A method and device for early fault identification of ship propulsion shafting

Abstract

This invention belongs to the field of fault diagnosis technology and discloses a method and device for early fault identification of ship propulsion shafting. The steps are as follows: (1) Constructing a differential Transformer autoencoder by combining a Transformer autoencoder and a Dense decoder; (2) Mapping the features extracted from the original vibration signal and original eddy current signal of the ship propulsion shafting into a latent space using the differential Transformer autoencoder, and mining the periodic features of the latent space using a spectral feature module; (3) Increasing the distribution difference between normal and abnormal samples in the periodic features through a memory update process and a query update process, and then reconstructing the latent space into the original signal using the differential Transformer autoencoder, thereby realizing the method for early fault identification of ship propulsion shafting. This invention improves the accuracy of early fault identification of ship propulsion shafting.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and more specifically, relates to a method and device for early fault identification of ship propulsion shafting. Background Technology

[0002] The propulsion shafting system is the power source of a ship, responsible for energy transmission and providing propulsion for navigation; it is often referred to as the "heart" of the ship. Due to the complex operating environment and constantly changing conditions of ships, as well as the adverse effects of variable loads, impacts, and seawater corrosion, the performance of the propulsion shafting system gradually degrades and eventually fails. This degradation often evolves from early-stage failures, causing its operating state to shift from normal to abnormal. Early failures in the propulsion shafting system are caused by minor damage such as slight fatigue, cracks, and corrosion, resulting in a gradual deviation from its normal operating state. Early identification of these early abnormalities in the propulsion shafting system can effectively prevent downtime, reduce maintenance costs, and ensure the safe operation of the ship.

[0003] In recent years, many scholars have dedicated themselves to developing signal analysis-based and artificial intelligence-based methods. However, during ship operation, due to the complex structure of the propulsion shafting system, the harsh working environment, and the extremely weak and easily masked early fault signals, current signal analysis-based methods struggle to effectively identify early anomalies in the propulsion shafting system. Furthermore, the data generated during the operation of the propulsion shafting system contains hidden periodic information, which can effectively reveal the state evolution process of the propulsion shafting system. Current AI-based feature extraction methods cannot uncover the periodic characteristics of the propulsion shafting system, making it difficult to identify abnormal states of the propulsion shafting system in advance. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the prior art, the present invention provides a method and device for early fault identification of ship propulsion shafting, which aims to solve the problem of low accuracy in early fault identification of ship propulsion shafting.

[0005] To achieve the above objectives, according to one aspect of the present invention, an early fault identification method for a ship's propulsion shafting system is provided, the method comprising the following steps:

[0006] (1) The differential Transformer autoencoder maps the features extracted from the original vibration signal and the original eddy current signal of the ship's propulsion shaft system into a latent space, and uses the spectral feature module to mine the periodic features of the latent space; wherein, the spectral feature module includes a global Fourier transform layer and a local Fourier transform layer; the differential Transformer autoencoder is constructed by combining the Transformer autoencoder and the Dense decoder.

[0007] (2) By increasing the distribution difference between normal and abnormal samples in the periodic features through the memory update process and the query update process, the differential Transformer autoencoder reconstructs the potential space into the original signal, thereby realizing an early fault identification method for ship propulsion shafting.

[0008] Furthermore, the frequency domain information of the latent space is obtained by using the DFT of the global Fourier transform layer, and the IDFT process is simulated by two linear layers to mine the hidden global periodic features in the latent space.

[0009] Furthermore, the local Fourier transform layer uses a sliding window to divide the latent space, obtains multiple time slices, and uses its DFT to obtain the frequency domain information of the latent space. It also uses two linear layers to simulate the IDFT process to mine hidden local periodic features in the latent space.

[0010] Furthermore, by using parallel global Fourier transform layers and local Fourier transform layers, the output global periodic features and local periodic features are added together to obtain the final periodic features.

