at the \(t\)-th iteration;

[0014] Based on the gradient descent method, the selected weight parameter update strategy, the multi-modal total loss the multi-modal classification loss update the weight parameter and the network parameter ; where represents the network parameter before update of the feature extraction network at the \(t\)-th iteration, is the initial value of the network parameter of the feature extraction network ;

[0015] S4: Iterate times for S1 - S3 to obtain the feature extraction network The final network parameters.

[0016] Preferably, the multi-modal classification loss is obtained based on the following formula:

[0017] ;

[0018] where denotes the transpose of.

[0019] Preferably, the predicted health class probability is obtained based on the following formula:

[0020] ;

[0021] ;

[0022] ;

[0023] where denotes the fused feature vector at the t-th iteration; F denotes the feature fusion operation; denotes the feature vector corresponding to the sensor data ; denotes the feature vector corresponding to the sensor data ; W and b denote the weight and bias of the fully connected layer in sequence; softmax denotes the softmax function; the softmax function and the fully connected layer constitute the classification layer.

[0024] Preferably, the multi-modal matching loss is calculated based on the following formula ;

[0025] ;

[0026] where ; denotes the similarity between the sensor data and the sensor data ; denotes the similarity between the sensor data and the sensor data ; denotes the similarity between the sensor data and the sensor data ; denotes the number of combinations obtained by selecting two different modalities from m modalities for combination; denotes the smoothing parameter.

[0027] Preferably, the similarity and the similarity And similarity The calculation formula is as follows:

[0028] ;

[0029] ;

[0030] ;

[0031] Wherein, represents the feature vector corresponding to the sensor data ; represents the feature vector corresponding to the sensor data ; represents the feature vector corresponding to the sensor data ; represents the feature vector corresponding to the sensor data ; T represents transpose; , , and respectively represent , , and 's L2 norm.

[0032] Preferably, if the selected weight parameter update strategy is a strategy based on the attenuation factor, S3 specifically includes the following steps:

[0033] S31: Calculate the multi-modal total loss based on the formula ;

[0034] S32: Minimize the multi-modal total loss using the gradient descent method, and update the network parameter to ; Wherein, represents the updated network parameter of the feature extraction network at the t-th iteration;

[0035] S33: Update the weight parameter based on the formula to ; Wherein, represents the updated weight parameter at the t-th iteration.

[0036] Preferably, if the selected weight parameter update strategy is a constraint-aware hybrid loss strategy, S3 specifically includes the following steps:

[0037] S31’: Calculate the multi-modal total loss based on the formula ;

[0038] S32’: Minimize the multi-modal total loss using the gradient descent method , and update the network parameters to ; where represents the updated network parameters of the feature extraction network at the t-th iteration;

[0039] S33’: Calculate the gradient of the multi-modal classification loss with respect to the weight parameter based on the multi-modal total loss and the multi-modal classification loss ; ;

[0040] S34’: Update the weight parameter based on the gradient to ; where represents the updated weight parameter at the t-th iteration.

[0041] Preferably, S33’ is implemented based on the following formula:

[0042] ;

[0043] where represents the Hessian matrix of the multi-modal total loss with respect to the network parameter and the weight parameter ; represents the Hessian matrix of the multi-modal total loss with respect to the network parameter ; represents the gradient of the multi-modal classification loss with respect to the network parameter ; represents the inverse matrix of.

[0044] Preferably, S34’ is implemented based on the following formula:

[0045] ;

[0046] where represents the learning rate.

[0047] Preferably, the feature extraction network is a convolutional neural network.

[0048] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment, which can effectively solve the modal gap problem existing in the identification of the health category of electromechanical equipment, thereby improving the identification accuracy and robustness of the multi-modal health category identification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0050] Figure 1 It is a flowchart of a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment provided by the present invention;

[0051] Figure 2 It is a performance comparison diagram of five methods in the embodiments provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] As Figure 1 shown, the embodiments of the present invention disclose a dynamic balance method for modal gaps in the identification of the health category of electromechanical equipment, including the following steps:

[0054] S1: Input the sensor data of the electromechanical equipment into the feature extraction network to obtain a feature vector (which is a d-dimensional feature vector); where represents the sensor data of the jth mode in the ith training sample at the tth iteration; represents the feature extraction network corresponding to the jth mode; represents the corresponding feature vector; t = 1, 2,..., ; i = 1, 2,..., n, j = 1, 2,..., m; represents the maximum number of iterations; n represents the number of training samples used in one iteration; m represents the number of modes included in a training sample;

[0055] In one embodiment, the feature extraction network is a convolutional neural network.

