CN-120579054-B - Port equipment operation and maintenance state identification method, system, terminal and medium

CN120579054BCN 120579054 BCN120579054 BCN 120579054BCN-120579054-B

Abstract

The invention relates to the technical field of multi-source monitoring, in particular to a port equipment operation and maintenance state identification method, a system, a terminal and a medium, which comprise the steps of obtaining a multi-source detection signal; the method comprises the steps of carrying out denoising treatment on a multisource detection signal by using an improved self-adaptive Kalman filter to obtain a sample signal, decomposing the sample signal into multiscale frequency domain features by wavelet transformation and Fourier transformation, carrying out convolution treatment on the multiscale frequency domain features to obtain multiscale time-frequency features, and identifying the residual service life, fault type and health state corresponding to the multiscale time-frequency features by using a multi-task learning model based on LSTM and CNN. The invention can effectively remove the noise of the multi-source detection signals, effectively utilize the detection signals, and improve the monitoring efficiency and the monitoring instantaneity.

Inventors

ZHANG XIAO
LI YONGCUI
LI BO
CHEN QIANG
LIU YAOHUI
Ge Xiubo
MA HUIJUAN
HAN RUI

Assignees

青岛港国际股份有限公司
青岛新前湾集装箱码头有限责任公司

Dates

Publication Date: 20260512
Application Date: 20250514

Claims (7)

