An intelligent monitoring method for fault diagnosis and prediction of underwater equipment

Abstract

The application discloses an intelligent monitoring method for underwater equipment fault diagnosis and prediction, relates to the technical field of equipment monitoring, and collects underwater equipment acoustic monitoring data, wherein labeled acoustic signal segments, synchronous working conditions and environmental parameter data and corresponding labels jointly constitute a training data set; nonlinear threshold noise reduction is performed on the acoustic monitoring data, and adaptive spectral equalization is performed in combination with the frequency spectrum envelope; a wavelet packet transform and an adaptive weight fusion strategy are adopted to perform multi-scale decomposition on the signal vector after spectral equalization and extract time-frequency features; an underwater equipment fault diagnosis model is constructed and trained; the underwater equipment fault diagnosis model is trained; intelligent monitoring of underwater equipment fault diagnosis and prediction; the nonlinear threshold noise reduction method combining local noise estimation and an adaptive threshold can suppress strong background noise while retaining transient fault characteristics; the adaptive spectral equalization technology based on a mel spectrum and a prior fault probability can dynamically enhance the energy of a fault-related frequency band.

Description

Technical Field

[0001] This invention relates to the field of equipment monitoring technology, and in particular to an intelligent monitoring method for underwater equipment fault diagnosis and prediction. Background Technology

[0002] The reliability and safety of underwater equipment are increasingly critical. However, the underwater environment is characterized by high noise, high pressure, and strong corrosiveness. Equipment operates in enclosed and harsh conditions for extended periods, leading to challenges such as delayed detection, difficulty in locating faults, and inaccurate risk assessments, especially in underwater production systems. Failures not only incur extremely high repair costs but can also trigger severe production stoppages, safety accidents, and even ecological disasters. Therefore, early and accurate fault diagnosis and prediction for underwater equipment, shifting from "reactive maintenance" to "predictive maintenance," has significant economic value.

[0003] Existing technologies have the following drawbacks: conventional bandpass filtering or fixed threshold noise reduction methods are difficult to adapt to the non-stationary characteristics of underwater noise and are prone to losing effective fault frequency components; single-scale time-frequency analysis methods cannot simultaneously capture fault features with different time-frequency resolutions, such as transient impacts and steady-state harmonics in acoustic signals; conventional fault diagnosis models usually rely only on acoustic signals and ignore key contextual factors such as equipment operating conditions and marine environment, resulting in a decrease in generalization ability when operating conditions change; conventional monitoring systems focus more on real-time fault diagnosis and lack continuous analysis of characteristic trends, making it difficult to achieve the leap from fault diagnosis to early prediction. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an intelligent monitoring method for underwater equipment fault diagnosis and prediction.

[0005] The technical solution adopted to solve the above-mentioned technical problems is: an intelligent monitoring method for underwater equipment fault diagnosis and prediction, including the following steps: S1 collects acoustic monitoring data from underwater equipment, divides the collected continuous acoustic signals into segments of fixed length, assigns a specific category label to each sample through manual data annotation, and combines the labeled acoustic signal segments, synchronized operating condition and environmental parameter data, and corresponding labels to form a training dataset.

[0006] S2 performs nonlinear threshold noise reduction on acoustic monitoring data to suppress background noise, and combines it with its spectral envelope for adaptive spectral equalization to balance the energy distribution of different frequency bands and enhance fault characteristics.

[0007] S3 employs a wavelet packet transform and adaptive weight fusion strategy to perform multi-scale decomposition on the spectral equalized signal vector and extract time-frequency features. Based on the discriminative nature of the sub-band features, it dynamically allocates fusion weights to generate fusion features that can characterize both the local details of the signal and its global spectral structure.

[0008] S4, Build and train an underwater equipment fault diagnosis model.

[0009] S5, Training of underwater equipment fault diagnosis model: Using the training dataset constructed in step S1, supervised training is performed on the gated multimodal context fusion network constructed in step S4.

[0010] S6, Intelligent monitoring for fault diagnosis and prediction of underwater equipment: After completing the model training in step S5 and obtaining the optimized fault diagnosis model, it is deployed in the online intelligent monitoring system of underwater equipment to realize real-time diagnosis of equipment operating status and early fault prediction.

[0011] Furthermore, the underwater equipment acoustic monitoring data in S1 includes equipment operating condition data and marine environmental parameters. The equipment operating condition data includes echo intensity and output power, while the marine environmental parameters include the water depth where the equipment is located and the surrounding seawater temperature.

[0012] Furthermore, the types of manual data annotations in S1 include: 1-normal state, 2-minor leak, 3-rupture leak, 4-sealing failure, and 5-pipeline corrosion.

[0013] Furthermore, the adaptive spectral equalization method in S2 includes the following steps: S201, Perform a short-time Fourier transform on the denoised time-domain signal vector to obtain the time spectrum. , Indicates the time frame index Frequency Index The short-time Fourier transform coefficients at that point, Represented as a time frame index, Indicates frequency index.

[0014] S202, will Mapping to the Mel scale, calculate the Mel spectrum. , Indicates the first The Mel frequency band in the first The amplitude of each time frame.

[0015] S203: Calculate the root mean square energy of each Mel band across all time frames based on its original spectral vector. Then, based on the probability of fault characteristics occurring in each band obtained statistically from prior fault data, calculate the equalization intensity factor. Using a reference gain constant as the target, adaptively scale the spectrum of each Mel band using a gain function to obtain the equalized Mel spectral vector. .

[0016] Furthermore, the adaptive spectral equalization method in S2 also includes the following steps: S204, equalizes the Mel spectrum The time-domain signal is reconstructed using inverse Mel transform and inverse short-time Fourier transform to obtain the spectrally equalized signal vector. .

