Intelligent detection method, system, terminal and medium for key equipment failure of ship

Abstract

The application provides a kind of ship key equipment failure intelligent detection method, system, terminal and medium, first acquisition ship key equipment's original vibration signal, and extract time-frequency domain statistical characteristics, and generate structured text prompt;Rely on lightweight signal encoder to complete vibration signal depth feature extraction, output signal embedding vector.Original vibration signal is converted into time-frequency image at the same time. Signal embedding vector, time-frequency image and field text prompt complete cross-modal feature alignment, combined with instruction fine-tuning strategy to build intelligent diagnosis model.In practical application, the vibration signal to be measured generates multi-modal representation input model, which can accurately output fault diagnosis result and explanatory text.The application deeply integrates multi-modal information, has cross-condition generalization and natural language interpretability, significantly improves the diagnosis accuracy and robustness, and has important engineering value for ensuring ship operation safety and reducing life cycle operation and maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of marine machinery fault detection technology, and in particular to an intelligent fault detection method, system, terminal and medium for key marine equipment. Background Technology

[0002] In complex marine environments, equipment is subjected to harsh conditions such as variable loads, high noise, and salt spray corrosion over long periods, making it prone to various failure modes, including bearing wear, gear pitting, rotor imbalance, and shaft misalignment. If these failures are not detected and diagnosed in a timely manner, they may lead to sudden equipment shutdown and loss of power.

[0003] Currently, fault diagnosis of critical ship equipment mainly employs traditional signal processing methods and conventional intelligent diagnostic techniques. Traditional methods, such as Fourier transform, short-time Fourier transform, and wavelet packet decomposition, rely on manually designed feature extractors and have limited capabilities in processing vibration signals under non-stationary and noisy backgrounds, making it difficult to simultaneously capture the low-frequency trend components and high-frequency impact components caused by faults. Conventional intelligent diagnostic techniques, such as support vector machines, random forests, and shallow convolutional neural networks, while achieving certain results in laboratory settings or under single operating conditions, suffer from insufficient model generalization ability and a sharp decline in cross-condition and cross-equipment transfer performance in actual ship applications due to significant differences in data distribution under different equipment, speeds, and loads. Furthermore, most existing intelligent diagnostic methods only output fault category labels and cannot provide diagnostic evidence. Ship maintenance personnel not only need to know "what the fault is," but also need to understand "why this fault is," in order to formulate accurate maintenance strategies. The uninterpretable nature of black-box models severely limits their practical deployment and personnel trust in the health management of critical ship equipment. On the other hand, existing methods often only utilize single-modal data, ignoring the complementary information between vibration signals, time-frequency images, and domain prior knowledge, failing to fully explore the potential of multimodal synergy, resulting in bottlenecks in diagnostic accuracy and robustness. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide an intelligent detection method, system, terminal and medium for critical ship equipment faults, which solves the technical problems of existing critical ship equipment fault diagnosis schemes, such as insufficient ability to process non-stationary strong noise signals under complex and harsh marine conditions, poor generalization and transferability of models across operating conditions and equipment, lack of interpretability, failure to integrate multimodal data and domain knowledge, and limitations in diagnostic accuracy, practicality and application.

[0005] To achieve the above and other related objectives, this invention provides an intelligent fault detection method for key ship equipment. The method includes: acquiring raw vibration signals from key ship equipment under different operating conditions to form multiple raw vibration signal samples labeled with fault categories; extracting time-frequency domain statistical features from each raw vibration signal sample and selecting features sensitive to fault types to construct structured text prompts; extracting features from each raw vibration signal sample using a lightweight signal encoder to obtain signal embedding vectors; wherein the lightweight signal encoder is pre-trained through supervised contrastive learning; performing time-frequency conversion on each raw vibration signal sample using continuous wavelet transform to obtain a time-frequency image; performing cross-modal embedding alignment on the signal embedding vector, time-frequency image, and structured text prompts of each raw vibration signal sample, and training a Qwen3-VL multimodal large model using instruction fine-tuning to construct the intelligent fault detection model; and generating fault diagnosis results and explanatory text based on the intelligent fault detection model, according to the signal embedding vector, time-frequency image, and structured text prompts obtained from the raw vibration signals to be detected.

[0006] In one embodiment of the present invention, the step of extracting time-frequency domain statistical features for each original vibration signal sample and selecting features sensitive to fault types to construct a structured text prompt includes: extracting time-domain and frequency-domain statistical features for each original vibration signal sample; wherein, the time-domain statistical features include: mean, standard deviation, variance, root mean square, kurtosis, skewness, peak-to-peak value, peak factor, morphological factor, and impulse factor; the frequency statistical features include: frequency mean, frequency variance, frequency standard deviation, root mean square frequency, frequency skewness, frequency kurtosis, regularity, coefficient of variation, eighth moment, and sixteenth moment; using the mutual information method to calculate the mutual information value between each statistical feature and the fault category label, selecting a preset number of statistical features with the largest mutual information value as features sensitive to fault types, and establishing a sensitive feature subset; generating key-value pairs of feature names and values ​​for each feature in the sensitive feature subset, and connecting all key-value pairs through a preset delimiter to form a structured text prompt.

[0007] In one embodiment of the present invention, the lightweight signal encoder includes: an input convolutional layer for preliminary feature extraction of the input raw vibration signal; a temporal convolutional network module comprising multiple sequentially connected temporal convolutional blocks, each temporal convolutional block comprising at least two causal dilated convolutional layers and a layer normalization layer, wherein the dilation rate of each temporal convolutional block increases with the block number, and a downsampling layer is provided between adjacent temporal convolutional blocks; a learnable temporal feature aggregation module connected to the output of the temporal convolutional network module for weighted aggregation of the multi-channel features output by the temporal convolutional network module through an attention mechanism to generate a fixed number of feature vectors; and a projection layer for mapping the feature vectors to the embedding dimension of the Qwen3-VL multimodal large model to obtain a signal embedding vector.

[0008] In one embodiment of the present invention, the pre-training method of the lightweight signal encoder includes: inputting each original vibration signal sample into the lightweight signal encoder; pooling the signal embedding vector output by the lightweight signal encoder along the vector dimension to obtain a global feature vector; mapping it to a low-dimensional space through a projection head and normalizing it; and optimizing the network parameters of the lightweight signal encoder and the projection head using a supervised contrastive loss function so that the distance between positive sample pairs in the low-dimensional feature space is less than the distance between negative sample pairs; wherein, original vibration signal samples belonging to the same fault category label are regarded as positive sample pairs, and original vibration signal samples not belonging to the same fault category label are regarded as negative sample pairs.

