Model training method for health state monitoring of large-scale port equipment

Abstract

The invention relates to a model training method for health state monitoring of large-scale equipment in a port. The method specifically comprises the following steps: collecting and marking detection data of the large-scale equipment in the port, and dividing a data set; adaptive time-frequency decomposition and feature reconstruction are carried out on the collected data, and an improved adaptive noise complete set empirical mode decomposition and dynamic spectrum kurtosis fusion method is adopted to generate a decomposition and reconstruction matrix; constructing a deep learning model, processing the sample data, and inputting the processed sample data into the model to train the model to obtain a trained model; and collecting detection data of the large-scale equipment in a port in real time, inputting the detection data into the trained model after processing, outputting a health state probability distribution result, and determining a current health state category of the equipment according to a maximum value of probability distribution. According to the method, adaptive time-frequency decomposition and feature reconstruction are carried out on the collected sample data, a deep learning model is constructed and trained, and real-time health detection and degradation evaluation can be carried out on the running state of the equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment status monitoring based on artificial intelligence, and in particular to a model training method for health status monitoring of large-scale port equipment. Background Art

[0002] With the development of port automation and scale, large-scale equipment such as bridge cranes and bucket-wheel stackers and reclaimers have become critical support for efficient port operations. This equipment is large in size and operates in a complex environment. It is constantly subjected to heavy loads, shocks, and the effects of marine weather, making it highly susceptible to failures such as bearing wear, gear fracture, and structural fatigue. Failure not only results in significant repair costs and production disruptions, but can also lead to serious safety incidents. Therefore, real-time health monitoring and degradation assessment of equipment operating status are crucial.

[0003] Existing equipment condition detection methods primarily rely on traditional vibration signal analysis and machine learning classification models. For example, time-frequency analysis methods such as short-time Fourier transforms and wavelet analysis generally employ fixed analysis windows or basis functions for feature extraction, making it difficult to simultaneously account for high-frequency transient shocks and low-frequency degradation modes. This can easily lead to insufficient feature extraction and fault mode aliasing. While empirical mode decomposition and its improved algorithms can adaptively decompose non-stationary signals, they still suffer from modal aliasing and over-decomposition in strong noise environments, resulting in low reconstructed signal-to-noise ratios and the obscuration of key shock features. Furthermore, traditional deep learning models typically directly input time-domain or frequency-domain features for classification when processing vibration signals, ignoring the gradual nature of the degradation process and the variability of multi-scale features. This often results in inaccurate identification of mild and moderate degradation states. The classification stage typically relies on a single fully connected layer to output class probabilities, failing to incorporate probabilistic corrections based on the evolution of the equipment's health status. This makes the predictions less accurate than those reflecting the continuous process of equipment degradation.

[0004] The existing technical solutions have the following objective shortcomings: the window length of methods such as short-time Fourier transform and wavelet transform is fixed, which makes it difficult to take into account both high-frequency transient impacts and low-frequency wear characteristics, and there are problems of insufficient resolution and feature loss; the use of fixed stopping criteria can easily cause modal aliasing or excessive decomposition, and the signal-to-noise ratio is low in a strong noise environment, resulting in the masking of key fault characteristics; ordinary convolutional neural networks or attention mechanisms are not sensitive enough to high-frequency impacts or medium-term degradation, and cannot model the gradual degradation process of equipment, resulting in a decrease in the classification performance of mild / moderate degradation; existing methods mostly use a single fully connected classifier, which does not distinguish the importance of features in different frequency bands, nor does it combine prior knowledge of degradation, resulting in the probability distribution being inconsistent with the actual degradation process, which is particularly prone to misjudgment of the medium-term state.

[0005] Therefore, the present invention proposes a model training method for health status monitoring of large-scale port equipment to solve the above problems. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention develops a model training method for health status monitoring of large-scale port equipment. The present invention performs adaptive time-frequency decomposition and feature reconstruction on the collected sample data, and constructs and trains a deep learning model to perform real-time health detection and degradation assessment of the equipment's operating status.

[0007] The technical solution to the technical problem of the present invention is a model training method for health status monitoring of large-scale port equipment, which is as follows: S1. Collect and label the inspection data of large-scale port equipment, and divide the labeled sample data into training set, validation set and test set in chronological order; Systematically collect and annotate the inspection data of large-scale port equipment. Vibration acceleration sensors installed on key parts of large-scale port equipment can collect vibration signals of the equipment in real time at a fixed sampling frequency. The collected sample data is labeled with health status categories, and the equipment health status is divided into five categories: healthy, slightly degraded, moderately degraded, severely degraded, and failed; The labeled sample data is divided into training set, validation set and test set in chronological order, which are used for model training respectively. The training set is specifically used for model parameter learning, the validation set is specifically used for model hyperparameter tuning, and the test set is specifically used for model final performance evaluation.

[0008] S2. Adaptively decompose and reconstruct the collected sample data in time and frequency, specifically using an improved adaptive noise complete set empirical mode decomposition and dynamic spectral kurtosis fusion method to generate a high-dimensional decomposition and reconstruction matrix and then reduce the dimension to a decomposition and reconstruction matrix; S2.1. Define the labeled sample data as the original vibration signal vector, perform the improved adaptive noise complete set empirical mode decomposition on the original vibration signal, and adopt the adaptive stopping criterion. The stopping criterion is defined as the change in the standard deviation of two adjacent iterations is less than the adaptive threshold. The eigenmode function obtained after decomposition by the improved adaptive noise complete set empirical mode decomposition algorithm has a total order of K, and the kth order eigenmode function is expressed as , , k is the index of the eigenmode 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 , then stop the iteration, otherwise continue the iteration to suppress the mode aliasing; in, represents the standard deviation of the mth iteration, Indicates the The standard deviation of the iterations; Represents a logarithmic function, with the default base being a natural constant; Represents the length of the original vibration signal vector; Indicates the number of iterations when extracting each order intrinsic mode function; S2.2. For each order of intrinsic mode function, convert the fourth power of the intrinsic mode function to the frequency domain amplitude through Fourier transform, and convert the quadratic power of the intrinsic mode function to the square of the frequency domain amplitude. Then calculate the ratio of the two to obtain the dynamic spectrum kurtosis value. S2.3. Based on the dynamic spectrum kurtosis of each order of the intrinsic mode function, calculate the average dynamic spectrum kurtosis of each order; Then, the adaptive threshold of each order is calculated based on the mean and standard deviation of the dynamic spectrum kurtosis of each order in the training set; The difference between the average dynamic spectrum kurtosis and the adaptive threshold is processed by a sign function to obtain a weight, which is then combined with the intrinsic mode function to construct a high-dimensional decomposition and reconstruction matrix. Finally, the high-dimensional decomposition and reconstruction matrix is ​​reduced in dimension by using the segmented averaging or downsampling method to obtain the decomposition and reconstruction matrix.

