Light-weight hydraulic pump fault diagnosis method applicable to all working conditions

Abstract

The invention relates to a light hydraulic pump fault diagnosis method applicable to all working conditions, and belongs to the technical field of fault diagnosis. The method comprises the steps that a resampling preprocessing technology combining cubic spline interpolation and nearest neighbor interpolation is adopted, unified length transformation is conducted on signals of each circle of a hydraulic pump, and high-frequency and low-frequency feature information of the signals is reserved; designing an order calibrator with global perception capability, wherein the order calibrator is used for stabilizing feature expression after resampling; a multi-cycle component fault feature decoupling module embedded with a prior decomposition factor is constructed, and key component features are extracted; an intrinsic task guiding strategy is provided, and the recognition capability of the model on key fault features is enhanced; interference of working condition characteristics such as rotating speed and pressure is inhibited through a multivariable working condition distribution removal method. The method has the advantages of being small in model parameter, high in calculation speed and the like while achieving full-working-condition high-precision diagnosis, and is suitable for real-time intelligent monitoring of the hydraulic pump in the industrial environment.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a fault diagnosis method for lightweight hydraulic pumps applicable to all working conditions. Background Technology

[0002] In the industrial sector, the long-term stability and safety of equipment are of paramount importance to actual production. These scenarios typically require systems with high stability and reliability. Failures can lead not only to high maintenance costs but also to serious safety accidents. Therefore, real-time fault diagnosis is necessary. However, the complex coupling of hydraulic pumps (gears, bearings, etc.) causes signal aliasing within the pump body, posing a challenge to real-time fault diagnosis.

[0003] In recent years, deep learning, with its powerful automatic feature extraction and pattern recognition capabilities, has replaced traditional methods relying on human experience and become a research hotspot in intelligent fault diagnosis. Numerous studies have applied deep neural networks to fault identification in hydraulic pumps, achieving good results. However, these methods generally assume that training and testing data are identically distributed. While they perform well under specific operating conditions, they struggle to cope with deviations in operating conditions caused by changes in pressure and speed during actual operation. This leads to a significant decrease in diagnostic performance under unknown operating conditions, revealing their insufficient generalization ability. Deep learning (DG), without the need for a target domain, can generalize to unknown operating conditions by training the model using only multiple source domain data, making it more suitable for fault diagnosis needs in the complex operating environment of hydraulic pumps.

[0004] Existing technologies have made some progress in fault diagnosis of hydraulic pumps under all operating conditions, but there are still areas for improvement, mainly in the following aspects:

[0005] (1) For hydraulic pumps with compact internal structure, highly coupled component features, indistinct initial fault characteristics and complex fault states, traditional methods are difficult to achieve high-precision diagnosis based on small parameters due to lack of physical relationship constraints.

[0006] (2) Multivariable working conditions (speed and pressure change simultaneously) result in a large range of working condition changes, leading to low accuracy of cross-working condition diagnosis.

[0007] (3) Existing methods stack too many network layers to improve cross-operating condition accuracy, which further leads to large network model parameters, poor real-time performance, and difficulty in meeting the real-time detection requirements of important equipment and the needs of embedded layout. Summary of the Invention

[0008] The purpose of this invention is to solve the problem that high precision and lightweight are difficult to achieve simultaneously in the prior art, and to provide a lightweight hydraulic pump fault diagnosis method applicable to all working conditions.

[0009] To achieve the above objectives, the technical solution of the present invention is: a lightweight hydraulic pump fault diagnosis method applicable to all working conditions, comprising:

[0010] Resampling preprocessing technology is used to perform uniform length transformation on the signal of the hydraulic pump for each cycle, preserving its high-frequency and low-frequency characteristic information;

[0011] Design an angle domain order calibration module with global perception capabilities to stabilize the feature representation after resampling;

[0012] Construct a multi-period, multi-scale component feature decoupler with embedded prior decomposition factors to extract key component features;

[0013] An intrinsic task-guided strategy is proposed, and the ability to extract key fault features is enhanced by a multivariate operating condition distribution removal method.