[0009] In one embodiment of the present invention, the step of using continuous wavelet transform to perform time-frequency conversion on each original vibration signal sample to obtain a time-frequency image includes: selecting Morlet wavelet as wavelet basis function to perform time-frequency conversion on each original vibration signal sample to obtain a two-dimensional time-frequency image.

[0010] In one embodiment of the present invention, the step of performing cross-modal embedding alignment on the signal embedding vector, time-frequency image, and structured text prompt of each original vibration signal sample, and training the Qwen3-VL multimodal large model using an instruction fine-tuning method to construct the fault intelligent detection model includes: inputting the time-frequency image of each original vibration signal sample into the visual encoder of the Qwen3-VL multimodal large model to extract a visual embedding; inputting the structured text prompt of each original vibration signal sample into the text embedding layer of the Qwen3-VL multimodal large model to extract a text embedding; concatenating the signal embedding vector, visual embedding, and text embedding of each original vibration signal sample along the sequence dimension to form a self-attention input sequence; and inputting the self-attention input sequence of each original vibration signal sample and the fault category label into the Qwen3-VL multimodal large model to train the fault intelligent detection model using an instruction fine-tuning method.

[0011] In one embodiment of the present invention, the step of generating fault diagnosis results and explanatory text based on the fault intelligent detection model, according to the signal embedding vector, time-frequency image, and structured text prompts obtained from the original vibration signal to be detected, includes: acquiring the original vibration signal to be detected; extracting time-frequency domain statistical features from the original vibration signal and selecting features sensitive to the fault type to construct structured text prompts; extracting features from the original vibration signal using a lightweight signal encoder to obtain a signal embedding vector; performing time-frequency conversion on the original vibration signal using continuous wavelet transform to obtain a time-frequency image; performing cross-modal embedding alignment on the signal embedding vector, time-frequency image, and structured text prompts to obtain a self-attention input sequence; and inputting the self-attention input sequence into the fault intelligent detection model to generate fault diagnosis results and explanatory texts.

[0012] To achieve the above and other related objectives, this invention provides an intelligent fault detection system for critical ship equipment. The system includes: a signal acquisition module for acquiring raw vibration signals from critical ship equipment under different operating conditions, forming multiple raw vibration signal samples labeled with fault categories; a structured text prompt generation module for extracting time-frequency domain statistical features from each raw vibration signal sample, filtering out features sensitive to fault types, and constructing structured text prompts; and a signal embedding vector generation module for extracting features from each raw vibration signal sample using a lightweight signal encoder to obtain a signal embedding vector. The lightweight signal encoder monitors... The system employs a supervised comparative learning pre-training method; a time-frequency image generation module, which performs time-frequency conversion on each original vibration signal sample using continuous wavelet transform to obtain a time-frequency image; a fault intelligent detection model construction module, which performs cross-modal embedding alignment on the signal embedding vector, time-frequency image, and structured text prompts of each original vibration signal sample, and trains the Qwen3-VL multimodal large model using instruction fine-tuning to construct the fault intelligent detection model; and a fault intelligent detection module, which generates fault diagnosis results and explanatory text based on the fault intelligent detection model, according to the signal embedding vector, time-frequency image, and structured text prompts obtained from the original vibration signal to be detected.

[0013] To achieve the above and other related objectives, the present invention provides an electronic terminal, comprising: one or more memories and one or more processors; the one or more memories are used to store a computer program; the one or more processors are connected to the memories and are used to run the computer program to execute the method.

[0014] To achieve the above and other related objectives, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0015] As described above, this invention provides an intelligent fault detection method, system, terminal, and medium for critical ship equipment, offering the following advantages: First, it collects raw vibration signals from critical ship equipment, extracts time-frequency domain statistical features, filters highly sensitive fault features, and generates structured text prompts. Then, it utilizes a lightweight signal encoder to extract deep features from the vibration signals, outputting a signal embedding vector. Simultaneously, continuous wavelet transform is employed to convert the raw vibration signals into time-frequency images. The signal embedding vector, time-frequency images, and domain text prompts are aligned across modal features. A Qwen3-VL multimodal large-scale model is trained using a command fine-tuning strategy to build an intelligent diagnostic model. In practical applications, the multimodal representation input model is generated from the vibration signal under test, enabling accurate output of fault diagnosis results and explanatory text. This invention deeply integrates multimodal information, possesses cross-condition generalization and natural language interpretability, significantly improving diagnostic accuracy and robustness, and has significant engineering value for ensuring ship operational safety and reducing life-cycle maintenance costs. Attached Figure Description

[0016] Figure 1 The diagram shown is a flowchart illustrating an intelligent fault detection method for key ship equipment according to an embodiment of the present invention.

[0017] Figure 2 The diagram shown is a schematic diagram of the intelligent fault detection process in one embodiment of the present invention.

[0018] Figure 3 The diagram shows the sorting of mutual information values ​​for 20 time-frequency domain statistical features in one embodiment of the present invention.

[0019] Figure 4 The image shown is a visualization of the t-SNE before and after using mutual information to filter features in one embodiment of the present invention.

[0020] Wherein, (a) is a visualization of the t-SNE after using mutual information to filter features in one embodiment of the present invention. (b) is a visualization of the t-SNE before using mutual information to filter features in one embodiment of the present invention.

[0021] Figure 5 The image shown is a visualization of the t-SNE of signal features after supervised contrastive learning pre-training in one embodiment of the present invention.

[0022] Figure 6 The diagram shown is a schematic of the instruction fine-tuning process based on the Qwen3-VL multimodal large model in one embodiment of the present invention.

[0023] Figure 7 The diagram shows a confusion matrix of the model on the test set in one embodiment of the present invention.

[0024] Figure 8 The diagram shown is a structural schematic of a ship critical equipment fault intelligent detection system according to an embodiment of the present invention.

[0025] Figure 9 The diagram shown is a structural schematic of an electronic terminal according to an embodiment of the present invention. Detailed Implementation

[0026] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0027] It should be noted that in the following description, reference is made to the accompanying drawings, which illustrate several embodiments of the invention. It should be understood that other embodiments may also be used, and changes in mechanical composition, structure, electrical system, and operation may be made without departing from the spirit and scope of the invention. The following detailed description should not be considered limiting, and the scope of the embodiments of the invention is defined only by the claims of the published patents. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. Spatially related terms, such as “upper,” “lower,” “left,” “right,” “below,” “below,” “lower part,” “above,” “upper part,” etc., may be used herein to illustrate the relationship between one element or feature shown in the figures and another element or feature.

