Gearbox fault diagnosis method, system, equipment and medium

Abstract

The invention discloses a gearbox fault diagnosis method, system and device and a medium, and the method comprises the steps: firstly obtaining a gearbox vibration signal, and forming a gearbox data set; carrying out denoising processing on a gearbox data set formed by the gearbox vibration signals by adopting a denoising method combining variational mode extraction and a wavelet half-soft-hard threshold value; performing continuous wavelet transform on the denoised gearbox vibration signal to obtain a two-dimensional time-frequency graph as an input sequence of a deep belief network (DBN); performing feature extraction on the input sequence through a DBN, and mining feature information in a two-dimensional time-frequency graph; and inputting the extracted feature information into a Crossform model, carrying out parameter optimization on the Crossform model by adopting an improved jellyfish algorithm, and obtaining a fault diagnosis result of the gearbox through the trained Crossform model. The method is high in diagnosis precision, is high in processing speed, and can effectively carry out the fault diagnosis of the gear box.

Description

Technical Field

[0001] The present invention belongs to the field of gearbox fault diagnosis, and in particular relates to a gearbox fault diagnosis method, system, equipment and medium. Background Art

[0002] Gearboxes, as a universal component that is essential for connecting and transmitting power in mechanical equipment, have been widely used in modern industrial equipment. They are widely used in metal cutting machine tools, aviation, power systems, agricultural machinery, transportation machinery, metallurgical machinery and other fields. With the development of science and technology, mechanical equipment is moving towards large-scale, automated and high-performance, and gear devices that transmit motion and power play an important role in this. However, due to the complex structure of gears and gearboxes and the harsh working environment, they are prone to damage and failure.

[0003] The operation status monitoring and fault diagnosis technology of gears and gearboxes is an interdisciplinary subject, which combines the essence of multiple disciplines such as mechanics, mechanics, electronics, mathematics, physics, and computers. This technology has the characteristics of strong engineering applicability, reliable technical foundation, and close relationship with the development of high technology. By monitoring and diagnosing the operation status of gears and gearboxes, potential problems can be discovered in time and preventive maintenance can be carried out, thereby ensuring the safe and reliable operation of the equipment and improving processing accuracy and production efficiency. With the development of technology, traditional fault diagnosis methods can no longer meet the current needs of complex systems. They can only extract appearance, function, behavior and other appearances from data representations, but cannot capture deeper information. Deep learning technology has developed rapidly in recent years. With its unique advantages, it has been widely used in various fields. In order to solve the problem that the traditional method of feature extraction is too cumbersome, Wan Anping et al. used convolutional neural network (CNN) to automatically extract signal features. Zhang Kai et al. proposed a new residual network method, which performs wavelet transform on the original signal to improve the residual network and focus on the prominent features of the wavelet coefficients. Hu Hao et al. proposed a gear fault diagnosis method based on deep belief network, which extracted the time domain signal and frequency domain signal of the gear and determined the optimal parameters of the model. The vibration data was then divided into training set and test set and input into the model to automatically extract the characteristics of the vibration signal and identify the fault. However, the operating environment of the gear is complex and changeable, and it is extremely difficult to obtain high-quality fault samples, resulting in a relative lack of actual training data, which makes the application of deep learning under small sample conditions have certain limitations. Summary of the invention

[0004] Purpose of the invention: The present invention discloses a gearbox fault diagnosis method, system, device and medium, which improve the utilization of gearbox vibration information and effectively improve the gearbox fault diagnosis efficiency.

[0005] Technical solution: A gearbox fault diagnosis method according to the present invention comprises the following steps:

[0006] (1) Acquire the gearbox vibration signal in advance to form a gearbox data set;

[0007] (2) The gearbox data set formed by the gearbox vibration signal is denoised using a denoising method combining variational mode extraction with wavelet semi-soft and hard thresholds;

[0008] (3) The denoised gearbox vibration signal is transformed by continuous wavelet transform to obtain a two-dimensional time-frequency graph as an input sequence of a deep belief network (DBN); the DBN is used to extract features from the input sequence and mine feature information in the two-dimensional time-frequency graph;

[0009] (4) The extracted feature information is input into the Crossformer model, and the improved jellyfish algorithm is used to optimize the parameters of the Crossformer model. The fault diagnosis results of the gearbox are obtained through the trained Crossformer model.

[0010] Furthermore, the gearbox data set described in step (1) includes vibration signals of the gearbox under four working conditions: normal, worn, pitting, and broken teeth.