[0009] S3: Build a deep learning model. After the sample data in the training set, validation set, and test set are processed in step S2, they are input into the model to train the model to obtain a trained model. The specific operations for building the model for training are as follows: Time-frequency dual-path feature recalibration: The decomposed and reconstructed matrix generates an attention mask through the time-frequency dual-path feature calibration mechanism. After fusion, the decomposed and reconstructed matrix is ​​adaptively recalibrated to generate a calibrated feature matrix. (1) Use depth-wise separable one-dimensional convolution operation to extract multi-scale temporal features of the decomposed reconstruction matrix, and then process the convolution output through a normalized exponential function to generate a temporal attention mask; (2) The decomposition and reconstruction matrix is ​​converted to the frequency domain through Fourier transform, and the frequency domain amplitude spectrum is obtained by taking the modulus. Then, the global energy distribution of the frequency domain features is learned using a multi-layer perceptron. The output is then converted back to the time domain through inverse Fourier transform to generate a frequency domain attention mask. (3) The time domain attention mask and the frequency domain attention mask are fused through a broadcast addition operation, and then the hyperbolic tangent function is used to constrain the fused weight range. Finally, they are recalibrated through element-by-element multiplication and decomposition and reconstruction matrix to output the calibration feature matrix.

[0010] Cascaded hole depth feature extraction: The calibrated feature matrix is ​​passed through the cascaded hole depth convolution module to extract multi-scale features, which are then spliced ​​to generate multi-scale fusion features; (1) The calibration feature matrix is ​​used as the input of the cascaded hole depth convolution module to initialize the input features of the first level convolution; (2) Use three-stage cascaded dilated depth-wise separable convolution to process input features. Each stage includes dilated depth-wise separable one-dimensional convolution, batch normalization, and Gaussian error linear unit activation function operations, gradually extracting the output features of each processing stage, and finally obtaining the output features of the three-stage processing stage of each layer of the network; (3) The output features of the three-level processing stage are spliced ​​along the channel dimension to form multi-scale fusion features.

[0011] Dynamic degradation-aware feature enhancement: Calculates feature offsets and degradation scores based on multi-scale fusion features, then dynamically adjusts channel responses to generate degradation-aware features. (1) Select healthy sample data from the training set and extract its baseline feature vector. Subtract the multi-scale fusion feature of the current sample data from the baseline feature vector to obtain the feature offset. (2) Perform global average pooling on the feature offset to obtain a global vector, which is then input into a multi-layer perceptron to learn the degradation trajectory and output a degradation score used to quantify the degree of device degradation; (3) Based on the degradation score and the learnable adaptation weight vector, the multi-scale fusion features are scaled and adjusted in the channel dimension to obtain the degradation perception features.

[0012] Gated multi-granularity feature fusion classification: A gated multi-branch structure is used to segment and fuse degradation-aware features to generate a predicted health status probability distribution result. (1) The degraded perceptual features are evenly divided into four feature sub-segments along the channel dimension, and each feature sub-segment has the same dimension; (2) For each feature sub-segment of the degraded perceptual feature, a nonlinear transformation is performed through the fully connected layer, and then the gate weight is calculated using the Sigmoid activation function; (3) For each feature sub-segment, dimensionality reduction and abstraction are performed through the fully connected layer, and then the rectified linear unit activation function is used to process it, and the gated transformation feature is obtained by combining the gated weight vector of each feature sub-segment; (4) The gated transformation features of each feature sub-segment are mapped to category scores through a fully connected layer, and then the category scores of all branches are summed and normalized by a normalized exponential function to obtain the predicted health status probability distribution.

[0013] Degradation perception probability correction: Generate a prior distribution based on the degradation score, correct the predicted health state probability distribution results, and generate the corrected health state probability; (1) Based on the degradation score, a prior probability distribution is generated through an exponential decay function. The center position of the distribution moves linearly with the degradation score. The calculation formula is as follows: , in, Indicates the The prior probability of the health state of the device; j represents the health state category index of the device, with values ​​of 1, 2, 3, 4, and 5, corresponding to the five states of health, slight degradation, moderate degradation, severe degradation, and failure, respectively; Indicates the health status category index of the device that is 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 through geometric mean to obtain the corrected probability. The calculation formula is as follows: , in, Indicates the Corrected probability of health-like state; The predicted Class health status probability; The predicted Class health status probability; Indicates the Prior probability of class health status; represents the correction strength parameter used to control the influence of the prior distribution.

[0014] Loss function calculation: Calculate the cross entropy loss and the consistency loss of the degradation degree; Calculate the classification cross entropy loss function based on the corrected probability and the true label; Calculate the consistency loss function of degradation based on the degradation score and the actual degradation degree; The degraded consistency loss function is weighted and added to the classification cross entropy loss function to obtain the total loss function.

[0015] Iteratively train the deep learning model and update the model parameters. Model training includes training phase, verification phase and testing phase. The operations of each phase are as follows: Training phase: With the training set as the core, the network parameters are continuously updated through iterative optimization algorithms to minimize the total loss function; Specifically, the mini-batch gradient descent method is used. Each time, a batch of samples is randomly selected from the training set and input into the constructed deep learning network. After forward propagation calculation, the predicted health status probability distribution and degradation score are obtained. The loss value of the current batch is then calculated based on the total loss function. During the backpropagation process, the gradient of the total loss function with respect to the parameters of each layer of the network is calculated, and the Adam optimizer is used to update the parameters and set the learning rate. Validation phase: After each complete traversal of the training set, the model performance is evaluated using the validation set. Input validation set samples into the model after the training phase, calculate the validation index, and dynamically adjust the training strategy based on the validation index. Use the early stopping strategy to prevent overfitting. If the validation set loss does not decrease for multiple consecutive traversal cycles, terminate the training early. Testing phase: Input the test set samples into the model that has passed the validation phase to obtain prediction results and measure the model's accuracy and adaptability to unknown data.

[0016] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects: The present invention adopts an improved adaptive noise complete set empirical mode decomposition and dynamic spectral kurtosis fusion method. Through the adaptive stopping criterion, it can avoid modal aliasing or over-decomposition caused by fixed iterations. At the same time, combined with dynamic spectral kurtosis, it can effectively separate high-frequency impact and low-frequency degradation modes in strong noise environments, avoiding the difficulty in simultaneously capturing high-frequency impact components and low-frequency degradation modes, which leads to insufficient feature extraction and modal aliasing. The present invention simultaneously uses a time-frequency dual-path feature calibration mechanism to utilize time-domain multi-scale convolution and frequency-domain energy modeling to generate an attention mask, enhance local impact and global energy features, avoid the loss of local impact features, and make them insensitive to subtle faults, thereby achieving refined feature recalibration. The present invention calculates feature offsets based on healthy samples and dynamically adjusts channel responses in combination with degradation scores to avoid the blurred boundary between healthy and degraded stage features, which can lead to insufficient sensitivity of the classification model to mid-term degradation states. Therefore, by enhancing the dynamic degradation perception features, it is possible to continuously quantify the gradual process of equipment from health to failure, thereby enhancing sensitivity to mid-term degradation states. The present invention maps feature segments to different frequency bands, combines a gating mechanism to achieve adaptive feature selection, and then corrects the classification probability through the degradation prior distribution. This can avoid the situation where the prediction probability of directly using a multi-granularity feature fusion classifier may not fully reflect the gradual degradation process, thereby leading to deviations in the judgment of the mid-term state, thereby improving the recognition rate and classification stability of mild / moderate degradation.

[0017] In summary, the present invention can perform real-time health detection and degradation assessment of the equipment operation status by adaptively decomposing and reconstructing the features of the collected sample data, and by constructing and training a deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0019] Figure 1Schematic diagram of the method of the present invention.

