A method for evaluating the state of a drive system of a port portal crane

Abstract

The application belongs to the field of crane fault diagnosis, and particularly relates to a driving system state evaluation method for a port portal crane. Through the design of a weighted convolution layer, the weighted extraction of bidirectional features is realized, through the design of an adaptive residual layer, feature redundancy is avoided, and through the improvement of the GRU gate structure, the historical average state is introduced, thereby realizing the efficient modeling and accurate diagnosis of multi-sensor time series data.

Description

Technical Field

[0001] This application belongs to the field of crane fault diagnosis, specifically relating to a method for assessing the status of a port gantry crane drive system. Background Technology

[0002] Gantry cranes are widely used in ports, docks, and other similar operations. Due to their long service life, complex working environment, and large variations in workload, their drive systems (composed of motors, gearboxes, and reducers) are prone to failure. Once a failure occurs, it can cause significant economic losses. Therefore, early fault diagnosis of gantry crane drive systems is crucial. Existing fault diagnosis methods lack sufficient modeling for long-term dependencies and suffer from feature redundancy, making it difficult to fully utilize multi-sensor data. Summary of the Invention

[0003] This application proposes a state assessment method for the drive system of a port gantry crane, enabling efficient modeling and accurate diagnosis of multi-sensor time-series data. The technical solution is as follows:

[0004] A method for assessing the condition of a port gantry crane drive system includes the following steps:

[0005] S1. Using historical monitoring data and test bench datasets, construct a fault dataset for gantry crane drive systems with balanced sample distribution across various sample types. , For the first m One original sample;

[0006] S2. Will The first weighted convolutional layer extracts spatial information, which is then processed to generate... ;

[0007] S3. After the first weighted convolutional layer, an improved adaptive residual layer is added to... With the original information New information fusion in the middle, generating ;

[0008] S4. Construction and Learnable parameter matrix of the same dimension As a positional encoding, the encoded signal then enters a second weighted convolutional layer to further extract features, generating... ;

[0009] S5. Will The improved GRU network is fed into the model, which incorporates historical average states, enabling the model to retrieve long-term temporal dependency features.

[0010] S6. Retrieve the hidden state of the last time step. , used for subsequent classification.

[0011] Preferably, step S1 involves stitching together the information from each sensor channel:

[0012] ;

[0013] in, For the first m Each channel of the original signal of each sample, C represent The number of channels, L represent Time steps for each channel.

[0014] Preferably, in step S2, Input the first weighted convolutional layer to extract spatial information:

[0015] ;

[0016] ;

[0017] ;

[0018] in, To parameterize and modify the activation function of the linear unit, For pointwise convolution, For depthwise convolution, This is the first dual-path fusion weight. The method employs a bidirectional design: path e prioritizes extracting intra-channel spatial features, while path f prioritizes fusing cross-channel information, using learnable weights. Adaptive fusion bidirectional output.

[0019] Preferably, the improved adaptive residual layer in step S3, from Extracting and Information about the differences, and comparing them with Fusion:

[0020] ;

[0021] in, These are trainable redundancy removal coefficients that control the adaptive removal of extracted features from the original signal. It is a trainable fusion balance coefficient that controls the fusion ratio of extracted features and new information.

[0022] Preferably, constructing and Learnable parameter matrix of the same dimension As a positional encoding, the encoded signal is then fed into a second weighted convolutional layer for further feature extraction.

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] in, For position encoding weights, The encoded output, This is the weight for the second dual-path fusion.

[0028] Preferably, The improved GRU network incorporates an improved update gate, an improved reset gate, improved candidate states, and a state update mechanism. This network introduces a long-term information memory mechanism, enabling the model to retrieve long-term temporal dependency features. The long-term information accumulation method is as follows:

[0029] ;

[0030] in, For the first m Each input sample in time i The hidden state vector.

[0031] Preferred, improved and updated door:

[0032] ;

[0033] in, It is the first m Input Samples In time t The input feature vector; It is the first m Each input sample in time t- Hidden state vector under 1, It is the first m Each sample from start to time t- The average value of the hidden state of 1 To update the gate input weight matrix, To update the hidden state weight matrix of the gate, To update the gate historical information weight matrix, To update the gate bias vector, This is the Sigmoid function.