[0028] Throughout this specification, when it is said that a part is "connected" to another part, this includes not only "direct connection" but also "indirect connection" by placing other elements in between. Furthermore, when it is said that a part "includes" a certain constituent element, unless otherwise stated otherwise, this does not exclude other constituent elements, but rather means that other constituent elements may also be included.

[0029] The terms "first," "second," and "third," etc., used herein are for the purpose of describing various parts, components, regions, layers, and / or segments, but are not limiting. These terms are used only to distinguish one part, component, region, layer, or segment from others. Therefore, the "first part," "component," "region," "layer," or "segment" described below may refer to a "second part," "component," "region," "layer," or "segment" without departing from the scope of this invention.

[0030] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are interpreted as inclusive, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition arise only when combinations of elements, functions, or operations are inherently mutually exclusive in some manner.

[0031] This invention provides an intelligent fault detection method for critical ship equipment. First, it collects the raw vibration signals of the critical ship equipment and extracts time-frequency domain statistical features. Highly sensitive fault features are then selected and structured text prompts are generated. A lightweight signal encoder is used to extract deep features from the vibration signals, outputting a signal embedding vector. Simultaneously, continuous wavelet transform is employed to convert the raw vibration signals into a time-frequency image. The signal embedding vector, time-frequency image, and domain text prompts are aligned across modal features. A Qwen3-VL multimodal large-scale model is trained using a command fine-tuning strategy to build an intelligent diagnostic model. In practical applications, the multimodal representation input model is generated from the vibration signal under test, enabling accurate output of fault diagnosis results and explanatory text. This invention deeply integrates multimodal information, possesses cross-condition generalization and natural language interpretability, significantly improving diagnostic accuracy and robustness. It has significant engineering value for ensuring ship operational safety and reducing life-cycle maintenance costs.

[0032] The present invention will now be described in detail with reference to the accompanying drawings, so that those skilled in the art can readily implement it. The present invention can be embodied in many different forms and is not limited to the embodiments described herein.

[0033] like Figure 1 This is a flowchart illustrating an intelligent fault detection method for key ship equipment according to an embodiment of the present invention.

[0034] The method includes:

[0035] Step S1: Acquire raw vibration signals of key ship equipment under different operating conditions and form multiple raw vibration signal samples labeled with fault categories.

[0036] In one embodiment, vibration sensors are deployed to collect raw vibration data of key ship equipment under normal operating conditions and various typical fault operating conditions. The raw vibration signals collected under a single operating condition are divided into independent raw vibration signal samples, and each sample is labeled with the corresponding fault category label to complete the initial construction of a labeled sample dataset.

[0037] Preferably, the vibration sensor is a piezoelectric accelerometer, which can stably adapt to complex vibration conditions of ships and has strong resistance to environmental interference. The signal sampling frequency is matched and set according to the actual operating speed of the ship's key equipment, reasonably covering the entire frequency range of equipment vibration, and can completely capture the low-frequency vibration component and high-frequency impact component corresponding to the fault characteristics, ensuring that the collected signal can fully carry the equipment operating status information.

[0038] Step S2: Extract time-frequency domain statistical features for each original vibration signal sample, and select features that are sensitive to fault types to construct structured text prompts.

[0039] In one embodiment, step S2 includes:

[0040] For each original vibration signal sample, time-domain and frequency-domain statistical features are extracted. The time-domain statistical features include: mean, standard deviation, variance, root mean square, kurtosis, skewness, peak-to-peak value, peak factor, morphological factor, and impulse factor. The frequency statistical features include: frequency mean, frequency variance, frequency standard deviation, root mean square frequency, frequency skewness, frequency kurtosis, regularity, coefficient of variation, eighth moment, and sixteenth moment.

[0041] The specific formulas and descriptions of the extracted time-frequency domain statistical features are shown in Table 1:

[0042] Table 1: Formulas and explanations for time-domain and frequency-domain statistical characteristics

[0043]

[0044] Where N is the total number of vibration signal sampling points in a single sample, n is the sampling point number, taking values ​​from 1, 2, 3…N, x(n) is the original vibration signal amplitude corresponding to the nth sampling point, K is the total number of frequency points after spectrum analysis, and k is the frequency point number, taking values ​​from 1, 2, 3…K. s(k) represents the actual frequency value corresponding to the k-th frequency point, and s(k) represents the spectral amplitude corresponding to the k-th frequency point.

[0045] After extracting all time-frequency domain statistical features, mutual information analysis is used to filter features, thereby eliminating redundant and invalid features and retaining features strongly correlated with faults.

[0046] The mutual information method is used to calculate the mutual information value between each statistical feature and the fault category label. The statistical features with the largest number of mutual information values ​​are selected as the features sensitive to the fault type, and a subset of sensitive features is established.

[0047] Specifically, the mutual information method is used to calculate the mutual information value between each statistical feature and the fault category label. The calculation formula is as follows:

[0048] ; (1)

[0049] Where X is a certain statistical feature, Y is the fault category label, p(x,y) is the joint probability distribution, and p(x) and p(y) are the marginal probability distributions, respectively.

[0050] Sort the mutual information values ​​corresponding to all features in descending order, and select the top K features to form a fault-sensitive feature subset, where K is a positive integer pre-set according to the actual detection accuracy requirements. Only the core features with high distinguishability of ship equipment fault categories are retained.

[0051] Generate key-value pairs for each feature name and value in the sensitive feature subset, and concatenate all key-value pairs using a preset delimiter to form a structured text prompt. Specifically, generate standard key-value pairs for each sensitive feature according to the uniform format of feature name = feature value, and then concatenate all key-value pairs sequentially using a preset uniform delimiter. Finally, generate a structured text prompt that is formatted correctly, semantically clear, and can be directly input into a multimodal model, thus completing this step of the processing flow.

[0052] Step S3: Use a lightweight signal encoder to extract features from each original vibration signal sample to obtain a signal embedding vector.

[0053] In one embodiment, the lightweight signal encoder includes:

[0054] The input convolutional layer is used to perform preliminary feature extraction on the input raw vibration signal;

[0055] The Temporal Convolutional Network (TCN) module contains multiple sequentially connected temporal convolutional blocks. Each temporal convolutional block contains at least two causal dilated convolutional layers and a layer normalization layer. The dilation rate of each temporal convolutional block increases with the block number. A downsampling layer is provided between adjacent temporal convolutional blocks.