[0017] Furthermore, the wavelet packet transform method in S3 is as follows: For the signal vector after spectral equalization... conduct Layer wavelet packet decomposition and extraction of the first wavelet packet. All layers The characteristics of each subband are obtained Each sub-carrying signal; let the first... Layer, First The wavelet packet coefficient vector of each sub-band is For each sub-band, two types of time-frequency characteristics are calculated: energy characteristics. Entropy characteristics energy characteristics Entropy characteristics By concatenating along the feature dimensions, a preliminary feature matrix for this sub-band is formed. , dimension .

[0018] Furthermore, the adaptive weighting method in S3 is as follows: S301, Calculate the discriminative score of the sub-band: Calculate the mean vector of the preliminary feature matrix of each sub-band on the fault class training sample set and the normal class training sample set respectively, and calculate the mean intra-class variance of the feature of the sub-band on the fault class samples and the mean intra-class variance on the normal class samples. Calculate the discriminative score of the feature of the sub-band based on the ratio of the square of the Euclidean distance between the two mean vectors to the sum of the mean intra-class variances of the two classes.

[0019] S302, Calculate the adaptive fusion weight: Based on the discriminative score of each sub-band feature, the scores of all sub-bands are non-linearly amplified using the natural exponential function, and the amplified results are summed and normalized, that is, an adaptive fusion weight is assigned to each sub-band.

[0020] S303, Weighted Stitching and Linear Projection: The preliminary feature matrix of each sub-band is multiplied by its corresponding adaptive fusion weight to obtain a weighted feature matrix. All weighted feature matrices are stitched together along the feature channel dimension according to the sub-band index to form a high-dimensional composite feature matrix. The composite feature matrix is ​​then linearly transformed by a learnable linear projection matrix to project it onto a fixed, lower-dimensional feature space, thereby obtaining a multi-resolution fusion feature matrix.

[0021] Furthermore, the method for constructing and training the underwater equipment fault diagnosis model in S4 is as follows: S401, collects equipment operating parameters synchronized with acoustic signals: echo intensity. Output power Marine environmental parameters: water depth where the equipment is located The surrounding seawater temperature This constitutes the original operating condition parameter vector. To eliminate the influence of dimensions and adapt to neural network processing, each parameter is standardized and nonlinearly encoded to generate a high-dimensional, learnable working condition context feature vector.

[0022] S402 employs a gated multimodal feature fusion strategy, which fuses the multi-resolution fusion feature matrix with the working condition context feature vector, and outputs the fault state probability through a pre-defined classification network.

[0023] The beneficial effects of the present invention are as follows: (1) The present invention adopts a nonlinear threshold denoising method that combines local noise estimation and adaptive threshold, which can retain transient fault characteristics while suppressing strong background noise.

[0024] (2) The present invention adopts an adaptive spectrum equalization technique based on Mel spectrum and prior fault probability, which can dynamically enhance the fault-related frequency band energy.

[0025] (3) The present invention adopts a multi-resolution time-frequency feature adaptive fusion mechanism based on wavelet packet decomposition and discriminative scoring to realize the optimized weighted fusion of multi-scale fault features.

[0026] (4) The present invention adopts a gated multimodal context fusion network. Through the gating mechanism driven by working conditions and environmental information, it realizes the adaptive fusion of acoustic features and context information, thereby improving the diagnostic robustness under varying working conditions. Attached Figure Description

[0027] Figure 1 This is a flowchart of an embodiment of the intelligent monitoring method for underwater equipment fault diagnosis and prediction of the present invention.

[0028] Figure 2 This is an experimental diagram of the original acoustic signal.

[0029] Figure 3 This is an experimental diagram of the signal after adaptive noise suppression.

[0030] Figure 4 This is the experimental diagram of the original signal spectrum.

[0031] Figure 5 This is the experimental spectrum diagram after spectral equalization.

[0032] Figure 6 This is a graph showing the change of the loss function during model training of the present invention, the non-spectral equalization method, the non-multi-resolution fusion method, and the non-gated fusion method.

[0033] Figure 7 This is a graph showing the impact of the number of wave packet decomposition layers on the final test accuracy of the model. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] like Figure 1 As shown, the intelligent monitoring method for underwater equipment fault diagnosis and prediction in this embodiment includes the following steps: S1 collects acoustic monitoring data from underwater equipment. The collected continuous acoustic signals are segmented into fixed-length segments. Each sample is assigned a specific category label through manual data annotation. The types of manual data annotation include: 1-normal state, 2-minor leakage, 3-rupture leakage, 4-sealing failure, and 5-pipeline corrosion. The labeled acoustic signal segments, synchronized operating condition and environmental parameter data, and corresponding labels together constitute the training dataset, which is divided into training set, validation set, and test set according to a preset ratio for subsequent model training, optimization, and evaluation.

[0036] High-sensitivity underwater acoustic sensors deployed at key locations on the target equipment (interconnectors, central manifolds, watertight compartments, signal acquisition modules, etc.) continuously collect broadband acoustic signals radiated by the equipment during operation. The underwater equipment acoustic monitoring data includes equipment operating condition data and marine environmental parameters. Equipment operating condition data includes echo intensity and output power, while marine environmental parameters include water depth and surrounding seawater temperature. The data acquisition system and equipment monitoring are synchronized with the external marine environmental monitoring network to ensure strict temporal alignment between acoustic signals and equipment operating parameters (such as echo intensity, output power, and oil pressure) and environmental parameters (such as water depth and surrounding seawater temperature). The acquisition process covers the normal operating status of the equipment under various load and pressurized operation combinations, as well as various typical fault states obtained through simulation or historical records, such as: 1-Normal state, 2-Minor leakage, 3-Rupture leakage, 4-Seal failure, 5-Pipeline corrosion.