[0034] Preferred, improved reset door:

[0035] ;

[0036] in, To reset the gate input weight matrix, To reset the weight matrix of the hidden state of the door, To reset the weight matrix of historical information of the gate, To reset the gate bias vector;

[0037] Preferred, improved candidate states:

[0038] ;

[0039] in, Input a weight matrix for the candidate states. Let the candidate state and hidden state weight matrix be... The candidate state historical information weight matrix is... This is the candidate state bias vector. It is the hyperbolic tangent function. Represents element-wise multiplication;

[0040] Status Update:

[0041] ;

[0042] Take the hidden state at the last time step Then, it enters the fully connected classification layer for classification.

[0043] Compared with the prior art, the beneficial effects of this application are as follows:

[0044] This application proposes a state assessment method for the drive system of a port gantry crane. By designing a weighted convolutional layer, it achieves weighted extraction of bidirectional features. By designing an adaptive residual layer, it avoids feature redundancy. By improving the GRU gate structure, it introduces historical average states, thereby achieving efficient modeling and accurate diagnosis of multi-sensor time-series data. Attached Figure Description

[0045] Figure 1 A schematic diagram of sensor measuring points in the drive system of a gantry crane;

[0046] Figure 2 A schematic diagram of the state assessment algorithm for a gantry crane drive system;

[0047] 1-Motor; 2-Gearbox; 3-Reduction gearbox; 4-Drum; 5-First acceleration sensor; 6-Second acceleration sensor; 7-Third acceleration sensor. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0049] A method for assessing the condition of a port gantry crane drive system includes the following steps:

[0050] 1. Set up multi-point vibration measurement for the drive system of the gantry crane: The drive system consists of a motor 1, a gearbox 2, a reduction gearbox 3, a drum 4, etc. Based on the structure and operating characteristics of the drive system, a first acceleration sensor 5 is set at the end face of the motor, a second acceleration sensor 6 is set at the end face of the gearbox, and a third acceleration sensor 7 is set at the end face of the reduction gearbox.

[0051] 2. Using historical monitoring data and test bench datasets, construct a balanced dataset of faults in the drive system of gantry cranes for various types of samples. , For the first m One original sample; It is a dataset One of the samples.

[0052] Combine the information from each sensor channel:

[0053] ;

[0054] in, For the first m Each channel of the original signal of each sample, C represent The number of channels, L represent Time steps for each channel.

[0055] Will Input the first weighted convolutional layer to extract spatial information:

[0056] ;

[0057] ;

[0058] ;

[0059] in, To parameterize and modify the activation function of the linear unit, For pointwise convolution, For depthwise convolution, This is the first dual-path fusion weight. A bidirectional design is adopted: path e prioritizes extracting intra-channel spatial features, while path f prioritizes fusing cross-channel information, using learnable weights. Adaptive fusion bidirectional output.

[0060] After the first weighted convolutional layer, to retain some of the original information and avoid the loss of important information, an improved adaptive residual layer is proposed. Extracting and Information about the differences, and comparing them with Fusion:

[0061] ;

[0062] in, These are trainable redundancy removal coefficients that control the adaptive removal of extracted features from the original signal. The fusion balancing coefficients are trainable and control the fusion ratio of extracted features and new information. Compared to ordinary residual connections that simply add together, this method... Removed from The inclusion of redundant components, along with the dynamic balancing of the contribution ratio between new and old information, enhances the effectiveness of information fusion.

[0063] To enhance the model's ability to perceive temporal order, a system is constructed that... Learnable parameter matrix of the same dimension As a location encoding, the encoded signal is then fed into a second weighted convolutional layer for further feature extraction.

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] in, For position encoding weights, The encoded output, This is the weight for the second dual-path fusion.

[0069] To further extract time series information, The improved GRU network incorporates an improved update gate, an improved reset gate, improved candidate states, and a state update mechanism. This network, by introducing historical average states, enables the model to retrieve long-term temporal dependency features. The long-term information accumulation method is as follows:

[0070] ;

[0071] in, For the first m Each input sample in time i The hidden state vector.

[0072] Improvement and Update Gate:

[0073] ;

[0074] in, It is the first m Input Samples In time t The input feature vector. It is the first m Each input sample in time t- Hidden state vector under 1, It is the first m Each sample from start to time t- The average value of the hidden state of 1 To update the gate input weight matrix, To update the hidden state weight matrix of the gate, To update the gate historical information weight matrix, To update the gate bias vector, This is the Sigmoid function.