[0056] The learnable temporal feature aggregation module (Token Learner) is connected to the output of the temporal convolutional network module and is used to perform weighted aggregation of the multi-channel features output by the temporal convolutional network module through an attention mechanism to generate a fixed number of feature vectors.

[0057] The projection layer is used to map the feature vectors to the embedding dimension of the Qwen3-VL multimodal large model to obtain the signal embedding vector.

[0058] Preferably, the lightweight signal encoder has the following specific structural parameters as shown in Table 2:

[0059] Table 2: Specific Model Parameters of Signal Embedding Module

[0060]

[0061] In one embodiment, in order to enhance the intra-class compactness and inter-class separability of signal features, supervised contrastive learning is used to pre-train the signal encoder.

[0062] The process of supervised contrastive learning pre-training includes:

[0063] Construct training batches. Define any two different vibration signal samples belonging to the same fault category as positive sample pairs, and define any two vibration signal samples belonging to different fault categories as negative sample pairs, thereby constructing the positive and negative sample constraint relationship required for supervised contrastive learning.

[0064] Each original vibration signal sample in the training batch is sequentially input into a lightweight signal encoder to obtain the corresponding signal embedding vector. Global pooling is performed on the signal embedding vector along the feature dimension to fuse temporal information and obtain a fixed-dimensional global feature vector. The global feature vector is input into a projection head network to be mapped to a low-dimensional contrast feature space, and normalization is performed on the mapped feature vector.

[0065] Using the supervised contrastive loss function as the optimization objective, the backpropagation algorithm is used to iteratively update the network parameters of the lightweight signal encoder and the projector. During training, the feature distribution objective is continuously constrained: to make the feature distance of positive sample pairs of the same fault category smaller and the features more compact in the low-dimensional feature space, and to make the feature distance of negative sample pairs of different fault categories larger and the boundaries clearer in the low-dimensional feature space. This allows the encoder to learn signal feature representations with strong discriminative ability for fault categories, thereby improving the recognition accuracy and generalization ability of subsequent fault detection models.

[0066] The supervised comparison loss function is defined as follows:

[0067] ; (2)

[0068] In the formula, N is the batch size, zi is the normalized projection vector of sample i, P(i) is the set of positive sample indices, A(i) is the set of all sample indices except itself, and τ is the temperature hyperparameter.

[0069] Step S4: Perform time-frequency conversion on each original vibration signal sample using continuous wavelet transform to obtain a time-frequency image.

[0070] In one embodiment, step S4 includes:

[0071] Morlet wavelet is selected as the wavelet basis function to complete the transformation of one-dimensional vibration signal to the time-frequency domain, realizing the joint representation of time-domain and frequency-domain information of non-stationary vibration signal. Among them, Morlet wavelet is a commonly used complex-valued wavelet basis function with excellent time-domain and frequency-domain localization characteristics. It is suitable for performing continuous wavelet transform on non-stationary signals such as ship vibration to generate time-frequency images and characterize the time-frequency energy features of faults.

[0072] The continuous wavelet transform uses Morlet wavelets as wavelet basis functions, and the transform formula is as follows:

[0073] (3)

[0074] In the formula, x(t) is the original vibration signal, ψ(t) is the Morlet wavelet basis function, a is the scaling factor, and b is the translation factor. By using a preset scaling range and translation step size, the one-dimensional vibration signal is mapped into a two-dimensional time-frequency image to preserve the time-frequency energy distribution characteristics of the fault signal.

[0075] It should be noted that steps S2, S3, and S4 can be performed simultaneously or sequentially, and this application does not impose any restrictions on this.

[0076] Step S5: Align the signal embedding vector, time-frequency image, and structured text prompts of each original vibration signal sample across modal embeddings, and train the Qwen3-VL multimodal large model using instruction fine-tuning to construct the fault intelligent detection model.

[0077] In one embodiment, step S5 includes:

[0078] The time-frequency image of each original vibration signal sample is input into the visual encoder of the Qwen3-VL multimodal large model. After visual feature encoding processing, the corresponding visual embedding is output. Among them, Qwen3-VL (Qwen 3 Vision-Language) supports image, text and sequence feature input at the same time, and can perform cross-modal understanding, fault classification, result generation and interpretable text output.

[0079] The structured text prompts for each original vibration signal sample are input into the text embedding layer of the Qwen3-VL multimodal large model to complete the text semantic encoding and extract the text embedding.

[0080] The signal embedding vector, visual embedding, and text embedding of each original vibration signal sample are concatenated along the sequence dimension to form a self-attention input sequence:

[0081] ; (4)

[0082] In the formula, For signal embedding vectors, For visual embedding, For text embedding, This is the unified input sequence that is ultimately fed into the backbone network of the Qwen3-VL multimodal large model.

[0083] The self-attention input sequence of each original vibration signal sample and the corresponding fault category label are input into the Qwen3-VL multimodal large model backbone network and trained by instruction fine-tuning to obtain the fault intelligent detection model.

[0084] Step S6: Based on the fault intelligent detection model, generate fault diagnosis results and explanatory text according to the signal embedding vector, time-frequency image and structured text prompts obtained from the original vibration signal to be detected.

[0085] In one embodiment, such as Figure 2 Step S6 includes:

[0086] Acquire raw vibration signals to be detected from key ship equipment;

[0087] Extract time-frequency domain statistical features from the original vibration signal, and filter out features sensitive to fault types to construct a structured text prompt; following the same processing method as step S2, extract various time-domain and frequency-domain statistical features of the vibration signal to be detected, filter out fault-sensitive features with the help of mutual information, and assemble them into a structured text prompt according to a predetermined format;

[0088] The original vibration signal is feature extracted using a lightweight signal encoder to obtain a signal embedding vector; the lightweight signal encoder pre-trained by supervised contrastive learning in step S3 is used to perform deep temporal feature extraction on the original vibration signal to be detected and output the corresponding signal embedding vector.

[0089] The original vibration signal is converted to a time-frequency image by continuous wavelet transform; this step is similar to step S4 above; following the transformation rules of step S4, the Morlet wavelet basis function is selected to perform continuous wavelet transform on the original vibration signal, converting the one-dimensional time domain signal into a two-dimensional time-frequency image that can reflect the energy distribution characteristics;

[0090] The signal embedding vector, time-frequency image, and structured text prompt are cross-modal embedded and aligned to obtain a self-attention input sequence. According to the cross-modal embedding alignment and sequence concatenation method in step S5, visual embedding and text embedding are extracted respectively, and then concatenated and combined with the signal embedding vector after unifying the dimension to generate a self-attention input sequence that adapts to the model input.