[0011] Furthermore, the query update process is used to generate synthetic memory terms, and the parameters of the latent space are updated by generating synthetic memory terms to achieve the mining of temporal representations; the memory update process obtains memory terms m by encoding the normal patterns hidden in the spectral feature module. i And use an incremental update strategy to adaptively adjust the acquired memory items, m i ∈R C , i = 1, ..., M.

[0012] Furthermore, the formula corresponding to the query update process is:

[0013]

[0014] In the formula, Let be the t-th sample in the input latent space. To synthesize memory terms, Let t be the t-th sample in the updated latent space; To check memory;

[0015] For the memory update process, an update gate θ is constructed for the query terms to determine how much information in the query terms is incorporated into the memory terms. The formula for the update gate is expressed as:

[0016]

[0017] In the formula, φ and Let σ be the parameter matrix, and σ be the Sigmoid activation function. Use conditions to query attention.

[0018] Furthermore, the reconstructed original signal is compared with the anomaly threshold, and the comparison results are used to identify early faults in the ship's propulsion shafting.

[0019] Furthermore, the state score of the sample is obtained based on the distance deviation between the acquired original vibration signal and the original eddy current signal and the potential space, and an anomaly rate of the state score is set, and then an anomaly threshold is determined based on the anomaly rate of the state score.

[0020] The present invention also provides an early fault identification system for a ship propulsion shafting system. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the early fault identification method for a ship propulsion shafting system as described above.

[0021] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the early fault identification method for ship propulsion shafting as described above.

[0022] In summary, compared with the prior art, the early fault identification method and equipment for ship propulsion shafting provided by the present invention have the following advantages:

[0023] 1. This invention combines a Transformer autoencoder and a Dense decoder to construct a differential Transformer autoencoder. The differential Transformer autoencoder can not only map the features extracted from the original vibration signal and the original eddy current signal of the ship's propulsion shaft system into a latent space, but also reconstruct the signal. At the same time, through the memory update process and the query update process, it increases the distribution difference between normal and abnormal samples in the periodic features, making the reconstructed signal closer to the actual signal. This is beneficial for realizing an early fault identification method for the ship's propulsion shaft system, while improving the accuracy of early fault identification of the ship's propulsion shaft system.

[0024] 2. The global Fourier transform layer and the local Fourier transform layer can effectively extract the global and local periodic features of the shaft system. By using a bilinear layer to simulate the inverse Fourier transform process, the periodicity representation capability of the original signal is enhanced, thereby improving the accuracy of fault identification.

[0025] 3. The query update process is used to generate synthetic memory terms and update the parameters of the latent space by generating synthetic memory terms, thereby realizing the mining of temporal representations, enhancing the temporal representation capability of the original signal, and thus improving the accuracy of fault identification.

[0026] 4. The state score of the sample is obtained based on the distance deviation between the acquired original vibration signal and the original eddy current signal and the potential space, and an anomaly rate of the state score is set. Then, an anomaly threshold is determined based on the anomaly rate of the state score. This constructs an anomaly threshold determination method, which is beneficial to improving the accuracy and stability of anomaly identification. Attached Figure Description

[0027] Figure 1 This is a flowchart of an early fault identification method for a ship propulsion shafting system provided by the present invention;

[0028] Figure 2 (a) and (b) in the text are Figure 1 A schematic diagram illustrating the principle of the spectral feature module involved in the early fault identification method for ship propulsion shafting.

[0029] Figure 3 yes Figure 1 An enhanced flowchart involving early fault identification methods for ship propulsion shafting;

[0030] Figure 4 This is an image showing the anomaly identification results obtained from an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0032] This invention provides an early fault identification method for ship propulsion shafting. The early fault identification method introduces a differential Transformer autoencoder to perform feature mining on the acquired raw signal and reconstruct the raw signal. It also introduces a spectral feature module to mine the global and local periodic features of the ship propulsion shafting and increases the distribution difference between normal and abnormal samples, thereby achieving early anomaly identification of the ship propulsion shafting.

[0033] Please see Figure 1 , Figure 2 and Figure 3 The early fault identification method mainly includes the following steps:

[0034] Step 1: Extract features from the original vibration and eddy current signals of the ship's propulsion shafting collected by multiple sensors; simultaneously, construct a differential Transformer autoencoder by combining a Transformer autoencoder and a Dense decoder.

