Model training method for port large equipment health state monitoring

Abstract

The present application relates to a kind of model training methods for port large equipment health state monitoring, specifically as follows: port large equipment detection data is collected, labeled and dataset division;The data collected is adaptively decomposed and feature reconstruction, improved adaptive noise complete set empirical mode decomposition and dynamic spectrum kurtosis fusion method are used, and decomposition reconstruction matrix is generated;Deep learning model is constructed, sample data is input into the model after processing to train the model, and the trained model is obtained;Real-time acquisition of port large equipment detection data is input into the trained model after processing, and the health state probability distribution result is output, and the maximum value of probability distribution is used to determine the current health state category of equipment.The present application can carry out real-time health detection and degradation evaluation to equipment operating state by adaptively decomposing and feature reconstruction to the sample data collected, constructing deep learning model and training it.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of device state monitoring based on artificial intelligence, and in particular to a model training method for port large device health state monitoring. BACKGROUND

[0002] With the development of port automation and scale, large equipment such as bridge cranes and bucket wheel stackers has become the key support for the efficient operation of the port. This kind of equipment is large in size and complex in operation environment, and is long-term under heavy load, impact and marine climate, which is prone to bearing wear, gear fracture, structural fatigue and other failures. Once failure occurs, not only will it cause huge maintenance cost and production interruption, but also may cause serious safety accidents, so real-time health detection and degradation evaluation of equipment operation state has great significance.

[0003] The existing device state detection method mainly relies on traditional vibration signal analysis and machine learning classification model. For example, time-frequency analysis methods such as short-time Fourier transform and wavelet analysis generally use fixed analysis window or basis function in feature extraction, which is difficult to consider high-frequency transient impact and low-frequency degradation mode at the same time, and is easy to cause feature extraction deficiency and fault mode aliasing. Empirical mode decomposition and its improved algorithm can adaptively decompose non-stationary signals, but still have the problems of mode aliasing and over-decomposition in strong noise background, and the reconstructed signal-to-noise ratio is low, which leads to the key impact features being covered. In addition, the traditional deep learning model mostly directly inputs the time domain or frequency domain features for classification when processing the vibration signal, ignoring the gradualness of the degradation process and the difference of multi-scale features, and the recognition accuracy of slight and moderate degradation state is often insufficient. The classification stage usually relies on a single fully connected layer to output class probability, and does not correct the probability combined with the evolution law of equipment health state, so that the prediction result is difficult to truly reflect the continuous process of equipment degradation.

[0004] The existing technical solutions have the following objective shortcomings: the window length of short-time Fourier transform, wavelet transform and other methods is fixed, which is difficult to consider high-frequency transient impact and low-frequency wear features, and there is a problem of insufficient resolution and feature loss; using fixed stopping criterion, easy to cause mode aliasing or over-decomposition, low signal-to-noise ratio in strong noise environment, leading to key fault features being covered; ordinary convolutional neural network or attention mechanism is not sensitive enough to high-frequency impact or intermediate degradation, which cannot model the gradual degradation process of equipment, resulting in decreased classification performance for slight / intermediate degradation; the existing methods mostly use a single fully connected classifier, which does not distinguish the importance of different frequency band features, and does not combine degradation priori knowledge, leading to inconsistent probability distribution and actual degradation process, especially prone to misjudge the intermediate state.

[0005] Therefore, the present application proposes a model training method for port large device health state monitoring to solve the above problems. SUMMARY

[0006] The present application aims at the deficiencies of the prior art, and provides a model training method for port large equipment health state monitoring.

[0007] The technical scheme for solving the technical problems of the present application is a model training method for port large equipment health state monitoring, which is as follows:

[0008] S1, collect and label port large equipment detection data, and divide the labeled sample data into a training set, a validation set and a test set in chronological order;

[0009] The port large equipment detection data is systematically collected and labeled, and the vibration signals of the equipment are collected in real time by the vibration acceleration sensor installed on the key parts of the large equipment in the port at a fixed sampling frequency.

[0010] The collected sample data is labeled for health state, and the equipment health state is divided into five categories: healthy, slight degradation, moderate degradation, severe degradation and failure.

[0011] The labeled sample data is divided into a training set, a validation set and a test set in chronological order, which are used for model training, wherein the training set is used for model parameter learning, the validation set is used for model hyperparameter optimization, and the test set is used for model final performance evaluation.

[0012] S2, adaptive time-frequency decomposition and feature reconstruction are performed on the collected sample data, specifically using an improved adaptive noise complete ensemble empirical mode decomposition and dynamic spectrum kurtosis fusion method, and the high-dimensional decomposition reconstruction matrix is reduced to a decomposition reconstruction matrix.

[0013] S2.1, define the labeled sample data as the original vibration signal vector, perform improved adaptive noise complete ensemble empirical mode decomposition on the original vibration signal, and use an adaptive stopping criterion, wherein the stopping criterion is defined as the standard deviation change of adjacent two iterations being less than an adaptive threshold; the intrinsic mode function obtained after decomposition of the improved adaptive noise complete ensemble empirical mode decomposition algorithm has a total order K, and the kth order intrinsic mode function is represented as , , k is the index of the intrinsic mode function.

[0014] Specifically, in the mth iteration, the standard deviation of the current intrinsic mode function and the standard deviation of the previous iteration are calculated, and if , the iteration is stopped, otherwise the iteration is continued, and the modal aliasing is suppressed.

[0015] wherein, denotes the standard deviation of the mth iteration, denotes the standard deviation of the mth iteration; denotes the standard deviation of the mth iteration; denotes the logarithm function, the default base is the natural constant; denotes the length of the original vibration signal vector; denotes the number of iterations when extracting each order of the intrinsic modal function;

[0016] S2.2, for each order of the intrinsic modal function, the fourth power of the intrinsic modal function is converted into the frequency domain amplitude by Fourier transform, and the square of the intrinsic modal function is converted into the square of the frequency domain amplitude, and then the ratio of the two is calculated to obtain the dynamic spectral kurtosis value;

[0017] S2.3, based on the dynamic spectral kurtosis of each order of the intrinsic modal function, the average dynamic spectral kurtosis of each order is calculated;

[0018] Then the adaptive threshold of each order is calculated according to the mean and standard deviation of each order of the dynamic spectral kurtosis in the training set;

[0019] The difference between the average dynamic spectral kurtosis and the adaptive threshold is processed by the sign function to obtain the weight, and then the intrinsic modal function is combined to construct a high-dimensional decomposition reconstruction matrix;

[0020] Finally, the high-dimensional decomposition reconstruction matrix is reduced in dimension by using the method of piecewise averaging or downsampling to obtain the decomposition reconstruction matrix.

[0021] S3, a deep learning model is constructed, and the sample data in the training set, the validation set and the test set are input into the model after being processed by step S2 to train the model, and a trained model is obtained. The specific operation of constructing and training the model is as follows:

[0022] Time-frequency dual-path feature re-calibration: the decomposition reconstruction matrix generates an attention mask through a time-frequency dual-path feature calibration mechanism, and after fusion, the decomposition reconstruction matrix is adaptively re-calibrated to generate a calibration feature matrix;

[0023] (1) a depth separable one-dimensional convolution operation is used to extract multi-scale time sequence features of the decomposition reconstruction matrix, and then a normalized exponential function is used to process the convolution output to generate a time domain attention mask;

[0024] (2) the decomposition reconstruction matrix is converted to the frequency domain by Fourier transform, and the frequency domain amplitude spectrum is obtained by taking the modulus, then a multi-layer perception is used to learn the global energy distribution of the frequency domain features, and then the output is converted back to the time domain by inverse Fourier transform to generate a frequency domain attention mask;

[0025] (3) The time domain attention mask and the frequency domain attention mask are fused by a broadcast addition operation, then the fused weight range is constrained using a hyperbolic tangent function, and finally the output calibration feature matrix is recalibrated by element-wise multiplication and a decomposition reconstruction matrix.