[0091] The constructed self-attention input sequence is fed into the trained intelligent fault detection model for inference operations. The model integrates multimodal features to complete state discrimination and fault category determination, and finally outputs a clear fault diagnosis result. At the same time, it automatically generates explanatory text with feature basis, fault cause and operation and maintenance reference value by combining various feature data, and completes online intelligent fault identification and interpretable diagnosis of key ship equipment.

[0092] To better describe the intelligent fault detection method for key ship equipment, it will now be explained in conjunction with specific embodiments.

[0093] Example 1: An intelligent detection method for bearing failure of a ship's main lubricating oil pump motor.

[0094] The method includes:

[0095] Step 1: Data Acquisition.

[0096] Piezoelectric accelerometers were installed at the bearing housings at the input and output ends of the main lubricating oil pump motor. The sampling frequency was set to 50kHz to collect vibration signals under the following conditions: normal motor operation, bearing inner ring fault, bearing outer ring fault, and bearing rolling element fault. Each sample had 2048 data points and was labeled with a fault category.

[0097] Step 2: Extraction and screening of time-frequency domain statistical features.

[0098] For each vibration signal sample collected, 20 time-frequency domain statistical features were extracted, including time-domain features (mean, standard deviation, variance, root mean square, kurtosis, skewness, peak-to-peak value, peak factor, morphological factor, impulse factor) and frequency-domain features (frequency mean, frequency variance, frequency standard deviation, frequency root mean square, frequency skewness, frequency kurtosis, regularity, coefficient of variation, eighth moment, sixteenth moment). The specific calculation formulas are shown in Table 1.

[0099] Then, the mutual information method is used to calculate the mutual information value between each statistical feature and the fault category label. These values ​​are then sorted from largest to smallest, and the top five features most sensitive to faults are selected to form a subset of sensitive features. For example... Figure 3As shown, the mutual information values ​​of the 20 time-frequency domain statistical features are ranked. To further verify the effectiveness of feature selection, all 20 features and the 5 selected features are subjected to t-SNE (t-distributed random neighborhood embedding) dimensionality reduction and visualization, as shown in the figure. Figure 4 As shown, the filtered features result in better intra-class compactness and inter-class separability for different fault categories in the low-dimensional space. The name and value of each filtered feature are generated as key-value pairs in the format "feature name=value", connected by commas to form structured text hints, such as "mean=0.023, kurtosis=5.672, root mean square frequency=0.458", thus injecting prior domain knowledge.

[0100] Step 3: Lightweight signal encoding and contrastive learning pre-training.

[0101] Design a lightweight signal encoder with the following structure: an input convolutional layer (Conv1d), three temporal convolutional network blocks (TCN Block 1, TCN Block 2, and TCN Block 3), with average pooling downsampling between each pair of blocks, and finally a learnable temporal feature aggregation module (TokenLearner) and a linear projection layer (Projection). The TCN Block employs causal dilated convolution (Conv1d), layer normalization (LayerNorm), and the SiLU activation function, with dilation rates of 1, 2, and 4 respectively, progressively increasing the receptive field. The TokenLearner generates attention weights through two 1×1 convolutions and Softmax, adaptively aggregating the 512-channel feature map into 32 feature vectors, each with 512 dimensions. The linear projection layer maps the feature vectors to 2560 dimensions, aligning with the embedding dimension of Qwen3-VL.

[0102] To enhance the intra-class compactness and inter-class separability of signal features, supervised contrastive learning was used to pre-train the signal encoder. During pre-training, vibration signals from the training set were input into the encoder, resulting in 32 feature vectors. These vectors were then averaged along their dimensions using pooling, and finally mapped to a 128-dimensional low-dimensional space using a projector. The supervised contrastive loss function used a temperature coefficient of 0.07. Different samples from the same fault category were treated as positive pairs, and samples from different fault categories were treated as negative pairs. The optimizer used was Adam, with a learning rate of 1e-3, a batch size of 32, and 20 training epochs. After each epoch, the feature quality was evaluated on the validation set, and the model parameters with the highest validation accuracy were saved.

[0103] like Figure 5As shown, the pre-trained signal features are visualized after t-SNE dimensionality reduction. Different fault categories form clearly separated clusters in the feature space, verifying that the signal encoder has good feature discrimination ability. After pre-training, the signal encoder can convert the original vibration signal into 32 2560-dimensional signal embedding vectors.

[0104] Step 4: Time-frequency image generation.

[0105] For each original vibration signal sample, a continuous wavelet transform is used to convert it into a two-dimensional time-frequency image. The Morlet wavelet is selected as the wavelet basis function, and the number of wavelet scales is set to 128 to balance frequency resolution and computational efficiency.

[0106] The continuous wavelet transform uses Morlet wavelets as wavelet basis functions, and the transform formula is as follows:

[0107] ; (3)

[0108] In the formula, x(t) is the original vibration signal, ψ(t) is the Morlet wavelet basis function, a is the scaling factor, and b is the translation factor.

[0109] Through this transformation, the one-dimensional vibration signal is mapped into a 224×224 pixel time-frequency image to meet the input requirements of the Qwen3-VL visual encoder. Each pixel in the image represents the magnitude of the transform coefficient at the corresponding scale and time, preserving the time-frequency energy distribution characteristics of the fault signal.

[0110] Step 5: Fine-tuning of multimodal large model instructions.

[0111] like Figure 6 As shown, a fine-tuning dataset containing multimodal inputs and text labels is constructed. Each training sample includes: the original vibration signal, the corresponding time-frequency image, the structured text prompt generated in step 2, and the fault category label.

[0112] First, the pre-trained signal encoder is loaded and its parameters are frozen; then the Qwen3-VL-4B-Instruct model is loaded, and the signal encoder is integrated into the model as a pre-module. Simultaneously, the tokenizer vocabulary is expanded, adding 32 special tokens.<SIGNAL_0> to<SIGNAL_31> These are used to place placeholder signal embedding vectors in text prompts. During model forward propagation, the signal encoder converts the raw vibration signal into 32 2560-dimensional signal embedding vectors, which replace the corresponding vectors in the input sequence.<SIGNAL_i> Positional embedding enables cross-modal alignment. Time-frequency images are extracted into visual embeddings using a Qwen3-VL-4B-Instruct visual encoder, and structured text cues are extracted into text embeddings using a text embedding layer.

