Fault diagnosis method and system based on deep space-time convolutional network ST-NN

Abstract

The invention belongs to the technical field of marine diesel engine fault diagnosis. The invention discloses a fault diagnosis method based on a deep space-time convolutional network ST-NN, and the method comprises the steps: obtaining original oil spectrum and / or ferrographic data, and an original vibration signal; performing dimensionality reduction on the original oil spectral and / or ferrographic data by adopting a rough set attribute reduction method, and removing redundant features to obtain a reduced oil and / or ferrographic feature set; performing time-frequency conversion on the original vibration signal by adopting continuous wavelet transform to obtain a time-frequency graph sequence; taking the reduced oil and / or ferrographic feature set as an additional feature channel, carrying out three-dimensional convolutional neural network fusion with a time-frequency graph sequence, and introducing a self-attention mechanism to carry out dynamic weighting on time steps to obtain a health index sequence; and according to the health index sequence, performing fault trend prediction by adopting a Gaussian process regression model to obtain a fault prediction result. The operation state of the diesel engine is effectively identified, and the accuracy and reliability of fault diagnosis are improved.

Description

Technical Field

[0001] The present application belongs to the technical field of marine diesel engine fault diagnosis, and in particular to a fault diagnosis method and system based on a deep spatiotemporal convolutional network (ST-NN). Background Art

[0002] In the field of marine diesel engine fault diagnosis, existing methods primarily rely on intelligent diagnostic models based on neural networks, such as BP neural networks and extreme learning machines (ELMs). However, these methods still have limitations when processing complex time series data. For example, BP neural networks are prone to falling into local optima and have slow convergence. While ELMs are computationally efficient, their randomly initialized weights can lead to large prediction errors. Furthermore, existing methods still have shortcomings in feature extraction and data dimensionality reduction, making it difficult for these models to fully utilize diesel engine oil monitoring data for accurate fault diagnosis. Summary of the Invention

[0003] Based on this, it is necessary to provide a fault diagnosis method and system based on deep spatiotemporal convolutional network ST-NN to address the above technical problems.

[0004] In a first aspect, the present application provides a fault diagnosis method based on a deep spatiotemporal convolutional network (ST-NN), comprising:

[0005] Obtaining raw oil spectrum and / or iron spectrum data, as well as raw vibration signals;

[0006] The original oil spectrum and / or iron spectrum data are subjected to dimensionality reduction using a rough set attribute reduction method to remove redundant features, thereby obtaining a simplified oil and / or iron spectrum feature set;

[0007] The original vibration signal is subjected to time-frequency conversion using continuous wavelet transform to obtain a time-frequency graph sequence;

[0008] The simplified oil and / or iron spectrum feature set is used as an additional feature channel to perform a three-dimensional convolutional neural network fusion with the time-frequency graph sequence, and a self-attention mechanism is introduced to dynamically weight the time steps to obtain a health indicator sequence;

[0009] According to the health indicator sequence, a Gaussian process regression model is used to predict the fault trend and obtain a fault prediction result.

[0010] In some practicable embodiments, the step of reducing the dimension of the original oil spectrum and / or ferrogram data using a rough set attribute reduction method to remove redundant features to obtain a simplified oil and / or ferrogram feature set includes:

[0011] The continuous attributes of the original oil spectrum and / or ferrogram data are discretized by equal-frequency binning to generate a discretized decision system S = (U, R, V, f), where U represents the object set, R = C∪D is the attribute set, C is the oil spectrum and / or ferrogram condition attribute subset, D is the fault state decision attribute subset, V represents the attribute value range, and f represents the information function that determines the attribute value of the sample, f = U × R → V;

[0012] Calculating all subsets of single attributes and key combination attributes in the discretized decision system S to obtain positive domains of all single attributes and key combination attributes, wherein the positive domains are obtained by dividing equivalence classes through indiscernibility relations;

[0013] According to the positive domains of all single attributes and key combination attributes, heuristic attribute reduction is performed to obtain the reduced oil and / or ferrogram feature set.

[0014] In some practicable manners, the step of performing time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency graph sequence includes:

[0015] Construct a basic wavelet function, wherein the scaling and translation calculation formula of the basic wavelet function ψ(t) is:

[0016]

[0017] Among them, ψ, represents the mother wavelet, k, represents the scale parameter, l, represents the translation parameter, and t represents time;

[0018] For the original vibration signal, continuous wavelet transform is used to perform time-frequency conversion to obtain a calculation result, wherein the calculation formula of the continuous wavelet transform is:

[0019]

[0020] in, represents the conjugate function of the wavelet function, CWT(s)(k, l) represents the continuous wavelet transform coefficient matrix, s(t) represents the original vibration signal, and dt represents the integral variable;

[0021] The calculation results are normalized to obtain a time-frequency graph sequence.

[0022] In some practicable embodiments, the step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes:

[0023] The simplified oil and / or iron spectrum feature set is used as an independent channel and expanded to the same resolution as the time-frequency graph sequence by a spatial interpolation method to form an additional feature channel;

[0024] The additional feature channel is fused with the time-frequency graph sequence to generate a fusion input tensor with a shape of (N, H, W, L, C), where N represents the number of samples, H represents the height of the time-frequency graph, W represents the width of the time-frequency graph, L represents the number of overlaps of the time-frequency graph sequence, and C represents the number of channels of the time-frequency graph sequence.

[0025] In some practicable embodiments, the step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes:

[0026] Construct at least three 3D convolutional layers, each using the ReLU activation function to perform a unified nonlinear transformation on the input time-frequency graph sequence and the simplified oil and / or ferrogram features to extract local correlation features in the time-frequency domain;

[0027] The fused input tensor is input into the three 3D convolutional layers. A ReLU activation function is used in each 3D convolutional layer to perform a nonlinear transformation on the fused input tensor. The multi-source features of adjacent time-frequency graphs are associated in the time dimension through the 3D convolutional layer to obtain associated multi-source fusion features, wherein the multi-source fusion features include vibration time-frequency graph, oil and iron spectrum features.