[0037] S2 performs nonlinear threshold noise reduction on acoustic monitoring data to suppress background noise, and combines it with its spectral envelope for adaptive spectral equalization to balance the energy distribution of different frequency bands and enhance fault characteristics.

[0038] Acoustic monitoring data of underwater equipment is subject to strong environmental noise interference, and the signal spectrum changes dynamically with the operating conditions of the equipment. This causes the marine background noise to mask potential fault characteristics, and the signal energy distribution is uneven under different fault modes. Conventional bandpass filtering or fixed threshold noise reduction methods will lose effective fault frequency components and cannot adapt to the spectral shift under different equipment conditions.

[0039] This embodiment of nonlinear threshold denoising calculates the original acoustic signal based on a local noise estimation factor and an adaptive threshold, enabling denoising processing at each sampling point to suppress background noise while preserving potential fault characteristics. This is expressed as follows:

[0040] In the above formula, Indicates the first The sound pressure amplitude after noise reduction at each sampling point is the time-domain signal vector after noise reduction. The nth element, in Pascals, has its background noise suppressed while potential transient fault characteristics are preserved; This represents the sampling point index, which is an integer and has a value range of 1. .

[0041] The sign function is used to preserve the polarity of the original signal; it outputs a positive value when the input is positive, a negative value when the input is negative, and a zero value when the input is zero. ; This function represents the maximum value and outputs the larger of the two input values. Indicates the first The original sound pressure amplitude at each sampling point, measured in Pascals, is acquired by an acoustic sensor and directly reflects the acoustic signal intensity at that moment.

[0042] Indicates the first The local noise estimation factor for each sampling point is a dimensionless scalar used to dynamically adjust the noise reduction intensity at that sampling point. A larger value indicates stronger noise reduction. The calculation method is expressed as follows: ; Represents the natural exponential function; This represents the sensitivity coefficient, a positive constant, which controls the steepness of the sigmoid function response curve, thus affecting the sensitivity of noise estimation to local energy changes. An example value is shown below. ; It represents a very small positive number.

[0043] Indicates the first The average energy of the signal within a local window centered on a sampling point represents the intensity of the local signal, and is calculated as follows: ; Let represent the half-window length for local energy calculation, which is an integer. The window radius for local energy estimation is defined as . The total length of the window is Example of a value: ; Indicates difference from The index of the sampling points; Indicates the first The original sound pressure amplitude of each sampling point is expressed in Pascals.

[0044] The global average energy of the entire input signal vector represents the overall signal strength, and is calculated as follows: ; Indicates difference from and Sampling point index: Indicates the first The original sound pressure amplitude of each sampling point is expressed in Pascals.

[0045] Indicates the first An adaptive threshold, measured in Pascals, is used to determine the amplitude boundary between noise and the signal. Components with amplitudes below this threshold are considered noise and suppressed. Its value is dynamically calculated based on the local signal statistical characteristics and is expressed as follows: ; This indicates a median operation, which outputs the median value of all elements in the input vector, used to robustly estimate the level of local noise. Indicates the first Centered on a sampling point, with a length of [number] points... The neighborhood vector contains the original sound pressure amplitude values ​​of all sampling points within that neighborhood; Represents the neighborhood half-width, which is an integer and defines the range of the local statistical region. Examples of possible values ​​are: ; This represents a logarithmic function, using the natural constant as the default. As the base; The length of the neighborhood vector is defined as follows: .

[0046] The time-domain signal vector after noise reduction is subjected to a short-time Fourier transform to obtain the time spectrum. The Mel spectrum is calculated, and an adaptive gain function is applied to equalize the Mel spectrum to enhance the energy of the fault characteristic frequency band and suppress irrelevant frequency bands. The time-domain signal is reconstructed through inverse transform. The adaptive spectral equalization method includes the following steps: S201, Perform a short-time Fourier transform on the denoised time-domain signal vector to obtain the time spectrum. , Indicates the time frame index Frequency Index The short-time Fourier transform coefficients at a given point are complex values. Their amplitude (modulus) represents the signal strength of that frequency component at that moment, and their phase represents the initial phase information of that frequency component. Represented as a time frame index, Indicates frequency index.

[0047] S202, will Mapping to the Mel scale, calculate the Mel spectrum. , Indicates the first The Mel frequency band in the first The amplitude of each time frame is a scalar quantity. It compresses high-frequency information and highlights the auditory sensitivity area of ​​mid and low frequencies, thus better representing the perceptual characteristics of sound. This indicates the Mel band index.

[0048] S203, Adaptive Gain Equalization: The root mean square energy of each Mel band across all time frames is calculated based on its original spectral vector. An equalization strength factor is calculated based on the probability of fault characteristics occurring in each band, statistically obtained from prior fault data. Using a reference gain constant as the target, the spectrum of each Mel band is adaptively scaled using a gain function to obtain the equalized Mel spectral vector. , is represented as:

[0049] In the above formula, Represents the equilibrium of the first Each Mel-band spectral vector is the equalized Mel spectrum. The Row vectors are sequences spanning all time frames, with a more uniform energy distribution and enhanced fault-related frequency bands.

[0050] Indicates the first The original spectral vector of each Mel band, spanning all time frames. sequence, that is: This characterizes the energy distribution of the frequency band over time. Indicates the first The Mel frequency band in the first The amplitude of each time frame.