[0026] Cascade hollow depth feature extraction: the calibration feature matrix is extracted by a cascade hollow depth convolution module to obtain multi-scale features, which are then spliced to generate multi-scale fusion features;

[0027] (1) The calibration feature matrix is used as the input of the cascade hollow depth convolution module, and the input feature of the first level convolution is initialized;

[0028] (2) The input feature is processed using a three-level cascade of hollow depth separable convolution, each level including a hollow depth one-dimensional separable convolution, batch normalization, and a Gaussian error linear unit activation function operation, and the output features of each processing stage are gradually extracted to obtain the output features of each level of the network at the third processing stage;

[0029] (3) The output features of the three processing stages are spliced along the channel dimension to form multi-scale fusion features.

[0030] Dynamic degradation perception feature enhancement: based on the multi-scale fusion features, the feature offset and the degradation score are calculated, and then the channel response is dynamically adjusted to generate the degradation perception feature;

[0031] (1) Select sample data in a healthy state from the training set and extract its baseline feature vector. The multi-scale fusion features of the current sample data are subtracted from the baseline feature vector to obtain the feature offset;

[0032] (2) The global vector is obtained by global average pooling of the feature offset, and then the multi-layer perception is input to learn the degradation trajectory, and the degradation score is output to quantify the degradation degree of the device;

[0033] (3) Based on the degradation score and the learnable adaptive weight vector, the multi-scale fusion features are scaled and adjusted in the channel dimension to obtain the degradation perception feature.

[0034] Gated multi-granularity feature fusion classification: the degradation perception feature is segmented and fused using a gated multi-branch structure to generate a predicted health state probability distribution result;

[0035] (1) The degradation perception feature is evenly divided into four feature segments along the channel dimension, and each feature segment has the same dimension;

[0036] (2) For each feature segment of the degradation perception feature, a non-linear transformation is performed through a fully connected layer, and then a Sigmoid activation function is used to calculate the gating weight;

[0037] (3) For each feature segment, dimensionality reduction and abstraction are performed through a fully connected layer, and then the modified linear unit activation function is used for processing. The gated transformation feature is obtained by combining the gated weight vector of each feature segment.

[0038] (4) The gated transformation features of each feature segment are mapped to class scores through a fully connected layer. Then, the class scores of all branches are summed and normalized by a normalized exponential function to obtain the predicted health status probability distribution.

[0039] Deterioration perception probability correction: Based on the degradation score, a prior distribution is generated, and the predicted health status probability distribution is corrected to generate the corrected health status probability.

[0040] (1) Based on the degradation score, a prior probability distribution is generated through an exponential decay function. The center of this distribution shifts linearly with the degradation score. The calculation formula is as follows:

[0041] ,

[0042] in, Indicates the first The prior probability of a health status category; j represents the index of the device health status category, with values ​​of 1, 2, 3, 4, and 5, corresponding to the five states of healthy, slightly deteriorated, moderately deteriorated, severely deteriorated, and failed, respectively; This represents the index of the device health status category that is distinct from j; This represents the sharpness parameter; Indicates the scaling parameter; Represents the natural exponential function; Indicates the degradation score;

[0043] (2) The prior distribution and the predicted probability are fused using the geometric mean to obtain the corrected probability. The calculation formula is as follows:

[0044] ,

[0045] in, Indicates the first The corrected probability of a health-like state; Indicates the predicted first Probability of a class of healthy states; Indicates the predicted first Probability of a class of healthy states; Indicates the first Prior probability of a health state; This represents the correction strength parameter used to control the influence of the prior distribution.

[0046] Loss function calculation: Calculate the cross-entropy loss and the consistency loss of the degree of degradation;

[0047] a classification cross-entropy loss function is calculated according to the correction probability and the true label;

[0048] a degradation consistency loss function is calculated based on the degradation score and the true degradation degree;

[0049] The degradation consistency loss function is weighted and added to the classification cross-entropy loss function to obtain a total loss function.

[0050] The deep learning model is iteratively trained and the model parameters are updated, and the training of the model includes a training phase, a validation phase and a test phase, and the operations of each phase are as follows:

[0051] Training phase: taking the training set as the core, the network parameters are continuously updated through iterative optimization algorithm to minimize the total loss function;

[0052] Specifically, the mini-batch gradient descent method is adopted, a batch of samples are randomly selected from the training set each time, input into the constructed deep learning network, and the predicted health state probability distribution and degradation score are obtained through forward propagation calculation; then the loss value of the current batch is calculated according to the total loss function; during the back propagation process, the gradient of the total loss function to the parameters of each layer of the network is calculated, and the Adam optimizer is used to update the parameters, and the learning rate is set;

[0053] Validation phase: the model performance is evaluated using the validation set after each complete traversal of the training set;

[0054] The validation set samples are input into the model after the training phase, the validation indicators are calculated, the training strategy is dynamically adjusted based on the validation indicators, the early stopping strategy is adopted to prevent overfitting, and if the validation set loss does not decrease for a plurality of traversal periods, the training is terminated in advance;

[0055] Test phase: the test set samples are input into the model after the validation phase, the prediction results are obtained, and the accuracy and the adaptation ability to unknown data of the model are measured.

[0056] The effects provided in the summary of the invention are only the effects of the embodiments, and not all the effects of the invention, and the above technical solutions have the following advantages or beneficial effects:

[0057] The improved adaptive noise complete ensemble empirical mode decomposition and dynamic spectral kurtosis fusion method is adopted, the adaptive stopping criterion is used to avoid mode aliasing or over-decomposition caused by fixed iteration, and the dynamic spectral kurtosis is combined to effectively separate high-frequency impact and low-frequency degradation mode in a strong noise environment, avoid the difficulty in simultaneously capturing high-frequency impact components and low-frequency degradation mode, and cause insufficient feature extraction and mode aliasing phenomenon;

[0058] The application simultaneously adopts time domain multi-scale convolution and frequency domain energy modeling through a time-frequency dual-path feature calibration mechanism to generate attention masks, enhance local impact and global energy features, and can avoid losing local impact features, resulting in insensitivity to subtle faults, and realize fine re-calibration of features;

[0059] The application calculates feature offsets based on healthy samples, dynamically adjusts channel responses in combination with degradation scores, avoids feature boundary blurring between healthy states and degradation stages, and thus the sensitivity of the classification model to the intermediate degradation state is insufficient, so that the dynamic degradation sensing features are enhanced to quantize the progressive process of the device from health to failure, and the sensitivity to the intermediate degradation state is strengthened;

[0060] The application maps the features to different frequency bands in combination with a gating mechanism to realize adaptive feature selection, and corrects the classification probability through a degradation prior distribution, which can avoid the prediction probability of the multi-granularity feature fusion classifier, which may not fully reflect the progressive degradation process, and thus cause deviation in the judgment of the intermediate state, and improve the recognition rate and classification stability of slight / moderate degradation.

[0061] To sum up, the application can realize real-time health detection and degradation evaluation of the device running state by adaptively decomposing and reconstructing the collected sample data, and by constructing and training a deep learning model. BRIEF DESCRIPTION OF DRAWINGS

[0062] The accompanying drawings are used to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation of the application.

[0063] Figure 1 The figure is a method flowchart of the application.

[0064] Figure 2 The figure is a training process diagram of the deep learning model.

[0065] Figure 3 The figure is a vibration signal comparison chart of the port large equipment in different health states.

[0066] Figure 4 The figure is a feature extraction effect comparison chart of different decomposition methods.