[0011] Furthermore, the implementation process of step (2) is as follows:

[0012] The variational mode extraction formula is as follows:

[0013]

[0014] Among them, f(t) represents the input signal; w d represents the approximate center frequency; f r (t) is the residual signal; α represents the quadratic penalty term; β(t) represents the impulse response of the filter; θ represents the Dirac distribution; u d (t) represents the expected mode;

[0015] The augmented Lagrangian function introduces the quadratic penalty term α and the Lagrangian multiplication operator λ to solve the reconstruction constraint problem. The formula is as follows:

[0016]

[0017] The variational problem is solved using the alternating direction multiplier algorithm and Fourier transform, alternating updates and λ n+1 The new expression is as follows:

[0018]

[0019] Approximate center frequency w d The penalty factor α determines the noise reduction performance of VME; the function formula of wavelet semi-soft and hard threshold is as follows:

[0020]

[0021] Among them, W is the wavelet coefficient, W m is the wavelet coefficient after threshold filtering, T is the threshold; is the proportionality coefficient, The vibration signal is first decomposed into high-frequency and low-frequency components using VME. The high-frequency IMF component is processed by wavelet semi-soft and hard thresholds and then fused to form a fault data set.

[0022] Furthermore, the process of optimizing the parameters of the Crossformer model using the improved jellyfish algorithm in step (4) is as follows:

[0023] S1: Set the objective function of the jellyfish algorithm to the diagnostic accuracy and initialize the preset parameters;

[0024] S2: Use reverse learning to build a new jellyfish algorithm initial population:

[0025]

[0026] Among them, a represents the upper limit of the population, b represents the lower limit of the population, γ represents the initial population individual, is the reverse population individual;

[0027] S3: Calculate the fitness value of each individual according to the objective function, and record the position of the individual with the highest current fitness value;

[0028] In order to regulate the movement of jellyfish between following the ocean current and moving within the jellyfish group, the time control mechanism includes the time control function C(t) and the constant C. The movement mode of the jellyfish group is determined by the value of the constant C. The formula of C(t) is as follows:

[0029]

[0030] Among them, T is the total time, t is the current time;

[0031] According to the value of C(t) and the random variable rand(0, 1), different methods are selected to update the position of each individual. Following the ocean current behavior, when C(t) is greater than 0.5, the position update formula is as follows:

[0032]

[0033] df=β×rand(0,1)×μ

[0034]

[0035] X i (t+1)=X i (t)+rand(0,1)×(X * -β×rand(0,1)×μ)

[0036] in, is the ocean current trend, n pop is the number of jellyfish; X * It is the jellyfish particle with the best identification result among the current population; c Is the factor that determines attractiveness; X i is the current jellyfish particle, X i (t) is the current jellyfish position, μ is the average position of all jellyfish; df is the difference between the current best position of the jellyfish and the average position of all jellyfish; X i (t+1) is the next generation of jellyfish particle position identification; β>0 is the distribution coefficient. After sensitivity analysis, β=3 was obtained;

[0037] X i (t+1)=X i +l×rand(0,1)×(U b -L b )

[0038] When (1-C(t)) is less than rand(0,1), the jellyfish actively moves; U b and L b are the upper and lower bounds of the search space, respectively: l>0 is a motion coefficient, which is related to the motion length around the jellyfish position; sensitivity analysis results, l = 0.1;

[0039]

[0040] When (1-C(t)) is greater than rand(0,1), the jellyfish moves passively; X i (t) is the position of the particle of the previous generation, is the movement step length of the particles of the previous generation;

[0041] S4: Send the iterative results of the jellyfish algorithm to the dimension learning for search. The calculation formula is as follows:

[0042] First, by using X i (t)’s current position and candidate position X i The radius R is calculated by the Euclidean distance between (t+1) i (t):

[0043] R i(t)=||X i (t+1)-X i (t)||

[0044] Then, N i (t) is X j The neighborhood of (t) is constructed by the following formula, where D i Yes X i (t) and X j (t) Euclidean distance between; X i (t), X j (t) are all random particles in the population;

[0045] N i (t) = {X j (t)|D i (X i (t),X j (t)≤R i (t)),X j (t)∈POP}

[0046] Once X is constructed i (t), then multi-neighborhood learning is performed by the following formula, where X DLH,d The d dimension (t+1) is obtained by using i A random neighborhood X selected from (t) n,d (t) d-dimensional data and random X selected from Pop r,d (t) is calculated as:

[0047] X DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t));

[0048] S5: The number of iterations is increased by 1 to determine whether the maximum number of iterations of the algorithm has been reached. If not, proceed to step S3; otherwise, terminate the operation, output the final iteration result, and obtain the optimal w of VME. d and α parameter.