[0028] In some practicable embodiments, the step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes:

[0029] After the at least three 3D convolutional layers, two fully connected layers are constructed, and a Dropout operation is performed between the two fully connected layers, wherein the fully connected layers include a first fully connected layer that uses a ReLU activation function and a second fully connected layer that uses a Sigmoid activation function to output a health indicator;

[0030] Inputting the associated multi-source fusion features into the first fully connected layer, performing dimensionality reduction and nonlinear activation using a ReLU activation function, and obtaining features processed by the first fully connected layer;

[0031] Perform a Dropout operation on the features processed by the first fully connected layer, randomly discarding neurons with a preset probability to obtain non-overfitted Dropout-processed features;

[0032] The Dropout processed features are input into the second fully connected layer, and the Sigmoid activation function is used to map the features into a health indicator sequence to obtain a health indicator sequence.

[0033] In some practicable embodiments, the step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes:

[0034] The associated multi-source fusion features are linearly transformed through a trainable parameter matrix and mapped into a Q / K / V matrix, wherein the shape of the associated multi-source fusion features is (N, T, d), N represents the number of samples, T represents the number of time steps, and d represents the feature dimension. The Q / K / V matrix includes a query matrix Q, a key matrix K, and a value matrix V. The formula of the Q / K / V matrix is:

[0035] Q=HI·W Q ;

[0036] K=HI·W K ;

[0037] V=HI·W v ;

[0038] in, W Q 、W K 、W v All are trainable parameters in the trainable parameter matrix;

[0039] By scaling the dot product, the influence weight of each time step on other time steps is calculated and normalized to generate a probability distribution weight matrix;

[0040] The probability distribution weight matrix is ​​used to perform weighted summation on the value matrix V to obtain the optimized feature representation, wherein the optimized feature representation formula is:

[0041] Z = A·V;

[0042] Among them, Z represents the optimized feature representation, A represents the probability distribution weight matrix;

[0043] The optimized feature representation is input into the fully connected layer to generate a health indicator sequence, and the trainable parameters of the three-dimensional convolutional neural network and the self-attention mechanism are jointly optimized through back propagation.

[0044] In some practicable embodiments, the step of using a Gaussian process regression model to predict a fault trend based on the health indicator sequence to obtain a fault prediction result includes:

[0045] Build a Gaussian process regression model;

[0046] The health indicator sequence is segmented according to the preset time window and normalized to obtain the segmentation results;

[0047] The segmentation results are input into a Gaussian process regression model to obtain a fault prediction result.

[0048] In a second aspect, the present application provides a fault diagnosis system based on a deep spatiotemporal convolutional network (ST-NN), applying the aforementioned fault diagnosis method based on a deep spatiotemporal convolutional network (ST-NN). The system includes:

[0049] an acquisition unit, configured to acquire raw oil spectrum and / or iron spectrum data, and raw vibration signals;

[0050] A reduction element is used to reduce the dimension of the original oil spectrum and / or iron spectrum data by using a rough set attribute reduction method to remove redundant features and obtain a reduced oil and / or iron spectrum feature set;

[0051] A time-frequency unit, configured to perform time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency graph sequence;

[0052] a fusion unit, configured to use the simplified oil and / or iron spectrum feature set as an additional feature channel, perform a three-dimensional convolutional neural network fusion with the time-frequency graph sequence, and introduce a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence;

[0053] The result unit is used to predict the fault trend based on the health indicator sequence using a Gaussian process regression model to obtain a fault prediction result.

[0054] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor executes the steps of the aforementioned method.

[0055] Beneficial effects: The present application provides a fault diagnosis method based on a deep spatiotemporal convolutional network ST-NN, which obtains original oil spectrum and / or iron spectrum data, as well as original vibration signals; the original oil spectrum and / or iron spectrum data are reduced in dimension using a rough set attribute reduction method, redundant features are removed, and a simplified oil and / or iron spectrum feature set is obtained; the original vibration signal is converted into time and frequency using a continuous wavelet transform to obtain a time-frequency graph sequence; the simplified oil and / or iron spectrum feature set is used as an additional feature channel and fused with the time-frequency graph sequence using a three-dimensional convolutional neural network, and a self-attention mechanism is introduced to dynamically weight the time steps to obtain a health indicator sequence; based on the health indicator sequence, a Gaussian process regression model is used to predict fault trends to obtain fault prediction results. Through the above method, rough set attribute reduction and deep learning technology are combined, and the rough set method is used to reduce the dimension of the oil spectrum and iron spectrum features of ship diesel engines, remove redundant information, optimize data input, and improve the computational efficiency of the model. The spatiotemporal characteristics of marine diesel engines are extracted through 3D-CNN and combined with Gaussian process regression to further predict fault trends. This method can effectively identify the operating status of diesel engines and improve the accuracy and reliability of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the conventional technology, the following briefly introduces the drawings required for use in the embodiments or the conventional technology descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 The figure is a flowchart of a fault diagnosis method based on a deep spatiotemporal convolutional network (ST-NN) in one embodiment.

[0058] Figure 2 2 is a schematic diagram of a fault diagnosis method based on a deep spatiotemporal convolutional network (ST-NN) in one embodiment. DETAILED DESCRIPTION

[0059] To facilitate understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The accompanying drawings provide embodiments of the present application. However, the present application may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein in the specification of this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The term "and / or" as used herein includes any and all couplings of one or more of the associated listed items.

[0061] It will be understood that the terms "first," "second," etc. used herein may be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish a first element from another element.

[0062] The following are some explanations of some terms involved in this application to facilitate understanding of this application:

[0063] Tensor is a basic concept in mathematics and computer science used to represent multidimensional data. It can be understood as a generalized form of multidimensional array.

[0064] Dropout is a regularization technique that randomly discards (sets to zero) the outputs of some neurons during training (usually with a ratio of 0.2-0.5) to prevent the model from overfitting.

[0065] Spatial interpolation is a method of estimating the value of an unknown location based on data from known discrete points.

[0066] ReLU (Rectured Linear Unit) is a widely used activation function and is very popular in deep learning due to its simplicity and efficiency.

[0067] A neuron is a mathematical model that simulates the behavior of biological neurons and is the basic computing unit of a network. It receives input signals and generates outputs through weighted summation and nonlinear transformation (activation function).

[0068] The Z-Score (standard score) is a statistical measure used to describe the degree to which a data point deviates from the mean of a dataset, expressed in units of standard deviation. Its core purpose is to standardize the values ​​in different datasets for easier comparison and analysis.