[0075] Improved reset door:

[0076] ;

[0077] in, To reset the gate input weight matrix, To reset the weight matrix of the hidden state of the door, To reset the weight matrix of historical information of the gate, To reset the gate bias vector.

[0078] Improve candidate states:

[0079] ;

[0080] in, Input a weight matrix for the candidate states. Let the candidate state and hidden state weight matrix be... The candidate state historical information weight matrix is... This is the candidate state bias vector. It is the hyperbolic tangent function. This represents element-wise multiplication.

[0081] Status Update:

[0082] ;

[0083] Take the hidden state at the last time step Then, it enters the fully connected classification layer for classification.

[0084] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for assessing the condition of a port gantry crane drive system, characterized in that, Includes the following steps: S1. Using historical monitoring data and test bench datasets, construct a historical dataset for various types of gantry crane drive systems with balanced sample distribution. , For the first m One original sample; S2. Will The first weighted convolutional layer extracts spatial information, which is then processed to generate... ; S3. After the first weighted convolutional layer, an improved adaptive residual layer is added to... With the original information New information fusion in the middle, generating ; S4. Construction and Learnable parameter matrix of the same dimension As a positional encoding, the encoded signal then enters a second weighted convolutional layer to further extract features, generating... ; S5. Will The model is fed into an improved GRU network, which incorporates historical average states, enabling it to retrieve long-term temporal dependency features. S6. Retrieve the hidden state of the last time step. , used for subsequent classification.

2. The method for assessing the state of a port gantry crane drive system according to claim 1, characterized in that, Step S1: Assemble the information from each sensor channel: ； in, For the first m Information of each channel of the original signal of each sample. C represent The number of channels, L represent Time steps for each channel.

3. The method for assessing the state of a port gantry crane drive system according to claim 1, characterized in that, In step S2, Input the first weighted convolutional layer to extract spatial information: ； ； ； in, To parameterize and modify the activation function of the linear unit, For pointwise convolution, For depthwise convolution, The first dual-path fusion weights are used; path e prioritizes extracting intra-channel spatial features, while path f prioritizes fusing cross-channel information, using learnable weights. Adaptive fusion bidirectional output.

4. The method for assessing the status of a port gantry crane drive system according to claim 1, characterized in that, The improved adaptive residual layer in step S3, from Extracting and Information about the differences, and comparing them with Fusion: in, These are trainable redundancy removal coefficients that control the adaptive removal of extracted features from the original signal. It is a trainable fusion balance coefficient that controls the fusion ratio of extracted features and new information.

5. The method for assessing the status of a port gantry crane drive system according to claim 1, characterized in that, Building and Learnable parameter matrix of the same dimension As a positional encoding, the encoded signal then enters the second weighted convolutional layer for further feature extraction. ； ； ； ； in, For position encoding weights, The encoded output, This is the weight for the second dual-path fusion.

6. The method for assessing the state of a port gantry crane drive system according to claim 1, characterized in that, Will The improved GRU network incorporates an improved update gate, an improved reset gate, improved candidate states, and a state update mechanism. By introducing historical average states, the network enables the model to retrieve long-term temporal dependency features. The long-term information accumulation method is as follows: ； in, For the first m Each input sample in time i The hidden state vector.

7. The method for assessing the state of a port gantry crane drive system according to claim 6, characterized in that, Improvement and Update Gate: ； in, It is the first m Input Samples In time t The input feature vector; It is the first m Each input sample in time t- Hidden state vector under 1, It is the first m Each sample from start to time t- The average value of the hidden state of 1 To update the gate input weight matrix, To update the hidden state weight matrix of the gate, To update the gate historical information weight matrix, To update the gate bias vector, This is the Sigmoid function.

8. The method for assessing the state of a port gantry crane drive system according to claim 7, characterized in that, Improved reset door: ； in, To reset the gate input weight matrix, To reset the weight matrix of the hidden state of the door, To reset the weight matrix of historical information of the gate, To reset the gate bias vector.

9. The method for assessing the state of a port gantry crane drive system according to claim 8, characterized in that, Improve candidate states: ； in, Input a weight matrix for the candidate states. Let the candidate state and hidden state weight matrix be... The candidate state historical information weight matrix is... This is the candidate state bias vector. It is the hyperbolic tangent function. Represents element-wise multiplication; Status Update: ； Take the last time step n Hidden state Then, it enters the fully connected classification layer for classification.

Images