[0051] This represents the reference gain constant, a dimensionless positive constant used to control the overall gain level. An example value is shown below. ; Indicates the time frame index; Represents the total number of time frames; it is an integer and characterizes the temporal resolution of the time spectrum. This represents a very small positive integer, used to prevent the denominator from being zero. Examples of its values ​​are: .

[0052] Indicates the first The equalization intensity factor for each mech band is a dimensionless scalar that controls the degree of gain adjustment in the spectrum of that band. A smaller value indicates a more lenient equalization operation for that band. It is calculated as follows: .

[0053] Represents the logarithmic function. This represents the first fault obtained from prior fault data. The average probability of a significant fault characteristic appearing in a mermaid frequency band is a dimensionless scalar, with a value ranging from... arrive Between these ranges, a higher probability indicates that the frequency band is more important for fault diagnosis; This represents the first fault obtained from prior fault data. The average probability of a significant fault characteristic appearing in a mermaid frequency band is a dimensionless scalar. A dummy variable representing the Mel band index, which is an integer and takes values ​​ranging from... arrive ; This represents the total number of Mel bands, which is an integer. Examples of possible values ​​are: .

[0054] It should be noted that, The calculation process involves operations on each time frame. and Both are vectors spanning time frames, and the radical result of the denominator is the scalar gain factor acting on the entire vector.

[0055] S204, equalizes the Mel spectrum The time-domain signal is reconstructed using inverse Mel transform and inverse short-time Fourier transform to obtain the spectrally equalized signal vector. With a dimension of 8192, it is a time-domain signal after noise suppression and spectral equalization. Its background noise is suppressed and the spectral energy distribution under different fault modes is more balanced, thus highlighting the potential fault characteristics.

[0056] S3 employs a wavelet packet transform and adaptive weight fusion strategy to perform multi-scale decomposition on the spectral equalized signal vector and extract time-frequency features. Based on the discriminative nature of the sub-band features, it dynamically allocates fusion weights to generate fusion features that can characterize both the local details of the signal and its global spectral structure.

[0057] Acoustic signals contain a wealth of fault information from low to high frequencies, but the characteristics of different faults may be manifested on different time scales and frequency resolutions. Conventional methods usually use time-frequency analysis on a single scale, which makes it difficult to capture transient impact characteristics and steady-state harmonic characteristics at the same time.

[0058] The wavelet packet transform method involves performing multi-level wavelet packet decomposition on the spectrally equalized signal vector to cover the entire frequency band, calculating the energy and entropy features for each sub-band signal, and concatenating them to form a preliminary feature matrix. Specifically, the method involves... conduct Layer wavelet packet decomposition and extraction of the first wavelet packet. All layers The characteristics of each subband are obtained The individual carries a signal.

[0059] Let the first Layer, First The wavelet packet coefficient vector of each sub-band is For each sub-band, two types of time-frequency characteristics are calculated: energy characteristics. Entropy characteristics energy characteristics Entropy characteristics By concatenating along the feature dimensions, a preliminary feature matrix for this sub-band is formed. , dimension .

[0060] in, Represents the sub-band coefficient vector The energy characteristics obtained from the calculation are derived from the sum of squares of all elements in the wavelet packet coefficient vector of the sub-band, which characterizes the energy intensity of the signal in the sub-band. Represents the sub-band coefficient vector The entropy feature calculated in the middle is obtained by calculating the Shannon entropy of the amplitude probability distribution of the sub-band coefficient vector, which characterizes the complexity or uncertainty of the sub-band signal. Indicates the first Layer The preliminary feature matrix of each sub-band is composed of energy features. Entropy characteristics The matrix formed by concatenating along the feature dimension has a dimension of ; Indicates the number of time segments.

[0061] The adaptive weighting method in this embodiment is as follows: The discriminative power of each sub-band feature in distinguishing between fault and normal states is evaluated. Based on this, adaptive fusion weights are calculated, and the features of all sub-bands are weighted, concatenated, and linearly projected to obtain a fixed-dimensional fusion feature matrix. The specific steps are as follows: S301, Calculate the discriminative score of the sub-band: Calculate the mean vector of the preliminary feature matrix of each sub-band on the fault class training sample set and the normal class training sample set, respectively. Calculate the mean intra-class variance of the sub-band feature on the fault class samples and the mean intra-class variance on the normal class samples. Based on the ratio of the squared Euclidean distance between the two mean vectors to the sum of the mean intra-class variances of the two classes, calculate the discriminative score of the sub-band feature, expressed as:

[0062] In the above formula, Indicates the first Layer The discriminative score of each subband is a dimensionless scalar. The higher the score, the stronger the ability of the subband feature to distinguish between faulty samples and normal samples. The layer index represents the wavelet packet decomposition; it is an integer, and its value ranges from... To the total number of floors ; The sub-band index is an integer, representing the index at the 1st digit of the sub-band. In the layer, The range of values ​​is from arrive ; It represents a very small positive number.

[0063] Indicates the first Layer The mean vector of the initial feature matrix of a sub-band across all fault class training samples represents the mean vector of the initial feature matrix of that sub-band across all fault class training samples, i.e. This characterizes the center position of the fault class sample on the feature of this sub-band; It represents the average value of the energy feature across all training samples of all fault classes, and is a scalar. The entropy feature represents the average value of the entropy feature across all training samples of the fault classes; it is a scalar.

[0064] Indicates the first Layer The mean vector of the initial feature matrix of a sub-band over all normal class training samples represents the mean vector of the initial feature matrix of that sub-band over all normal class training samples, i.e. This represents the center position of the normal class sample on this sub-band feature; It represents the average value of the energy feature across all normal class training samples, and is a scalar. It represents the average value of the entropy feature across all normal class training samples, and is a scalar. This represents the L2 norm.

