Search

CN-121834419-B - Port gantry crane driving system state evaluation method

CN121834419BCN 121834419 BCN121834419 BCN 121834419BCN-121834419-B

Abstract

The application belongs to the field of crane fault diagnosis, and particularly relates to a state evaluation method for a driving system of a port portal crane. The weighted convolution layer is designed to realize the weighted extraction of the bidirectional features, the self-adaptive residual layer is designed to avoid the redundancy of the features, and the GRU gate structure is improved to introduce the historical average state, so that the high-efficiency modeling and the accurate diagnosis of the multi-sensor time sequence data are realized.

Inventors

  • YANG XU
  • ZHANG ZHEN
  • SUN HUA
  • QI BIN
  • HAO WEIJIAN
  • Luan ning
  • SUN HOUJIANG
  • FU DAOCHENG
  • LI HOUDE
  • ZHANG SHIZHEN

Assignees

  • 山东陆海装备集团青岛有限公司
  • 山东大学
  • 青岛港国际股份有限公司

Dates

Publication Date
20260512
Application Date
20260311

Claims (9)

  1. 1. The method for evaluating the state of the driving system of the port portal crane is characterized by comprising the following steps of: s1, forming a portal crane driving system historical data set with balanced distribution of various samples by utilizing historical monitoring data and a test bed data set , Is the mth original sample; S2, will The first weighted convolution layer is input to extract space information, and the space information is generated after processing ; S3, after the action of the first weighted convolution layer, adding an improved self-adaptive residual layer, and then With the original information New information fusion in (1) is generated ; S4, construction and application Learner parameter matrix of the same dimension As position coding, the coded signal enters a second weighted convolution layer to further extract characteristics to generate ; S5, will Sending into an improved GRU network, wherein the network introduces a historical average state, so that the model can call long-term time sequence dependence characteristics; s6, taking the hidden state of the last time step For subsequent classification.
  2. 2. The method for evaluating the state of a driving system of a portal crane for a port according to claim 1, wherein the step S1 is to splice information of each sensor channel: ; Wherein, the Each channel information of the original signal for the mth sample, C represents L represents the number of channels Number of time steps for each channel.
  3. 3. The port gantry crane driving system state evaluation method according to claim 1, wherein in step S2 will be Inputting a first weighted convolution layer to extract spatial information: ; ; ; Wherein, the The linear cell activation function is modified for parameterization, In order to convolve a point by point, In the case of a deep convolution, The path e preferentially extracts the spatial characteristics in the channel, the path f preferentially fuses the cross-channel information, and the weight can be learned And (5) self-adaptive fusion and bidirectional output.
  4. 4. The method for evaluating the state of a driving system of a portal crane for a port according to claim 1, wherein the modified adaptive residual layer of step S3 is selected from the group consisting of Extraction of the Chinese medicine Information of the difference and compare it with Fusion: Wherein, the For trainable redundancy removal coefficients, control adaptively removes extracted features from the original signal, And controlling the fusion proportion of the extracted features and the new information for the trainable fusion balance coefficient.
  5. 5. The port gantry crane driving system state evaluation method according to claim 1, wherein the method comprises the steps of constructing and Learner parameter matrix of the same dimension As a position code, the coded signal is then fed into a second weighted convolution layer for further feature extraction: ; ; ; ; Wherein, the For the position-coding weights, For the output after the encoding, And the second two-way fusion weight.
  6. 6. The port gantry crane driving system state evaluation method according to claim 1, wherein the following is performed An improved update gate, an improved reset gate, an improved candidate state, a state update, is entered into an improved GRU network that enables a model to invoke long-term timing dependent features by introducing historical average states, a long-term information accumulation method as follows: ; Wherein, the Is the hidden state vector of the mth input sample at time i.
  7. 7. The method for evaluating the state of a driving system of a portal crane for a harbor according to claim 6, wherein, The improvement update door: ; Wherein, the Is the mth input sample Inputting a feature vector at time t; Is the hidden state vector for the mth input sample at time t-1, Is the average value of the hidden states of the mth sample from the beginning to time t-1, In order to update the gate input weight matrix, To update the gate hidden state weight matrix, In order to update the gate history information weight matrix, In order to update the gate bias vector, Is a Sigmoid function.
  8. 8. The port gantry crane drive system status assessment method of claim 7, wherein the reset gate is modified to: ; Wherein, the In order to reset the gate input weight matrix, To reset the gate hidden state weight matrix, To reset the gate history information weight matrix, To reset the gate bias vector.
  9. 9. The port gantry crane driving system state evaluation method according to claim 8, wherein the candidate state is improved: ; Wherein, the A weight matrix is input for the candidate states, The state weight matrix is hidden for the candidate states, For the candidate state history information weight matrix, As a candidate state-biasing vector, As a function of the hyperbolic tangent, Representing element-by-element multiplication; And (5) updating the state: ; Taking the hidden state of the last time step n And entering a full-connection classification layer for classification.