[0069] The scaled dot product is the core operation in the Attention Mechanism, used to calculate the relevance score between a query and a key. The core idea is to measure similarity using the dot product and introduce a scaling factor to stabilize the gradient, preventing excessively large dot product values ​​from causing the Softmax function's gradient to vanish.

[0070] In practical applications, marine diesel engines are subject to prolonged exposure to harsh environments such as high temperatures, high loads, and high vibration, leaving their core components vulnerable to corrosion and wear. Existing diagnostic methods often rely on oil testing and vibration signal analysis to monitor the health of diesel engines. Oil testing primarily relies on spectral analysis and ferrography, analyzing the concentration of metal particles in the oil to determine mechanical wear. Vibration signal analysis, on the other hand, utilizes accelerometers to collect vibration data from diesel engines during operation and employs signal processing techniques such as fast Fourier transforms (FFTs) and wavelet transforms to extract characteristic parameters.

[0071] However, these traditional methods have some significant shortcomings. First, the diagnostic method based on BP neural network is prone to falling into local optimality when dealing with complex nonlinear problems, and due to the gradient vanishing problem of the backpropagation algorithm, the training process converges slowly. Secondly, although the ELM method has a fast calculation speed, its randomly initialized weights may lead to large prediction errors, resulting in poor model stability. In addition, traditional oil detection and vibration signal analysis methods rely on manually set feature extraction rules, lack adaptability and generalization capabilities, and are difficult to accurately identify complex fault modes. When faced with a complex and changing combat environment, the diagnostic accuracy and real-time performance of these methods are difficult to meet the needs of ship diesel engine health monitoring.

[0072] like Figure 1 and Figure 2 As shown, in the first aspect, the present application provides a fault diagnosis method based on a deep spatiotemporal convolutional network (ST-NN), comprising:

[0073] S100, acquiring original oil spectrum and / or iron spectrum data, and original vibration signal.

[0074] Among them, for the collection of original oil spectrum and / or iron spectrum data, the original oil spectrum and / or iron spectrum data can be collected at the key nodes of the diesel engine lubricating oil circulation pipeline (main oil channel outlet / before the filter). For example, spectral data: an atomic emission spectrometer is used to monitor the concentrations of 15 metal elements such as Fe, Cu, and Al online (ppm level accuracy); iron spectrum data: the concentration ratio (DL / DS) of large abrasive particles (>5μm) and small abrasive particles (1-2μm) is obtained by a direct-reading iron spectrometer; sampling frequency: a complete set of oil data every 10 minutes.

[0075] For the collection of the original vibration signal, the collection device can be placed at a location such as the main bearing seat. This application does not limit how to obtain the original oil spectrum and / or iron spectrum data, and the original vibration signal, and the location arrangement and collection can be carried out according to actual conditions.

[0076] S200 , performing dimensionality reduction on the original oil spectrum and / or iron spectrum data using a rough set attribute reduction method, removing redundant features, and obtaining a reduced oil and / or iron spectrum feature set.

[0077] Specifically, obtaining the simplified oil and / or ferrographic feature set may include the following steps:

[0078] S201 , discretizing the continuous attributes of the original oil spectrum and / or iron spectrum data into equal-frequency bins to generate a discretized decision system S=(U, R, V, f).

[0079] Where U represents the object set, U={x1,x2…x n}, corresponding to n oil samples x; R = C∪D = {r1, r2, ...r m} is the attribute set, C is the oil spectrum and / or iron spectrum condition attribute subset, D is the fault state decision attribute subset; V, represents V r is the value range of the attribute r∈R, f can be regarded as the information function between U and R, which determines the attribute value of the sample, f: U×R→V, f(x, r)∈V r .

[0080] Specifically, first, the raw oil spectrum and / or iron spectrum data can be formed into a dataset (n×m matrix, n samples×m monitoring indicators). This method can convert the raw data into a normalized form that is computable, scalable, and interpretable.

[0081] Next, build the decision table:

[0082] Conditional attribute subset C: Quantified oil iron spectrum characteristics (Fe concentration, Cu concentration, DL / DS ratio, etc.).

[0083] Decision attribute subset D: device status label (0 normal / 1 fault).

[0084] Discretization of continuous attributes:

[0085] The equal-frequency binning method (the number of samples in each bin is the same) can be used to process each continuous attribute (such as Fe concentration) separately and divide it equally according to the sample frequency. For example, the Fe concentration is divided into 3 bins, each containing approximately n / 3 samples.

[0086] It should be noted that the indiscernibility relation (IND) of rough sets requires that attribute values ​​be finite discrete values, and continuous values ​​must be converted into categorical variables through discretization before equivalence classes can be calculated.

[0087] After obtaining U, R, V, and f in the above steps, the discretized decision system S can be output.

[0088] S202, calculating all single attributes and key combination attribute subsets in the discretized decision system S, and obtaining positive domains of all single attributes and key combination attributes.

[0089] Among them, the positive domain is obtained by dividing the equivalence classes through the indiscernibility relation.

[0090] Specifically, if X is a subset of the set of U objects, then For each X and the unclear relationship Define the lower approximation set B of X * (X), based on knowledge B (knowledge B represents the ability to classify or distinguish objects in the domain U), determine the set of objects in U that must belong to X X represents any subset of U. When B is a single attribute, the equivalence class and positive domain of the single conditional attribute are calculated. The equivalence relation of all conditional attributes includes the results of independent calculations for each single attribute. By combining the key attribute subset B, multiple attributes can be used to improve classification capabilities.

[0091] The upper approximation set B of X * (X), the set of objects in U that may belong to X according to knowledge B Y represents any object in U.

[0092] For conditional attributes, if the decision values ​​of objects in the indistinguishable set are the same, the decision system is compatible; otherwise, it is incompatible. All conditional attribute equivalence relations in the compatible decision system are calculated. For each equivalence relation, the upper and lower approximations relative to the decision value are obtained by calculation according to the heuristic function:

[0093] λ=(|B * (X)|-|B * (X)|) / |U|.