[0067] Figure 5 The figure is a comparison chart of the effect difference between the adaptive time-frequency decomposition method of the application and the conventional short-time Fourier transform in processing non-stationary vibration signals.

[0068] Figure 6 The figure is a classification accuracy comparison chart of different models in five key health states.

[0069] Figure 7 Feature space distribution map of different health states extracted by the method of the present application.

[0070] Figure 8 Trend chart of score change of different methods during the device degradation acceleration test. DETAILED DESCRIPTION

[0071] In order to clearly illustrate the technical features of the present application, the present application will be described in detail below with specific embodiments, and in conjunction with the accompanying drawings.

[0072] Example 1

[0073] As shown in the figure, a model training method for port large equipment health state monitoring includes the following steps: Figure 1

[0074] S1, collect and label the port large equipment detection data, and divide the labeled sample data into a training set, a validation set and a test set in chronological order;

[0075] S2, perform adaptive time-frequency decomposition and feature reconstruction on the collected sample data, specifically using an improved adaptive noise complete ensemble empirical mode decomposition and dynamic spectrum kurtosis fusion method, and after generating a high-dimensional decomposition reconstruction matrix, reduce the dimension to a decomposition reconstruction matrix;

[0076] S3, construct a deep learning model, input the sample data in the training set, the validation set and the test set after step S2 into the model to train the model, and obtain a trained model, as shown in the figure, the model includes the following operations: Figure 2

[0077] Time-frequency dual-path feature re-labeling: the decomposition reconstruction matrix generates an attention mask through a time-frequency dual-path feature calibration mechanism, and after fusion, the decomposition reconstruction matrix is adaptively re-labeled to generate a calibration feature matrix;

[0078] Cascade hollow deep feature extraction: the calibration feature matrix extracts multi-scale features through a cascade hollow deep convolution module, and then performs splicing to generate multi-scale fusion features;

[0079] Dynamic degradation perception feature enhancement: based on the multi-scale fusion features, the feature offset and the degradation score are calculated, and then the channel response is dynamically adjusted to generate a degradation perception feature;

[0080] Gated multi-granularity feature fusion classification: the degradation perception feature is segmented and processed and fused using a gated multi-branch structure to generate a predicted health state probability distribution result;

[0081] ​​Deterioration-aware probability correction: based on the deterioration score to generate a prior distribution, correct the predicted health state probability distribution result, and generate a corrected health state probability;

[0082] Loss function calculation: calculate the cross-entropy loss and the consistency loss of the deterioration degree;

[0083] Iterative training of the deep learning model and updating of the model parameters;

[0084] S4, real-time acquisition of port large equipment detection data, after processing by steps S1 and S2, input into the trained model, output health state probability distribution result, according to the maximum value of the probability distribution to determine the current health state category of the equipment.

[0085] In specific embodiments, S1 is as follows:

[0086] Systematic acquisition and labeling of monitoring data, through the installation of vibration acceleration sensors in key parts of large equipment such as bridge cranes, bucket wheel stackers and reclaimers in the port, such as motor bearings, gearboxes, walking mechanisms, etc., to collect the vibration signals of the equipment in real time at a fixed sampling frequency;

[0087] At the same time, in order to construct the training set required by supervised learning, the health state category of each sample needs to be labeled, and the labeling process is determined by the equipment maintenance experts according to the actual operation state and historical maintenance records of the equipment. The health state of the equipment is divided into five categories: healthy, slight deterioration, moderate deterioration, severe deterioration and failure.

[0088] The labeled samples are divided into training set, validation set and test set in chronological order, wherein the training set is used for model parameter learning, the validation set is used for hyperparameter optimization, and the test set is used for final performance evaluation.

[0089] Optionally, in order to enhance the generalization ability of the model, data enhancement techniques such as adding random Gaussian noise and time stretching can be used to simulate the complex noise environment of the port site.

[0090] As shown in FIG. 1, the vibration signal features of the port large equipment in different health states are set as follows: Figure 3 The sampling frequency is 10 kHz, the original vibration signal collected is a continuous time domain waveform, each sample is cut to a fixed length of 4096 sampling points, corresponding to the vibration response of the equipment within 0.4096 seconds, forming an original vibration signal vector; Figure 3 The horizontal coordinate is time, in seconds, and the vertical coordinate is vibration amplitude, in meters per second squared;

[0091] As shown in FIG. 2, the horizontal coordinate is time, in seconds, and the vertical coordinate is vibration amplitude, in meters per second squared; Figure 3It can be seen that the vibration signal waveform of the healthy state is relatively regular and smooth, mainly showing regular vibration composed of the fundamental frequency and harmonics of the equipment, and the background noise amplitude is small.

[0092] With the aggravation of the deterioration degree, the waveform begins to change obviously: in the slight deterioration state, sporadic and small amplitude impact pulses have begun to appear; in the moderate deterioration state, not only the amplitude and number of impact pulses have increased, but also the overall signal amplitude modulation phenomenon can be observed, that is, the envelope of the waveform begins to fluctuate periodically; in the severe deterioration state, strong and periodic impact pulses become the most prominent feature; and in the failure state, the signal has completely lost regularity, showing strong non-stationary characteristics, dominated by large amplitude random impact and noise, and accompanied by obvious low frequency drift.

[0093] In the specific implementation, S2 is specifically as follows:

[0094] The port equipment vibration data faces the challenges of strong noise interference and multi-scale impact characteristics in health state detection, and the conventional processing methods such as short-time Fourier transform use a fixed window length, which is difficult to capture high-frequency impact components and low-frequency deterioration modes at the same time, resulting in insufficient feature extraction and modal aliasing phenomenon.

[0095] The improved adaptive noise complete ensemble empirical mode decomposition and dynamic spectrum kurtosis fusion method is adopted to realize multi-scale adaptive decomposition and reconstruction of the vibration signal, and the specific steps are as follows:

[0096] 1) Improved adaptive noise decomposition:

[0097] The original vibration signal is subjected to improved adaptive noise complete ensemble empirical mode decomposition, and the conventional adaptive noise complete ensemble empirical mode decomposition algorithm before improvement uses a fixed stopping criterion, such as a fixed number of iterations or a fixed standard deviation threshold, which may lead to modal aliasing or over-decomposition;

[0098] The improved adaptive noise complete ensemble empirical mode decomposition algorithm of the application adopts an adaptive stopping criterion, that is, the stopping criterion is defined as the standard deviation change of adjacent two iterations being less than an adaptive threshold, that is, ;

[0099] Specifically, in the mth iteration, the standard deviation of the current intrinsic mode function and the standard deviation of the previous iteration are calculated, and if , the iteration is stopped, otherwise the iteration is continued, thereby suppressing modal aliasing;

[0100] The original vibration signal vector is defined as , the dimension is 4096, representing the vibration amplitude of the equipment sampling point;

[0101] The kth order intrinsic mode function is defined as , which represents the kth intrinsic mode function obtained by the improved adaptive noise complete ensemble empirical mode decomposition algorithm, is used to describe the oscillation mode of different time scales in the signal;

[0102] wherein k is the index of the intrinsic mode function, and the value range is 1 to K; K is the total order of the intrinsic mode function, and the value is 8; is the length of the original vibration signal vector, ; is the iteration number when each intrinsic mode function is extracted; represents the standard deviation of the mth iteration; represents the standard deviation of the mth iteration; is a logarithmic function, and the default base is the natural constant;

[0103] It should be noted that, like the conventional adaptive noise complete ensemble empirical mode decomposition algorithm, the iteration process includes calculating the upper and lower envelope lines of the signal, averaging, and subtracting the mean value step until the stopping criterion is met.