[0065] This represents the average intra-class variance of the subband features on the fault class samples. It is a scalar that characterizes the dispersion of features within the fault class. It is obtained by collecting the energy and entropy feature values ​​of all fault class training samples on the subband, calculating the variance of each feature dimension (energy and entropy) separately, and then taking the average of these two variances.

[0066] This represents the average intra-class variance of the feature in the normal class samples. It is a scalar that characterizes the dispersion of features within the normal class. It is obtained by collecting the energy and entropy feature values ​​of all normal class training samples in the sub-band, calculating the variance of each feature dimension (energy and entropy) separately, and then taking the average of these two variances.

[0067] In practical implementation, and This is obtained by collecting the energy and entropy eigenvalues ​​of all fault-type training samples in this sub-band, and calculating the arithmetic mean of these eigenvalues. and It is obtained by collecting the energy eigenvalues ​​and entropy eigenvalues ​​of all normal class training samples on this subband, and calculating the arithmetic mean of these eigenvalues.

[0068] S302, Calculate the adaptive fusion weights: Based on the discriminative score of each sub-band feature, the scores of all sub-bands are non-linearly amplified using the natural exponential function. The amplified results are then summed and normalized, thus assigning an adaptive fusion weight to each sub-band, expressed as:

[0069] In the above formula, Indicates the first Layer The adaptive fusion weight of each subband is a dimensionless scalar. The larger the adaptive fusion weight, the greater the contribution of the subband feature to the final fused feature. Represents the natural exponential function; This represents the temperature coefficient, a positive constant, used to control the sharpness of the weight distribution. The larger the value, the more concentrated the weights are on a few sub-bands with high discriminative scores. Example values ​​are shown below. ; Indicates the first Layer The discriminant score of each band is a dimensionless scalar. The dummy variable representing the layer index of the wavelet packet decomposition is an integer, with a value ranging from... arrive ; This represents a sub-band indexed dummy variable, which is an integer and takes values ​​ranging from... arrive ; Indicates the first Layer The discriminant score of a subband is a dimensionless scalar.

[0070] S303, Weighted Concatenation and Linear Projection: The preliminary feature matrix of each sub-band is multiplied by its corresponding adaptive fusion weight to obtain a weighted feature matrix. All weighted feature matrices are concatenated along the feature channel dimension according to the sub-band index to form a high-dimensional composite feature matrix. This composite feature matrix is ​​then linearly transformed using a learnable linear projection matrix, projecting it onto a fixed, lower-dimensional feature space, thus obtaining a multi-resolution fusion feature matrix, represented as:

[0071] In the above formula, This represents a multi-resolution fusion feature matrix with dimension . It is a fixed-dimensional feature representation obtained after multi-scale decomposition, feature extraction and adaptive fusion. It integrates time-frequency information from different time scales from fine to coarse, and dynamically weights it according to the ability of each sub-band to distinguish fault states, so as to more comprehensively and robustly characterize the operating status of underwater equipment. The time dimension length is equal to the number of time segments in the initial feature matrix. ; This represents the fixed dimension of the fused feature. It is a hyperparameter that represents the number of channels in the final fused feature. It can be set to 128, 256, or 512. Represents a linear projection matrix with dimension . These are trainable parameters responsible for mapping high-dimensional concatenated features to fixed parameters. 3D space, in which It is the total number of feature channels after splicing; Indicates a splicing operation; Indicates the first Layer The adaptive fusion weights of each subband are dimensionless scalars. Indicates the first Layer The preliminary feature matrix of each sub-band is composed of energy features. Entropy characteristics The matrix formed by concatenating along the feature dimension has a dimension of .

[0072] S4, Build and train an underwater equipment fault diagnosis model.

[0073] Multi-resolution acoustic fusion features integrate multi-scale time-frequency information, but the failure modes of underwater equipment are not only related to acoustic features, but also affected by the coupling influence of equipment operating conditions and the external marine environment. Conventional diagnostic models only take acoustic features as input and ignore these key contextual information, which leads to a decrease in the model's generalization ability when operating conditions change, and makes it difficult to distinguish between false anomalies caused by environmental disturbances and real equipment failures.

[0074] This embodiment employs a gated multimodal context fusion network to extract operating condition context feature vectors from equipment sensor logs and marine environmental monitoring data. An attention-based gated fusion module adaptively fuses the acoustic fusion feature matrix with the operating condition context feature vectors. A classifier then outputs the probability distribution of fault categories. The specific steps are as follows: S401, collects equipment operating parameters synchronized with acoustic signals: echo intensity. Output power Marine environmental parameters: water depth where the equipment is located The surrounding seawater temperature This constitutes the original operating condition parameter vector. ,in, Let be the total number of parameters, with an example value of 4. To eliminate the influence of dimensions and adapt to neural network processing, each parameter is standardized and non-linearly encoded to generate a high-dimensional, learnable working condition context feature vector, represented as:

[0075] In the above formula, This represents the feature vector of the working condition context, with dimension . It is a joint representation of the equipment operating state and the environmental state after encoding. Its high-dimensional nonlinear mapping can capture the complex relationship between operating parameters and potential failure modes. This indicates a modified linear unit activation function, introducing nonlinearity into the network; This represents the second-layer encoding weight matrix, with dimension . , are trainable parameters; This represents the first-layer encoding weight matrix, with dimension 1. , are trainable parameters used to map standardized operating condition parameters to the hidden layer space.

[0076] Indicates the first The original operating condition or environmental parameter value is a scalar, for example, including echo intensity. (Unit: Megapascal) Output power (Unit: kilowatts) Water depth where the equipment is located (Unit: meters) Surrounding seawater temperature (Unit: degrees Celsius); This represents the index of the operating condition parameter, which is an integer and has a value range of 1. ; Represents the original operating condition parameter vector The mean vector, with dimension . ; Represents the original operating condition parameter vector The standard deviation vector, with dimension . .