[0113] The three components are concatenated along the sequence dimension and then input into the backbone network of the large model. Fine-tuning is performed using the LoRA (Low-Rank Adaptation) parameter-efficient fine-tuning method. The core principle of LoRA is to add trainable parameters of low-rank decomposition next to the weight matrix of the pre-trained model. For an original weight matrix... Its update amount can be decomposed into the product of two low-rank matrices:

[0114] ; (5)

[0115] In the formula, , Since the rank r is much smaller than d and k, the original weights W0 are frozen during fine-tuning, and only matrices A and B participate in training.

[0116] The target modules are q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj, with a LoRA rank of 16, a scaling factor of 32, and dropout of 0.05. Training parameters include a batch size of 16, gradient accumulation steps of 2, a learning rate of 2e-5, and the AdamW optimizer. Training is performed for 5 epochs. After each epoch, the loss is evaluated on the validation set, and the best model is saved.

[0117] Step 6: Model testing and performance comparison.

[0118] After fine-tuning the instructions, the model's diagnostic accuracy is tested using a test set. The original vibration signals, time-frequency images, and structured text prompts of the test samples are processed according to the steps described above and input into the fine-tuned Qwen3-VL multimodal large model. The model outputs the fault category and a natural language explanation. For example... Figure 7 As shown, the confusion matrix illustrates the model's prediction results for various fault types. The average accuracy, recall, precision, and F1 score on the test set reached 99.54%. To further verify the effectiveness of this invention, a comparison with existing methods was conducted, and the results are shown in Table 3. The table shows that the method of this invention significantly outperforms the comparative methods in all evaluation metrics, demonstrating that the intelligent diagnostic method driven by multimodal information alignment and a large model has higher diagnostic accuracy and robustness.

[0119] Table 3: Comparison Results of Different Diagnostic Methods

[0120] SVM 0.8749 0.8749 0.8690 0.8693 TCN 0.9886 0.9886 0.9886 0.9886 Fine-tuning Qwen3-VL using only time-frequency graphs 0.8205 0.8132 0.8206 0.8122 The proposed method 0.9954 0.9954 0.9954 0.9954

[0121] Based on the comparative experiments described above, it can be further observed that although traditional methods have achieved high performance on this dataset, they rely on single-modal features and are unable to fully represent the multi-source information of signals under complex operating conditions. Furthermore, when fine-tuning large models using only time-frequency plots, the lack of structured statistical features and constraints from the original time-series information leads to biases in the model's capture of key information, resulting in a significant decrease in overall performance. In contrast, the proposed method achieves collaborative modeling of multimodal information by fusing the original vibration signal, time-frequency image, and statistical feature cues, enabling the model to have stronger discriminative ability among different fault types.

[0122] In summary, the proposed method not only achieves excellent performance across multiple evaluation metrics but also maintains stable output under complex conditions, demonstrating strong robustness and generalization ability. Furthermore, the multimodal fusion mechanism effectively compensates for the insufficient information from a single data source, enabling the model to better adapt to different types of inputs and providing reliable support for subsequent engineering applications.

[0123] Similar in principle to the above embodiments, the present invention provides an intelligent fault detection system for key ship equipment.

[0124] The following specific embodiments are provided in conjunction with the accompanying drawings:

[0125] like Figure 8 This diagram illustrates the structure of an intelligent fault detection system for critical ship equipment according to an embodiment of the present invention. The system includes:

[0126] Signal acquisition module 1 is used to acquire raw vibration signals of key ship equipment under different operating conditions and form multiple raw vibration signal samples labeled with fault categories.

[0127] The structured text prompt generation module 2 is connected to the signal acquisition module 1 and is used to extract time-frequency domain statistical features for each original vibration signal sample, and filter out features that are sensitive to the fault type to construct structured text prompts.

[0128] The signal embedding vector generation module 3 is connected to the signal acquisition module 1 and is used to extract features from each original vibration signal sample using a lightweight signal encoder to obtain a signal embedding vector; wherein, the lightweight signal encoder is pre-trained through supervised contrastive learning.

[0129] The time-frequency image generation module 4 is connected to the signal acquisition module 1 and is used to perform time-frequency conversion on each original vibration signal sample using continuous wavelet transform to obtain a time-frequency image.

[0130] The fault intelligent detection model construction module 5 is connected to the structured text prompt generation module 2, the signal embedding vector generation module 3, and the time-frequency image generation module 4. It is used to perform cross-modal embedding alignment of the signal embedding vector, time-frequency image, and structured text prompt of each original vibration signal sample, and to train the Qwen3-VL multimodal large model by means of instruction fine-tuning to construct the fault intelligent detection model.

[0131] The fault intelligent detection module 6 is connected to the fault intelligent detection model construction module 5, and is used to generate fault diagnosis results and explanatory text based on the fault intelligent detection model, according to the signal embedding vector, time-frequency image and structured text prompts obtained from the original vibration signal to be detected.

[0132] Since the implementation principle of the intelligent fault detection system for critical ship equipment has been described in the aforementioned embodiments, it will not be repeated here.

[0133] The intelligent fault detection method for critical ship equipment provided in this invention can be implemented on the terminal side or the server side. Regarding the hardware structure of the electronic terminal, please refer to [link to relevant documentation]. Figure 9 This is a schematic diagram of an optional hardware structure of an electronic terminal 1000 provided in an embodiment of the present invention. The electronic terminal 1000 can be a mobile phone, computer device, tablet device, personal digital processing device, factory back-end processing device, etc. The electronic terminal 1000 includes: at least one processor 1001, a memory 1002, at least one network interface 10010, and a user interface 1009. The various components in the device are coupled together through a bus system 1005. It is understood that the bus system 1005 is used to realize the connection and communication between these components. In addition to a data bus, the bus system 1005 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 9 All buses are labeled as bus systems.

[0134] The user interface 1009 may include a monitor, keyboard, mouse, trackball, clicker, button, touchpad, or touch screen.

[0135] It is understood that memory 1002 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memory.

[0136] In this embodiment of the invention, the memory 1002 is used to store various types of data to support the operation of the electronic terminal 1000. Examples of this data include: any executable program for operation on the electronic terminal 1000, such as the operating system 10021 and application program 10022; the operating system 10021 contains various system programs, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and handling hardware-based tasks. The application program 10022 may contain various applications, such as media players, browsers, etc., for implementing various application services. The intelligent fault detection method for key ship equipment provided in this embodiment of the invention can be included in the application program 10022.