[0020] Figure 2 Schematic diagram of the training process for building a deep learning model.

[0021] Figure 3 This is a comparison chart of vibration signals of large port equipment in different health states.

[0022] Figure 4 A comparison chart of feature extraction effects of different decomposition methods.

[0023] Figure 5 A comparison chart of the effects of the adaptive time-frequency decomposition method of the present invention and the conventional short-time Fourier transform in processing non-stationary vibration signals.

[0024] Figure 6 The following is a comparison of the classification accuracy of different models for five key health conditions.

[0025] Figure 7 This is the feature space distribution diagram of different health states extracted by the method of the present invention.

[0026] Figure 8 This is a trend chart of score changes for different methods during the equipment degradation accelerated test. DETAILED DESCRIPTION

[0027] In order to clearly illustrate the technical features of this solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0028] Example 1 like Figure 1 As shown in FIG, a model training method for health status monitoring of large-scale port equipment includes the following steps: S1. Collect and label the inspection data of large-scale port equipment, and divide the labeled sample data into training set, validation set and test set in chronological order; S2. Adaptively decompose and reconstruct the collected sample data in time and frequency, specifically using an improved adaptive noise complete set empirical mode decomposition and dynamic spectral kurtosis fusion method to generate a high-dimensional decomposition and reconstruction matrix and then reduce the dimension to a decomposition and reconstruction matrix; S3, build a deep learning model, input the sample data in the training set, validation set and test set after processing in step S2 into the model to train the model, and obtain a trained model, such as Figure 2 As shown, the model includes the following operations: Time-frequency dual-path feature recalibration: The decomposed and reconstructed matrix generates an attention mask through the time-frequency dual-path feature calibration mechanism. After fusion, the decomposed and reconstructed matrix is ​​adaptively recalibrated to generate a calibrated feature matrix. Cascaded hole depth feature extraction: The calibrated feature matrix is ​​passed through the cascaded hole depth convolution module to extract multi-scale features, which are then spliced ​​to generate multi-scale fusion features; Dynamic degradation-aware feature enhancement: Calculates feature offsets and degradation scores based on multi-scale fusion features, then dynamically adjusts channel responses to generate degradation-aware features. Gated multi-granularity feature fusion classification: A gated multi-branch structure is used to segment and fuse degradation-aware features to generate a predicted health status probability distribution result. Degradation perception probability correction: Generate a prior distribution based on the degradation score, correct the predicted health state probability distribution results, and generate the 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; S4. Real-time collection of large-scale equipment inspection data at the port, after processing in steps S1 and S2, is input into the trained model, and the health status probability distribution result is output. The current health status category of the equipment is determined based on the maximum value of the probability distribution.

[0029] In a specific embodiment, S1 is as follows: Systematically collect and annotate monitoring data. Vibration acceleration sensors installed on key parts of large equipment such as bridge cranes and bucket wheel stackers at ports, such as motor bearings, gearboxes, and travel mechanisms, collect vibration signals of the equipment in real time at a fixed sampling frequency. At the same time, in order to construct the training set required for supervised learning, each sample needs to be labeled with the health status category. The labeling process is carried out by equipment maintenance experts based on the actual operating status of the equipment and historical maintenance records. The equipment health status is divided into five categories: healthy, slightly degraded, moderately degraded, severely degraded, and failed.

[0030] The labeled samples are divided into training set, validation set and test set in chronological order. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation.

[0031] Optionally, to enhance the generalization ability of the model, the training set can also adopt data enhancement techniques, such as adding random Gaussian noise, time stretching, etc., to simulate the complex noise environment of the port site.

[0032] like Figure 3The figure shows the vibration signal characteristics of large-scale port equipment in different health states. The sampling frequency is set to 10kHz, and the collected original vibration signal is a continuous time domain waveform. Each sample is intercepted with a fixed length of 4096 sampling points, corresponding to the vibration response of the equipment within 0.4096 seconds, forming the original vibration signal vector. Figure 3 The horizontal axis is time in seconds, and the vertical axis is vibration amplitude in meters per second squared. Depend on Figure 3 It can be seen that the vibration signal waveform in a healthy state is relatively regular and smooth, mainly showing a regular vibration composed of the equipment's fundamental frequency and its harmonics, and the background noise amplitude is relatively small; As the degradation level increases, the waveform begins to change significantly: sporadic, small-amplitude shock pulses begin to appear in the mildly degraded state; in the moderately degraded state, not only do the amplitude and number of shock pulses increase, but amplitude modulation can also be observed in the overall signal, that is, the envelope of the waveform begins to show periodic fluctuations; in the severely degraded state, strong, periodic shock pulses become the most prominent feature; and in the failure state, the signal has completely lost its regularity, showing strong non-stationary characteristics, dominated by large-amplitude random shocks and noise, and accompanied by obvious low-frequency drift.

[0033] In a specific implementation manner, S2 is specifically as follows: Vibration data of port equipment faces the challenges of strong noise interference and multi-scale impact characteristics in health status detection. Conventional processing methods such as short-time Fourier transform use a fixed window length, which makes it difficult to simultaneously capture high-frequency impact components and low-frequency degradation modes, resulting in insufficient feature extraction and modal aliasing.

[0034] The present invention adopts an improved adaptive noise complete set empirical mode decomposition and dynamic spectrum kurtosis fusion method to achieve multi-scale adaptive decomposition and reconstruction of vibration signals. The specific steps are as follows: 1) Improved adaptive noise decomposition: An improved adaptive noise complete ensemble empirical mode decomposition (EMD) is performed on the original vibration signal. Before the improvement, the conventional adaptive noise complete ensemble EMD algorithm uses a fixed stopping criterion, such as a fixed number of iterations or a fixed standard deviation threshold, which may lead to mode aliasing or over-decomposition. The improved adaptive noise complete set empirical mode decomposition algorithm of the present invention adopts an adaptive stopping criterion, that is, the stopping criterion is defined as the change in the standard deviation of two adjacent iterations is less than the adaptive threshold, that is, ; Specifically, in the mth iteration, the standard deviation of the current eigenmode function and the standard deviation of the previous iteration are calculated. If , then stop the iteration, otherwise continue the iteration to suppress the mode aliasing; The original vibration signal vector is defined as , the dimension is 4096, representing the vibration amplitude of the equipment sampling point; At the same time, the k-th order eigenmode function is defined as , characterizes the kth order intrinsic mode function obtained after decomposition by the improved adaptive noise complete set empirical mode decomposition algorithm, which is used to characterize the oscillation modes of different time scales in the signal; Where k is the index of the intrinsic mode function, ranging from 1 to K; K is the total order of the intrinsic mode function, with a value of 8; is the length of the original vibration signal vector, ; is the number of iterations when extracting each eigenmode function; represents the standard deviation of the mth iteration; Indicates the The standard deviation of the iterations; For logarithmic functions, the default base is a natural constant; It should be noted that, similar to the conventional adaptive noise complete set empirical mode decomposition algorithm, the iterative process includes calculating the upper and lower envelopes of the signal, finding the mean, and subtracting the mean until the stopping criterion is met.