[0077] This represents the first-layer encoding bias vector, with dimension . , are trainable parameters; This represents the number of hidden units in the first coding layer; it is a hyperparameter with an example value of 32. This represents the second-layer encoding bias vector, with dimension . , are trainable parameters; The dimension of the feature vector representing the operating condition context is a hyperparameter that can be set to 32 or 64.

[0078] In practical implementation, the original operating condition parameter vector mean vector and standard deviation vector These are statistics pre-calculated from historical datasets (containing various operating conditions) used for model training. During the deployment phase, these fixed statistics are used to standardize the newly collected operating condition parameter vectors.

[0079] S402 employs a gated multimodal feature fusion strategy, fusing the multi-resolution fusion feature matrix with the operating condition context feature vector, and outputting the fault state probability through a pre-defined classification network, expressed as:

[0080] In the above formula, The fault state probability vector is the fault state probability vector predicted by the model, with dimension 1. , The element Indicates that the sample belongs to the first The predicted probability of each fault category (or normal category); This represents the category index, which is an integer with a value range of 100. ; The total number of categories (including normal state) is 5 in one embodiment, including: 1-normal state, 2-minor leak, 3-rupture leak, 4-sealing failure, and 5-pipeline corrosion. This represents the Softmax function, which normalizes the network output into a probability distribution. This represents the classifier weight matrix, with dimension 1. , are trainable parameters; This indicates the flattening operation, which reshapes a two-dimensional matrix into a one-dimensional vector. This represents the hidden layer dimension of the fully connected layer; it is a hyperparameter with an example value of 128. This indicates a modified linear unit activation function; This represents the weight matrix of the fully connected layer, with dimension 1. , are trainable parameters; This represents the bias vector of the fully connected layer, with dimension . , are trainable parameters; This represents the classifier bias vector, with dimension 1. , are trainable parameters.

[0081] This represents the multimodal fusion feature matrix with dimension . It also includes the device's acoustic characteristics and operating environment context information, enabling a more comprehensive characterization of the device's state under specific operating conditions. The calculation method is expressed as follows: ; Indicated Transpose; This represents the weight matrix for linear transformation of the extended operating condition characteristics, with dimension 1. , are trainable parameters; The bias vector represents the linear transformation of the extended operating condition features, with dimension . , are trainable parameters.

[0082] This represents the expanded working condition context feature matrix, since yes The vector, and It is To perform element-by-element addition on a matrix, it is necessary to... Extending this to the time dimension, it can be represented as: ; express Transpose of; The dimension is A column vector of all 1s.

[0083] This represents the modulated acoustic feature matrix, with dimension 1. It is the result of channel scaling (i.e., weighting) of the multi-resolution fused feature matrix after gating vectors. It represents the acoustic features most relevant to fault diagnosis after being filtered and emphasized according to the current operating context. The calculation method is expressed as follows: ; Represents the gated vector, with dimension . Each element has a value between 0 and 1, used to adaptively control the throughput of each acoustic feature channel. A larger value indicates a more important acoustic feature of the corresponding channel. The calculation method is expressed as follows: ; This represents a sigmoid function, also known as the sigmoid activation function, which compresses the output to... interval; This represents a global average pooling operation, which averages the input matrix along the time dimension (the first dimension). ( Compressed to ; This represents the gate weight matrix, with dimension 1. , are trainable parameters; Indicates a splicing operation; Represents the gated bias vector, with dimension . , are trainable parameters; The weight matrix represents the linear transformation of the acoustic features, with dimension 1. , are trainable parameters; The bias vector represents the linear transformation of the acoustic features, with dimension 1. , are trainable parameters; This represents element-wise multiplication; This represents the multi-resolution fusion feature matrix.

[0084] It should be noted that the gating vector The generation also depends on the acoustic features themselves (through global average pooling). The obtained channel-level statistics and operating condition context features This allows the model to dynamically determine which acoustic feature channels are important for fault diagnosis and which can be suppressed under specific conditions such as current equipment pressure, load, water depth, and temperature, thereby achieving condition-dependent feature selection and fusion.

[0085] S5, Training of underwater equipment fault diagnosis model: Using the training dataset constructed in step S1, supervised training is performed on the gated multimodal context fusion network constructed in step S4.

[0086] The training process aims to optimize the trainable parameters in the network so that the model can accurately map the input multimodal features to the correct fault categories. During training, samples are read in batches from the training set. Each sample contains the multi-resolution fused feature matrix obtained after steps S2 and S3, the working condition context feature vector extracted in step S4, and its true class label.

[0087] After forward propagation, the model outputs a predicted failure state probability vector. The difference between the model's predicted probability distribution and the true label distribution is calculated using a standard cross-entropy loss function, which serves as the loss value for this training batch. The gradient of the loss function with respect to all trainable parameters is calculated using the backpropagation algorithm, and stochastic gradient descent or its variants (such as the Adam optimizer) are used to update the network parameters based on the gradient to minimize the loss function.

[0088] The training process iterates multiple times on the complete training set. To prevent overfitting and determine the optimal stopping point, after each training epoch, the model's performance is evaluated using an independent validation set, calculating the classification accuracy and loss value on the validation set. The stopping criterion is set as follows: when the loss value on the validation set no longer decreases but begins to increase within several consecutive preset training epochs, or when the accuracy no longer shows significant improvement, the model is considered to have reached its optimal generalization ability. At this point, training is stopped, and the model parameters with the best performance on the validation set are saved as the final trained underwater equipment fault diagnosis model.