[0094] Here, λ is the classification uncertainty measure for the current attribute subset B with respect to the set X. A smaller value indicates a more accurate classification. By optimizing λ, the most critical features for fault diagnosis can be automatically selected, balancing model accuracy and efficiency. |U| normalizes the proportion of the fuzzy region, making λ a relative uncertainty measure (ranging between 0 and 1).

[0095] Retain the attribute with the smallest λ value. Use the lower approximation value for the given domain as the basis for pruning. Remove redundant features (such as the concentration of irrelevant metals in the oil spectrum) through simplification. Calculate the λ value for each candidate attribute subset B (a single attribute is the smallest unit of the candidate subset, and its λ value needs to be calculated separately as the benchmark for the combined attribute). Select B with the smallest λ as the simplification result to ensure that the retained features can distinguish between "normal" and "faulty" states to the greatest extent. Reduce model complexity while retaining key diagnostic information. Improve the training efficiency and generalization ability of subsequent models such as 3D-CNN.

[0096] The role of λ in attribute reduction is to prioritize the attribute subset B that minimizes λ. When λ approaches 0, the fuzzy region disappears, and B can completely and accurately classify X. The λ value of the combined attribute must be significantly lower than that of a single attribute (for example, a single attribute λ ≥ 0.3, while the combined λ = 0.05), thus demonstrating the necessity of multi-attribute collaboration. When λ approaches 1, the λ value of a single attribute is too high, the fuzzy region disappears, and B cannot provide effective classification information. The attribute has poor classification ability alone and may be eliminated.

[0097] S203: performing heuristic attribute reduction based on the positive domains of all single attributes and key combination attributes to obtain a reduced oil and / or ferrogram feature set.

[0098] Specifically, the attribute with the smallest lambda value is retained. Given a set of objects, the lower approximation is used as the pruning basis:

[0099] U=U-POS red {D};

[0100] Among them, POS red {D} represents the set of objects whose conditional attributes for the reduced set red (the reduced oil and / or ferrographic feature set red) can be included in the decision attribute D in the domain U. This represents the positive domain (the union of lower approximation sets) of the reduced set red for the decision attribute D. This process is repeated repeatedly, selecting the remaining attributes. Each time, the retained attributes are combined with the attributes in the reduced set red to form a new equivalence relation. This recursive operation is terminated when the given domain is empty. The recursion terminates when the remaining attributes cannot further reduce the λ value, or when U is empty.

[0101] S300 , performing time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency graph sequence.

[0102] Specifically, obtaining the time-frequency graph sequence may include the following steps:

[0103] S301, constructing a basic wavelet function.

[0104] The scaling and translation calculation formula of the basic wavelet function ψ(t) is:

[0105]

[0106] Among them, ψ, represents the mother wavelet, k, represents the scale parameter (such as k∈[2, 256] distributed logarithmically), which can evenly cover the signal frequency band. The specific range is set according to the characteristic frequency of the equipment fault, l, represents the translation parameter, and t represents time;

[0107] For example,

[0108] Among them, π -1 / 4 , represents the normalization coefficient, e, represents the base of the natural logarithm, i, represents the imaginary unit, w o , represents the center frequency of the wavelet, t, represents the time variable, Represents the Gaussian envelope function. For example, the center frequency of the Morlet wavelet w o =6, which satisfies the permissible condition.

[0109] S302: Perform time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a calculation result.

[0110] The calculation formula of the continuous wavelet transform is:

[0111]

[0112] in, represents the conjugate function of the wavelet function, CWT(s)(k, l) represents the continuous wavelet transform coefficient matrix, s(t) represents the original vibration signal, and dt represents the integral variable. The CWT continuous wavelet transform coefficient matrix is ​​converted into a grayscale image (the CWT coefficients are complex numbers, such as the modulus value |CWT(s)(k, l)|), for example, (with a size of H×W×1, which serves as the time input channel for 3D-CNN spatiotemporal modeling).

[0113] S303: Normalize the calculation results to obtain a time-frequency graph sequence.

[0114] Normalization is to normalize each scale k independently, and the method can be a conventional normalization method. This application does not limit how to implement normalization. For example, the CWT coefficients are normalized to the maximum and minimum and mapped to the interval [0, 1].

[0115] The continuous wavelet transform (CWT) performs multi-scale analysis of vibration signals through expansion and shift functions, effectively extracting time-frequency domain information. CWT has low computational complexity and good time-frequency localization properties.

[0116] S400: Using the simplified oil and / or iron spectrum feature set as an additional feature channel, performing a three-dimensional convolutional neural network fusion with the time-frequency graph sequence, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence.

[0117] Specifically, obtaining a health indicator sequence may include the following steps:

[0118] S401: The simplified oil and / or iron spectrum feature set is used as an independent channel and expanded to the same resolution as the time-frequency graph sequence by a spatial interpolation method to form an additional feature channel.

[0119] Specifically, spatial interpolation expands additional features, such as the conventional bilinear interpolation method, which expands the one-dimensional time series into a two-dimensional spatial feature map.

[0120] For example, the temporal features of the oil / iron spectrum are mapped to the time-frequency graph space to capture their spatial correlation with the vibration energy. For example, areas with high metal concentrations may correspond to high-frequency energy concentrations in the time-frequency graph.

[0121] Oil characteristics: shape (N, L, C1);

[0122] Ferrographic characteristics: shape (N, L, C2);

[0123] After spatial interpolation:

[0124] Oil extended characteristics: (N, H, W, L, C1);

[0125] Ferrospectral extended features: (N, H, W, L, C2).

[0126] S402: Fuse the additional feature channel with the time-frequency graph sequence to generate a fusion input tensor with a shape of (N, H, W, L, C), where N represents the number of samples, H represents the height of the time-frequency graph, W represents the width of the time-frequency graph, L represents the number of overlaps of the time-frequency graph sequence, and C represents the number of channels of the time-frequency graph sequence.

[0127] Specifically, the fused input tensor can be generated by splicing along the channel dimension. In the channel dimension, each channel represents the spatiotemporal characteristics of a different modality (e.g., channel 1 is vibration energy, channels 2-4 are oil metal concentration, and channels 5-6 are ferrographic wear parameters). In the spatial dimension, the (H, W) of the time-frequency graph corresponds to the frequency-time resolution, and the expanded oil / ferrographic features are aligned with it, allowing the convolution kernel to capture local cross-modal correlations.