[0014] Furthermore, the method includes the following steps:

[0015] Step S1: Use a data acquisition card to acquire the sound pressure or vibration signal of the hydraulic pump equipment, as well as the pressure and speed signals under the current operating conditions;

[0016] Step S2: Based on the rotation speed, the collected sound pressure or vibration signal is resampled using dual characteristics, retaining the high-frequency and low-frequency signals respectively;

[0017] Step S3: Design an angle domain order calibration module to remove resampling errors and high-frequency artifacts;

[0018] Step S4: Construct a multi-cycle, multi-scale component feature decoupler driven by prior decomposition factors, which captures the features of the dominant order and serves as a stable cross-condition feature extractor.

[0019] Step S5: Based on the cause and location of the fault, the main task is decomposed into two sub-tasks as guiding items to improve the ability to extract key fault features of the main task.

[0020] Step S6: Design a DANN-based domain discriminator branch to suppress the representation of working condition information and encourage the feature extractor to retain the purest fault features.

[0021] Step S7: Construct a multi-task loss function, which includes a main fault diagnosis task, a fault factor task, a fault location subtask, a speed discriminator, and a pressure discriminator.

[0022] Furthermore, the implementation method of step S1 is as follows: according to the sensor arrangement method suitable for the hydraulic pump being measured, select a sound pressure or vibration sensor as the model signal input, and synchronously collect or interact with peripheral equipment to obtain real-time speed and pressure transient values.

[0023] Furthermore, step S2 is implemented as follows: the signal obtained in step S1 is subjected to nearest neighbor interpolation and cubic spline interpolation based on the instantaneous rotational speed, and then uniformly resampled to the highest rotational speed supported by the device; the two interpolation results are combined into a dual-channel signal, which is used as the input timing signal of the model. The specific implementation is as follows:

[0024] S21. Perform nearest neighbor interpolation and cubic spline interpolation on the signal obtained in step S1 according to the instantaneous rotation speed, and uniformly resample to the highest rotation speed supported by the device, so that each rotation cycle under all working conditions is uniformly represented by two sequences of length M, and the angular interval between adjacent sampling points is constant at (360 / M)°.

[0025]

[0026] In Equation 1, Win represents the dynamic window length, fs represents the sampling rate, and rpm(t) represents the rotational speed.

[0027] S22. Combine the two interpolation results into a dual-channel signal and use it as the input timing signal of the model.

[0028] Furthermore, step S3 specifically includes the following steps:

[0029] S31. Perform FFT decomposition on the two channel signals respectively to obtain the real and imaginary parts;

[0030] S32. Convolution feature extraction is performed on the real and imaginary parts respectively. Through multiple iterations, pseudo-features, including resampling errors and artifacts, will be gradually removed during the iteration process.

[0031] S33. After convolution learning, the real and imaginary parts are merged back into a complex structure, and then iFFT inverse transformation is performed to convert it back into a time-series signal.

[0032] S34. Perform residual connection between the signal obtained in step S33 and the resampled signal of the original input, retaining local features and learned global features.

[0033] Furthermore, step S4 specifically includes the following steps:

[0034] S41. Perform FFT analysis on the dual-channel resampled signal obtained in step S2, normalize the maximum and minimum values ​​of the angle domain amplitude, and obtain the first K dominant orders as multi-period decomposition factors for subsequent signals.

[0035] freq order =Amp(FFF(X) ID (2)

[0036]

[0037] A = Avg(MinMaxNorm(freq) order (4)

[0038]

[0039] In Equation 2-6, X ID This represents the one-dimensional time-series signal of the dual-channel resampled signal obtained in step S2, where FFF(·) represents the Fast Fourier Transform of the one-dimensional order-domain sequence, Amp represents the angular domain amplitude, and freq represents the angular domain amplitude. order The obtained order spectrum is represented by MinMaxNorm(·), which represents the maximum and minimum value normalization of the signal, Avg(·) represents the mean of the health state signal under different operating conditions, and A represents the average order spectrum after averaging the normalized order spectra. This represents the k principal orders selected from the orders, {f1, ..., fk}. k} corresponds to the first k principal orders, {p1, ..., p k} corresponds to the order length of the first k main orders; through the above calculation process, the main orders and order lengths of the possible working conditions throughout the entire life cycle of the equipment are finally obtained, and this prior knowledge is injected into the subsequent component decoupling unit. The prior knowledge helps the model find the general feature expression of different working conditions.