[0049] Furthermore, the CrossFormer model in step (4) adopts a pyramid structure to divide the model into four stages; each stage consists of a cross-scale embedding layer CEL and several CrossFormer blocks; CEL receives the output or input image of the last level as input or output and generates a cross-scale embedding; several CrossFormer blocks are established after CEL, each block involving long-short distance attention and dynamic position bias; a specialized head appears after the last stage for accounting for a specific task.

[0050] A gearbox fault diagnosis system described in the present invention comprises: a data acquisition module, which pre-acquires a gearbox vibration signal to form a gearbox data set; a data processing module, which adopts a denoising method combining VME with wavelet semi-soft and hard thresholds to denoise the gearbox vibration signal; a data conversion module, which uses continuous wavelet transform to obtain a two-dimensional time-frequency graph from the denoised gearbox vibration signal as an input sequence of a deep belief network DBN; extracts features from the input sequence through the DBN to mine feature information in the two-dimensional time-frequency graph; and a model construction and optimization module, which inputs the extracted feature information into a Crossformer model, and uses an improved jellyfish algorithm to optimize the parameters of the Crossformer model, and obtains a gearbox fault diagnosis result through the trained Crossformer model.

[0051] A device described in the present invention includes a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the gearbox fault diagnosis method described above when running the computer program.

[0052] A storage medium according to the present invention stores a computer program, and when the computer program is executed by at least one processor, the steps of the gearbox fault diagnosis method described above are implemented. .

[0053] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: the present invention adopts reverse learning OBL to optimize the initialization population of the jellyfish algorithm. During the iteration process, OBL initializes the population and generates reverse solutions, providing more opportunities for exploring potential solutions, thereby improving the diversity of the population and the global search capability of the algorithm. In order to improve the local search capability and exploration capability of the jellyfish optimization algorithm, it is necessary to strengthen the selection of particles during the search process, and increase the dimensional learning search strategy. The present invention has high recognition accuracy and fast processing speed, and can effectively diagnose gear faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of a gearbox fault diagnosis method;

[0055] Figure 2 This is the signal denoising flow chart. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0057] like Figure 1 As shown, the present invention proposes a gearbox fault diagnosis method, comprising the following steps:

[0058] Step 1: Collect the gearbox vibration signal to form a gearbox data set, including the vibration signals of the gearbox under four working conditions: normal, worn, pitting, and broken teeth.

[0059] Step 2: If Figure 2 As shown in the figure, the gearbox data set formed by the gearbox vibration signal is denoised by using a denoising method combining variational mode extraction (VME) with wavelet semi-soft and hard thresholds.

[0060] The VME constrained variational model formula is as follows:

[0061]

[0062] Among them, f(t) represents the input signal; w d represents the approximate center frequency; f r (t) is the residual signal; α represents the quadratic penalty term; β(t) represents the impulse response of the filter; θ represents the Dirac distribution; u d (t) represents the expected mode.

[0063] The augmented Lagrangian function introduces the quadratic penalty term α and the Lagrangian multiplication operator λ to solve the reconstruction constraint problem. The formula is as follows:

[0064]

[0065] The above variational problem is solved using the alternating direction multiplier algorithm and Fourier transform, alternating updates and λ n+1 The new expression can be obtained as follows:

[0066]

[0067] Approximate center frequency w d and the penalty factor α determine the noise reduction performance of VME. Therefore, choosing the right w d and α are the key to the accurate decomposition of the signal by the VME algorithm.

[0068] The function formula of wavelet semi-soft and hard threshold is as follows:

[0069]

[0070] Among them, W is the wavelet coefficient, W m is the wavelet coefficient after threshold filtering, T is the threshold; is the proportionality coefficient, When 0 and 1 are taken respectively, the functions correspond to the hard threshold function and the soft threshold function respectively; when When , the function is called wavelet semi-soft and hard threshold function. First, the vibration signal is decomposed into high-frequency and low-frequency components using VME. The high-frequency IMF component is processed by wavelet semi-soft and hard thresholds and then fused to form a fault data set.