[0035] 2) Dynamic spectral kurtosis calculation: For each order intrinsic mode function, the fourth power of the intrinsic mode function is converted into the frequency domain amplitude through Fourier transform, and the quadratic power of the intrinsic mode function is converted into the square of the frequency domain amplitude. Then the ratio of the two is calculated to obtain the dynamic spectrum kurtosis value, which quantifies the intensity of the impact characteristics of the intrinsic mode function in the frequency domain and is expressed as: , Where, represents Fourier transform, which is used to convert signals from the time domain to the frequency domain; It represents the dynamic spectrum kurtosis value of the k-th order eigenmode function at frequency f, which is used to characterize the intensity of the impact feature at this frequency; represents the fourth power of the k-th order eigenmode function; represents the quadratic value of the k-th order eigenmode function; It should be noted that The term characterizing the intrinsic mode function amplifies the transient impact component and high peak value in the signal through the fourth power operation, making it more significant in the frequency domain. The power spectrum of the signal is obtained by quadratic operation based on the term characterizing the intrinsic mode function, which is used for normalization processing. The ratio of the two is the dynamic spectrum kurtosis value, which can effectively quantify the intensity of the impact feature in the frequency domain. 3) Adaptive threshold filtering reconstruction: Based on the dynamic spectrum kurtosis of each order eigenmode function, the average dynamic spectrum kurtosis of each order is calculated and expressed as: , Where, is the number of frequency points; is the k-th order average dynamic spectrum kurtosis; It should be noted that the summation operation Only covers from 0 to The positive frequency range, is the sampling frequency. Since the length of the original vibration signal is 4096, and the Fourier transform results in 4096 frequency points including positive and negative frequencies, but in view of the spectrum symmetry of the real signal, only positive frequency points are usually used in actual calculations to avoid redundancy. Therefore, the number of frequency points is The value is 2049.

[0036] Then, the adaptive threshold is calculated based on the mean and standard deviation of the dynamic spectrum kurtosis of this order in the training set, and the adaptive threshold corresponding to the k-th order intrinsic mode function is calculated. , which is calculated from the mean and standard deviation of the kurtosis of the order spectrum in the training set, and is used to distinguish significant shocks from noise, expressed as ; in, is the mean value of the dynamic spectrum kurtosis value of the k-th order eigenmode function at frequency f in the training set; is the standard deviation of the dynamic spectrum kurtosis value of the k-th order eigenmode function at frequency f in the training set; Next, the difference between the average dynamic spectrum kurtosis and the adaptive threshold is processed by the sign function to obtain the weight, and then combined with the intrinsic mode function to construct a high-dimensional decomposition and reconstruction matrix, which is expressed as: , Where, Reconstruct the matrix for high-dimensional decomposition, the dimension is ; is the K-th order eigenmode function, is the first-order eigenmode function, is the second-order eigenmode function; is the K-th order average dynamic spectrum kurtosis, is the first-order average dynamic spectrum kurtosis, is the second-order average dynamic spectrum kurtosis; is a sign function, which outputs +1 when its input is greater than 0, retaining the intrinsic mode function component, and outputs -1 when its input is less than 0, suppressing the intrinsic mode function component; is the adaptive threshold corresponding to the K-th order intrinsic mode function, is the adaptive threshold corresponding to the first-order intrinsic mode function, is the adaptive threshold corresponding to the second-order intrinsic mode function; is the transpose operation; Finally, the high-dimensional decomposition and reconstruction matrix is ​​reduced in dimension. The dimensionality reduction method can be adopted by segmented averaging or downsampling to obtain dimensional feature matrix, which is defined as the decomposition and reconstruction matrix , the dimension is ; It should be noted that the spectral kurtosis distribution law of each order intrinsic mode function is learned through the training set statistics to make the adaptive threshold It can adapt to the noise floor of different devices and build a high-dimensional decomposition and reconstruction matrix by combining the hard decision mechanism of the sign function. , which not only retains the significant impact component but also completely suppresses the dominant component of noise, can solve the compatibility problem between high-frequency impact and low-frequency wear characteristics under strong background noise in ports. The high-frequency impact component is compressed and enhanced in the time domain, while the low-frequency wear pattern has a more concentrated energy distribution in the frequency domain. like Figure 4 As shown in the figure, a comparative analysis of the feature extraction effects of different decomposition methods is carried out to verify the feature extraction ability of the improved adaptive noise complete set empirical mode decomposition algorithm proposed in the present invention in a strong noise environment; the experiment compares the performance of conventional empirical mode decomposition, set empirical mode decomposition, adaptive noise decomposition and the method of the present invention in reconstructing the signal-to-noise ratio index; from Figure 4 It can be seen that the horizontal axis represents the four decomposition methods, and the vertical axis is in decibels (dB), which represents the signal-to-noise ratio of the reconstructed signal to the original signal; The column height corresponding to the method of the present invention is significantly higher than that of the other three methods, and the error bar range is the smallest, indicating that its reconstruction signal-to-noise ratio mean is the highest and the stability is the best. Figure 4 The middle scatter points represent the original measurement values ​​of 5 independent repeated experiments. The scatter point distribution shows that the data points of the method of the present invention are concentrated in the high signal-to-noise ratio area, while the data points of the conventional method are scattered in the lower range of 3-4dB, indicating that the adaptive stopping criterion and dynamic spectral kurtosis fusion mechanism adopted by the present invention effectively suppress the modal aliasing phenomenon in the strong noise environment of the port, so that the high-frequency impact components and low-frequency degradation modes are more completely preserved.

[0037] like Figure 5 As shown in the figure, the adaptive time-frequency decomposition results are analyzed to compare the effects of the adaptive time-frequency decomposition method proposed in the present invention and the conventional short-time Fourier transform in processing non-stationary vibration signals. The experiment uses a signal in a moderately degraded state as input, and the processing results of the two methods are shown from top to bottom. Figure 5The upper middle image shows the result of a conventional short-time Fourier transform (STFT), which uses a fixed time window. While the frequency resolution in the low-frequency region (bottom of the image) is acceptable, the temporal resolution in the high-frequency region (top of the image) is insufficient. This causes the high-frequency transient impulse components (the yellow vertical lines in the high-frequency region) to appear blurred and diffuse in time, with unclear details. Figure 5 The lower middle figure shows the adaptive time-frequency decomposition result of the present invention. It can be observed that the time-frequency energy distribution of the signal is more concentrated and clear across the entire time-frequency plane. In particular, in the high-frequency region, the transient component representing the critical fault impact can be more accurately located at the time of occurrence. At the same time, good frequency aggregation is maintained throughout the entire analysis frequency band, indicating that the present invention can dynamically adjust the analysis parameters according to the local characteristics of the signal, thereby more effectively separating the impact component hidden in strong background noise.

[0038] In a specific implementation, the time-frequency dual-path feature recalibration is as follows: Since the characteristic energy distribution of vibration signals undergoes nonlinear shifts during equipment degradation, conventional channel attention mechanisms use global pooling operations, which lose local impact features and make them insensitive to subtle faults. Therefore, the present invention generates an attention mask through a time-frequency dual-path feature calibration mechanism and adaptively recalibrates the features. The specific steps are as follows: 1) Temporal Attention Mask Generation: The depth-wise separable one-dimensional convolution operation is used to extract the multi-scale temporal features of the decomposed and reconstructed matrix. The convolution output is then processed by a normalized exponential function to generate a temporal attention mask, thereby enhancing the response of the key local impact area, which is expressed as: , Where, Denotes a depth-wise separable one-dimensional convolution operation with kernel sizes of 3, 5, and 7, and dilation rates of 1, 3, and 5, respectively, for extracting multi-scale temporal features. is the normalized exponential function, The term is used to map the convolution output into an attention weight in the form of a probability distribution; Represents a temporal attention mask for enhancing the response of key local impact regions.