[0040] S42. Transform the signal according to the dominant order, convert the time series into a two-dimensional tensor with multiple periods, and embed the intra-period and inter-period variations into columns and rows respectively;

[0041]

[0042] In Equation 7, This represents the signal obtained by performing the dimensionality transformation on the one-dimensional time series signal for the i-th time step. This indicates the order p obtained from S41. i With order length f i This transforms a one-dimensional time series signal into a two-dimensional order relation, actively exposing the order relation to the model for learning. Padding(·) performs zero-expansion of the time series along the time dimension, making it consistent with... Compatible, where p i and f i These represent the number of rows and columns of the transformed 2D tensor, respectively. This represents the two-dimensional order relationship derived from the i-th order.

[0043] S43. Design a shared feature extractor to automatically extract features for the dominant order, suppress other non-order features, and focus on retaining the essential features of the order that have generalizability.

[0044] S44. After extracting the corresponding order features, restore the two-dimensional order feature map to a one-dimensional sequence and remove the padding introduced by the previous splicing operation.

[0045]

[0046] In Equation 8, This represents the i-th order relation graph after network learning, and Trunc(·) represents the truncation function. This represents the result of converting the i-th order relation back into a one-dimensional feature.

[0047] S45. Use average pooling to extract features that are translation invariant.

[0048] S46. Design a SelfNorm module suitable for one-dimensional feature sequences to alleviate the distribution difference between the training and testing phases; the SelfNorm module uses the mean and standard deviation within the channel as the statistical basis, and uses lightweight convolution to adaptively generate scaling and offset factors to dynamically adjust the feature distribution.

[0049]

[0050] In Equation 9, A and A' represent the original feature and the adjusted feature, respectively, μ A and σ A 、μ' A and σ' A These represent the mean and variance before and after adjustment, respectively.

[0051] S47. After the feature decoupling module of each component, a set of shared one-dimensional convolutional layers is used to realize feature convergence. In the feature convergence process, average pooling operation is introduced to compress the feature length, reduce redundant information, and contribute to the feature convergence layer so that the domain-independent gradient can more fully act on the shallow feature extractor during backpropagation and suppress the expression of working condition features.

[0052] S48. Design a residual connection method to align the length of the feature output of the previous layer through convolution and pooling, and connect it to the next layer as a residual term. This alleviates the gradient vanishing problem and promotes the efficient transfer of deep features, thereby improving the training stability and expressive power of the model in a multi-layer structure.

[0053] Furthermore, step S44 specifically includes the following steps:

[0054] S441. The detail enhancement convolution module DEconv, which uses a multi-directional difference mechanism, enhances the model's ability to model key-order local features, makes up for the shortcomings of traditional convolution in structure perception, and improves the robustness and generalization ability of the model under complex conditions.

[0055]

[0056] In Equation 10, F in F represents the input to the convolution kernel. out This represents the output of the convolution kernel, DEConv(·) represents the enhancement-of-details convolution operation, and K... i＝1:5 These represent the kernels of VC, CDC, ADC, HDC, and VDC, respectively. * indicates convolution operation, and K... cvt This represents the transformation kernel equivalent to parallel convolution.

[0057] S442. Multi-scale wavelet convolution WTConv is used to enhance the extraction of key low-frequency features and improve the multi-scale modeling capability for non-stationary signals.

[0058] Furthermore, the implementation method of step S5 is as follows: design a fault state master classifier driven by the collaborative sub-tasks of fault type and fault location, and decouple feature type and fault location information through a fully connected layer to improve the attention to key fault features.

[0059]

[0060] X2 = X - X1 (12)

[0061] In Equations 11-12, W represents the weight of the linear layer, b represents the bias of the linear layer, X represents the input of the module, σ(W) represents the spectral norm of W, X1 and X2 represent the outputs of the nonlinear decoupler, and are the inputs of the classifiers for fault type and fault location.