[0104] 2) Dynamic spectral kurtosis calculation:

[0105] For each order intrinsic mode function, the fourth power of the intrinsic mode function is converted to the frequency domain amplitude by Fourier transform, and the square of the intrinsic mode function is converted to the square of the frequency domain amplitude, and then the ratio of the two is calculated to obtain the dynamic spectral kurtosis value, which quantifies the strength of the impact feature in the frequency domain of the intrinsic mode function, and is represented as:

[0106] ,

[0107] In the formula, represents the Fourier transform, which is used to convert the signal from the time domain to the frequency domain; represents the dynamic spectral kurtosis value of the kth intrinsic mode function at f frequency, which is used to characterize the strength of the impact feature at the frequency; represents the fourth power of the kth intrinsic mode function; represents the square of the kth intrinsic mode function;

[0108] It should be noted that, the term represents that the intrinsic mode function amplifies the transient impact component in the signal through the fourth power operation, has a high peak value, and makes it more significant in the frequency domain, the term represents that the intrinsic mode function obtains the power spectrum of the signal through the square operation, which is used for normalization processing, and the ratio of the two obtains the dynamic spectral kurtosis value, which can effectively quantify the strength of the impact feature in the frequency domain;

[0109] 3) Adaptive threshold filtering reconstruction: ​

[0110] Based on the dynamic spectral kurtosis of each order modal function, the average dynamic spectral kurtosis of each order is calculated, denoted as:

[0111] ,

[0112] In the formula, is the number of frequency points; is the k-th order average dynamic spectral kurtosis;

[0113] It should be noted that the summation operation only covers the positive frequency range from 0 to , and is the sampling frequency. Since the length of the original vibration signal is 4096, and after Fourier transform, 4096 frequency points including positive and negative frequencies are obtained, but considering the spectral symmetry of real signals, only positive frequency points are usually used in actual calculation to avoid redundancy, so the number of frequency points is 2049.

[0114] Then, the adaptive threshold is calculated according to the mean and standard deviation of the dynamic spectral kurtosis of the order in the training set, and the adaptive threshold corresponding to the k-th order modal function is , which is calculated from the mean and standard deviation of the spectral kurtosis of the order in the training set, and is used to distinguish significant impacts from noise, denoted as ;

[0115] In the formula, is the mean of the dynamic spectral kurtosis value of the k-th order modal function at frequency f in the training set; is the standard deviation of the dynamic spectral kurtosis value of the k-th order modal function at frequency f in the training set;

[0116] Then, the difference between the average dynamic spectral kurtosis and the adaptive threshold is processed by the sign function to obtain the weight, and then the high-dimensional decomposition reconstruction matrix is constructed by combining the modal functions, denoted as:

[0117] ,

[0118] In the formula, is the high-dimensional decomposition reconstruction matrix, with dimensions ; is the K-th order modal function, is the first order modal function, is the second order modal function; is the K-th order average dynamic spectral kurtosis, is the first order average dynamic spectral kurtosis, is the second order average dynamic spectral kurtosis; is a symbol function, whose input is greater than 0, output +1, reserve the eigenmode function component, whose input is less than 0, output -1, suppress the eigenmode function component; is the adaptive threshold value corresponding to the Kth order eigenmode function, is the adaptive threshold value corresponding to the 1st order eigenmode function, is the adaptive threshold value corresponding to the 2nd order eigenmode function; is a transpose operation;

[0119] Finally, the high-dimensional decomposition reconstruction matrix is reduced in dimension, and the dimension reduction method can adopt the method of piecewise averaging or downsampling to obtain a characteristic matrix of dimension, which is defined as the decomposition reconstruction matrix , and the dimension is ;

[0120] It should be noted that the spectral kurtosis distribution law of each order eigenmode function is learned by the training set, so that the adaptive threshold can adapt to the noise basis of different devices, and the high-dimensional decomposition reconstruction matrix is constructed by combining the hard decision mechanism of the symbol function, which not only retains the significant impact component but also completely suppresses the noise dominant component, and can solve the compatibility problem of high-frequency impact and low-frequency wear characteristics under strong background noise in the port; the high-frequency impact component is compressed and enhanced in the time domain, and the low-frequency wear pattern is more concentrated in the frequency domain energy distribution;

[0121] As shown in Figure 4 , the feature extraction effect comparison and analysis of different decomposition methods are carried out to verify the feature extraction capability of the improved adaptive noise complete ensemble empirical mode decomposition algorithm proposed by the application in a strong noise environment; the performances of the conventional empirical mode decomposition, the ensemble empirical mode decomposition, the adaptive noise decomposition and the method of the application in the reconstruction signal-to-noise ratio index are compared; from Figure 4 it can be seen that the abscissa represents four kinds of decomposition methods, and the ordinate unit is decibel (dB), which represents the signal-to-noise ratio value of the reconstructed signal and the original signal;

[0122] The column height corresponding to the method of the application is significantly higher than that of the other three methods, and the error line range is the smallest, indicating that the reconstructed signal-to-noise ratio mean is the highest and the stability is the best, Figure 4 the original measurement values of 5 independent repeated experiments are represented by the scattered points, and the scattered point distribution shows that the data points of the method of the application are concentrated in the high signal-to-noise ratio area, while the data points of the conventional method are scattered in the lower interval of 3-4 dB, indicating that the adaptive stopping criterion and the dynamic spectral kurtosis fusion mechanism adopted by the application effectively suppress the modal aliasing phenomenon in the strong noise environment of the port, so that the high-frequency impact component and the low-frequency degradation mode are more completely retained.

[0123] As shown in Figure 5As shown, the adaptive time-frequency decomposition result analysis is performed, and the difference in effect of the adaptive time-frequency decomposition method proposed in the application and the conventional short-time Fourier transform in processing non-stationary vibration signals is contrastively displayed; the experiment uses a signal in a moderate degradation state as input, and the processing results of the two methods are displayed from top to bottom respectively; Figure 5 The upper middle graph is the result of the conventional short-time Fourier transform, which adopts a fixed time window length, and although the frequency resolution in the low-frequency region (the bottom of the graph) is acceptable, the time resolution in the high-frequency region (the top of the graph) is insufficient, resulting in that the transient impact component (the yellow vertical line appearing in the high-frequency region of the graph) in the high-frequency region is blurred and diffused in time, and the details are unclear; Figure 5 The lower middle graph is the adaptive time-frequency decomposition result of the application, and it can be observed that the time-frequency energy distribution of the signal is more concentrated and clear on the whole time-frequency plane, especially in the high-frequency region, the transient component representing the key fault impact can be more accurately positioned at the time point of occurrence, and at the same time, good frequency concentration is maintained in the whole analysis frequency band, indicating that the application can dynamically adjust the analysis parameters according to the local characteristics of the signal, so as to more effectively separate the impact component hidden in the strong background noise.

[0124] In the specific implementation, the time-frequency dual-path feature recalibration is specifically as follows:

[0125] Because the characteristic energy distribution of the vibration signal is nonlinearly shifted in the equipment degradation process, the conventional channel attention mechanism uses a global pooling operation, which will lose the local impact features, resulting in insensitivity to subtle faults; therefore, the application generates an attention mask through the time-frequency dual-path feature calibration mechanism to adaptively recalibrate the features, and the specific steps are as follows:

[0126] 1) Time domain attention mask generation:

[0127] A depth separable one-dimensional convolution operation is used to extract multi-scale time sequence features of the decomposition reconstruction matrix, and then a normalized exponential function is used to process the convolution output to generate a time domain attention mask, which in turn enhances the response of the key local impact area, and is expressed as:

[0128] ,

[0129] In the formula, represents a depth separable one-dimensional convolution operation, and the convolution kernel size includes 3, 5 and 7, and the expansion rate is 1, 3 and 5 respectively, which is used to extract multi-scale time sequence features; is a normalized exponential function, The term is used to map the convolution output to the attention weight in the form of a probability distribution; represents the time domain attention mask, which is used to enhance the response of the key local impact area.