[0035] Feature extraction is performed on the original vibration signal and original eddy current signal of the ship propulsion shaft system collected by multiple sensors to obtain six frequency domain features of the original vibration signal: average frequency amplitude, frequency gravity, frequency mean square, frequency root mean square, frequency variance, and frequency standard deviation. The radial runout value of the original eddy current signal is also obtained.

[0036] The Transformer autoencoder is used to map extracted features to a latent space, and the corresponding formula is:

[0037]

[0038] In the formula, DAS represents the differential attention score, and Q... a With Q b For the query vector, K a With K b Let V be the key vector, and d be the value vector. k Let be the transformation matrix, `head` represent the multi-head attention output, `Swish` and `GN` be the Swish function and group normalization, `h` be the latent space, and `ε`, `W0`, `W1`, `W2`, and `W` be the values ​​of the transform matrix. A This is the weight matrix.

[0039] In one implementation, the features extracted from the original vibration signal and the original eddy current signal of the ship's propulsion shafting acquired from multiple sensors are defined as... Using the input features as inputs, the Differential Transformer autoencoder uses linear layers to transform the input features into a query vector Q. a With Q b Key vector K a With K b The transformation process corresponding to the value vector V can be expressed as:

[0040]

[0041] In the formula, W Q W K W V Let be the parameter matrix. A multi-head attention mechanism is used to obtain the latent space of the input features:

[0042]

[0043] In the formula, DAS represents the differential attention score, and Q... a With Q bFor the query vector, K a With K b Let V be the key vector, and d be the value vector. k Let be the transformation matrix, head represent the output of multi-head attention, Swish and GN represent the Swish function and group normalization, respectively, h be the latent space, and ε, W0, W1, W2, and W A This is the weight matrix.

[0044] The Dense decoder uses two dense layers to decode the latent space and obtain the reconstructed signal.

[0045] The root mean square error is used to construct the reconstruction error of the differential Transformer autoencoder, which describes the difference between the reconstructed signal and the original signal:

[0046]

[0047] Step 2: The differential Transformer autoencoder maps the extracted features to a latent space and uses a spectral feature module to mine the periodic features of the latent space; wherein, the spectral feature module includes a global Fourier transform layer and a local Fourier transform layer.

[0048] Frequency domain information of the latent space is obtained using the DFT of a global Fourier transform layer, and the IDFT process is simulated using two linear layers to mine hidden global periodic features in the latent space. The IDFT process of the global Fourier transform layer can be represented as:

[0049]

[0050] In the formula, P c For the compensated phase value, W g and b g These are the weight matrix and bias term of the residual block, φ. g O is the original phase angle. g1 With O g2 These are the outputs of the first and second linear layers in a bilinear layer, respectively, where cos(·) is the cosine activation function, and ω... g For the global angular frequency, A g This represents the global amplitude spectrum.

[0051] The Local Fourier Transform (LDT) layer uses a sliding window to segment the latent space, obtaining multiple time slices. It then uses the Discrete Fourier Transform (DFT) to acquire the frequency domain information of the latent space and employs a bilinear layer to simulate the Indirect IDFT process to uncover hidden local periodic features in the latent space. The IDFT process of the LDT layer can be represented as:

[0052]

[0053] In the formula, P′ c For the compensated phase value, W l and b l Here are the weight matrix and bias terms of the residual block, φ l The original phase angle, and These are the outputs of the first and second linear layers in a bilinear layer, respectively, where cos(·) is the cosine activation function, and ω... l For local angular frequency, A l This is a local amplitude spectrum.

[0054] By using parallel global Fourier transform layers and local Fourier transform layers, the output global periodic features and local periodic features are added together to obtain the final periodic feature output.