[0137] The methods disclosed in the above embodiments of the present invention can be applied to or implemented by the processor 1001. The processor 1001 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 1001 or by instructions in the form of software. The processor 1001 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 1001 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor 1001 may be a microprocessor or any conventional processor, etc. The steps of the accessory optimization method provided in the embodiments of the present invention can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, which is located in a memory. The processor reads the information in the memory and combines it with its hardware to complete the steps of the aforementioned method.

[0138] In an exemplary embodiment, the electronic terminal 1000 may be used by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), or complex programmable logic devices (CPLDs) to execute the aforementioned method.

[0139] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0140] In the embodiments provided in this application, the computer-readable and writable storage medium may include read-only memory, random access memory, EEPROM, CD-ROM or other optical disc storage devices, disk storage devices or other magnetic storage devices, flash memory, USB flash drive, portable hard drive, or any other medium capable of storing desired program code in the form of instructions or data structures and accessible by a computer. Additionally, any connection may be appropriately referred to as a computer-readable medium. For example, if instructions are transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. However, it should be understood that computer-readable and writable storage media and data storage media do not include connections, carrier waves, signals, or other transient media, but are intended for non-transient, tangible storage media. The disks and optical discs used in the application include compact optical discs (CDs), laser optical discs, optical discs, digital multifunction optical discs (DVDs), floppy disks, and Blu-ray discs, where disks typically copy data magnetically, while optical discs use lasers to copy data optically.

[0141] Compared with the prior art, the present invention has the following advantages:

[0142] 1. High diagnostic accuracy: By simultaneously utilizing the temporal features of vibration signals, the visual features of time-frequency images, and the textual semantics of statistical features, the deep collaboration of three modal information is achieved through the self-attention mechanism of the Qwen3-VL large model, resulting in a diagnostic accuracy significantly higher than that of single-modal methods.

[0143] 2. Lightweight and efficient encoding: The present invention is a signal encoder based on TCN and learnable token aggregation. It has a compact structure and few parameters. Supervised contrastive learning pre-training enhances the separability of features and is suitable for edge computing or resource-constrained scenarios.

[0144] 3. Enhance generalization ability: This invention uses mutual information to filter sensitive features and constructs structured text prompts, combining physical priors with data-driven approaches to improve the model's generalization ability under limited samples and varying conditions.

[0145] 4. Enhanced interpretability: This invention utilizes the text generation capabilities of the Qwen3-VL large model to output diagnostic reasons in natural language form, making it easier for maintenance personnel to understand and make decisions.

[0146] 5. Strong engineering applicability: This invention is designed for key ship equipment and can be extended to other rotating machinery fault diagnosis scenarios, with broad application prospects.

[0147] In summary, the intelligent fault detection method, system, terminal, and medium for key ship equipment of the present invention first collects the original vibration signals of key ship equipment, extracts time-frequency domain statistical features, filters highly sensitive fault features, and generates structured text prompts. It then uses a lightweight signal encoder to complete deep feature extraction of the vibration signals and outputs a signal embedding vector. Simultaneously, continuous wavelet transform is used to convert the original vibration signals into time-frequency images. The signal embedding vector, time-frequency images, and domain text prompts are aligned across modal features. A Qwen3-VL multimodal large-scale model is trained using a command fine-tuning strategy to build an intelligent diagnostic model. In practical applications, the multimodal representation input model of the vibration signal under test can accurately output fault diagnosis results and explanatory text. This invention deeply integrates multimodal information, possesses cross-condition generalization and natural language interpretability, significantly improving diagnostic accuracy and robustness, and has significant engineering value for ensuring ship operational safety and reducing life-cycle maintenance costs. Therefore, this invention effectively overcomes the various shortcomings of existing technologies and has high industrial application value.

[0148] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for intelligent detection of faults in key ship equipment, characterized in that, The method includes: The raw vibration signals of key ship equipment under different operating conditions were acquired and generated into multiple raw vibration signal samples labeled with fault categories. Time-frequency domain statistical features are extracted for each original vibration signal sample, and features sensitive to fault type are selected to construct structured text prompts; A lightweight signal encoder is used to extract features from each original vibration signal sample to obtain a signal embedding vector; wherein, the lightweight signal encoder is pre-trained through supervised contrastive learning. Continuous wavelet transform was used to perform time-frequency conversion on each original vibration signal sample to obtain a time-frequency image; The signal embedding vector, time-frequency image, and structured text prompts of each original vibration signal sample are aligned across modal embeddings, and the Qwen3-VL multimodal large model is trained by instruction fine-tuning to build an intelligent fault detection model. Based on the aforementioned intelligent fault detection model, fault diagnosis results and explanatory text are generated according to the signal embedding vector, time-frequency image, and structured text prompts obtained from the original vibration signal to be detected. The process of extracting time-frequency domain statistical features for each original vibration signal sample and selecting features sensitive to fault types to construct structured text prompts includes: extracting time-domain and frequency-domain statistical features for each original vibration signal sample; where time-domain statistical features include: mean, standard deviation, variance, root mean square, kurtosis, skewness, peak-to-peak value, peak factor, morphological factor, and impulse factor; frequency statistical features include: frequency mean, frequency variance, frequency standard deviation, root mean square frequency, frequency skewness, frequency kurtosis, regularity, coefficient of variation, eighth moment, and sixteenth moment; using the mutual information method to calculate the mutual information value between each statistical feature and the fault category label, selecting the statistical features with the largest preset number of mutual information values ​​as features sensitive to fault types, and establishing a sensitive feature subset; generating key-value pairs of feature names and values ​​for each feature in the sensitive feature subset, and connecting all key-value pairs through a preset delimiter to form a structured text prompt; The lightweight signal encoder includes: an input convolutional layer for preliminary feature extraction of the input raw vibration signal; a temporal convolutional network module containing multiple sequentially connected temporal convolutional blocks, each temporal convolutional block containing at least two causal dilated convolutional layers and a layer normalization layer, wherein the dilation rate of each temporal convolutional block increases with the block number, and a downsampling layer is provided between adjacent temporal convolutional blocks; a learnable temporal feature aggregation module connected to the output of the temporal convolutional network module, used to perform weighted aggregation of the multi-channel features output by the temporal convolutional network module through an attention mechanism to generate a fixed number of feature vectors; and a projection layer used to map the feature vectors to the embedding dimension of the Qwen3-VL multimodal large model to obtain a signal embedding vector. The pre-training method of the lightweight signal encoder includes: inputting each original vibration signal sample into the lightweight signal encoder; pooling the signal embedding vector output by the lightweight signal encoder along the vector dimension to obtain a global feature vector; mapping it to a low-dimensional space through a projection head and normalizing it; and optimizing the network parameters of the lightweight signal encoder and the projection head using a supervised contrastive loss function so that the distance between positive sample pairs in the low-dimensional feature space is less than the distance between negative sample pairs. Specifically, original vibration signal samples belonging to the same fault category label are considered as positive sample pairs, and original vibration signal samples not belonging to the same fault category label are considered as negative sample pairs.