[0130] 2) Frequency domain attention mask generation:

[0131] The decomposition reconstruction matrix is converted to the frequency domain by Fourier transform, the frequency domain amplitude spectrum is obtained by taking the modulus, and then the global energy distribution of the frequency domain features is learned using a multi-layer perception. The output is converted back to the time domain by inverse Fourier transform to generate a frequency domain attention mask reflecting the global energy distribution characteristics, denoted as:

[0132] ,

[0133] In the formula, represents the modulus operation for obtaining the frequency domain amplitude spectrum; represents a multi-layer perception, including two hidden layers, each with a dimension of 64, for learning the global energy distribution of the frequency domain features, and the output layer has a dimension of 512; represents an inverse Fourier transform for converting the frequency domain weight back to the time domain form; represents the frequency domain attention mask for reflecting the global energy distribution characteristics in the vibration signal;

[0134] 3) Feature re-calibration output:

[0135] The time domain attention mask and the frequency domain attention mask are fused by broadcast addition operation, then the range of the fused weight is constrained using the hyperbolic tangent function, and finally the re-calibration matrix is output by element-wise multiplication with the decomposition reconstruction matrix, denoted as:

[0136] ,

[0137] In the formula, represents the broadcast addition operation for aligning and adding the time domain attention mask and the frequency domain attention mask by dimension; is the hyperbolic tangent activation function, the term is used to constrain the fused weight as much as possible within the interval to prevent over-amplification of the features; represents Hadamard product, i.e. element-wise multiplication; represents the calibration feature matrix with a dimension of ;

[0138] It should be noted that the value range of the hyperbolic tangent activation function is , but since the function input value of the term is usually not extreme, the actual output is mostly concentrated in the range of , thereby effectively preventing the features from being over-amplified; the broadcast addition operation automatically expands the tensor with smaller dimensions to the tensor with larger dimensions, e.g. the dimension of the time domain attention mask is ​And the frequency domain attention mask has a dimension of The frequency domain attention mask will be copied along the first dimension times, and the dimension becomes Then an element-wise addition operation is implemented with the time domain attention mask The time domain attention mask adopts a multi-scale depthwise separable convolution, and the hierarchical receptive field accidentally captures the impact transmission attenuation law specific to the port equipment, the small convolution kernel focuses on the gear meshing transient impact, and the large expansion rate convolution perceives the long-period resonance caused by bearing wear, the frequency domain attention mask The frequency-time domain conversion is realized through Fourier transform and inverse Fourier transform, and at the same time, the multi-layer perception structure adopts a bottleneck design with a hidden layer dimension of 64, forcing the network to learn the compressed representation of frequency energy, when the local gear fracture and the overall bearing wear coexist, the time domain path enhances the fracture impact pulse, and the frequency domain path amplifies the bearing characteristic frequency sideband, which can effectively improve the separability of the composite fault features.

[0139] In the embodiment, the cascaded hollow depth feature extraction is specifically as follows:

[0140] Because the conventional convolutional neural network has a large amount of parameters in vibration signal processing, is not sensitive to high-frequency features, is difficult to capture multi-scale impact features, and has low feature extraction efficiency and limited receptive field, therefore, the cascaded hollow depth convolution module is used to extract multi-scale features, expand the receptive field and reduce the amount of parameters, and the specific steps are as follows:

[0141] 1) Feature initialization:

[0142] The calibration feature matrix is taken as the input of the cascaded hollow depth convolution module, the input feature of the first stage convolution is initialized, and is represented as ;

[0143] Wherein, represents the initial input feature of the lth layer network; is the network layer index of the cascaded hollow depth convolution module;

[0144] 2) Cascaded hollow convolution processing:

[0145] The input feature is processed by using a three-stage cascaded hollow depth separable convolution, each stage includes a hollow depth one-dimensional convolution, batch normalization and Gaussian error linear unit activation function operation, and the output features of each processing stage are gradually extracted, and are represented as:

[0146] ,

[0147] In the formula, The index of the processing stage is 1, 2, 3; The output feature of the third processing stage of the lth network is represented as The output feature of the third processing stage of the lth network is represented as The output feature of the third processing stage of the lth network is represented as The batch normalization operation is represented as The Gaussian error linear unit activation function is represented as The output feature of the first processing stage of the lth network is represented as The output feature of the first processing stage of the lth network is represented as The output feature of the first processing stage of the lth network is represented as The output feature of the first processing stage of the lth network is represented as The output feature of the first processing stage of the lth network is represented as The output feature of the first processing stage of the lth network is represented as

[0148] It should be noted that the processing stage refers to three consecutive processing steps inside the cascaded dilated depth convolution module, each level including dilated depth separable one-dimensional convolution, batch normalization and Gaussian error linear unit activation function activation operation; the dilated rate is 2, 4, 2, the first layer dilated rate 2 captures short period impacts such as gear engagement, the middle layer dilated rate 4 perceives the device structure resonance frequency band, and the last layer returns to 2 to restore the feature resolution. This "expansion-contraction" structure forms a time-frequency focusing ability similar to wavelet analysis.

[0149] 3) Multi-scale feature splicing:

[0150] The output features of the three processing stages are spliced along the channel dimension to form multi-scale fusion features, thereby expanding the receptive field and covering the impact period in the vibration signal, which is represented as

[0151] ,

[0152] In the formula, The splicing operation of the three feature segments along the channel dimension is represented as The output feature of the first processing stage of the lth network is represented as The output feature of the second processing stage of the lth network is represented as The output feature of the third processing stage of the lth network is represented as The multi-scale fusion feature is represented as

[0153] The specific implementation of the dynamic deterioration perception feature enhancement is as follows:

[0154] Because the vibration characteristics of port equipment undergo nonlinear manifold distortion due to material fatigue and structural wear during long-term operation, conventional feature extraction methods cannot explicitly model the degradation process. The boundary between the healthy state and the degradation stage is blurred, resulting in insufficient sensitivity of classification models to the intermediate degradation state. Therefore, this invention dynamically adjusts the feature response to perceive the degradation process by calculating feature offset and degradation score. The specific steps are as follows:

[0155] 1) Feature offset calculation:

[0156] A baseline feature vector is extracted from healthy samples in the training set. The difference between the multi-scale fused features of the current sample and the baseline feature vector is calculated to obtain the feature offset, which quantifies the degree of deviation relative to the health state, and is expressed as:

[0157] ,

[0158] In the formula, This represents the feature offset, used to quantify the degree to which the current sample's features deviate from its health status. The baseline feature vector representing health status is obtained by calculating the mean feature value of the top 5% of healthy samples in the training set.

[0159] Specifically, the selection of the top 5% of healthy samples in the training set is based on the sorting of amplitude variance or signal-to-noise ratio. Specifically, the samples in the training set are sorted according to the amplitude variance or signal-to-noise ratio of the vibration signal, and the top 5% of samples with the smallest variance or the highest signal-to-noise ratio and labeled as healthy are selected.