[0055] In one implementation, the global Fourier transform layer first performs Fourier decomposition of the latent space h(t) using the DFT to obtain the frequency domain signal. The expression for the DFT is:

[0056]

[0057] In the formula, Re[·] represents the real part of the DFT, and Im[·] represents the imaginary part of the DFT. ω g,k Let be the angular frequency of the k-th frequency component. j is the imaginary unit, j 2 =-1. n represents the total number of sampling points in the latent space, k = 0, 1, ..., n-1, and t represents the t-th sampling point. The amplitude spectrum and phase spectrum of the signal can be analyzed using DFT, expressed as:

[0058]

[0059] In the formula, A g,k With φ g,k Let represent the amplitude and phase spectra of the k-th frequency component. Since the input signal is a real signal, according to the conjugate symmetry of the Fourier transform, only the first half of the frequency domain components needs to be considered, i.e., k = 0, 1, ..., n / 2, thus reducing the computational load. After obtaining the amplitude and phase spectra, IDFT is used to restore the frequency domain signal to the time domain signal, which is represented as:

[0060]

[0061] Next, a bilinear layer is used to simulate the IDFT operation, and phase angle compensation is achieved through phase adaptation. This is accomplished by introducing a residual connection structure to compensate for the original phase angle φ. g Perform nonlinear correction to implement the IDFT process:

[0062]

[0063] In the formula, P c For the compensated phase value, W g and b g These are the weight matrix and bias term of the residual block, φ. g This is the original phase angle; and These are the outputs of the first and second linear layers in a bilinear layer, respectively, where cos(·) is the cosine activation function, and ω... g For the global angular frequency, A g This represents the global amplitude spectrum.

[0064] The local Fourier transform layer uses a sliding window to divide the latent space h, obtaining multiple time slices h. s Similarly, for time slice h s Perform DFT and IDFT operations to reconstruct time-domain information:

[0065]

[0066] In the formula, ω l , φ l and A l The spectral information is obtained from time slices. The amplitude spectrum and phase spectrum of the k-th frequency component are represented as A. l,k , φ l,k .

[0067] Next, a bilinear layer structure is used to simulate the IDFT operation process, and phase angle compensation is achieved through phase adaptation. This is accomplished by introducing a residual connection structure to compensate for the original phase angle φ. g Perform nonlinear correction to implement the IDFT process:

[0068]

[0069] In the formula, P′ c For the compensated phase value, W l and b l Here are the weight matrix and bias terms of the residual block, φ l The original phase angle, and These are the outputs of the first and second linear layers in a bilinear layer, respectively, where cos(·) is the cosine activation function, and ω... l For local angular frequency, A l This is a local amplitude spectrum.

[0070] After obtaining the output of the second linear layer, a dense layer is used to linearly map the IDFT output to obtain the time-series slice set H. l :

[0071]

[0072] Considering that the last sample point of each time slice may not provide sufficient temporal information, a self-attention mechanism is used to focus on the last sample point of the time slice to improve temporal representation. For the last time slice H... l It is mapped to query vector O l Key vector K l Sum vector V l For each time series in each time series slice set, the output of the self-attention mechanism is represented as:

[0073]

[0074] A dense layer is used to perform dimensionality transformation on the output of the self-attention mechanism to obtain updated local periodic features:

[0075] h l =Dense(Attn)

[0076] Finally, the extracted local periodic features and global periodic features are added together to enhance the representational power of the periodic features, and the periodic feature output q of the spectral feature module is obtained:

[0077]

[0078] Step 3: The distribution difference between normal and abnormal samples in the periodic features is increased through the memory update process and the query update process, and then the differential Transformer autoencoder reconstructs the latent space into the original signal.

[0079] The query-update process is primarily used to generate synthetic memory terms, and then uses these terms to update the parameters of the latent space, thereby enabling the mining of temporal representations. The query-update process can be represented as:

[0080]

[0081] In the formula, Let be the t-th sample in the input latent space. To synthesize memory terms, Let t be the t-th sample in the updated latent space; To check memory.

[0082] During the memory update phase, define the query term q. s By encoding the normal patterns hidden in the spectral feature module, the memory term m is obtained. i ∈R C(i = 1, ..., M) and adaptively adjust these memory entries using an incremental update strategy. For the query item of the t-th sample... Its conditional query attention It can be represented as:

[0083]

[0084] In the formula, λ is a temperature parameter used to control the concentration of the query item distribution.