Description

Port gantry crane driving system state evaluation method Technical Field The application belongs to the field of crane fault diagnosis, and particularly relates to a state evaluation method for a driving system of a port portal crane. Background The portal crane is widely applied to the operation scenes such as ports and wharfs, and the driving system (comprising a motor, a gearbox, a speed reducer and the like) is easy to break down due to long service period, complex working environment and large working load change. Once a failure occurs, serious economic losses may occur, and therefore, early failure diagnosis of the gantry crane drive system is important. The existing fault diagnosis method has the problems of insufficient modeling of long-term dependence, characteristic redundancy and the like, and is difficult to fully utilize multi-sensor data. Disclosure of Invention The application provides a state evaluation method for a driving system of a port portal crane, which realizes efficient modeling and accurate diagnosis of multi-sensor time sequence data. The technical proposal is as follows: a port portal crane driving system state evaluation method comprises the following steps: S1, utilizing historical monitoring data and a test bed data set to form a gantry crane driving system fault data set with balanced distribution of various samples ,Is the mth original sample; S2, will The first weighted convolution layer is input to extract space information, and the space information is generated after processing; S3, after the action of the first weighted convolution layer, adding an improved self-adaptive residual layer, and thenWith the original informationNew information fusion in (1) is generated; S4, construction and applicationLearner parameter matrix of the same dimensionAs position coding, the coded signal enters a second weighted convolution layer to further extract characteristics to generate; S5, willSending into an improved GRU network, wherein the network introduces a historical average state, so that the model can call long-term time sequence dependence characteristics; s6, taking the hidden state of the last time step For subsequent classification. Preferably, step S1 splices the information of each sensor channel: ; Wherein, the For each channel of the original signal of the mth sample, C representsL represents the number of channelsNumber of time steps for each channel. Preferably, in step S2Inputting a first weighted convolution layer to extract spatial information: ; ; ; Wherein, the The linear cell activation function is modified for parameterization,In order to convolve a point by point,In the case of a deep convolution,Is the first two-way fusion weight. The method adopts bidirectional design, the path e preferentially extracts the spatial characteristics in the channel, the path f preferentially fuses the cross-channel information, and the path f passes through the learnable weightAnd (5) self-adaptive fusion and bidirectional output. Preferably, the step S3 of improving the adaptive residual layer is performed byExtraction of the Chinese medicineInformation of the difference and compare it withFusion: ; Wherein, the For trainable redundancy removal coefficients, control adaptively removes extracted features from the original signal,And controlling the fusion proportion of the extracted features and the new information for the trainable fusion balance coefficient. Preferably, construction andLearner parameter matrix of the same dimensionAs position coding, the coded signal enters a second weighted convolution layer to further extract characteristics; ; ; ; ; Wherein, the For the position-coding weights,For the output after the encoding,And the second two-way fusion weight. Preferably, willAn improved update gate, an improved reset gate, an improved candidate state, a state update, are entered into an improved GRU network. The network enables the model to invoke long-term timing dependency features by introducing a long-term information memory mechanism. The long-term information accumulation method is as follows: ; Wherein, the Is the hidden state vector of the mth input sample at time i. Preferably, the retrofit door is modified: ; Wherein, the Is the mth input sampleInputting a feature vector at time t; Is the hidden state vector for the mth input sample at time t-1, Is the average value of the hidden states of the mth sample from the beginning to time t-1,In order to update the gate input weight matrix,To update the gate hidden state weight matrix,In order to update the gate history information weight matrix,In order to update the gate bias vector,Is a Sigmoid function. Preferably, the reset gate is modified: ; Wherein, the In order to reset the gate input weight matrix,To reset the gate hidden state weight matrix,To reset the gate history information weight matrix,To reset the gate bias vector; preferably, the candidate state is improved: ; Wherein, the A weight matrix i