[0089] S6, Intelligent monitoring for fault diagnosis and prediction of underwater equipment: After completing the model training in step S5 and obtaining the optimized fault diagnosis model, it is deployed in the online intelligent monitoring system of underwater equipment to realize real-time diagnosis of equipment operating status and early fault prediction.

[0090] The intelligent monitoring process is as follows: The system continuously collects raw acoustic monitoring data, equipment operating parameters, and marine environmental parameters of the target equipment in real time. For newly acquired raw acoustic signal segments, adaptive noise suppression and spectral equalization preprocessing are first performed strictly according to step S2. Then, multi-resolution time-frequency feature fusion is performed according to the method in step S3 to generate the corresponding acoustic fusion feature matrix.

[0091] Simultaneously, the synchronously acquired operating condition and environmental parameters are converted into operating condition context feature vectors according to the encoding method in step S4. The real-time generated acoustic fusion feature matrix and the operating condition context feature vectors are input together into the pre-trained gated multimodal context fusion network. The network performs forward propagation, adaptively combining acoustic features and operating condition context information through its internal gating fusion mechanism, ultimately outputting a probability distribution vector representing the current equipment state belonging to various fault categories (including normal state). Based on this probability vector, the diagnostic module typically selects the category with the highest probability as the equipment state diagnostic result for the current moment, and displays or records it in real time.

[0092] For fault prediction, the system not only outputs instantaneous diagnostic results but also continuously records and analyzes the state probability sequence or characteristic trend over a period of time (such as several hours or days). By analyzing the upward trend of the probability of specific fault categories, the slow deterioration of characteristic indicators, or the occurrence of periodic anomalies, combined with the equipment's operating history, the system can issue early warnings, indicating that a certain type of fault may be developing. This achieves an intelligent monitoring upgrade from "fault diagnosis" to "fault prediction," providing decision support for predictive maintenance of equipment.

[0093] The experiment in this embodiment uses the following method: like Figures 2 to 5 As shown, the actual effects of adaptive noise suppression and spectral equalization processing techniques are intuitively demonstrated through time-domain and frequency-domain comparative analysis. The experiment uses acoustic signals containing leakage fault characteristics to compare the changes in the original signal, the signal after noise reduction, and the spectrum before and after processing. The horizontal axis is time in seconds, the sound pressure amplitude is in Pascals, the frequency is in Hertz, and the power spectral density is in decibels.

[0094] like Figure 2 The image shows the time-domain waveform of the original acoustic signal. It can be seen that the signal is severely contaminated by strong noise, completely masking the fault characteristics. For example... Figure 3 As shown, the signal after adaptive noise suppression processing is effectively suppressed, and the periodic impulse characteristics begin to appear. Figure 4 As shown, the original signal spectrum is displayed. The energy in the fault characteristic frequency band (100 to 500 Hz) is submerged by background noise and cannot be clearly identified. Figure 5As shown, the spectrum after spectral equalization is displayed. The energy of the fault characteristic frequency band is significantly enhanced, highlighted in orange, while background noise is suppressed, and the fault-related spectral components become clearly distinguishable. The comparison of the four sub-figures demonstrates the effectiveness of the preprocessing technique of this invention. Adaptive noise suppression can suppress background noise without losing fault characteristics, while spectral equalization can balance the energy distribution of different frequency bands and enhance weak fault characteristics.

[0095] like Figure 6 As shown in one embodiment, the loss function change curves of different methods during model training are analyzed, and the convergence performance of the complete method proposed in this invention is compared with that of several key technical modules that are missing. The horizontal axis represents the number of training rounds, with the unit being rounds, indicating the number of iterations (the first 100 rounds) in which the model completes one full forward and backward propagation on the training set. The vertical axis represents the loss value, which is dimensionless and reflects the cross-entropy error between the model prediction and the true label. The lower the value, the better the model fit. The figure compares four configurations: the method of this invention fully retains the adaptive noise suppression and spectral equalization, multi-resolution time-frequency fusion, and gated multimodal context fusion modules; the spectral equalization-free method omits the spectral equalization step and only performs nonlinear threshold noise reduction; the multi-resolution fusion-free method only uses time-frequency features of a single scale and does not perform wavelet packet multi-scale decomposition and adaptive weight fusion; the gated fusion-free method simply concatenates acoustic features with the working context and does not use a gating mechanism for condition-dependent feature selection.

[0096] Experimental results show that the method of the present invention reduces the loss most rapidly in the initial stage and eventually stabilizes near the minimum value. In contrast, the loss values ​​of the non-spectral equalization and non-multi-resolution fusion methods are always high and fluctuate significantly due to insufficient feature extraction or insufficient noise suppression. The non-gated fusion method, on the other hand, ignores the modulation effect of the working environment on the acoustic features, resulting in a slow convergence process and oscillations in the later stage.

[0097] like Figure 7 As shown, in one embodiment, the impact of the number of wavelet packet decomposition layers on the final test accuracy of the model is analyzed, aiming to determine the optimal decomposition depth in multi-resolution feature fusion. The horizontal axis represents the number of wavelet packet decomposition layers, in layers, indicating the number of layers for multi-level wavelet packet decomposition of the preprocessed signal. Each additional layer doubles the number of sub-bands, and the frequency resolution is also refined accordingly. The vertical axis represents the test accuracy, in percentage, measuring the proportion of samples correctly classified by the model on the independent test set.