[0128] S403: construct at least three 3D convolutional layers, each of which uses a ReLU activation function to perform a unified nonlinear transformation on the input time-frequency graph sequence and the simplified oil and / or iron spectrum features to extract local correlation features in the time-frequency domain.

[0129] For example, the kernel sizes of the 3D convolutional layers are 7×7×2, 3×3×2, and 3×3×2, respectively, and the stride is set to 2×2×1.

[0130] Spatial dimensions (H, W): The large kernel (7×7) captures global frequency domain patterns (such as broadband resonance of bearing faults), and the small kernel (3×3) extracts local details (such as high-frequency harmonics of gear meshing).

[0131] Time dimension (L): The kernel size is 2, which correlates features of adjacent time steps and models the evolution of faults (e.g., a gradual change in oil concentration causing an increase in vibration energy).

[0132] It should be noted that the ReLU activation function is applied immediately after the convolution operation of each 3D convolutional layer to perform an element-by-element nonlinear transformation on the output feature map. In order to break the linear limitation, the relationship between the time-frequency map (vibration signal) and the oil / ferrogram characteristics can be highly nonlinear. For example, a sudden increase in the metal concentration of the oil may be related to the nonlinear growth of the high-frequency energy of the vibration, and a simple linear convolution cannot capture such a complex pattern. In addition, multimodal feature fusion, ReLU allows the model to learn nonlinear interactive features across modalities (vibration, oil, ferrogram), for example, the joint impact of oil parameter changes and vibration energy diffusion.

[0133] For example, consider the case of early wear in a bearing. The iron particle concentration in the oil (an oil characteristic) is nonlinearly positively correlated with the high-frequency energy in the vibration time-frequency graph (a vibration characteristic). ReLU uses nonlinear mapping to enable the model to fit this complex relationship, rather than relying solely on linear combinations.

[0134] It's also worth noting that, through 3D convolutional layers and ReLU activation functions, the features of different modes (vibration time-frequency diagrams, oil, and ferrograms) are uniformly encoded into the same nonlinear space. This way, regardless of whether the input is vibration, oil, or ferrogram features, the same nonlinear transformation (ReLU) is used to extract cross-modal correlation features, enabling unified processing. Features from all modes share the same set of convolution kernel parameters and activation functions, ensuring consistent feature transformation rules rather than independent processing, facilitating subsequent calculations.

[0135] S404: Input the fused input tensor into the three 3D convolutional layers. In each 3D convolutional layer, a ReLU activation function is used to perform a nonlinear transformation on the fused input tensor. The multi-source features of adjacent time-frequency graphs are associated in the time dimension through the 3D convolutional layer to obtain associated multi-source fusion features.

[0136] The multi-source fusion features include vibration time-frequency diagram, oil and iron spectrum features.

[0137] Specifically, the 3D convolution kernel slides along the time axis, performing a weighted fusion of multimodal features from adjacent time steps to capture the evolution of fault characteristics. The convolution kernel then performs a nonlinear transformation on the convolution results using the ReLU function, suppressing irrelevant responses and enhancing cross-modal correlation features. For example, if the time kernel size is 2 (i.e., each convolution operation associates two consecutive time steps, such as t and t+1), and the oil concentration suddenly increases at time step t, and the vibration energy increases at t+1, the convolution kernel automatically learns this causal relationship and outputs a tensor shape.

[0138] In the 3D convolution kernel structure, the feature map of each convolution layer is connected not only to the current time-frequency map (TFR) but also to multiple adjacent TFRs, thereby extracting more temporal information. In contrast, traditional 2D convolution can only extract local features from a single TFR and cannot fully utilize the temporal dependencies and spatial characteristics of diesel engine vibration signals. By constructing a TFR cube with consecutive time steps, 3D convolution not only extracts the spatial features of a single TFR but also captures the temporal characteristics between adjacent TFRs, improving the accuracy of fault mode recognition.

[0139] For example, after three layers of convolution processing, the fused input tensor (N, 128, 256, 10, 6) gradually evolves to (N, 64, 128, 9, 64), (N, 32, 64, 8, 128), and (N, 32, 64, 7, 256). Each channel in the final output represents a cross-modal correlation pattern of vibration, oil, and iron spectrum characteristics.

[0140] It should be noted that each 3D convolution kernel contains learnable weight parameters, which extract local feature patterns in the spatial dimension (H×W) and learn the weighted relationship between features at adjacent time steps in the temporal dimension (L). These weight parameters are automatically optimized during the training process, enabling the model to capture the spatiotemporal evolution of fault characteristics.

[0141] During the convolution process, the vibration time-frequency map, oil characteristics, and ferrographic characteristics are uniformly calculated using shared convolution kernel parameters, achieving cross-modal feature fusion. Each output channel comprehensively reflects the correlation pattern of multi-source features.

[0142] For example, when the first convolution kernel is processed at time step t=5, it will simultaneously consider:

[0143] The value of the oil characteristic at t = 5 (e.g. iron concentration 0.9);

[0144] The response of the vibration characteristics at t = 6 (such as high-frequency energy 0.8);

[0145] Through training, the features of these two time steps are given a strong correlation weight (such as 0.7), thereby encoding the delayed effect of the fault.

[0146] S405: After the at least three 3D convolutional layers, construct two fully connected layers, and perform a Dropout operation between the two fully connected layers.

[0147] The fully connected layers include a first fully connected layer that uses the ReLU activation function and a second fully connected layer that uses the Sigmoid activation function to output health indicators. A fully connected layer consists of multiple neurons (computing units), each of which receives all inputs from the previous layer and performs a linear transformation using a weight matrix and bias.

[0148] In the previous steps, the associated multi-source fusion features are obtained, with a shape example of (N, 32, 64, 7, 256), where N is the number of samples. Global average pooling is performed to compress the spatial dimension to a single point, retaining the time step and channel information. The shape after pooling is (N, 1, 1, 7, 256). The redundant spatial dimensions (H = 1, W = 1) are removed, and the output shape becomes (N, 7, 256). Next, the time step and channel dimensions are merged to obtain a two-dimensional feature matrix (shape example: (N, 7×256 = 1792)).