[0085] Then, an update gate θ is constructed for the query terms to determine how much information from the query terms is incorporated into the memory term. The formula for the update gate is expressed as:

[0086]

[0087] In the formula, φ and Let be the parameter matrix, and σ be the sigmoid activation function. The memory update phase is performed only during the training phase.

[0088] During the query update phase, the Softmax function is used to compute the memory term and the query term, and to obtain conditional memory attention.

[0089]

[0090] Next, by analyzing the memory item m i Conditioned memory and attention Weighted fusion is performed to form a synthetic memory item.

[0091]

[0092] Finally, the memory item is synthesized by splicing. and the original query item Get the updated query items The query term is then passed as input to the Dense decoder.

[0093] Step four involves comparing the reconstructed original signal with the anomaly threshold, and using the comparison results to identify early faults in the ship's propulsion shafting.

[0094] The state score of the sample is obtained based on the distance deviation between the acquired original vibration signal and the original eddy current signal and the potential space, and an anomaly rate of the state score is set. Then, an anomaly threshold is determined based on the anomaly rate of the state score.

[0095] The state score of the sample is obtained based on the distance deviation between the acquired original vibration signal and the original eddy current signal and the potential space. The formula for the state score of the sample at time t is:

[0096]

[0097] In the formula, DLS represents the latent space, and DIS represents the input space. This represents the memory item most recently associated with the query term. and S represents the sample in the original signal at time t and the sample in the reconstructed signal at time t, respectively; i The state score.

[0098] In one implementation, a bias-based early anomaly identification criterion is constructed using the latent space and the acquired raw signals. The bias of the latent space (DLS) is defined as the difference between a query item and its nearest memory item at time step t. The distance between the normal and abnormal states. The memory term contains the prototype of the normal state; therefore, the DLS in the abnormal state is greater than the DLS in the normal state. Furthermore, the deviation of input space (DLS) is defined as the distance between the input signal and the reconstructed signal at time step t. The product of DIS and DLS increases the distance between the normal and abnormal states. Wherein, the state fraction S... i It can be represented as:

[0099]

[0100] Set the anomaly rate for the state scores, and use the percentile method to determine the anomaly threshold θ:

[0101] θ=Q(S i ,δ)

[0102] In the formula, Q(·) represents the percentile of the state score corresponding to the training data, and δ is the anomaly rate. The state scores S of the test samples are compared... te The state of the current sample is determined by a threshold θ.

[0103]

[0104] The present invention will be further described in detail below with reference to specific embodiments.

[0105] The proposed method (PETMA) was validated using experimental data from a ship's propulsion shafting system. The propulsion shafting system mainly comprises the port and starboard propulsion shafting systems, each consisting primarily of the main engine, gearbox, propulsion shaft, intermediate shaft, stern shaft, sliding bearings, and propeller. Daily navigation was simulated at sea, averaging approximately 12 hours per day. The port and starboard propulsion shafting systems operated at five speeds: 100 rpm, 134 rpm, 166 rpm, 188 rpm, and 210 rpm. The vibration sensors (VS) were sampled at a frequency of 12.8 kHz, while the eddy current sensors (ECS) were sampled at a frequency of 10 kHz. The vibration sensors collected vibration signals in the x, y, and z directions, while the eddy current sensors collected eddy current signals in the x and y directions. The sampling interval for each sensor was one hour, and each sampling time was 10 seconds. On day 122 of the experiment, a 40Hz frequency split gradually appeared in the stern shaft, indicating that the starboard stern shaft had transitioned from a normal state to a degraded state. The experiment lasted for 245 days, and finally, the ship's propulsion shafting suffered severe wear and could no longer operate normally. Therefore, the starboard stern shaft was selected for early anomaly identification.