[0039] 2) Frequency domain attention mask generation: The decomposition and reconstruction matrix is ​​converted to the frequency domain through Fourier transform, and the frequency domain amplitude spectrum is obtained by taking the modulus. Then, the global energy distribution of the frequency domain features is learned using a multi-layer perceptron. The output is then converted back to the time domain through inverse Fourier transform to generate a frequency domain attention mask that reflects the global energy distribution characteristics, which is expressed as: , Where, Express Modulo operation, used to obtain the frequency domain amplitude spectrum; Represents a multilayer perceptron, which consists of two hidden layers, each with a dimension of 64, used to learn the global energy distribution of frequency domain features, and an output layer with a dimension of 512; represents the inverse Fourier transform, which is used to convert the frequency domain weights back to the time domain form; represents the frequency domain attention mask, which is used to reflect the global energy distribution characteristics in the vibration signal; 3) Feature recalibration output: The time domain attention mask and the frequency domain attention mask are fused through a broadcast addition operation, and then the hyperbolic tangent function is used to constrain the fused weight range. Finally, the calibration is performed by element-by-element multiplication and decomposition and reconstruction matrix, and the calibration feature matrix is ​​output, which is expressed as: , Where, represents the broadcast addition operation, which is used to align the time domain attention mask and the frequency domain attention mask by dimension and add them together; is the hyperbolic tangent activation function, The term is used to constrain the fusion weights to be within the interval as much as possible. Internally, prevent over-amplification of features; represents the Hadamard product, that is, element-wise multiplication; Represents the calibration feature matrix, with dimension ; It should be noted that the range of the hyperbolic tangent activation function is But due to The function input value of the term is usually not extreme, and the actual output is mostly concentrated in range, thereby effectively preventing the feature from being over-amplified; the broadcast addition operation automatically expands the tensor with smaller dimensions to a tensor with larger dimensions, such as the time domain attention mask The dimension is , while the frequency domain attention mask The dimension is , then the frequency domain attention mask Will be copied along the first dimension times, the dimension becomes , and then implement the temporal attention mask Element-wise addition operation; temporal attention mask Using multi-scale depth-wise separable convolution, its layered receptive field unexpectedly captures the unique impact transmission attenuation law of port equipment. Small convolution kernels focus on the instantaneous impact of gear meshing, large expansion rate convolution perceives the long-period resonance caused by bearing wear, and frequency domain attention mask Frequency-to-time domain conversion is achieved through Fourier transform and inverse Fourier transform. At the same time, the multi-layer perceptron structure adopts a bottleneck design with a hidden layer dimension of 64, forcing the network to learn the compressed representation of frequency domain energy. When local gear fracture and overall bearing wear coexist, the time domain path enhances the fracture impact pulse, while the frequency domain path amplifies the bearing characteristic frequency sidebands, which can effectively improve the separability of composite fault features.

[0040] In a specific implementation, the cascaded hole depth feature extraction is as follows: Since conventional convolutional neural networks have a large number of parameters in vibration signal processing and are insensitive to high-frequency features, it is difficult to capture multi-scale impact features, resulting in low feature extraction efficiency and limited receptive field. Therefore, the present invention uses cascaded hole deep convolution modules to extract multi-scale features, expand the receptive field and reduce the number of parameters. The specific steps are as follows: 1) Feature initialization: The calibration feature matrix is ​​used as the input of the cascaded hole depth convolution module to initialize the input features of the first level convolution, which is expressed as ; in, Represents the initial input features of the l-th layer network; The network layer index of the cascaded dilated depth convolution module; 2) Cascaded dilated convolution processing: The input features are processed using three-stage cascaded dilated depthwise separable convolutions. Each stage includes dilated depthwise separable one-dimensional convolution, batch normalization, and Gaussian error linear unit activation function operations, and the output features of each processing stage are gradually extracted, which can be expressed as: , Where, is the index of the processing stage, with a value range of 1, 2, and 3; The void ratio is The hole depth of can separate the one-dimensional convolution; The dilation rate is used to control the expansion degree of the convolution kernel, and its values ​​are 2, 4, and 2 respectively; Represents the batch normalization operation, which is used to accelerate training and improve stability; is the Gaussian error linear unit activation function; Indicates the first layer of the network - output features of 1 processing stage, , corresponding to the initial input features of the l-th layer network ; Indicates the first layer of the network Output features of each processing stage; It should be noted that the processing stage refers to three consecutive processing steps within the cascaded dilated depth convolution module. Each stage includes dilated depth separable one-dimensional convolution, batch normalization, and Gaussian error linear unit activation function activation operations; the dilation rate takes values ​​of 2, 4, and 2, respectively. The first layer dilation rate of 2 captures short-period shocks such as gear meshing, the middle layer dilation rate of 4 senses the resonant frequency band of the device structure, and the last layer drops back to 2 to restore the feature resolution. This "expansion-contraction" structure forms a time-frequency focusing capability similar to wavelet analysis. 3) Multi-scale feature stitching: The output features of the three-stage processing stage are spliced ​​along the channel dimension to form a multi-scale fusion feature, thereby expanding the receptive field and covering the impact period in the vibration signal, which is expressed as: , Where, Represents the concatenation operation of three feature sub-segments along the channel dimension; Represents the output features of the first processing stage of the l-th layer network; Represents the output features of the second processing stage of the l-th layer network; Represents the output features of the third processing stage of the l-th layer network; Represents multi-scale fusion features with a channel dimension of 512.

[0041] The specific implementation method of the dynamic degradation perception feature enhancement is as follows: Since 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 characteristics is blurred, resulting in insufficient sensitivity of the classification model to the mid-term degradation state. Therefore, the present invention dynamically adjusts the feature response to perceive the degradation process by calculating the feature offset and degradation score. The specific steps are as follows: 1) Feature offset calculation: Extract the baseline feature vector from the healthy samples in the training set, and subtract the multi-scale fusion feature of the current sample from the baseline feature vector to obtain the feature offset, which quantifies the degree of deviation from the healthy state and is expressed as: , Where, Represents the feature offset, which is used to quantify the degree of deviation of the current sample feature from the healthy state; Represents the baseline feature vector of health status, which is obtained by calculating the feature mean of the first 5% healthy samples in the training set; Specifically, the top 5% healthy samples in the training set are selected based on the amplitude variance or signal-to-noise ratio sorting. Specifically, the samples in the training set are sorted by the amplitude variance or signal-to-noise ratio of the vibration signal, and the top 5% of the samples with the smallest variance or the highest signal-to-noise ratio and labeled as healthy are selected. 2) Degradation score generation: The feature offset is globally averaged and pooled to obtain a global vector, which is then input into a multi-layer perceptron to learn the degradation trajectory. The output degradation score is used to quantify the degree of device degradation and is expressed as: , Where, Degradation score, ranging from 0 to 1, is used to quantify the degree of degradation of the current state of the device; Represents the global average pooling operation, which is used to compress the feature offset into a global vector of fixed dimension; Represents a three-hidden-layer multilayer perceptron with hidden layer dimensions of 128, 64, and 32, respectively. It is used to learn the nonlinear mapping from feature offset to degradation degree. Its output layer uses the Sigmoid activation function to convert the degradation score The constraint is between 0 and 1, where 0 represents perfect health and 1 represents complete failure; It should be noted that by constructing the health status benchmark feature vector As the "zero drift anchor point" of the feature space, the multi-layer perceptron adopts a dimensionality reduction structure design with hidden layer dimensions of 128, 64, and 32 respectively. Its bottleneck layer forces the network to extract the core degradation factors such as bearing wear and gear backlash, and output the degradation score Have clear physical meaning; 3) Dynamic adjustment of channel response: Based on the degradation score and the learnable adaptive weight vector, the multi-scale fusion feature is scaled and adjusted in the channel dimension to obtain the degradation-aware feature, which enhances the response to the degradation-sensitive channel, expressed as: , Where, is the degradation perception feature, which means 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 adaptation weight vector, with dimension , is a trainable parameter; Multi-scale fusion features The number of channels.