[0071] Step 3: Use continuous wavelet transform (CWT) to obtain a two-dimensional time-frequency graph as the input sequence of the deep belief network (DBN); perform feature extraction on the input sequence through DBN to mine the feature information in the two-dimensional time-frequency graph.

[0072] The Morlet wavelet is selected, and the denoised gearbox vibration signal is transformed through continuous wavelet transform to obtain a two-dimensional time-frequency graph and normalized as the input sequence. The deep belief network DBN consists of two RBMs and a BP classifier.

[0073] Step 4: Input the extracted feature information into Crossformer and combine it with the intelligent optimization algorithm to optimize the parameters. The fault diagnosis results of the gearbox are obtained through the trained Crossformer model.

[0074] The optimization process steps of the intelligent optimization algorithm for the Crossforme fault diagnosis model are as follows:

[0075] S1: Set the objective function of the jellyfish algorithm to diagnostic accuracy and initialize the preset parameters.

[0076] S2: Use reverse learning to build a new jellyfish algorithm initial population:

[0077]

[0078] Among them, a represents the upper limit of the population, b represents the lower limit of the population, γ represents the initial population individual, For reverse population individuals.

[0079] S3: Calculate the fitness value of each individual according to the objective function, and record the position of the individual with the highest current fitness value;

[0080] In order to regulate the movement of jellyfish between following the ocean current and moving within the jellyfish group, the time control mechanism includes the time control function C(t) and the constant C. The value of the constant C is used to determine the movement mode of the jellyfish group, including following the ocean current, active movement, and passive movement. The formula of C(t) is as follows:

[0081]

[0082] Among them, T is the total time and t is the current time.

[0083] According to the value of C(t) and the random variable rand(0, 1), different methods are selected to update the position of each individual. The position update formula following the ocean current behavior (when C(t) is greater than 0.5) is as follows:

[0084]

[0085] df=β×rand(0,1)×μ

[0086]

[0087] X i (t+1)=X i (t)+rand(0,1)×(X * -β×rand(0,1)×μ)

[0088] in, is the ocean current trend, n pop is the number of jellyfish; X * It is the jellyfish particle with the best identification result among the current population; c Is the factor that determines attractiveness; X i is the current jellyfish particle, X i (t) is the current jellyfish position, μ is the average position of all jellyfish; df is the difference between the current best position of the jellyfish and the average position of all jellyfish; X i (t+1) is the position of the jellyfish particle identified by the next generation; β>0 is the distribution coefficient. After sensitivity analysis, β=3 was obtained.

[0089] X i (t+1)=X i +l×rand(0,1)×(U b -L b )

[0090] When (1-C(t)) is less than rand(0,1), the jellyfish moves actively; b and L b are the upper and lower bounds of the search space, respectively: l>0 is a motion coefficient, which is related to the motion length around the jellyfish position; sensitivity analysis results, l = 0.1;

[0091]

[0092] When (1-C(t)) is greater than rand(0,1), the jellyfish moves passively; i (t) is the position of the particle of the previous generation, is the movement step of the previous generation of particles.

[0093] S4: Send the iterative results of the jellyfish algorithm to the dimension learning for search. The calculation formula is as follows:

[0094] First, by using X i (t)’s current position and candidate position X i The radius R is calculated by the Euclidean distance between (t+1) i (t):

[0095] R i (t)=||X i (t+1)-X i (t)||

[0096] Then, N i (t) is X j The neighborhood of (t) is constructed by the following formula, where D i Yes X i (t) and X j (t), X i (t), X j (t) are all random particles in the population:

[0097] N i (t) = {X j (t)|D i (X i (t),X j (t)≤R i (t)),X j (t)∈POP}

[0098] Once X is constructed i (t), then multi-neighborhood learning is performed by the following formula, where X DLH,d The d dimension (t+1) is obtained by using i A random neighborhood X selected from (t) n,d (t) d-dimensional data and random X selected from Pop r,d (t) is calculated as:

[0099] X DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t))

[0100] S5: The number of iterations is increased by 1 to determine whether the maximum number of iterations of the algorithm has been reached. If not, proceed to step S3; otherwise, terminate the operation, output the final iteration result, and obtain the optimal w of VME. d and α parameter.