[0149] It should be noted that the role of the first fully connected layer is to reduce the feature from 1792 dimensions to 512 dimensions and use the ReLU activation function to filter out non-linear negative values.

[0150] ReLU activation function:

[0151] ReLU(v)=max(0,v);

[0152] Where v is the input value.

[0153] ReLU(v) output value: when v>0, output v, otherwise output 0.

[0154] Sigmoid activation function:

[0155]

[0156] Among them, v is the input value, e is a natural constant, and σ(v) is the output value, ranging from (0, 1), which is the probability output of the binary classification problem.

[0157] v=W·X+b;

[0158] Among them, W is the weight of the fully connected layer, b is the bias, and X is the pooled feature (that is, the two-dimensional feature result after the multi-source fusion feature output of the 3D convolution is compressed).

[0159] The function of the Dropout layer is to randomly block neurons with a probability of 50% during training to prevent the model from overfitting.

[0160] The role of the second fully connected layer is to map the 512-dimensional features to the target dimension (such as predicting the next 5 time steps) and output the health indicator (value range [0, 1]) through the Sigmoid function.

[0161] S406 , inputting the associated multi-source fusion features into the first fully connected layer, performing dimensionality reduction and nonlinear activation using a ReLU activation function, and obtaining features processed by the first fully connected layer.

[0162] Specifically, the reshaped two-dimensional features (e.g., (N, 1792)) are input and linearly transformed using a weight matrix and bias term. The resulting features are nonlinearly activated using the ReLU function to suppress negative responses, outputting 512-dimensional features (shape: (N, 512)). The goal is to extract high-level, abstract features across modalities, such as the correlation between sudden increases in vibration energy and metal concentration in the oil.

[0163] S407: Perform a Dropout operation on the features processed by the first fully connected layer, randomly discard neurons with a preset probability, and obtain non-overfitted Dropout-processed features.

[0164] Specifically, during the training phase, features processed by the first fully connected layer (composed of multiple neuron outputs) are randomly set to zero with a preset probability (e.g., 50%), allowing the model to learn redundant feature pathways. During the inference phase, all neurons are retained, but the weights are scaled by probability to maintain output stability. For example, if a neuron is dedicated to the iron spectrum concentration feature of oil, dropout forces the model to also rely on vibration features for prediction.

[0165] S408: Input the Dropout processed features into the second fully connected layer, and use a Sigmoid activation function to map the features into a health indicator sequence to obtain a health indicator sequence.

[0166] Specifically, the features after Dropout processing are input into the second fully connected layer, and the output is mapped to the [0, 1] interval through the Sigmoid function.

[0167] For example, the output matrix is ​​a (N, 5) matrix representing the health index for the next five time steps, such as [0.1, 0.3, 0.5, 0.7, 0.9]. Values ​​closer to 1 indicate a higher risk of failure. The health index comprehensively reflects the combined status of vibration, oil, and ferrography. For example, 0.9 indicates that immediate maintenance is required.

[0168] It should be noted that a dropout layer is added after the last convolutional layer to prevent overfitting. In addition, the last layer uses the Sigmoid activation function to fit the health indicator (HI), and the other layers use the ReLU activation function to avoid the gradient vanishing problem.

[0169] Through the collaboration of fully connected layers and Dropout, the model gradually maps multi-source fusion features into an intuitive sequence of health indicators, capturing the complex patterns of fault evolution while avoiding excessive reliance on training data, thus meeting the robustness requirements of industrial scenarios.

[0170] S500 , performing fault trend prediction using a Gaussian process regression model according to the health indicator sequence to obtain a fault trend prediction.

[0171] Specifically, obtaining a fault trend prediction may include the following steps:

[0172] S501, establish a Gaussian process regression model.

[0173] The Gaussian process regression model (GPR) formula is:

[0174] f(x)~GP(m(x),k(x,x'));

[0175] Among them, m(x) is the mean function, which represents the baseline trend of the health index (HI), and k(x, x') is the covariance function, which measures the similarity of HI sequences in different time windows.

[0176] S502 , segmenting the health indicator sequence according to a preset time window and normalizing it to obtain a segmentation result.

[0177] Specifically, the health indicator sequence is divided according to a preset time window. For example, the health indicator sequence is divided into input-output pairs according to a time window (such as a window size k=10 and a step size s=1).

[0178] Next, the input-output pairs are subjected to data standardization to eliminate differences and complete normalization. The specific normalization method may be Z-Score or other conventional methods.

[0179] S503: Input the segmentation result into a Gaussian process regression model to obtain a fault prediction result.

[0180] Specifically, the covariance matrix is ​​calculated and the input-output pairs are fed into a trained GPR model. The GPR model then outputs predictions for several future time windows. These predictions are then averaged to form a predicted mean, reflecting future trends. The GPR model also outputs the prediction variance, which indicates the confidence level of the prediction (the smaller the variance, the higher the confidence level).

[0181] The prediction results are judged for their credibility through the prediction variance, thus obtaining the fault prediction results.

[0182] The input (training data) consists of segments of the historical health indicator (HI) series, segmented into fixed time windows. The sliding window moves forward one time step at a time, with a step size of 1, generating multiple input-output pairs. The output (predicted target) is the HI value m time steps into the future.

[0183] In the Gaussian process regression model, the mean function m(x) describes the baseline trend of HI, and the covariance function k(x, x') measures the similarity of different time windows.

[0184] For example, the training phase of the Gaussian process regression model: learn the parameters of the covariance function from historical input-output pairs, and establish the time-dependent pattern of the HI sequence. Prediction phase: input the HI sequence of the current time window, covariance calculates the covariance matrix of the current window and the historical window, and generates HI prediction values ​​for future time steps by weighting historical similar patterns, such as [0.85, 0.90, 0.95, 0.98, 1.0], reflecting the rising trend of fault risk. Prediction variance (credibility), the smaller the variance value, the more reliable the prediction result; a sudden increase in variance may indicate unknown interference. Application of prediction results: take the moving average of the predicted mean sequence (such as the mean of the next 5 steps is 0.94) as a comprehensive indicator of fault risk. If the mean value continues to exceed the threshold (such as 0.9), an early warning is triggered.