[0098] Experimental results show that when the number of layers increases from 1 to 4, the accuracy increases and reaches its peak at 4 layers. However, when the number of layers increases to 5 or 6, the accuracy decreases slightly. This is because shallow decomposition cannot capture high-frequency transient impact features and low-frequency modulation information, while deep decomposition introduces too many detail subbands containing a lot of noise or redundant information, which interferes with the classifier's decision-making.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. An intelligent monitoring method for fault diagnosis and prediction of underwater equipment, characterized in that, Includes the following steps: S1, collect acoustic monitoring data of underwater equipment, divide the collected continuous acoustic signal into segments of fixed length, assign a specific category label to each sample through manual data annotation, and combine the labeled acoustic signal segments, synchronized operating condition and environmental parameter data and corresponding labels to form a training dataset. S2 performs nonlinear threshold noise reduction on acoustic monitoring data to suppress background noise, and combines it with its spectral envelope for adaptive spectral equalization to balance the energy distribution of different frequency bands and enhance fault characteristics. S3 employs a wavelet packet transform and adaptive weight fusion strategy to perform multi-scale decomposition on the spectral equalization signal vector and extract time-frequency features. Based on the discriminative nature of the sub-band features, it dynamically allocates fusion weights to generate fusion features that can characterize both the local details of the signal and its global spectral structure. S4, Build and train an underwater equipment fault diagnosis model; S5, Training of underwater equipment fault diagnosis model: Using the training dataset constructed in step S1, supervised training is performed on the gated multimodal context fusion network constructed in step S4. S6, Intelligent monitoring for fault diagnosis and prediction of underwater equipment: After completing the model training in step S5 and obtaining the optimized fault diagnosis model, it is deployed in the online intelligent monitoring system of underwater equipment to realize real-time diagnosis of equipment operating status and early fault prediction.

2. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 1, characterized in that: The underwater equipment acoustic monitoring data in S1 includes equipment operating condition data and marine environmental parameters. The equipment operating condition data includes echo intensity and output power, while the marine environmental parameters include the water depth where the equipment is located and the surrounding seawater temperature.

3. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 1, characterized in that: The types of manually labeled data in S1 include normal state, minor leak, rupture leak, pipeline corrosion, and seal failure.

4. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 1, characterized in that, The adaptive spectral equalization method in S2 includes the following steps: S201, Perform a short-time Fourier transform on the denoised time-domain signal vector to obtain the time spectrum. , Indicates the time frame index Frequency Index The short-time Fourier transform coefficients at that point, Represented as a time frame index, Indicates frequency index; S202, will Mapping to the Mel scale, calculate the Mel spectrum. , Indicates the first The Mel frequency band in the first The amplitude of each time frame; S203: Calculate the root mean square energy of each Mel band across all time frames based on its original spectral vector. Then, based on the probability of fault characteristics occurring in each band obtained statistically from prior fault data, calculate the equalization intensity factor. Using a reference gain constant as the target, adaptively scale the spectrum of each Mel band using a gain function to obtain the equalized Mel spectral vector. .

5. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 4, characterized in that, The adaptive spectral equalization method in S2 further includes the following steps: S204, equalizes the Mel spectrum The time-domain signal is reconstructed using inverse Mel transform and inverse short-time Fourier transform to obtain the spectrally equalized signal vector. .

6. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 1, characterized in that, The wavelet packet transform method in S3 is as follows: For the signal vector after spectral equalization... conduct Layer wavelet packet decomposition and extraction of the first wavelet packet. All layers The characteristics of each subband are obtained Each one carries a signal; Let the first Layer, First The wavelet packet coefficient vector of each sub-band is For each sub-band, calculate two types of time-frequency features: Energy characteristics Entropy characteristics energy characteristics Entropy characteristics By concatenating along the feature dimensions, a preliminary feature matrix for this sub-band is formed. , dimension .

7. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 1, characterized in that, The adaptive weighting method in S3 is as follows: S301, Calculate the discriminative score of the sub-band: Calculate the mean vector of the preliminary feature matrix of each sub-band on the fault class training sample set and the normal class training sample set respectively, and calculate the mean intra-class variance of the feature of the sub-band on the fault class samples and the mean intra-class variance on the normal class samples. Calculate the discriminative score of the feature of the sub-band based on the ratio of the square of the Euclidean distance between the two mean vectors to the sum of the mean intra-class variances of the two classes. S302, Calculate the adaptive fusion weight: Based on the discriminative score of each sub-band feature, the scores of all sub-bands are non-linearly amplified using the natural exponential function, and the amplified results are summed and normalized, that is, an adaptive fusion weight is assigned to each sub-band. S303, Weighted Stitching and Linear Projection: The preliminary feature matrix of each sub-band is multiplied by its corresponding adaptive fusion weight to obtain a weighted feature matrix. All weighted feature matrices are stitched together along the feature channel dimension according to the sub-band index to form a high-dimensional composite feature matrix. The composite feature matrix is ​​then linearly transformed by a learnable linear projection matrix to project it onto a fixed, lower-dimensional feature space, thereby obtaining a multi-resolution fusion feature matrix.

8. The intelligent monitoring method for underwater equipment fault diagnosis and prediction according to claim 1, characterized in that, The method for constructing and training the underwater equipment fault diagnosis model in S4 is as follows: S401, collects equipment operating parameters synchronized with acoustic signals: echo intensity. Output power Marine environmental parameters: water depth where the equipment is located The surrounding seawater temperature This constitutes the original operating condition parameter vector. To eliminate the influence of dimensions and adapt to neural network processing, each parameter is standardized and nonlinearly encoded to generate a high-dimensional, learnable working condition context feature vector. S402 employs a gated multimodal feature fusion strategy, which fuses the multi-resolution fusion feature matrix with the working condition context feature vector, and outputs the fault state probability through a pre-defined classification network.

Images