[0101] The main feature of CrossFormer is that it uses a pyramid structure to divide the CrossFormer model into four stages. Each stage consists of a cross-scale embedding layer CEL and multiple CrossFormer blocks. These CEL layers receive the output or input image of the last level as input or output, and then generate cross-scale embeddings. Each CrossFormer block can process the spatial information and channel information in the input data and establish cross-scale interactions between different scales of the feature map. By using this structure, the CrossFormer model can effectively learn the spatial and channel features of the input data and perform cross-scale interactions on features of different scales, thereby improving the performance of the model; several CrossFormer blocks are established after CEL, each of which involves long-short distance attention and dynamic position bias; a specialized head appears after the final stage of accounting for specific tasks.

[0102] The present invention also provides a gearbox fault diagnosis system, comprising: a data acquisition module, which pre-acquires a gearbox vibration signal to form a gearbox data set; a data processing module, which uses a denoising method combining VME with wavelet semi-soft and hard thresholds to denoise the gearbox vibration signal; a data conversion module, which uses continuous wavelet transform to obtain a two-dimensional time-frequency graph from the denoised gearbox vibration signal as an input sequence of a deep belief network DBN; extracts features from the input sequence through the DBN to mine feature information in the two-dimensional time-frequency graph; a model construction and optimization module, which inputs the extracted feature information into a Crossformer model, and uses an improved jellyfish algorithm to optimize the parameters of the Crossformer model, and obtains the gearbox fault diagnosis result through the trained Crossformer model.

[0103] The present invention also provides a device, comprising a memory and a processor, wherein: the memory is used to store a computer program that can be run on the processor; the processor is used to execute the steps of the gearbox fault diagnosis method as described above when running the computer program.

[0104] The present invention also provides a storage medium having a computer program stored thereon, and when the computer program is executed by at least one processor, the steps of the gearbox fault diagnosis method as described above are implemented.

[0105] The present invention is not limited to the above-mentioned specific implementation cases. Within the knowledge scope of those skilled in the art, any equivalent replacement or change of the technical solution and the inventive concept according to the present invention should be included in the protection scope of the present invention.

Claims

1. A gearbox fault diagnosis method, characterized in that: The following steps are involved: (1) Acquire the gearbox vibration signal in advance to form a gearbox data set; (2) The gearbox data set formed by the gearbox vibration signal is denoised using a denoising method combining variational mode extraction with wavelet semi-soft and hard thresholds; (3) The denoised gearbox vibration signal is transformed by continuous wavelet transform to obtain a two-dimensional time-frequency graph as an input sequence of a deep belief network (DBN); the DBN is used to extract features from the input sequence and mine feature information in the two-dimensional time-frequency graph; (4) The extracted feature information is input into the Crossformer model, and the improved jellyfish algorithm is used to optimize the parameters of the Crossformer model. The fault diagnosis results of the gearbox are obtained through the trained Crossformer model.

2. A gearbox fault diagnosis method according to claim 1, characterized in that: The gearbox data set described in step (1) includes vibration signals of the gearbox under four working conditions: normal, worn, pitting, and broken teeth.

3. A gearbox fault diagnosis method according to claim 1, characterized in that: The implementation process of step (2) is as follows: The variational mode extraction formula is as follows: Among them, f(t) represents the input signal; w d represents the approximate center frequency; f r (t) is the residual signal; α represents the quadratic penalty term; β(t) represents the impulse response of the filter; θ represents the Dirac distribution; u d (t) represents the expected mode; The augmented Lagrangian function introduces the quadratic penalty term α and the Lagrangian multiplication operator λ to solve the reconstruction constraint problem. The formula is as follows: The variational problem is solved using the alternating direction multiplier algorithm and Fourier transform, alternating updates and λ n+1 The new expression is as follows: Approximate center frequency w d The penalty factor α determines the noise reduction performance of VME; the function formula of wavelet semi-soft and hard threshold is as follows: Among them, W is the wavelet coefficient, W m is the wavelet coefficient after threshold filtering, T is the threshold; is the proportionality coefficient, The vibration signal is first decomposed into high-frequency and low-frequency components using VME. The high-frequency IMF component is processed by wavelet semi-soft and hard thresholds and then fused to form a fault data set.