[0056] The convolutional neural network from input to output includes a 5*5 convolutional layer, a ReLU activation function layer, a batch normalization layer, a 2*2 max pooling layer, a 7*7 convolutional layer, a ReLU activation function layer, a batch normalization layer, a 2*2 max pooling layer, a 9*9 convolutional layer, a ReLU activation function layer, a batch normalization layer, and a global average pooling layer connected in sequence. The network structures of the convolutional neural networks adopted by each modality are the same except for the network parameters.

[0057] S2: After fusing the respective feature vectors input them into the classification layer to obtain the predicted probability of the healthy class ; where represents the predicted probability of the healthy class of the i-th training sample at the t-th iteration;

[0058] Based on the predicted probability of the healthy class and the true healthy class label calculate the multi-modal classification loss ; where represents the true healthy class label of the i-th training sample at the t-th iteration; represents the multi-modal classification loss at the t-th iteration;

[0059] In one embodiment, the multi-modal classification loss is obtained based on the following formula:

[0060] ;

[0061] where represents the transpose of.

[0062] In one embodiment, the predicted probability of the healthy class is obtained based on the following formula:

[0063] ;

[0064] ;

[0065] ;

[0066] where represents the fused feature vector at the t-th iteration (which is a d*m-dimensional feature vector); F represents the feature fusion operation; represents the sensor data The corresponding feature vector; Represents sensor data The corresponding feature vector; W and b represent the weights and biases of the fully connected layer in sequence (W is a matrix of C*d; b is a matrix of C*m; C represents the number of health categories); softmax represents the softmax function; the softmax function and the fully connected layer constitute the classification layer.

[0067] It can be understood that: predicting the probability of the health category Includes predicting the probabilities of C (the number of health categories) health categories respectively;

[0068] It can be understood that: the true health category label Is represented by a One-Hot vector, that is, a binary vector of length C (the number of health categories); only one element in each binary vector is 1, indicating that the current sample belongs to this category, and the remaining elements are all 0, indicating that the current sample does not belong to the remaining categories.

[0069] It can be understood that: each feature extraction network 、The fully connected layer and the softmax function constitute a multi-modal health category identification model, and the multi-modal health category identification model identifies the health category of the electromechanical device based on the sensor data of the electromechanical device. Taking a three-phase motor as an example, the sensor data of the three-phase motor Includes three-phase vibration acceleration signals, three-phase current signals, rotational speed signals and acoustic signals; the health categories of the three-phase motor include normal, rolling element fault, inner race fault, outer race fault, cage fault, rotor imbalance, rotor bending, rotor broken bar, motor phase loss, voltage imbalance, short circuit.

[0070] Based on the feature vector Calculate the multi-modal matching loss ; where Represents the multi-modal matching loss at the t-th iteration;

[0071] In one embodiment, the multi-modal matching loss is calculated based on the following formula ;

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] Where ; Represents sensor data and the similarity of the sensor data ; represents the sensor data and the similarity of the sensor data ; represents the sensor data and the sensor data ; represents the number of combinations obtained by selecting two different modalities from m modalities for combination; represents the smoothing parameter; represents the sensor data corresponding feature vector; represents the sensor data corresponding feature vector; represents the sensor data corresponding feature vector; represents the sensor data corresponding feature vector; T represents transpose; , , and successively represent , , and L2 norm of.

[0077] It can be understood that:

[0078] represents the matching loss between two modalities; represents the total modality matching loss.

[0079] S3: Based on the weight parameter weight the multi-modal classification loss and the multi-modal matching loss to obtain the multi-modal total loss ; where represents the weight parameter before update at the t-th iteration, is the initial value of the weight parameter (this initial value is a manually set value); represents the multi-modal total loss at the t-th iteration;

[0080] Based on the gradient descent method, the selected weight parameter update strategy, the multi-modal total loss the multi-modal classification loss update the weight parameter and the network parameter ; where represents the feature extraction network at the t-th iteration The network parameters before the update, are the initial values of the network parameters for the feature extraction network (the initial values are artificially set values);

[0081] In one embodiment, if the selected weight parameter update strategy is a strategy based on a decay factor, then S3 specifically includes the following steps:

[0082] S31: Based on the formula calculate to obtain the multimodal total loss ;

[0083] S32: Use the gradient descent method to minimize the multimodal total loss , and update the network parameters to be ; where represents the network parameters after the update of the feature extraction network at the t-th iteration;

[0084] S33: Based on the formula , update the weight parameter to be ; where represents the updated weight parameter at the t-th iteration.