[0042] In a specific implementation, the gated multi-granularity feature fusion classification is as follows: In the classification of port equipment health status, different degradation modes correspond to different characteristic segments of the vibration signal. A single fully connected classifier cannot effectively capture local subtle abnormal features. Moreover, the importance of features in different degradation stages varies. Global uniform processing can easily reduce the differentiation between mid-term degradation states and local impact features. Therefore, the present invention adopts a gated multi-branch structure to segment and fuse features to enhance classification sensitivity. The specific steps are as follows: 1) Uniform segmentation of features: The degraded perceptual features are evenly divided into four feature sub-segments along the channel dimension, expressed as: , Where, The first feature sub-segment representing the degradation perception feature is It is evenly divided into 4 sub-segments along the channel dimension, and each sub-segment dimension is 128 dimensions. Similarly, The second feature sub-segment representing the degradation perception feature, The third feature sub-segment representing the degradation perception feature, The fourth feature sub-segment representing the degradation perception feature, each feature sub-segment has a dimension of 128; Represents the concatenation operation of 4 feature sub-segments along the channel dimension; 2) Gating weight generation: For each feature sub-segment of the degraded perceptual feature, a nonlinear transformation is performed through the fully connected layer, and then the gate weight is calculated using the Sigmoid activation function to evaluate the importance of the feature sub-segment to the classification decision, which is expressed as: , Where, The first one represents the degraded perceptual feature feature sub-segments; The feature sub-segment index takes values ​​of 1, 2, 3, or 4. is the Sigmoid activation function; Indicates the The gating weight vector corresponding to each feature sub-segment has a value range of 0 to 1 and is used to regulate the contribution of the feature of the segment; represents a fully connected layer with an output dimension of 128, which is used to perform nonlinear transformation on the features of each sub-segment; 3) Gated transform feature calculation: For each feature sub-segment, dimensionality reduction and abstraction are performed through the fully connected layer, and then the rectified linear unit activation function is used to process it. Combined with the gating weight vector, the gated transformation feature is obtained to achieve feature selection and enhancement, which is expressed as: , Where, Indicates the The gated transformation features of feature sub-segments; Represents a fully connected layer with an output dimension of 64, used for feature dimensionality reduction and abstraction; Represents the rectified linear unit activation function, which is used to enhance the nonlinear expression ability of the model; Represents the Hadamard product, that is, element-by-element multiplication, which is used to achieve the scaling of the gating weight to the eigenvalue; 4) Multi-branch fusion classification: The gated transformation features of each feature sub-segment are mapped to category scores through a fully connected layer, and then the category scores of all branches are summed and normalized by a normalized exponential function to obtain the predicted health state probability distribution, which is expressed as: , Where, Represents the predicted health state probability distribution, with a dimension of 5, corresponding to the five states of healthy, slightly deteriorated, moderately deteriorated, severely deteriorated, and failed; represents a fully connected layer with an output dimension of 5, which is used to map the features of each branch into category scores; It should be noted that the 512-dimensional degraded perception feature The system is evenly segmented into four 128-dimensional feature subsegments, not simply segmented. Its physical meaning corresponds to the frequency band division of the vibration signal. Each subsegment independently performs feature selection using a gated weight vector. For example, a broken gear fault primarily activates the high-frequency subsegment, while bearing wear dominates the mid-frequency subsegment. The sigmoid function outputs a continuous value between [0, 1] rather than a hard threshold, preserving the gradual variation of feature contribution. Furthermore, the gated transform feature calculation uses a multi-branch score summation approach, enabling classification decisions to incorporate multi-band evidence, improving the recognition rate of moderately degraded images. When multiple subsegments give weak positive signals, their joint judgment is more reliable than that of a single fully connected layer.

[0043] In a specific implementation, the degradation perception probability correction is as follows: In the classification of port equipment health status, the prediction probability directly using a multi-granularity feature fusion classifier may not fully reflect the gradual degradation process, resulting in a bias in the judgment of the mid-term status. Therefore, the present invention generates a prior distribution based on the degradation score and corrects the classification probability to reflect the degree of degradation. The specific steps are as follows: 1) Calculation of degradation prior distribution: Based on the degradation score, the prior probability distribution is generated by an exponential decay function. The center position of the distribution moves linearly with the degradation score, tending to a higher degradation state as the degradation score increases, thereby reflecting the degree of degradation, which is expressed as: , Where, Indicates the The prior probability of the health state of the equipment is 1, 2, 3, 4, and 5, which correspond to the five states of health, slight degradation, moderate degradation, severe degradation, and failure, respectively. is the device health status category index that is different from j; is the sharpness parameter, which controls the concentration of the prior distribution and takes a value of 10; is a scaling parameter that maps the degradation score to a category index range, preferably with a value of 4 such that , corresponding to category indexes 1 to 5; represents the natural exponential function; It should be noted that the prior distribution The center position of Linear movement, when When the center is (Health), when When the center is (failure), thus reflecting the degree of deterioration.

[0044] 2) Probability correction: The prior distribution and the predicted probability are fused through geometric mean to obtain the corrected probability, thereby incorporating degradation knowledge into the classification decision and improving the classification performance of the medium-term degradation state, which is expressed as: , Where, Indicates the Corrected probability of health-like state; Indicates the Corrected probability of health-like state; Indicates the Prior probability of class health status; It is a calibration intensity parameter used to control the influence of the prior distribution, with a value ranging from 0 to 1 and a value of 0.3.

[0045] In a specific implementation, the loss function is calculated as follows: 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, and the degradation consistency loss term is calculated based on the degradation score and the true degradation degree, ensuring that the degradation score is consistent with the true degradation degree. It is expressed as: , Where, is the loss function; is the cross entropy loss term, which is calculated as follows: , For the One-hot encoding of the true label of the class; is the degradation consistency loss term, and its calculation method is expressed as: , is the loss weight parameter, which is used to balance the two losses and has a value of 0.5; is the true degree of degradation, according to the true label mapping, if the true label is healthy, then , slightly deteriorated, then , moderate degradation, then , severe degradation, then , fails, then .