[0160] 2) Degradation score generation:

[0161] Global average pooling is performed on the feature offsets to obtain a global vector, which is then input into a multilayer perceptron to learn the degradation trajectory. The output degradation score is used to quantify the degree of device degradation, and is expressed as:

[0162] ,

[0163] In the formula, This represents the degradation score, with a value ranging from 0 to 1, used to quantify the degree of degradation of the current state of the equipment; This represents the global average pooling operation, used to compress feature offsets into a global vector of fixed dimensions. This represents a three-hidden-layer multilayer perceptron with hidden layer dimensions of 128, 64, and 32, used to learn a non-linear mapping from feature shifts to degradation levels. Its output layer uses the sigmoid activation function to assign degradation scores. The constraint is between 0 and 1, where 0 represents complete health and 1 represents complete failure;

[0164] It should be noted that by constructing a baseline feature vector of health status... As the "zero drift anchor point" of the feature space, the multi-layer perceptron adopts a dimension reduction structure design with hidden layer dimensions of 128, 64, and 32 respectively. The bottleneck layer forces the network to extract the core factors of bearing wear and gear backlash, and outputs the degradation score with clear physical meaning;

[0165] 3) Channel response dynamic adjustment:

[0166] Based on the degradation score and the learnable adaptive weight vector, the channel dimension of the multi-scale fusion feature is scaled and adjusted to obtain the degradation perception feature, which enhances the response of the degradation-sensitive channel, and is expressed as:

[0167] ,

[0168] In the formula, is the degradation perception feature, which represents the feature output after degradation perception adjustment, and the channel dimension is 512; represents the channel scaling operation, that is, multiplying each channel of the feature vector by the corresponding scaling coefficient; is the adaptive weight vector, and the dimension is is a trainable parameter; is the multi-scale fusion feature .

[0169] In the specific implementation, the gated multi-granularity feature fusion classification is as follows:

[0170] In the classification of the health status of port equipment, different degradation modes correspond to different characteristic segments of the vibration signal. A single fully connected classifier cannot effectively capture local subtle abnormal features, and the importance of features at different degradation stages is different. Global uniform processing can easily reduce the discrimination of mid-term degradation state and local impact features. Therefore, the present application adopts a gated multi-branch structure to process and fuse the features, enhances the classification sensitivity, and the specific steps are as follows:

[0171] 1) Uniform segmentation of features:

[0172] The degradation perception feature is uniformly divided into four feature segments along the channel dimension, and is expressed as:

[0173] ,

[0174] In the formula, represents the first feature segment of the degradation perception feature, which is obtained by uniformly dividing the degradation perception feature into 4 segments along the channel dimension, and each segment has a dimension of 128. Similarly, represents the second feature segment of the degradation perception feature, the 3rd feature sub-segment representing the degradation-aware feature, the 4th feature sub-segment representing the degradation-aware feature, each feature sub-segment with a dimension of 128; representing a concatenation operation along the channel dimension on the 4 feature sub-segments;

[0175] 2) Gating weight generation:

[0176] For each feature sub-segment of the degradation-aware feature, a non-linear transformation is performed through a fully connected layer, and then a Sigmoid activation function is used to calculate the gating weight for evaluating the importance of the feature sub-segment to the classification decision, represented as:

[0177] ,

[0178] wherein, represents the 1st feature sub-segment of the degradation-aware feature; is a feature sub-segment index, taking values 1, 2, 3, 4 is a Sigmoid activation function; represents the gating weight vector corresponding to the 1st feature sub-segment, with a value range of 0 to 1, used to regulate the contribution degree of the feature sub-segment; represents a fully connected layer with an output dimension of 128, used for non-linear transformation of the features of each sub-segment; 3) Gating transformed feature calculation:

[0179] For each feature sub-segment, dimension reduction and abstraction are performed through a fully connected layer, and then a rectified linear unit activation function is used for processing, combined with the gating weight vector to obtain the gating transformed feature, realizing feature selection and enhancement, represented as:

[0180]

[0181] ,

[0182] wherein, represents the gating transformed feature of the 1st feature sub-segment; represents a fully connected layer with an output dimension of 64, used for feature dimension reduction and abstraction; represents a rectified linear unit activation function, used to enhance the non-linear expression ability of the model; represents a Hadamard product, i.e., element-wise multiplication, used to realize scaling of the feature values by the gating weight; 4) Multi-branch fusion classification:

[0183]

[0184] ​​​The gated transform features of each feature segment are mapped to class scores through a fully connected layer. Then, the class scores of all branches are summed and normalized using a normalized exponential function to obtain the predicted health status probability distribution, as follows:

[0185] ,

[0186] In the formula, This represents the predicted probability distribution of health status, with a dimension of 5, corresponding to five states: healthy, slightly deteriorated, moderately deteriorated, severely deteriorated, and failed. This indicates a fully connected layer with an output dimension of 5, used to map the features of each branch to class scores;

[0187] It should be noted that the 512-dimensional degradation perception features The feature is uniformly divided into four 128-dimensional feature segments, which is not a simple segmentation. Its physical meaning corresponds to the frequency band division of the vibration signal. Each segment independently selects features through a gating weight vector. For example, gear tooth breakage mainly activates the high-frequency segment, while bearing wear dominates the mid-frequency segment. At the same time, the Sigmoid function outputs continuous values ​​[0,1] instead of a hard threshold, preserving the gradual change characteristics of feature contribution. In addition, the gating transformation feature calculation adopts a multi-branch fraction summation method, which enables the classification decision to integrate multi-frequency band evidence and improve the recognition rate of moderate degradation. When multiple segments give weak positive signals, their joint judgment is more reliable than a single fully connected layer.

[0188] In a specific implementation, the degradation perception probability correction is as follows:

[0189] In port equipment health status classification, the predicted probability directly using a multi-granularity feature fusion classifier may not adequately reflect the gradual degradation process, leading to biases in the judgment of intermediate status. Therefore, this invention generates a prior distribution based on degradation scores and corrects the classification probability to reflect the degree of degradation. The specific steps are as follows:

[0190] 1) Calculation of prior distribution of degradation:

[0191] Based on the degradation score, a prior probability distribution is generated using an exponential decay function. The center of this distribution shifts linearly with the degradation score, tending towards higher degradation states as the degradation score increases, thus reflecting the degree of degradation. This distribution is expressed as:

[0192] ,

[0193] In the formula, Indicates the first The prior probability of a health status class; j is the index of the equipment health status class, with values ​​of 1, 2, 3, 4, and 5, corresponding to the five states of healthy, slightly deteriorated, moderately deteriorated, severely deteriorated, and failed, respectively; To distinguish it from the device health status category index of j; This is the sharpness parameter, which controls the degree of concentration of the prior distribution, and its value is 10. As a scaling parameter, the degradation score is mapped to a category index range, with a preferred value of 4, making... The corresponding category indices are 1 to 5; Represents the natural exponential function;

[0194] It should be noted that the prior distribution The central position follows Linear movement, when At that time, the center was (Health), when At that time, the center was (Failure), thus reflecting the degree of degradation.

[0195] 2) Probability correction:

[0196] By fusing the prior distribution and predicted probability using geometric mean, a corrected probability is obtained. This incorporates degradation knowledge into the classification decision, improving the classification performance for intermediate degradation states. This can be expressed as:

[0197] ,

[0198] In the formula, Indicates the first The corrected probability of a health-like state; Indicates the first The corrected probability of a health-like state; Indicates the first Prior probability of a health status; The correction intensity parameter is used to control the influence of the prior distribution. Its value ranges from 0 to 1, and the value is 0.3.

[0199] In a specific implementation, the loss function is calculated as follows:

[0200] The loss function consists of a classification cross-entropy loss term and a degradation consistency loss term. The classification cross-entropy loss term is calculated based on the corrected probability and the true label, while the degradation consistency loss term is calculated based on the degradation score and the true degradation level, ensuring that the degradation score is consistent with the true degradation level. This is expressed as:

[0201] ,

[0202] In the formula, The loss function; The cross-entropy loss term is calculated as follows:

[0203] ,

[0204] For the first class of real label one-hot encoding;

[0205] For the degradation consistency loss term, the calculation method is represented as:

[0206]

[0207] For the loss weight parameter, used to balance the two loss terms, the value is 0.5; For the real degradation degree, according to the real label mapping, if the real label is healthy, then , slight degradation, then , moderate degradation, then , severe degradation, then , failure, then .