[0085] In one embodiment, if the selected weight parameter update strategy is a constrained-aware hybrid loss strategy, then S3 specifically includes the following steps:

[0086] S31’: Based on the formula calculate to obtain the multimodal total loss ;

[0087] S32’: Use the gradient descent method to minimize the multimodal total loss , and update the network parameters to be ; where represents the network parameters after the update of the feature extraction network at the t-th iteration;

[0088] S33’: Based on the multimodal total loss , the multimodal classification loss calculate the gradient of the multimodal classification loss with respect to the weight parameter ;

[0089] In one embodiment, S33’ is implemented based on the following formula:

[0090] ;

[0091] Among them, represents the multi-modal total loss with respect to the network parameters and the weight parameters Hessian matrix (i.e., the multi-modal total loss for the network parameters and the weight parameters second-order partial derivative); represents the multi-modal total loss with respect to the network parameters Hessian matrix (i.e., the multi-modal total loss for the network parameters second-order partial derivative); represents the gradient of the multi-modal classification loss with respect to the network parameters (i.e., the multi-modal classification loss for the network parameters first-order partial derivative); represents inverse matrix.

[0092] S34’: Update the weight parameter based on the gradient to be ; Among them, represents the updated weight parameter at the t-th iteration.

[0093] In a certain embodiment, S34’ is implemented based on the following formula:

[0094] ;

[0095] Among them, represents the learning rate.

[0096] S4: Repeat S1 - S3 for T times to obtain the final network parameters of the feature extraction network .

[0097] It can be understood that: during the iteration process, the parameters of the fully connected layer are updated by minimizing the multi-modal classification loss at the same time.

[0098] Finally, taking a three-phase motor in a mechanical and electrical equipment as an example, the present invention simulated 11 different states (i.e., health categories) of the three-phase motor, including normal state, 4 bearing health categories (rolling element fault, inner race fault, outer race fault, cage fault), 3 rotor health categories (rotor imbalance, rotor bending, rotor bar breakage), and 3 electrical health categories (motor phase loss, voltage imbalance, short circuit); and collected sensor data of 4 modes in various health states, including three-phase vibration acceleration signals, three-phase current signals, rotational speed signals, and acoustic signals, with a sampling frequency of 25.6 kHz for all. 500 samples were collected for each health category, and each sample included sensor data of 8 channels (i.e., vibration acceleration signals in the X direction, Y direction, and Z direction, current signals of phase A, phase B, phase C, rotational speed signal, and acoustic signal), and the sensor data of each mode contained 2048 data points.

[0099] Based on the collected data, the method of the present invention was compared with four advanced multi-modal imbalance learning methods (PMR, OGM, AGM, MLA). The method of the present invention, PMR, OGM, AGM, and MLA all adopted the same feature extraction network (as shown in Table 1) and hyperparameter settings (as shown in Table 2), and specifically used accuracy (ACC), mean average precision (MAP), and macro F1 index to evaluate the performance of each method (see Table 3 and Figure 2 )). Through Table 3 and Figure 2 It can be seen that the health identification effect of the method of the present invention is significantly better than that of PMR, OGM, AGM, and MLA, and the model of the present invention adopting the constraint-aware hybrid loss strategy performs better than the model adopting the strategy based on the attenuation factor. The method of the present invention effectively solves the modal gap problem in the health category identification of mechanical and electrical equipment and improves the health category identification accuracy.

[0100] Table 1. Network structure of the feature extraction network

[0101]

[0102] Table 2. Hyperparameter settings

[0103] Table 3. Performance comparison of different methods

[0104]