[0046] In a specific implementation, the deep learning model is iteratively trained and the model parameters are updated. The model training includes a training phase, a verification phase, and a testing phase. The operations in each phase are as follows: Training phase: With the training set as the core, the network parameters are continuously updated through iterative optimization algorithms to minimize the total loss function; Specifically, the mini-batch gradient descent method is used. Each time, a batch of samples is randomly selected from the training set. The batch size can usually be 32 or 64. These samples are input into the constructed deep learning network. After forward propagation calculation, the predicted health state probability distribution and degradation score are obtained. The loss value of the current batch is then calculated according to the total loss function. The loss function consists of a weighted classification cross entropy loss and degradation consistency loss. The cross entropy loss measures the difference between the predicted probability and the true label, and the degradation consistency loss ensures the consistency of the degradation score with the actual degradation degree. During the backpropagation process, the gradient of the total loss function with respect to the parameters of each layer of the network is calculated. The parameters are updated using the Adam optimizer, and the learning rate is set to 0.001. Validation phase: After each complete traversal of the training set, the model performance is evaluated using the validation set. Input validation set samples into the model after the training phase, calculate the validation index, and dynamically adjust the training strategy based on the validation index. Use the early stopping strategy to prevent overfitting. If the validation set loss does not decrease for multiple consecutive traversal cycles, terminate the training early. Testing phase: Input the test set samples into the model that has passed the validation phase to obtain prediction results and measure the model's accuracy and adaptability to unknown data.

[0047] In a specific implementation manner, S4 is specifically as follows: After model training is complete, it can be used to perform health status detection on new large-scale port equipment vibration data. Specifically, vibration signals are collected from the vibration sensors of the equipment to be tested in real time, and sample segments of 4096 lengths are intercepted using the same preprocessing method. Then, adaptive time-frequency decomposition and feature reconstruction are performed. The data processed by adaptive time-frequency decomposition and feature reconstruction is input into a trained deep learning model. After going through the steps of time-frequency dual-path feature recalibration, cascaded hole depth feature extraction, dynamic degradation perception feature enhancement, gated multi-granularity feature fusion classification and degradation perception probability correction, the health status probability distribution and degradation score of the sample are output. Finally, the current health status category of the equipment is determined according to the maximum value of the probability distribution, including healthy, slightly degraded, moderately degraded, severely degraded or failed, to achieve online real-time health status detection and early warning.

[0048] Example 2 like Figure 6 As shown, the classification accuracy of different models in five health status categories was analyzed to evaluate the classification performance of the five models (support vector machine, random forest, one-dimensional convolutional neural network, long short-term memory network, and the method of the present invention) for the five health statuses of the equipment. The horizontal axis of the grouped bar chart represents the health status category (healthy, slightly degraded, moderately degraded, severely degraded, and failed), and the vertical axis is in percentage (%), representing the classification accuracy. The column of the method of the present invention maintains the highest position in all health status categories, especially in the moderate degradation and slight degradation categories, where its accuracy advantage is most significant. The performance of conventional models declines significantly in the moderate degradation state. However, the method of the present invention effectively captures the weak features of the early stage of equipment degradation through the time-frequency dual-path feature recalibration mechanism and gated multi-granularity feature fusion. Figure 6 It can be seen that the column height of the method of the present invention has the smallest fluctuation between different states, indicating that it has a balanced recognition ability for each degradation stage, solving the problem of insufficient sensitivity of conventional models to mid-term degradation states.

[0049] like Figure 7 As shown, a visual analysis of the feature distribution of different health states is performed, and the distribution pattern of the features extracted by the method of the present invention in two-dimensional space is displayed by t-SNE dimensionality reduction technology; Figure 7 In the figure, samples in five health states are marked with different colors: healthy (blue), slightly degraded (green), moderately degraded (yellow), severely degraded (orange), and failed (red). The horizontal and vertical axes are two feature dimensions, representing the abstract feature space after dimensionality reduction. Figure 7Three key phenomena can be seen: 1) The healthy state (blue) and the failed state (red) are clustered on opposite sides of the space, forming a clear separation; 2) The three degradation states (green, yellow, and orange) exhibit a continuous transition distribution, consistent with the progressive degradation characteristics of equipment; 3) Moderate degradation (yellow) has moderate overlap with adjacent states but is largely separable. These distribution characteristics stem from the adaptive adjustment of the characteristic manifold by the dynamic degradation perception mechanism, enabling the model to accurately quantify the gradual transition from healthy to failed. Figure 7 No abnormal mixed clusters appeared in the data, which proved that the method of the present invention effectively solved the problem of blurred boundaries between healthy and degraded states.

[0050] Example 3 like Figure 8 As shown in the figure, the trend analysis of the score change during the accelerated degradation experiment of the equipment was carried out. Through the accelerated degradation experiment, the change pattern of the degradation score during the 100-day operation of the equipment from healthy to failure was simulated. The horizontal axis of the line chart is the operation time (days), and the vertical axis is the degradation score (range 0-1); Figure 8 It contains three curves: the true degree of degradation (black dotted line), the score of the method of the present invention (cyan solid line), and the score of the conventional method (purple solid line); the background color block indicates the partition threshold of the five types of health status; the experimental results show that: 1) the curve of the present invention (cyan) is highly consistent with the true curve (black), and accurately tracks the trend changes at key nodes such as 30 days (start of mild degradation), 50 days (moderate degradation turning point), and 70 days (severe degradation transition); 2) the conventional method (purple) lags significantly in the early and middle stages (20-60 days) and fails to detect the initial degradation in time; 3) the actual detection points marked by scatter points show that the scoring error of the present invention at the sampling point is smaller; the experimental results show that the degradation consistency loss function and the dynamic feature offset calculation enable the model to perceive the nonlinear feature drift caused by material fatigue, Figure 8 The abnormal decline of the normal curve after 80 days exposes the risk of misjudging compound faults using conventional methods.

[0051] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A model training method for health status monitoring of large-scale port equipment, characterized by: The following steps are involved: S1. Collect and label the inspection data of large-scale port equipment, and divide the labeled sample data into training set, validation set and test set in chronological order; S2. Adaptively decompose and reconstruct the collected sample data in time and frequency, specifically using an improved adaptive noise complete set empirical mode decomposition and dynamic spectral kurtosis fusion method to generate a high-dimensional decomposition and reconstruction matrix and then reduce the dimension to a decomposition and reconstruction matrix; S3. Build a deep learning model, input the sample data in the training set, validation set, and test set processed in step S2 into the model to train the model, and obtain a trained model; S4. Real-time collection of large-scale equipment inspection data at the port, after processing in steps S1 and S2, is input into the trained model, and the health status probability distribution result is output. The current health status category of the equipment is determined based on the maximum value of the probability distribution.

2. The model training method for health status monitoring of large-scale port equipment according to claim 1 is characterized in that: The operations in building a deep learning model are as follows: Time-frequency dual-path feature recalibration: The decomposed and reconstructed matrix generates an attention mask through the time-frequency dual-path feature calibration mechanism. After fusion, the decomposed and reconstructed matrix is ​​adaptively recalibrated to generate a calibrated feature matrix. Cascaded hole depth feature extraction: The calibrated feature matrix is ​​passed through the cascaded hole depth convolution module to extract multi-scale features, which are then spliced ​​to generate multi-scale fusion features; Dynamic degradation-aware feature enhancement: Calculates feature offsets and degradation scores based on multi-scale fusion features, then dynamically adjusts channel responses to generate degradation-aware features. Gated multi-granularity feature fusion classification: A gated multi-branch structure is used to segment and fuse degradation-aware features to generate a predicted health status probability distribution result. Degradation perception probability correction: Generate a prior distribution based on the degradation score, correct the predicted health state probability distribution results, and generate the 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 model training method for health status monitoring of large-scale port equipment according to claim 1 is characterized in that: S2 is as follows: S2.