[0208] In the specific implementation, the deep learning model is iteratively trained and the model parameters are updated. The training of the model includes training phase, validation phase and test phase, and the specific operations of each phase are as follows:

[0209] Training phase: taking the training set as the core, the network parameters are constantly updated through iterative optimization algorithm to minimize the total loss function;

[0210] Specifically, the mini-batch gradient descent method is adopted, a batch of samples are randomly selected from the training set each time, the batch size can usually be 32 or 64, input into the constructed deep learning network, and the predicted health state probability distribution and degradation score are obtained through forward propagation calculation; Then calculate the loss value of the current batch according to the total loss function, the loss function is composed of classification cross-entropy loss and degradation consistency loss weighted, wherein the cross-entropy loss measures the difference between the predicted probability and the real label, and the degradation consistency loss ensures the consistency of the degradation score and the real degradation degree; In the process of back propagation, the gradient of the total loss function to the parameters of each layer of the network is calculated, and the Adam optimizer is used to update the parameters, and the learning rate is set to 0.001;

[0211] Validation phase: each time the training set is completely traversed, the model performance is evaluated using the validation set;

[0212] The validation set samples are input into the model after the training phase, the validation indicators are calculated, and the training strategy is dynamically adjusted based on the validation indicators, the early stopping strategy is adopted to prevent overfitting, and if the validation set loss does not decrease for multiple traversal periods, the training is terminated in advance;

[0213] Test phase: input the test set samples into the model after the validation phase, get the prediction results, and measure the accuracy of the model and the adaptation ability to unknown data.​

[0214] In the detailed description, S4 is specifically as follows:

[0215] After the model training is completed, it can be used for health state detection of new port large equipment vibration data. Specifically, vibration signals are collected in real time from vibration sensors of the equipment to be detected, a sample segment with a length of 4096 is intercepted according to the same preprocessing mode, and adaptive time-frequency decomposition and feature reconstruction processing are performed;

[0216] The data after adaptive time-frequency decomposition and feature reconstruction processing are input into the trained deep learning model, and sequentially pass through time-frequency double-path feature re-scaling, cascaded hollow deep feature extraction, dynamic degradation perception feature enhancement, gated multi-granularity feature fusion classification and degradation perception probability correction steps, and output the health state probability distribution and degradation score of the sample;

[0217] Finally, according to the maximum value of the probability distribution, the current health state category of the equipment is determined, including health, slight degradation, moderate degradation, severe degradation or failure, realizing online real-time health state detection and early warning.

[0218] Example 2

[0219] As shown in Figure 6 , the classification accuracy of different models on five health states is analyzed, and the classification performance of five models (support vector machine, random forest, one-dimensional convolutional neural network, long short-term memory network and the method of the application) on five health states of the equipment is evaluated. The horizontal coordinate of the grouped column chart represents the health state category (health, slight degradation, moderate degradation, severe degradation, failure), and the vertical coordinate unit is percentage (%), representing the classification accuracy; the column body of the method of the application maintains the highest position in all health state categories, especially in the moderate degradation and slight degradation categories, and its accuracy advantage is most significant. The conventional model shows a significant decline in moderate degradation state, while the method of the application effectively captures the weak features in the early stage of equipment degradation through time-frequency double-path feature re-scaling mechanism and gated multi-granularity feature fusion. Figure 6 As can be seen from the figure, the column body of the method of the application fluctuates the least among different states, indicating that it has balanced recognition ability for each degradation stage, solving the problem of insufficient sensitivity of the conventional model to the moderate degradation state.

[0220] As shown in Figure 7 , the feature distribution visualization analysis of different health states is performed, and the distribution rule of the features extracted by the method of the application in two-dimensional space is displayed through t-SNE dimension reduction technology. Figure 7In the diagram, samples of five health states are labeled with different colors: healthy (blue), slightly deteriorated (green), moderately deteriorated (yellow), severely deteriorated (orange), and failed (red); the horizontal and vertical axes are two feature dimensions, representing the abstract feature space after dimensionality reduction. Figure 7 Three key phenomena are visible: 1) Healthy state (blue) and failed state (red) are clustered on opposite sides of the space, forming a clear separation; 2) The three deterioration states (green, yellow, orange) are distributed in a continuous transition, which is consistent with the gradual deterioration characteristics of equipment; 3) Moderate deterioration (yellow) has a certain degree of overlap with adjacent states, but the main components are still distinguishable. The distribution characteristics are derived from the adaptive adjustment of the feature manifold by the dynamic deterioration perception mechanism, which enables the model to accurately quantify the gradual process from healthy to failed. Figure 7 No abnormal mixed clusters were found, proving that the method of the present invention effectively solves the problem of blurred boundaries between healthy and deteriorated states.

[0221] Example 3

[0222] like Figure 8 As shown, the trend of score changes during the accelerated degradation test of the equipment is analyzed. Through the accelerated degradation test, the change pattern of degradation score during 100 days of operation from healthy to failure of the equipment is simulated. The horizontal axis of the line graph is the operating time (days), and the vertical axis is the degradation score (0-1 interval). Figure 8 The model includes three curves: the actual degree of degradation (black dashed line), the score of the method of this invention (cyan solid line), and the score of the conventional method (purple solid line). Background color blocks indicate the thresholds for five health status categories. Experimental results show that: 1) the curve of this invention (cyan) closely matches the actual curve (black), accurately tracking trend changes at key nodes such as 30 days (start of slight degradation), 50 days (transition to moderate degradation), and 70 days (transition to severe degradation); 2) the conventional method (purple) lags significantly in the early to mid-stages (20-60 days), failing to detect initial degradation in a timely manner; 3) the actual detection points marked by scatter plots show that the scoring error of this invention is smaller at the sampling points. Experimental results indicate that the degradation consistency loss function and dynamic feature offset calculation enable the model to perceive nonlinear feature drift caused by material fatigue. Figure 8 The abnormal drop in the conventional curve after 80 days exposes the risk of misjudging complex faults using conventional methods.

[0223] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for model training for port large equipment health state monitoring, characterized in that, The method comprises the following steps: S1, collecting and labeling the detection data of large-scale equipment in the port, and dividing the labeled sample data into a training set, a validation set and a test set in chronological order; S2, adaptively decomposing and reconstructing the collected sample data in time and frequency, specifically using an improved adaptive noise complete ensemble empirical mode decomposition and dynamic spectral kurtosis fusion method, generating a high-dimensional decomposition and reconstruction matrix, and then reducing the dimension to a decomposition and reconstruction matrix; S2 is specifically as follows: S2.1, defining the labeled sample data as an original vibration signal vector, performing improved adaptive noise complete ensemble empirical mode decomposition on the original vibration signal, and using an adaptive stopping criterion, the stopping criterion being defined as the standard deviation of adjacent two iterations being less than an adaptive threshold; The improved adaptive noise complete set empirical mode decomposition algorithm decomposes to obtain an intrinsic mode function, the total order of the intrinsic mode function is K, and the kth order intrinsic mode function is represented as , , k represents an index of the intrinsic mode function. Specifically, in the mth iteration, the standard deviation of the current eigenmode function and the standard deviation of the previous iteration are calculated, if the iteration is stopped, otherwise the iteration is continued, and further the mode aliasing is suppressed. wherein, denotes the standard deviation of the mth iteration, denotes the standard deviation of the mth iteration; denotes the standard deviation of the mth iteration; denotes the logarithm function with the natural constant as the default base; denotes the length of the original vibration signal vector; denotes the number of iterations when extracting each order of the intrinsic modal function. S2.2, for each intrinsic mode function, converting the fourth power of the intrinsic mode function into a frequency domain amplitude through Fourier transform, and converting the square of the intrinsic mode function into a frequency domain amplitude square, then calculating the ratio of the two to obtain a dynamic spectral kurtosis value; S2.3, based on the dynamic spectral kurtosis of each intrinsic mode function, calculating the average dynamic spectral kurtosis of each order; Then calculate the adaptive threshold of each order according to the mean and standard deviation of the dynamic spectral kurtosis in the training set; Then the difference between the average dynamic spectral kurtosis and the adaptive threshold is processed by a sign function to obtain a weight, and then a high-dimensional decomposition and reconstruction matrix is constructed by combining the intrinsic mode function; Finally, the high-dimensional decomposition and reconstruction matrix is reduced in dimension by using a piecewise average or downsampling method to obtain a decomposition and reconstruction matrix; S3, constructing a deep learning model, inputting the sample data in the training set, the validation set and the test set processed by step S2 into the model to train the model, and obtaining a trained model; S4, real-time collection of port large-scale equipment detection data, inputting the processed data in steps S1 and S2 into the trained model, outputting a health state probability distribution result, and determining the current health state category of the equipment according to the maximum value of the probability distribution.