[0106] Feature extraction is performed on the vibration and eddy current signals of the stern shaft to obtain the characteristics of the shaft system, and the resulting anomaly identification results are as follows: Figure 4 As shown, PETMA can accurately identify early anomalies in the shaft system.

[0107] To verify the superiority of the proposed method, a comparative analysis was conducted using nine existing conventional methods, including IFOrest, AKSC, GMSVM, IIMEmAE, VAE, RSTRN, LRAE, FCVAE, and VQRAE. Each method was trained and tested using the same samples, and ten experiments were performed, with the average value calculated. Based on the experimental data, the same verification process as described above was used for validation. The specific verification results of this invention are shown in Table 1.

[0108] Table 1

[0109]

[0110] Table 1 shows the comparison results of their diagnostic accuracy. As can be seen from Table 1, the early fault identification accuracy of the present invention is 96%, which is significantly higher than the other 9 comparative analysis methods. This also shows that the present invention has strong creative and engineering application value.

[0111] The present invention also provides an early fault identification system for a ship propulsion shafting system. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the early fault identification method for a ship propulsion shafting system as described above.

[0112] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the early fault identification method for ship propulsion shafting as described above.

[0113] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of early fault identification of a marine propulsion shafting, characterized by, The method includes the following steps: (1) The differential Transformer autoencoder maps the features extracted from the original vibration signal and the original eddy current signal of the ship's propulsion shaft system into a latent space, and uses the spectral feature module to mine the periodic features of the latent space; wherein, the spectral feature module includes a global Fourier transform layer and a local Fourier transform layer; the differential Transformer autoencoder is constructed by combining the Transformer autoencoder and the Dense decoder. (2) By increasing the distribution difference between normal and abnormal samples in the periodic features through the memory update process and the query update process, the differential Transformer autoencoder reconstructs the latent space into the original signal, thereby realizing the early fault identification method of the ship propulsion shaft system. The query updating process is used to generate a synthetic memory item and update the parameters of the latent space by generating the synthetic memory item; and the memory updating process obtains the memory item by encoding the normal mode hidden in the spectral feature module and adaptively adjusts the obtained memory item in combination with an incremental updating strategy ; The formula corresponding to the query update process is: wherein is the i-th sample in the input latent space, t is the i-th sample in the input latent space, is the synthesized memory item, is the i-th sample in the updated latent space; t is the memory attention; is the memory attention; For the memory updating process, the update gate is constructed for the query term The update gate formula is given as follows for determining how much information in the query term is fused into the memory term In the formula, and For the parameter matrix, Use the Sigmoid activation function; Use conditions to query attention.

2. The method for early fault identification of ship propulsion shafting as described in claim 1, characterized in that: Frequency domain information of the latent space is obtained by using the DFT of the global Fourier transform layer, and the IDFT process is simulated by the bilinear layer to mine the hidden global periodic features in the latent space.

3. The method for early fault identification of ship propulsion shafting as described in claim 2, characterized in that: The local Fourier transform layer uses a sliding window to divide the latent space, obtains multiple time slices, and uses its DFT to obtain the frequency domain information of the latent space. The bilinear layer is used to simulate the IDFT process to mine the hidden local periodic features in the latent space.

4. The method for early fault identification of ship propulsion shafting as described in claim 3, characterized in that: By using parallel global Fourier transform layers and local Fourier transform layers, the output global periodic features and local periodic features are added together to obtain the final periodic features.

5. The method for early fault identification of ship propulsion shafting as described in claim 1, characterized in that: The reconstructed original signal is compared with the anomaly threshold, and the early fault identification of the ship's propulsion shaft system is realized based on the comparison result.

6. The method for early fault identification of ship propulsion shafting as described in claim 5, characterized in that: The state score of the sample is obtained based on the distance deviation between the acquired original vibration signal and the original eddy current signal and the potential space, and an anomaly rate of the state score is set. Then, an anomaly threshold is determined based on the anomaly rate of the state score.

7. An early fault identification system for a ship's propulsion shafting system, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the early fault identification method for the ship propulsion shafting as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the early fault identification method for ship propulsion shafting as described in any one of claims 1-6.

Images