2. The intelligent fault detection method for key ship equipment according to claim 1, characterized in that, The step of performing time-frequency conversion on each original vibration signal sample using continuous wavelet transform to obtain a time-frequency image includes: Morlet wavelet is selected as the wavelet basis function to perform time-frequency transformation on each original vibration signal sample to obtain a two-dimensional time-frequency image.

3. The intelligent fault detection method for key ship equipment according to claim 1, characterized in that, The process of aligning the signal embedding vector, time-frequency image, and structured text prompts of each original vibration signal sample across modalities, and training the Qwen3-VL multimodal large model using a command fine-tuning method to construct an intelligent fault detection model includes: The time-frequency image of each original vibration signal sample is input into the visual encoder of the Qwen3-VL multimodal large model and extracted as a visual embedding. The structured text prompts for each original vibration signal sample are input into the text embedding layer of the Qwen3-VL multimodal large model and extracted as text embeddings. The signal embedding vector, visual embedding, and text embedding of each original vibration signal sample are concatenated along the sequence dimension to form a self-attention input sequence; The self-attention input sequence and fault category label of each original vibration signal sample are input into the Qwen3-VL multimodal large model and trained by instruction fine-tuning to obtain the fault intelligent detection model.

4. The intelligent fault detection method for key ship equipment according to claim 1, characterized in that, The generation of fault diagnosis results and explanatory text based on the fault intelligent detection model, according to the signal embedding vector, time-frequency image, and structured text prompts obtained from the original vibration signal to be detected, includes: Acquire the raw vibration signal to be detected; Time-frequency domain statistical features are extracted from the original vibration signal, and features sensitive to fault types are selected to construct structured text prompts; A lightweight signal encoder is used to extract features from the original vibration signal to obtain a signal embedding vector; The original vibration signal was converted to time-frequency using continuous wavelet transform to obtain a time-frequency image; The signal embedding vector, time-frequency image, and structured text prompt are cross-modal embedded and aligned to obtain a self-attention input sequence; The self-attention input sequence is input into the fault intelligent detection model to generate fault diagnosis results and explanatory text.

5. An intelligent fault detection system for critical ship equipment, characterized in that, The system includes: The signal acquisition module is used to acquire raw vibration signals of key ship equipment under different operating conditions and form multiple raw vibration signal samples labeled with fault categories. The structured text prompt generation module is used to extract time-frequency domain statistical features from each original vibration signal sample, and filter out features that are sensitive to fault types to construct structured text prompts. The signal embedding vector generation module is used to extract features from each original vibration signal sample using a lightweight signal encoder to obtain a signal embedding vector; wherein, the lightweight signal encoder is pre-trained through supervised contrastive learning; The time-frequency image generation module is used to perform time-frequency conversion on each original vibration signal sample using continuous wavelet transform to obtain a time-frequency image. The fault intelligent detection model construction module is used to perform cross-modal embedding alignment of the signal embedding vector, time-frequency image and structured text prompts of each original vibration signal sample, and to train the Qwen3-VL multimodal large model by means of instruction fine-tuning to construct the fault intelligent detection model. The fault intelligent detection module is used to generate fault diagnosis results and explanatory text based on the fault intelligent detection model, according to the signal embedding vector, time-frequency image and structured text prompts obtained from the original vibration signal to be detected; The process of extracting time-frequency domain statistical features for each original vibration signal sample and selecting features sensitive to fault types to construct structured text prompts includes: extracting time-domain and frequency-domain statistical features for each original vibration signal sample; where time-domain statistical features include: mean, standard deviation, variance, root mean square, kurtosis, skewness, peak-to-peak value, peak factor, morphological factor, and impulse factor; frequency statistical features include: frequency mean, frequency variance, frequency standard deviation, root mean square frequency, frequency skewness, frequency kurtosis, regularity, coefficient of variation, eighth moment, and sixteenth moment; using the mutual information method to calculate the mutual information value between each statistical feature and the fault category label, selecting the statistical features with the largest preset number of mutual information values ​​as features sensitive to fault types, and establishing a sensitive feature subset; generating key-value pairs of feature names and values ​​for each feature in the sensitive feature subset, and connecting all key-value pairs through a preset delimiter to form a structured text prompt; The lightweight signal encoder includes: an input convolutional layer for preliminary feature extraction of the input raw vibration signal; a temporal convolutional network module containing multiple sequentially connected temporal convolutional blocks, each temporal convolutional block containing at least two causal dilated convolutional layers and a layer normalization layer, wherein the dilation rate of each temporal convolutional block increases with the block number, and a downsampling layer is provided between adjacent temporal convolutional blocks; a learnable temporal feature aggregation module connected to the output of the temporal convolutional network module, used to perform weighted aggregation of the multi-channel features output by the temporal convolutional network module through an attention mechanism to generate a fixed number of feature vectors; and a projection layer used to map the feature vectors to the embedding dimension of the Qwen3-VL multimodal large model to obtain a signal embedding vector. The pre-training method of the lightweight signal encoder includes: inputting each original vibration signal sample into the lightweight signal encoder; pooling the signal embedding vector output by the lightweight signal encoder along the vector dimension to obtain a global feature vector; mapping it to a low-dimensional space through a projection head and normalizing it; and optimizing the network parameters of the lightweight signal encoder and the projection head using a supervised contrastive loss function so that the distance between positive sample pairs in the low-dimensional feature space is less than the distance between negative sample pairs. Specifically, original vibration signal samples belonging to the same fault category label are considered as positive sample pairs, and original vibration signal samples not belonging to the same fault category label are considered as negative sample pairs.

6. An electronic terminal, characterized in that, include: One or more memories and one or more processors; The one or more memories are used to store computer programs; The one or more processors are connected to the memory and are used to run the computer program to perform the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

Images