[0105] In the present specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0106] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can 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 broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic balance method for modal clearance in the identification of the health category of electromechanical equipment, characterized in that, It includes the following steps: S1: Input the sensor data of the electromechanical equipment into the feature extraction network to obtain the feature vector ; where represents the sensor data of the j-th modality in the i-th training sample at the t-th iteration; represents the feature extraction network corresponding to the j-th modality; represents the corresponding feature vector; t = 1, 2,..., ; i = 1, 2,..., n, j = 1, 2,..., m; represents the maximum number of iterations; n represents the number of training samples used in one iteration; m represents the number of modalities included in one training sample; S2: Input each feature vector into the classification layer after feature fusion to obtain the predicted probability of the health category ; where represents the predicted probability of the health category of the i-th training sample at the t-th iteration; Based on the predicted health category probabilities and the true health category labels calculate the multi-modal classification loss ; where represents the true health category label of the i-th training sample at the t-th iteration; represents the multi-modal classification loss at the t-th iteration; Based on the feature vector Calculate the multi-modal matching loss ; where represents the multi-modal matching loss at the t-th iteration; S3: Based on the weight parameter weight the multi-modal classification loss and the multi-modal matching loss to obtain the weighted sum as the multi-modal total loss ; where represents the weight parameter before update at the t-th iteration, is the initial value of the weight parameter; represents the multi-modal total loss at the t-th iteration; Based on the gradient descent method, the selected weight parameter update strategy, and the multi-modal total loss and the multi-modal classification loss update the weight parameters and network parameters ; where represents the network parameters of the feature extraction network before the update at the t-th iteration , and is the initial value of the network parameters of the feature extraction network ; S4: Iteration Perform steps S1 - S3 for times to obtain the final network parameters of the feature extraction network.

2. The dynamic balance method of modal clearance in the identification of the health category of a mechanical and electrical device according to claim 1, wherein The multi-modal classification loss is obtained based on the following formula: ； Among them, denotes transpose of.

3. The dynamic balance method of modal clearance in the health category identification of a mechanical and electrical device according to claim 2, characterized in that, The predicted probability of the health category is obtained based on the following formula: ； ； ； Among them, represents the fused feature vector at the t-th iteration; F represents the feature fusion operation; represents the sensor data corresponding feature vector; represents the sensor data corresponding feature vector; W and b represent the weight and bias of the fully connected layer in sequence; softmax represents the softmax function; the softmax function and the fully connected layer constitute the classification layer.

4. A dynamic balance method for modal clearance in the identification of the health category of a mechanical and electrical device according to claim 3, characterized in that Calculate the multimodal matching loss based on the following formula ; ； Among them, ; represents the similarity between sensor data and sensor data ; represents the similarity between sensor data and sensor data ; represents the similarity between sensor data and sensor data ; represents the number of combinations obtained by selecting two different modalities from m modalities for combination; represents the smoothing parameter.

5. A dynamic balance method for modal clearance in the identification of the health category of a mechanical and electrical device according to claim 4, characterized in that Similarity and Similarity and Similarity are calculated by the following formula: ； ； ； Among them, represents the sensor data corresponding eigenvector; represents the sensor data corresponding eigenvector; represents the sensor data corresponding eigenvector; represents the sensor data corresponding eigenvector; T represents transpose; 、 、 and represent in sequence 、 、 and the L2 norms of.

6. The dynamic balance method of modal clearance in the identification of the health category of a mechanical and electrical device according to claim 5, characterized in that, If the selected weight parameter update strategy is a strategy based on the decay factor, then S3 specifically includes the following steps: S31: Based on the formula calculate to obtain the multimodal total loss ; S32: Minimize the multimodal total loss using the gradient descent method , and update the network parameters as ; where represents the updated network parameters of the feature extraction network at the t-th iteration; S33: Update the weight parameter based on the formula to ; where represents the updated weight parameter at the t-th iteration.

7. A dynamic balance method for modal clearance in the identification of the health category of a mechanical and electrical device according to claim 5, characterized in that If the selected weight parameter update strategy is a constraint-aware hybrid loss strategy, then S3 specifically includes the following steps: S31’: Obtained based on the formula Calculate the total multimodal loss ; S32’: Minimize the multimodal total loss using the gradient descent method , and update the network parameters where ; among which represents the updated network parameters of the feature extraction network at the t-th iteration; S33’: Based on the multi-modal total loss , the multi-modal classification loss Calculate the multi-modal classification loss For the weight parameter Gradient of ; S34’: Based on the gradient Update the weight parameter as ; where represents the updated weight parameter at the t-th iteration.

8. A dynamic balance method for modal clearance in the identification of the health category of electromechanical equipment according to claim 7, characterized in that S33’ is implemented based on the following formula: ； Among them, represents the multimodal total loss with respect to the network parameters and the weight parameters Hessian matrix; represents the multimodal total loss with respect to the network parameters Hessian matrix; represents the gradient of the multimodal classification loss with respect to the network parameters ; represents the inverse matrix of 9. The dynamic balance method of modal clearance in the identification of the health category of a mechanical and electrical device according to claim 8, characterized in that S34’ is implemented based on the following formula: ； Among them, represents the learning rate.

10. A dynamic balance method for modal clearance in the identification of the health category of a mechanical and electrical device according to claim 1, characterized in that, The feature extraction network is a convolutional neural network.

Images