4. A gearbox fault diagnosis method according to claim 1, characterized in that: In step (4), the improved jellyfish algorithm is used to optimize the parameters of the Crossformer model as follows: S1: Set the objective function of the jellyfish algorithm to the diagnostic accuracy and initialize the preset parameters; S2: Use reverse learning to build a new jellyfish algorithm initial population: Among them, a represents the upper limit of the population, b represents the lower limit of the population, γ represents the initial population individual, is the reverse population individual; S3: Calculate the fitness value of each individual according to the objective function, and record the position of the individual with the highest current fitness value; In order to regulate the movement of jellyfish between following the ocean current and moving within the jellyfish group, the time control mechanism includes the time control function C(t) and the constant C. The movement mode of the jellyfish group is determined by the value of the constant C. The formula of C(t) is as follows: Among them, T is the total time, t is the current time; According to the value of C(t) and the random variable rand(0, 1), different methods are selected to update the position of each individual. Following the ocean current behavior, when C(t) is greater than 0.5, the position update formula is as follows: df=β×rand(0,1)×μ X i (t+1)=X i (t)+rand(0,1)×(X * -β×rand(0,1)×μ) in, is the ocean current trend, n pop is the number of jellyfish; X * It is the jellyfish particle with the best identification result among the current population; c Is the factor that determines attractiveness; X i is the current jellyfish particle, X i (t) is the current jellyfish position, μ is the average position of all jellyfish; df is the difference between the current best position of the jellyfish and the average position of all jellyfish; X i (t+1) is the next generation of jellyfish particle position identification; β>0 is the distribution coefficient. After sensitivity analysis, β=3 was obtained; X i (t+1)=X i +l×rand(0,1)×(U b -L b ) When (1-C(t)) is less than rand(0,1), the jellyfish actively moves; U b and L b are the upper and lower bounds of the search space, respectively: l>0 is a motion coefficient, which is related to the motion length around the jellyfish position; sensitivity analysis results, l = 0.1; When (1-C(t)) is greater than rand(0,1), the jellyfish moves passively; X i (t) is the position of the particle of the previous generation, is the movement step of the particles of the previous generation; S4: Send the iterative results of the jellyfish algorithm to the dimension learning for search. The calculation formula is as follows: First, by using X i (t)’s current position and candidate position X i The radius R is calculated by the Euclidean distance between (t+1) i (t): R i (t)=||X i (t+1)-X i (t)|| Then, N i (t) is X j The neighborhood of (t) is constructed by the following formula, where D i Yes X i (t) and X j (t) Euclidean distance between; X i (t), X j (t) are all random particles in the population; N i (t)={X j (t)|D i (X i (t),X j (t)≤R i (t)),X j (t)∈POP} Once X is constructed i (t), then multi-neighborhood learning is performed by the following formula, where X DLH,d The d dimension (t+1) is obtained by using i A random neighborhood X selected from (t) n,d (t) d-dimensional data and random X selected from Pop r,d (t) is calculated as: X DLH,d (t+1)=X i,d (t)+rand×(X n,d (t)-X r,d (t)); S5: The number of iterations is increased by 1 to determine whether the maximum number of iterations of the algorithm has been reached. If not, proceed to step S3; otherwise, terminate the operation, output the final iteration result, and obtain the optimal w of VME. d and α parameter.

5. A gearbox fault diagnosis method according to claim 1, characterized in that: The CrossFormer model in step (4) adopts a pyramid structure to divide the model into four stages; each stage consists of a cross-scale embedding layer CEL and several CrossFormer blocks; CEL receives the output or input image of the last level as input or output and generates a cross-scale embedding; several CrossFormer blocks are established after CEL, each of which involves long-short distance attention and dynamic position bias; a specialized head appears after the last stage for accounting for specific tasks.

6. A gearbox fault diagnosis system using the method according to any one of claims 1 to 5, characterized in that: include: A data acquisition module is used to obtain the gearbox vibration signal in advance and form a gearbox data set; A data processing module is used to denoise the gearbox vibration signal by using a denoising method combining VME with wavelet semi-soft and hard thresholds; The data conversion module uses continuous wavelet transform to obtain a two-dimensional time-frequency graph of the gearbox vibration signal after denoising as an input sequence of the deep belief network DBN; the DBN is used to extract features of the input sequence and mine feature information in the two-dimensional time-frequency graph; The model building and optimization module inputs the extracted feature information into the Crossformer model, and uses the improved jellyfish algorithm to optimize the parameters of the Crossformer model. The fault diagnosis results of the gearbox are obtained through the trained Crossformer model.

7. A device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of the gearbox fault diagnosis method as claimed in any one of claims 1 to 5 when running the computer program.

8. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by at least one processor, the steps of the gearbox fault diagnosis method according to any one of claims 1 to 5 are implemented.

Images