[0062] Furthermore, the implementation method of step S6 is as follows: Since the signal is affected by rotational speed and pressure, a gradient inversion layer is introduced to invert the gradient during backpropagation, which prompts the feature extractor to extract shared features that are independent of the operating conditions, thereby improving the generalization ability of fault diagnosis.

[0063] Furthermore, step S7 is implemented by simultaneously optimizing two objectives: minimizing the class loss function L. y And maximizing the neighborhood classification loss function L d This guides the model to extract general features that are independent of operating conditions.

[0064] R λ (x)=x (13)

[0065]

[0066] Equations 13 and 14 indicate that R λ (x) indicates that no changes are made during forward propagation. λ represents the gradient reversal during backpropagation, λ represents the gradient reversal strength, and I represents the identity mapping of the identity matrix.

[0067]

[0068] Where, d i Labels for the source and target domains of the operating conditions. The feature sequence output by the feature extractor Loss is the domain label predicted by the model. domain This indicates a loss in the domain.

[0069]

[0070] Where q is the number of categories, y i Let i be the true label of the i-th sample. Let ω be the probability of the i-th prediction made by the model. i Let be the weight of the category corresponding to the i-th sample.

[0071] Loss DANTNet = loss main·task +λ1·loss cause +λ2·loss location + λ3·Loss rpm +λ4·Loss bar (17)

[0072] Where λ1 and λ2 are the weights corresponding to the two auxiliary tasks, λ3 and λ4 are the weights corresponding to the two domain removal tasks, and loss... main·task The loss function for the main task of fault classification is loss cause Let loss be the fault cause loss function. location Loss is the loss function at the point of failure. rpm To remove the loss function in the speed domain, Loss bar To remove the loss function from the stress domain, Loss DANTNet This represents the total loss of the classifier.

[0073] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions executable by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of any of the methods described above.

[0074] Compared with the prior art, the present invention has the following advantages: while achieving high-precision diagnosis under all working conditions, the present invention has the advantages of small model parameters and fast calculation speed, and is suitable for real-time intelligent monitoring of hydraulic pumps in industrial environments. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the overall process described in a preferred embodiment of the present invention;

[0076] Figure 2This is a schematic diagram of the angle domain calibration module described in a preferred embodiment of the present invention;

[0077] Figure 3 This is the order waterfall diagram described in the preferred embodiment of the present invention;

[0078] Figure 4 This is a schematic diagram of the multi-cycle, multi-scale component feature decoupling device described in a preferred embodiment of the present invention;

[0079] Figure 5 This is a schematic diagram illustrating the domain generalization described in a preferred embodiment of the present invention. Detailed Implementation

[0080] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0081] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0082] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0083] This invention provides a fault diagnosis method for lightweight hydraulic pumps applicable to all operating conditions, including:

[0084] Resampling preprocessing technology is used to perform uniform length transformation on the signal of the hydraulic pump for each cycle, preserving its high-frequency and low-frequency characteristic information;

[0085] Design an angle domain order calibration module with global perception capabilities to stabilize the feature representation after resampling;

[0086] Construct a multi-period, multi-scale component feature decoupler with embedded prior decomposition factors to extract key component features;

[0087] An intrinsic task-guided strategy is proposed, and the ability to extract key fault features is enhanced by a multivariate operating condition distribution removal method.

[0088] The following is a detailed implementation process of the present invention.

[0089] This example illustrates a fault diagnosis method for lightweight hydraulic pumps applicable to all operating conditions. Figure 1 As shown, the main steps are as follows:

[0090] Step S1: Use a data acquisition card to acquire the sound pressure or vibration signal of the hydraulic pump equipment, as well as the pressure and speed signals under the current operating conditions;

[0091] Step S2: Based on the rotational speed, the collected sound pressure or vibration signal is resampled using dual characteristics, retaining the high-frequency and low-frequency signals respectively;

[0092] Step S3: Design an angle domain order calibration module to remove resampling errors and high-frequency artifacts;

[0093] Step S4: Construct a multi-period, multi-scale component feature decoupler driven by prior decomposition factors, which captures the features of the dominant order and serves as a stable cross-condition feature extractor.