1. Define the labeled sample data as the original vibration signal vector, perform an improved adaptive noise complete set empirical mode decomposition on the original vibration signal, and adopt an adaptive stopping criterion, which is defined as the change in the standard deviation of two consecutive iterations being less than an adaptive threshold; The eigenmode function obtained after decomposition by the improved adaptive noise complete set empirical mode decomposition algorithm has a total order of K. The k-th order eigenmode function is expressed as , , k represents the index of the eigenmode 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 , then stop the iteration, otherwise continue the iteration to suppress the mode aliasing; in, represents the standard deviation of the mth iteration, Indicates the The standard deviation of the iterations; Represents a logarithmic function, with the default base being a natural constant; Represents the length of the original vibration signal vector; Indicates the number of iterations when extracting each order intrinsic mode function; S2.

2. For each order of intrinsic mode function, convert the fourth power of the intrinsic mode function to the frequency domain amplitude through Fourier transform, and convert the quadratic power of the intrinsic mode function to the square of the frequency domain amplitude. Then calculate the ratio of the two to obtain the dynamic spectrum kurtosis value. S2.

3. Based on the dynamic spectrum kurtosis of each order of the intrinsic mode function, calculate the average dynamic spectrum kurtosis of each order; Then, the adaptive threshold of each order is calculated based on the mean and standard deviation of the dynamic spectrum kurtosis of each order in the training set; The difference between the average dynamic spectrum kurtosis and the adaptive threshold is processed by a sign function to obtain a weight, which is then combined with the intrinsic mode function to construct a high-dimensional decomposition and reconstruction matrix. Finally, the high-dimensional decomposition and reconstruction matrix is ​​reduced in dimension by using the segmented averaging or downsampling method to obtain the decomposition and reconstruction matrix.

4. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: The operation of time-frequency dual-path feature recalibration is as follows: (1) Use depth-wise separable one-dimensional convolution operation to extract multi-scale temporal features of the decomposed reconstruction matrix, and then process the convolution output through a normalized exponential function to generate a temporal attention mask; (2) The decomposition and reconstruction matrix is ​​converted to the frequency domain through Fourier transform, and the frequency domain amplitude spectrum is obtained by taking the modulus. Then, the global energy distribution of the frequency domain features is learned using a multi-layer perceptron. The output is then converted back to the time domain through inverse Fourier transform to generate a frequency domain attention mask. (3) The time domain attention mask and the frequency domain attention mask are fused through a broadcast addition operation, and then the hyperbolic tangent function is used to constrain the fused weight range. Finally, they are recalibrated through element-by-element multiplication and decomposition and reconstruction matrix to output the calibration feature matrix.

5. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: The operation of cascade hole depth feature extraction is as follows: (1) The calibration feature matrix is ​​used as the input of the cascaded hole depth convolution module to initialize the input features of the first level convolution; (2) Use three-stage cascaded dilated depth-wise separable convolution to process input features. Each stage includes dilated depth-wise separable one-dimensional convolution, batch normalization, and Gaussian error linear unit activation function operations, gradually extracting the output features of each processing stage, and finally obtaining the output features of the three-stage processing stage of each layer of the network; (3) The output features of the three-level processing stage are spliced ​​along the channel dimension to form multi-scale fusion features.

6. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: The operations for dynamic degradation perception feature enhancement are as follows: (1) Select healthy sample data from the training set and extract its baseline feature vector. Subtract the multi-scale fusion feature of the current sample data from the baseline feature vector to obtain the feature offset. (2) Perform global average pooling on the feature offset to obtain a global vector, which is then input into a multi-layer perceptron to learn the degradation trajectory and output a degradation score used to quantify the degree of device degradation; (3) Based on the degradation score and the learnable adaptation weight vector, the multi-scale fusion features are scaled and adjusted in the channel dimension to obtain the degradation perception features.

7. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: The operation of gated multi-granularity feature fusion classification is as follows: (1) The degraded perceptual features are evenly divided into four feature sub-segments along the channel dimension, and each feature sub-segment has the same dimension; (2) For each feature sub-segment of the degraded perceptual feature, a nonlinear transformation is performed through the fully connected layer, and then the gate weight is calculated using the Sigmoid activation function; (3) For each feature sub-segment, dimensionality reduction and abstraction are performed through the fully connected layer, and then the rectified linear unit activation function is used to process it, and the gated transformation feature is obtained by combining the gated weight vector of each feature sub-segment; (4) The gated transformation features of each feature sub-segment are mapped to category scores through a fully connected layer, and then the category scores of all branches are summed and normalized by a normalized exponential function to obtain the predicted health status probability distribution.

8. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: The operation of degradation perception probability correction is as follows: (1) Based on the degradation score, a prior probability distribution is generated through an exponential decay function. The center position of the distribution moves linearly with the degradation score. The calculation formula is as follows: ， in, Indicates the The prior probability of the health state of the device; j represents the health state category index of the device, with values ​​of 1, 2, 3, 4, and 5, corresponding to the five states of health, slight degradation, moderate degradation, severe degradation, and failure, respectively; Indicates the health status category index of the device that is 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 through geometric mean to obtain the corrected probability. The calculation formula is as follows: ， in, Indicates the Corrected probability of health-like state; The predicted Class health status probability; The predicted Class health status probability; Indicates the Prior probability of class health status; represents the correction strength parameter used to control the influence of the prior distribution.

9. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: The specific operations for calculating the loss function are as follows: Calculate the classification cross entropy loss function based on the corrected probability and the true label; Calculate the consistency loss function of degradation based on the degradation score and the actual degradation degree; The degraded consistency loss function is weighted and added to the classification cross entropy loss function to obtain the total loss function.

10. The model training method for health status monitoring of large-scale port equipment according to claim 2 is characterized in that: Iteratively train the deep learning model and update the model parameters. Model training includes training phase, verification phase and testing phase. The operations of each phase are as follows: Training phase: With the training set as the core, the network parameters are continuously updated through iterative optimization algorithms to minimize the total loss function; Specifically, the mini-batch gradient descent method is used. Each time, a batch of samples is randomly selected from the training set and input into the constructed deep learning network. After forward propagation calculation, the predicted health status probability distribution and degradation score are obtained. The loss value of the current batch is then calculated based on the total loss function. During the backpropagation process, the gradient of the total loss function with respect to the parameters of each layer of the network is calculated, and the Adam optimizer is used to update the parameters and set the learning rate. Validation phase: After each complete traversal of the training set, the model performance is evaluated using the validation set. Input validation set samples into the model after the training phase, calculate the validation index, and dynamically adjust the training strategy based on the validation index. Use the early stopping strategy to prevent overfitting. If the validation set loss does not decrease for multiple consecutive traversal cycles, terminate the training early. Testing phase: Input the test set samples into the model that has passed the validation phase to obtain prediction results and measure the model's accuracy and adaptability to unknown data.

Images