2. The method of claim 1, wherein the method further comprises: The operation of constructing a deep learning model is specifically as follows: Time-frequency dual-path feature re-labeling: the decomposition and reconstruction matrix generates an attention mask through a time-frequency dual-path feature calibration mechanism, and after fusion, the decomposition and reconstruction matrix is adaptively re-labeled to generate a calibration feature matrix; Cascade hollow deep feature extraction: the calibration feature matrix extracts multi-scale features through a cascade hollow deep convolution module, and then performs splicing to generate multi-scale fusion features; Dynamic degradation perception feature enhancement: based on the multi-scale fusion features, a feature offset and a degradation score are calculated, and then the channel response is dynamically adjusted to generate a degradation perception feature; Gated multi-granularity feature fusion classification: the degradation perception feature is segmented and fused by using a gated multi-branch structure to generate a predicted health state probability distribution result; Degradation perception probability correction: based on the degradation score, a prior distribution is generated to correct the predicted health state probability distribution result to generate a corrected health state probability; Loss function calculation: calculate the cross-entropy loss and the consistency loss of the degradation degree; Iteratively train the deep learning model and update the model parameters.

3. The method of claim 2, wherein the method further comprises: The operation of time-frequency dual-path feature re-labeling is specifically as follows: (1) The multi-scale time sequence features of the decomposition and reconstruction matrix are extracted using a depth separable one-dimensional convolution operation, and then the convolution output is processed by a normalized exponential function to generate a time domain attention mask; (2) The decomposition and reconstruction matrix is converted to the frequency domain by Fourier transform, the frequency domain amplitude spectrum is obtained by taking the modulus, and then the global energy distribution of the frequency domain features is learned using a multi-layer perceptron, and the output is converted back to the time domain by inverse Fourier transform to generate a frequency domain attention mask; (3) The time domain attention mask and the frequency domain attention mask are fused by broadcast addition operation, then the range of the fused weight is constrained by hyperbolic tangent function, and finally the calibrated feature matrix is output by re-calibrating the decomposition and reconstruction matrix through element-wise multiplication.

4. The port-oriented large equipment health state monitoring model training method according to claim 2, characterized in that, The operation of cascaded hollow depth feature extraction is as follows: (1) The calibrated feature matrix is used as the input of the cascaded hollow depth convolution module to initialize the input feature of the first level convolution; (2) The input feature is processed using a three-level cascaded hollow depth separable convolution, each level including a hollow depth separable one-dimensional convolution, a batch normalization, and a Gaussian error linear unit activation function operation, and the output features of each processing stage are gradually extracted to finally obtain the output features of each layer network at the three processing stages; (3) The output features of the three processing stages are spliced along the channel dimension to form a multi-scale fusion feature.

5. The method of claim 2, wherein the method further comprises: The operation of dynamic deterioration perception feature enhancement is as follows: (1) Select sample data in a healthy state from the training set and extract its baseline feature vector, then subtract the multi-scale fusion feature of the current sample data from the baseline feature vector to obtain a feature offset; (2) The global vector is obtained by performing global average pooling on the feature offset, and then the multi-layer perceptron is used to learn the deterioration trajectory, and the deterioration score for quantifying the deterioration degree of the device is output; (3) Based on the deterioration score and the learnable adaptive weight vector, the multi-scale fusion feature is scaled and adjusted in the channel dimension to obtain the deterioration perception feature.

6. The method of claim 2, wherein the method further comprises: The operation of gated multi-granularity feature fusion classification is as follows: (1) The deterioration perception feature is evenly divided into four feature segments along the channel dimension, and each feature segment has the same dimension; (2) For each feature segment of the deterioration perception feature, a non-linear transformation is performed through a fully connected layer, and then a Sigmoid activation function is used to calculate the gating weight; (3) For each feature segment, dimension reduction and abstraction are performed through a fully connected layer, and then a rectified linear unit activation function is used for processing, combined with the gating weight vector of each feature segment to obtain a gated transformed feature; (4) The gated transformed feature of each feature segment is mapped to a class score through a fully connected layer, then the class scores of all branches are summed, and a normalized exponential function is used for normalization to obtain the predicted health state probability distribution.

7. The method of claim 2, wherein the method further comprises: The operation of deterioration perception probability correction is as follows: (1) Based on the deterioration score, a prior probability distribution is generated by an exponential decay function, and the center position of the distribution moves linearly with the deterioration score, the calculation formula is as follows: ， wherein, represents the prior probability of the jth health state class; j represents the index of the device health state class, taking values of 1, 2, 3, 4, 5, corresponding to the five states of healthy, slight degradation, moderate degradation, severe degradation, and failure, respectively; represents the index of the device health state class different from j; represents the sharpness parameter; represents the scaling parameter; represents the natural exponential function; represents the degradation score; (2) The prior distribution and the predicted probability are fused by geometric mean to obtain the corrected probability, the calculation formula is as follows: ， wherein, represents the first class health state; represents the predicted first class health state probability; represents the predicted first class health state probability; represents the first class health state; represents a correction strength parameter for controlling the influence of the prior distribution.

8. The port-oriented large equipment health state monitoring model training method according to claim 2, characterized in that, The operation of loss function calculation is as follows: A classification cross-entropy loss function is calculated according to the correction probability and the real label; A degradation consistency loss function is calculated based on the degradation score and the real degradation degree; The degradation consistency loss function is weighted and added to the classification cross-entropy loss function to obtain a total loss function.

9. The port-oriented large equipment health state monitoring model training method according to claim 2, characterized in that, The deep learning model is iteratively trained and the model parameters are updated, and the training of the model includes a training phase, a validation phase and a test phase, and the operations of each phase are as follows: Training phase: taking the training set as the core, the network parameters are continuously updated by iterative optimization algorithm to minimize the total loss function; Specifically, the mini-batch gradient descent method is adopted, a batch of samples are randomly selected from the training set each time, input into the constructed deep learning network, the predicted health state probability distribution and degradation score are obtained through forward propagation calculation, then the loss value of the current batch is calculated according to the total loss function; during the back propagation process, the gradient of the total loss function to the parameters of each layer of the network is calculated, and the Adam optimizer is used to update the parameters, and the learning rate is set; Validation phase: the model performance is evaluated using the validation set after each complete traversal of the training set; The validation set samples are input into the model after the training phase, the validation indicators are calculated, the training strategy is dynamically adjusted based on the validation indicators, the early stopping strategy is adopted to prevent overfitting, and if the validation set loss does not decrease for multiple traversal periods, the training is terminated in advance; Test phase: the test set samples are input into the model after the validation phase, the prediction results are obtained, and the accuracy of the model and the adaptability to unknown data are measured.

Images