[0094] Step S5: Based on the cause and location of the fault, decompose the main task into two sub-tasks as guiding terms to help the main task acquire sufficiently generalized features.

[0095] Step S6: Design a DANN-based domain discriminator branch. Due to the influence of multiple variables such as rotational speed and pressure in actual working conditions, the distribution suppresses the expression of working condition information, and promotes the feature extractor to retain the purest fault features.

[0096] Step S7: Construct a multi-task loss function, which includes a main fault diagnosis task, a fault factor task, a fault location subtask, a speed discriminator, and a pressure discriminator.

[0097] Further, step S1 is as follows: according to the sensor arrangement suitable for the hydraulic pump being measured, a sound pressure or vibration sensor is selected as the model signal input, and real-time speed and pressure transient values ​​are obtained synchronously or through interaction with peripheral devices.

[0098] Furthermore, step S2 specifically includes the following steps:

[0099] S21: Perform nearest neighbor interpolation and cubic spline interpolation on the signal obtained in step S1 based on the instantaneous rotational speed, and uniformly resample to the highest rotational speed supported by the equipment. This ensures that each rotational cycle under all operating conditions is uniformly represented by two sequences of length M, with the angular interval between adjacent sampling points remaining constant at (360 / M)°.

[0100]

[0101] In Equation 1, Win represents the dynamic window length, fs represents the sampling rate, and rpm(t) represents the rotational speed.

[0102] S22: Combine the two interpolation results into a dual-channel signal, which is used as the input timing signal of the model.

[0103] Further, step S3 is as follows Figure 2 As shown, the specific steps include:

[0104] S31: Perform FFT decomposition on the two channel signals respectively to obtain the real and imaginary parts;

[0105] S32: Convolution feature extraction is performed on the real and imaginary parts respectively. Through multiple iterations, resampling errors and other pseudo-features such as artifacts will be gradually removed as the iteration process proceeds.

[0106] S33: After convolution learning, the real and imaginary parts are merged back into a complex structure, and then the inverse iFFT is performed to convert it back into a time-series signal.

[0107] S34: Perform a residual connection with the resampled signal of the original input to preserve local features and the learned global features.

[0108] Further, step S4 is as follows Figure 3 As shown, the specific steps include:

[0109] S41: Perform FFT analysis on the resampled signal, normalize the angular domain amplitude to obtain the first K dominant orders as multi-period decomposition factors for subsequent signals, such as... Figure 4 As shown;

[0110] freq order =Amp(FFF(X) ID (2)

[0111]

[0112] A = Avg(MinMaxNorm(freq) order (4)

[0113]

[0114] In Equation 2-6, FFF(·) represents the Fast Fourier Transform of the one-dimensional domain sequence. Amp represents the amplitude in the angular domain. MinMaxNorm(·) represents the maximum-minimum normalization of the signal, and Avg(·) represents the averaging of the health state signal under different operating conditions. This represents the k principal orders selected from the order. {f1, ..., f k} corresponds to the first k principal orders, {p1, ..., p k} corresponds to the order length of the first k principal orders.

[0115] S42: Transform the signal according to the dominant order, convert the time series into a two-dimensional tensor with multiple periods, and embed intra-period and inter-period variations into columns and rows respectively;

[0116]

[0117] In Equation 7, Padding(·) is the zero-expansion of the time series along the time dimension, making it consistent with the time series. Compatible, where p i and f i These represent the number of rows and columns of the transformed 2D tensor, respectively.

[0118] S43: Design a shared feature extractor to automatically extract features for the dominant order, suppress other non-order features, and focus on retaining the essential features of the order that have generalizability.

[0119] S44: After extracting the features of this order, restore the two-dimensional feature map to a one-dimensional sequence and remove the padding introduced by the previous splicing operation.

[0120]

[0121] In Equation 8, Trunc(·) represents the truncation function.

[0122] S45: Use average pooling to extract feature representations with translation invariance;

[0123] S46: A SelfNorm module suitable for one-dimensional feature sequences was designed to alleviate the distribution difference between the training and testing phases. This module uses the mean and standard deviation within the channel as the statistical basis and uses lightweight convolution to adaptively generate scaling and offset factors to dynamically adjust the feature distribution.