1. A method for identifying the operation and maintenance states of port equipment, comprising the following steps: Acquiring a multi-source detection signal; Denoising the multi-source detection signal by using an improved self-adaptive Kalman filter to obtain a sample signal; Decomposing the sample signal into multi-scale frequency domain features through wavelet transformation and Fourier transformation, and carrying out convolution processing on the multi-scale frequency domain features to obtain multi-scale time-frequency features; identifying residual life, fault type and health state corresponding to the multi-scale time-frequency characteristics by utilizing a multi-task learning model based on LSTM and CNN; The adaptive Kalman filter includes: Adaptive process noise covariance : Wherein, the Is a smoothing factor which is used to smooth the image, Is the state matrix of the sensor at time k, Is the variance of the current sensor state matrix; Adaptive measurement of noise covariance : Wherein, the Is a smoothing factor which is used to smooth the image, Is a matrix of detection signals generated at time k, Is the variance of the current detection signal matrix; Denoising the multi-source detection signal by using an improved adaptive Kalman filter to obtain a sample signal, wherein the method comprises the following steps of: Presetting smoothing factor values corresponding to various working conditions and detecting signal value ranges; Matching the multisource detection signals with the detection signal value ranges corresponding to each working condition to determine the current working condition; The value of the smoothing factor corresponding to the current working condition is called, and the value of the smoothing factor corresponding to the current working condition is substituted into the self-adaptive Kalman filter; decomposing the sample signal into multi-scale frequency domain features through wavelet transformation and Fourier transformation, and carrying out convolution processing on the multi-scale frequency domain features to obtain multi-scale time-frequency features, wherein the method comprises the following steps: Wherein, the Is the output fusion frequency domain characteristic; Is the frequency; is a scale for adjusting the degree of stretching of the wavelet function; the method comprises the steps of translating, namely controlling the position of a wavelet function on a time axis, wherein N is the number of samples and indicates the total number of time domain samples; Is a sample of the state matrix output by the adaptive Kalman filter in the time domain; J is an imaginary unit; deep learning feature fusion is carried out on the fusion frequency domain features by utilizing a pre-training convolutional neural network model, so as to obtain multi-scale time-frequency features; The method comprises the steps of arranging fused frequency domain features obtained by wavelet transformation and Fourier transformation into a format suitable for CNN input, arranging features under different scales and frequencies into a two-dimensional matrix, and expanding the two-dimensional matrix into three-dimensional tensors to serve as input data of CNN, wherein a first dimension represents the number of samples, a second dimension represents the feature scale, and a third dimension represents a frequency range.
2. The method of claim 1, wherein acquiring the multi-source detection signal comprises: a detection signal of a sensor combination of a plurality of detection points which are pre-deployed on the harbour site is obtained, wherein the sensor combination comprises a temperature sensor, a vibration sensor, a current sensor, a voltage sensor, a pressure sensor and a displacement sensor.
3. The method of claim 1, wherein the LSTM and CNN based multitasking learning model comprises: the sharing layer is used for constructing a full-connection layer, inputting the size into the dimension of the data characteristic, and outputting the size into the size of the appointed hidden layer; The LSTM layer is used for constructing an LSTM network, the input size is the output size of the sharing layer, and the hidden layer size is the designated LSTM hidden layer size; the CNN layer is used for constructing a one-dimensional convolution layer, the number of input channels is the output size of the sharing layer, the number of output channels is the designated CNN channel number, the convolution kernel size is 3, and the filling is 1; defining an output layer of the task, which is used for outputting a plurality of task results; The residual service life prediction layer is used for constructing a full connection layer, wherein the input size is the size after the output splicing of the LSTM layer and the CNN layer, and the output size is 1; The fault type classification layer is used for constructing a full-connection layer, wherein the input size is the size of the LSTM layer and the CNN layer after output splicing, and the output size is the class number of the fault type; the health state evaluation layer is used for constructing a full connection layer, wherein the input size is the size of the LSTM layer and the CNN layer after output splicing, and the output size is the category number of the health state; Combining the model, namely combining the sharing layer, the LSTM layer, the CNN layer and the output layer of the task into a multi-task model; During forward propagation, data firstly passes through a sharing layer, then is respectively input into an LSTM layer and a CNN layer, outputs of the two layers are spliced, and then are respectively input into output layers of all tasks, so that a prediction result of each task is obtained.
4. A method according to claim 3, wherein the training method of the LSTM and CNN based multitasking learning model comprises: setting the total number of training wheels and the weight of the multi-task loss; for each round of training: Acquiring a batch of input data, a real residual service life, a fault type and a health state from a training data loader; The input data are transmitted into a model to obtain a prediction result of each task; Respectively calculating the loss of each task, and adding the loss to a loss list; Calculating total loss according to the multi-task loss function; Updating model parameters using an optimizer, comprising: Calculating gradients of the loss of each task respectively; Calculating the average value of all task gradients as a reference gradient ; Let K be the number of tasks, Is a parameter vector of the model; for the ith task, its loss function is The corresponding gradient is The update vector to be calculated is d; the problem of calculating the update vector d is expressed as a constraint optimization problem: Wherein, the Is a super parameter for controlling the angular constraint of the update vector d and the reference gradient; constructing a Lagrangian function to solve the constraint optimization problem: Wherein, the Is the Lagrangian multiplier; The optimization problem is concave about d due to the constraint, and when When the schlieren condition is satisfied, the maximum and minimum operators are exchanged, and the dual problem is obtained: Wherein, the W is on probabilistic simplex, i.e And is also provided with ; At a given point And w, the optimal update vector can be obtained by deriving the Lagrangian function with respect to d and making it zero Is represented by the expression: Wherein, the Is an m×m identity matrix; Calculation of : Is provided with ; First calculate ; Recalculating Marking the calculation result as D; finally according to SMW formula Calculation of ; Order the Then, the optimal update vector is obtained by optimizing w in the dual problem, which is simplified into: Solving the simplified dual problem by using a gradient descent method to obtain an optimal w; substituting the optimal w into And then again Substitution into Calculating to obtain a final update vector d, wherein the update vector d is a new gradient obtained by combining the reference gradient and the current task gradient; Assigning the new gradient back to the parameters of the model; Model parameters are updated using an optimizer.
5. A port equipment operation and maintenance behavior recognition system, comprising: The acquisition module is used for acquiring the multi-source detection signals; the denoising module is used for denoising the multi-source detection signal by utilizing an improved self-adaptive Kalman filter to obtain a sample signal; The processing module is used for decomposing the sample signal into multi-scale frequency domain features through wavelet transformation and Fourier transformation, and carrying out convolution processing on the multi-scale frequency domain features to obtain multi-scale time-frequency features; the identification module is used for identifying the residual life, the fault type and the health state corresponding to the multi-scale time-frequency characteristics by utilizing a multi-task learning model based on LSTM and CNN; The adaptive Kalman filter includes: Adaptive process noise covariance : Wherein, the Is a smoothing factor which is used to smooth the image, Is the state matrix of the sensor at time k, Is the variance of the current sensor state matrix; Adaptive measurement of noise covariance : Wherein, the Is a smoothing factor which is used to smooth the image, Is a matrix of detection signals generated at time k, Is the variance of the current detection signal matrix; Denoising the multi-source detection signal by using an improved adaptive Kalman filter to obtain a sample signal, wherein the method comprises the following steps of: Presetting smoothing factor values corresponding to various working conditions and detecting signal value ranges; Matching the multisource detection signals with the detection signal value ranges corresponding to each working condition to determine the current working condition; The value of the smoothing factor corresponding to the current working condition is called, and the value of the smoothing factor corresponding to the current working condition is substituted into the self-adaptive Kalman filter; decomposing the sample signal into multi-scale frequency domain features through wavelet transformation and Fourier transformation, and carrying out convolution processing on the multi-scale frequency domain features to obtain multi-scale time-frequency features, wherein the method comprises the following steps: Wherein, the Is the output fusion frequency domain characteristic; Is the frequency; is a scale for adjusting the degree of stretching of the wavelet function; the method comprises the steps of translating, namely controlling the position of a wavelet function on a time axis, wherein N is the number of samples and indicates the total number of time domain samples; Is a sample of the state matrix output by the adaptive Kalman filter in the time domain; J is an imaginary unit; deep learning feature fusion is carried out on the fusion frequency domain features by utilizing a pre-training convolutional neural network model, so as to obtain multi-scale time-frequency features; The method comprises the steps of arranging fused frequency domain features obtained by wavelet transformation and Fourier transformation into a format suitable for CNN input, arranging features under different scales and frequencies into a two-dimensional matrix, and expanding the two-dimensional matrix into three-dimensional tensors to serve as input data of CNN, wherein a first dimension represents the number of samples, a second dimension represents the feature scale, and a third dimension represents a frequency range.
6. A terminal, comprising: The memory is used for storing a port equipment operation and maintenance state identification program; A processor for implementing the steps of the method for identifying the operation and maintenance states of the port equipment according to any one of claims 1 to 4 when executing the operation and maintenance state identification program of the port equipment.
7. A computer readable storage medium storing a computer program, characterized in that the readable storage medium stores a port device operation and maintenance status identifying program, which when executed by a processor, implements the steps of the port device operation and maintenance status identifying method according to any one of claims 1-4.