[0185] In one embodiment, the step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes:

[0186] In the fault diagnosis of marine diesel engines, data is input into feature engineering and divided into test set, validation set and training set by feature engineering. The training set is used to train time-frequency analysis, 3D-CNN spatiotemporal modeling, and to construct health indicators. This health indicator is also used as a self-attention mechanism in the validation set to perform fault prediction and assist in adjusting the parameters of 3D-CNN based on the fault prediction results. After the model is finally formed, the data in the test set is sent to the model test to obtain the prediction results. The associated multi-source fusion features are linearly transformed through the trainable parameter matrix and mapped into a Q / K / V matrix, wherein the shape of the associated multi-source fusion features is (N, T, d), N represents the number of samples, T represents the number of time steps, d represents the feature dimension, and the Q / K / V matrix includes the query matrix Q, the key matrix K and the value matrix V. The formula of the Q / K / V matrix is:

[0187] Q=HI·W Q ;

[0188] K=HI·W K ;

[0189] V=HI·W v ;

[0190] in, W Q 、W K 、W v are all trainable parameters in the trainable parameter matrix.

[0191] Perform linear transformation on the associated multi-source fusion features to generate a query matrix Q = HI·W Q , bond matrix K = HI·W K , value matrix V = HI·W v .

[0192] By scaling the dot product, the influence weight of each time step on other time steps is calculated and normalized to generate a probability distribution weight matrix.

[0193] The formula and normalization of the scaled dot product are conventional formulas and methods, and this application does not limit them.

[0194] The probability distribution weight matrix is ​​used to perform weighted summation on the value matrix V to obtain the optimized feature representation, wherein the optimized feature representation formula is:

[0195] Z = A·V;

[0196] Among them, Z represents the optimized feature representation, and A represents the probability distribution weight matrix.

[0197] The value matrix is ​​weighted and summed using the probability distribution weight matrix to obtain the optimized feature representation Z, where the feature of each time step is a weighted combination of the features of other time steps.

[0198] The optimized feature representation is input into the fully connected layer to generate a health indicator sequence, and the trainable parameters of the three-dimensional convolutional neural network and the self-attention mechanism are jointly optimized through back propagation.

[0199] The self-attention mechanism is modular and easy to adjust.

[0200] In summary, the core advantage of the ST-CNN architecture lies in its ability to simultaneously extract feature information in both temporal and spatial dimensions. As a spatiotemporal modeler, 3D-CNN is able to more comprehensively capture the operating status of marine diesel engines and accurately identify wear fault modes compared to traditional two-dimensional convolutional networks (2D-CNNs). Furthermore, combined with the self-attention mechanism (SAM), this approach adaptively assigns weights to different time steps, improving the extraction of key features and avoiding the loss of diagnostic accuracy often associated with traditional methods due to inappropriate feature selection. Finally, the GPR method is used to predict the health status of diesel engines, quantifying the uncertainty of the prediction results and improving the reliability and robustness of fault diagnosis.

[0201] In a second aspect, the present application provides a fault diagnosis system based on a deep spatiotemporal convolutional network (ST-NN), applying the aforementioned fault diagnosis method based on a deep spatiotemporal convolutional network (ST-NN). The system includes:

[0202] an acquisition unit, configured to acquire raw oil spectrum and / or iron spectrum data, and raw vibration signals;

[0203] A reduction element is used to reduce the dimension of the original oil spectrum and / or iron spectrum data by using a rough set attribute reduction method to remove redundant features and obtain a reduced oil and / or iron spectrum feature set;

[0204] A time-frequency unit, configured to perform time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency graph sequence;

[0205] a fusion unit, configured to use the simplified oil and / or iron spectrum feature set as an additional feature channel, perform a three-dimensional convolutional neural network fusion with the time-frequency graph sequence, and introduce a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence;

[0206] The result unit is used to predict the fault trend based on the health indicator sequence using a Gaussian process regression model to obtain a fault prediction result.

[0207] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor executes the steps of the aforementioned method.

[0208] Example

[0209] For the information system DS = (U, C ∪ D, V, f), the introduced heuristic attribute reduction algorithm process is as follows:

[0210] 1) Initialize the reduction set red to (empty set), C is initialized to all conditional attributes;

[0211] 2) Calculate the λ value of each conditional attribute in C using formula (1);

[0212] 3) Sort the λ values ​​in descending order and select the attribute with the smallest λ value. If the attributes have the same λ value, the attribute with the smallest value can be selected. If the input values ​​are the same and the attribute values ​​are the same, a random attribute can be selected.

[0213] 4) Select the attribute a i Put it into the reduced set red and delete attribute a from C i , execute formula (2);

[0214] 5) If U is Go to step 7, otherwise, go to step 6;

[0215] 6) Combine the attributes in the reduced set red with the remaining attributes in C, and repeat steps 2 to 5;

[0216] 7) Output the reduced set red, and the reduction calculation is completed.

[0217] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0218] The various embodiments in the present disclosure are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0219] The scope of protection of the present disclosure is not limited to the above-described embodiments. Obviously, those skilled in the art may make various modifications and variations to the present disclosure without departing from the scope and spirit of the present disclosure. If such modifications and variations fall within the scope of the claims of the present disclosure and their equivalents, the present disclosure is intended to include such modifications and variations.

Claims

1. A fault diagnosis method based on deep spatiotemporal convolutional network ST-NN, characterized in that the method include: Obtaining raw oil spectrum and / or iron spectrum data, as well as raw vibration signals; The original oil spectrum and / or iron spectrum data are subjected to dimensionality reduction using a rough set attribute reduction method to remove redundant features, thereby obtaining a simplified oil and / or iron spectrum feature set; The original vibration signal is subjected to time-frequency conversion using continuous wavelet transform to obtain a time-frequency graph sequence; The simplified oil and / or iron spectrum feature set is used as an additional feature channel to perform a three-dimensional convolutional neural network fusion with the time-frequency graph sequence, and a self-attention mechanism is introduced to dynamically weight the time steps to obtain a health indicator sequence; According to the health indicator sequence, a Gaussian process regression model is used to predict the fault trend and obtain a fault prediction result.

2. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 1 is characterized in that: The step of reducing the dimension of the original oil spectrum and / or ferrogram data by using a rough set attribute reduction method to remove redundant features to obtain a simplified oil and / or ferrogram feature set includes: The continuous attributes of the original oil spectrum and / or ferrogram data are discretized by equal-frequency binning to generate a discretized decision system S = (U, R, V, f), where U represents the object set, R = C∪D is the attribute set, C is the oil spectrum and / or ferrogram condition attribute subset, D is the fault state decision attribute subset, V represents the attribute value range, and f represents the information function that determines the attribute value of the sample, f = U × R → V; Calculating all subsets of single attributes and key combination attributes in the discretized decision system S to obtain positive domains of all single attributes and key combination attributes, wherein the positive domains are obtained by dividing equivalence classes through indiscernibility relations; According to the positive domains of all single attributes and key combination attributes, heuristic attribute reduction is performed to obtain the reduced oil and / or ferrogram feature set.

3. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 1 is characterized in that: The step of performing time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency graph sequence includes: Construct a basic wavelet function, wherein the scaling and translation calculation formula of the basic wavelet function ψ(t) is: Among them, ψ, represents the mother wavelet, k, represents the scale parameter, l, represents the translation parameter, and t represents time; For the original vibration signal, continuous wavelet transform is used to perform time-frequency conversion to obtain a calculation result, wherein the calculation formula of the continuous wavelet transform is: in, represents the conjugate function of the wavelet function, CWT(s)(k, l) represents the continuous wavelet transform coefficient matrix, s(t) represents the original vibration signal, and dt represents the integral variable; The calculation results are normalized to obtain a time-frequency graph sequence.

4. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 1, characterized in that: The step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes: The simplified oil and / or iron spectrum feature set is used as an independent channel and expanded to the same resolution as the time-frequency graph sequence by a spatial interpolation method to form an additional feature channel; The additional feature channel is fused with the time-frequency graph sequence to generate a fusion input tensor with a shape of (N, H, W, L, C), where N represents the number of samples, H represents the height of the time-frequency graph, W represents the width of the time-frequency graph, L represents the number of overlaps of the time-frequency graph sequence, and C represents the number of channels of the time-frequency graph sequence.

5. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 4 is characterized in that: The step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes: Construct at least three 3D convolutional layers, each using the ReLU activation function to perform a unified nonlinear transformation on the input time-frequency graph sequence and the simplified oil and / or ferrogram features to extract local correlation features in the time-frequency domain; The fused input tensor is input into the three 3D convolutional layers. A ReLU activation function is used in each 3D convolutional layer to perform a nonlinear transformation on the fused input tensor. The multi-source features of adjacent time-frequency graphs are associated in the time dimension through the 3D convolutional layer to obtain associated multi-source fusion features, wherein the multi-source fusion features include vibration time-frequency graph, oil and iron spectrum features.

6. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 5 is characterized in that: The step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes: After the at least three 3D convolutional layers, two fully connected layers are constructed, and a Dropout operation is performed between the two fully connected layers, wherein the fully connected layers include a first fully connected layer that uses a ReLU activation function and a second fully connected layer that uses a Sigmoid activation function to output a health indicator; Inputting the associated multi-source fusion features into the first fully connected layer, performing dimensionality reduction and nonlinear activation using a ReLU activation function, and obtaining features processed by the first fully connected layer; Perform a Dropout operation on the features processed by the first fully connected layer, randomly discarding neurons with a preset probability to obtain non-overfitted Dropout-processed features; The Dropout processed features are input into the second fully connected layer, and the Sigmoid activation function is used to map the features into a health indicator sequence to obtain a health indicator sequence.

7. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 6 is characterized in that: The step of using the simplified oil and / or iron spectrum feature set as an additional feature channel, fusing it with the time-frequency graph sequence through a three-dimensional convolutional neural network, and introducing a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence includes: The associated multi-source fusion features are linearly transformed through a trainable parameter matrix and mapped into a Q / K / V matrix, wherein the shape of the associated multi-source fusion features is (N, T, d), N represents the number of samples, T represents the number of time steps, and d represents the feature dimension. The Q / K / V matrix includes a query matrix Q, a key matrix K, and a value matrix V. The formula of the Q / K / V matrix is: Q=HI·W Q ; K=HI·W K ; V=HI·W v ; in, W Q 、W K 、W v All are trainable parameters in the trainable parameter matrix; By scaling the dot product, the influence weight of each time step on other time steps is calculated and normalized to generate a probability distribution weight matrix; The probability distribution weight matrix is ​​used to perform weighted summation on the value matrix V to obtain the optimized feature representation, wherein the optimized feature representation formula is: Z = A·V; Among them, Z represents the optimized feature representation, A represents the probability distribution weight matrix; The optimized feature representation is input into the fully connected layer to generate a health indicator sequence, and the trainable parameters of the three-dimensional convolutional neural network and the self-attention mechanism are jointly optimized through back propagation.

8. The fault diagnosis method based on deep spatiotemporal convolutional network (ST-NN) according to claim 7 is characterized in that: The step of using a Gaussian process regression model to predict fault trends based on the health indicator sequence to obtain a fault prediction result includes: Build a Gaussian process regression model; The health indicator sequence is segmented according to the preset time window and normalized to obtain the segmentation results; The segmentation results are input into a Gaussian process regression model to obtain a fault prediction result.

9. A fault diagnosis system based on deep spatiotemporal convolutional network (ST-NN), characterized by: The fault diagnosis method based on the deep spatiotemporal convolutional network (ST-NN) according to any one of claims 1 to 8 is applied, and the system includes: an acquisition unit, configured to acquire raw oil spectrum and / or iron spectrum data, and raw vibration signals; A reduction element is used to reduce the dimension of the original oil spectrum and / or iron spectrum data by using a rough set attribute reduction method to remove redundant features and obtain a reduced oil and / or iron spectrum feature set; A time-frequency unit, configured to perform time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency graph sequence; a fusion unit, configured to use the simplified oil and / or iron spectrum feature set as an additional feature channel, perform a three-dimensional convolutional neural network fusion with the time-frequency graph sequence, and introduce a self-attention mechanism to dynamically weight the time steps to obtain a health indicator sequence; The result unit is used to predict the fault trend based on the health indicator sequence using a Gaussian process regression model to obtain a fault prediction result.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images