Description

Port equipment operation and maintenance state identification method, system, terminal and medium Technical Field The invention relates to the technical field of multisource monitoring, in particular to a port equipment operation and maintenance state identification method, a port equipment operation and maintenance state identification system, a port equipment operation and maintenance state identification terminal and a port equipment operation and maintenance state identification medium. Background The core of the state monitoring of the harbour site is to effectively process the multi-source sensor signals. However, in actual operation, signal characteristics generated by different sensors are quite different, a unified processing mode is difficult to be adopted, and noise interference in signals also increases processing difficulty. In conventional kalman filtering applications, it is generally assumed that noise and system dynamics are fixed. However, in complex harbour sites, the dynamic and noise characteristics of the system may be dynamically changing due to environmental factors, equipment state changes, etc. Therefore, the traditional Kalman filtering cannot effectively denoise dynamic multi-source sensor signals. The conventional threshold comparison monitoring method has a plurality of limitations. On the one hand, the monitoring precision is limited, and useful information in the sensor signal cannot be fully mined, so that the judgment on the equipment state is not accurate enough. On the other hand, if multiple monitoring results are to be obtained, multiple sets of judgment logics are required to be set, so that redundant calculation is caused, precious calculation resources are wasted, data processing efficiency is reduced, the requirement of real-time monitoring is difficult to meet, and potential faults of equipment are not easy to discover in time. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a port equipment operation and maintenance state identification method, a port equipment operation and maintenance state identification system, a port equipment operation and maintenance state identification terminal and a port equipment operation and maintenance state identification medium, so as to solve the technical problems. In a first aspect, the present invention provides a method for identifying operation and maintenance states of port equipment, including: Acquiring a multi-source detection signal; Denoising the multi-source detection signal by using an improved self-adaptive Kalman filter to obtain a sample signal; Decomposing the sample signal into multi-scale frequency domain features through wavelet transformation and Fourier transformation, and carrying out convolution processing on the multi-scale frequency domain features to obtain multi-scale time-frequency features; And identifying the residual life, fault type and health state corresponding to the multi-scale time-frequency characteristic by using a multi-task learning model based on LSTM and CNN. In an alternative embodiment, acquiring the multi-source detection signal includes: a detection signal of a sensor combination of a plurality of detection points which are pre-deployed on the harbour site is obtained, wherein the sensor combination comprises a temperature sensor, a vibration sensor, a current sensor, a voltage sensor, a pressure sensor and a displacement sensor. In an alternative embodiment, the adaptive kalman filter includes: Adaptive process noise covariance : Wherein, the Is a smoothing factor which is used to smooth the image,Is the state matrix of the sensor at time k,Is the variance of the current sensor state matrix; Adaptive measurement of noise covariance : Wherein, the Is a smoothing factor which is used to smooth the image,Is a matrix of detection signals generated at time k,Is the variance of the current detection signal matrix. In an alternative embodiment, denoising the multi-source detection signal using a modified adaptive kalman filter to obtain a sample signal, including: Presetting smoothing factor values corresponding to various working conditions and detecting signal value ranges; Matching the multisource detection signals with the detection signal value ranges corresponding to each working condition to determine the current working condition; And calling the smoothing factor value corresponding to the current working condition, and substituting the smoothing factor value corresponding to the current working condition into the adaptive Kalman filter. In an alternative embodiment, the decomposing the sample signal into multi-scale frequency domain features through wavelet transformation and fourier transformation, and performing convolution processing on the multi-scale frequency domain features to obtain multi-scale time-frequency features, including: Wherein, the Is the output fusion frequency domain characteristic; Is the frequency; is a scale for adju