[0124]

[0125] In Equation 9, A and A' represent the original feature and the adjusted feature, respectively, μ A and σ A 、μ' A and σ' A These represent the mean and variance before and after adjustment, respectively.

[0126] S47: After the feature decoupling module of each component, a set of shared one-dimensional convolutional layers is used to realize feature convergence. In the feature convergence process, average pooling operation is introduced to compress the feature length, reduce redundant information, and contribute to the feature convergence layer so that the domain-independent gradient can play a more fully role in the shallow feature extractor during backpropagation, and suppress the expression of working condition features.

[0127] S48: Design a residual connection method to align the length of the feature output of the previous layer through simple convolution and pooling, and connect it to the next layer as a residual term. This alleviates the gradient vanishing problem and promotes the efficient transfer of deep features, thereby improving the training stability and expressive power of the model in a multi-layer structure.

[0128] Specifically, step S44 includes the following steps:

[0129] S441: DEconv, a detail enhancement convolution module using a multi-directional difference mechanism, enhances the model's ability to model key-order local features, makes up for the shortcomings of traditional convolution in structure perception, and improves the robustness and generalization ability of the model under complex conditions.

[0130]

[0131] In Equation 10, DEConv(·) represents the detail enhancement convolution operation, and K i＝1:5 These represent the kernels of VC, CDC, ADC, HDC, and VDC, respectively. * indicates convolution operation, and K... cvt This represents the transformation kernel equivalent to parallel convolution.

[0132] S442: Employs multi-scale wavelet convolution WTConv to enhance the extraction of key low-frequency features and improve the multi-scale modeling capability for non-stationary signals;

[0133] Furthermore, step S5 specifically includes designing a fault state master classifier driven collaboratively by fault type and fault location subtasks, and decoupling feature type and fault location information through a fully connected layer to improve the focus on key features;

[0134]

[0135] X2 = X - X1 (12)

[0136] In Equations 11-12, X1 and X2 are used as classifiers for fault type and fault location, respectively. σ(W) represents the spectral norm of W.

[0137] Furthermore, step S6 specifically includes introducing a gradient inversion layer to invert the gradient during backpropagation, since the signal is affected by rotational speed and pressure, thereby prompting the feature extractor to extract shared features that are independent of the operating conditions, thus improving the generalization ability of fault diagnosis.

[0138] Furthermore, the overall domain generalization framework of step S7 is as follows: Figure 5 As shown, this specifically includes simultaneously optimizing two objectives: minimizing the class loss function L. y And maximizing the neighborhood classification loss function L d This guides the model to extract general features that are independent of operating conditions;

[0139] R λ (x)=x (13)

[0140]

[0141] Equations 13 and 14 indicate that the forward propagation does not change the gradient, while the backward propagation reverses the gradient.

[0142]

[0143] Where, d i Labels for the source and target domains of the operating conditions. The feature sequence output by the feature extractor Loss is the domain label predicted by the model. domain This indicates a loss in the domain.

[0144]

[0145] Where q is the number of categories, y i Let i be the true label of the i-th sample. Let ω be the probability of the i-th prediction made by the model. i Let be the weight of the category corresponding to the i-th sample.

[0146] Loss DANTNet =loss main·task +λ1·loss cause +λ2·loss location +λ3·Loss rpm +λ4·Loss bar (17)

[0147] Where λ1 and λ2 are the weights corresponding to the two auxiliary tasks, λ3 and λ4 are the weights corresponding to the two domain removal tasks, and loss... main·task The loss function for the main task of fault classification is loss cause Let loss be the fault cause loss function. location Loss is the loss function at the point of failure. rpm To remove the loss function in the speed domain, Loss bar To remove the loss function from the stress domain, Loss DANTNet This represents the total loss of the classifier.

[0148] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A fault diagnosis method for lightweight hydraulic pumps applicable to all working conditions, characterized in that, include: Resampling preprocessing technology is used to perform uniform length transformation on the signal of the hydraulic pump for each cycle, preserving its high-frequency and low-frequency characteristic information; Design an angle domain order calibration module with global perception capabilities to stabilize the feature representation after resampling; Construct a multi-period, multi-scale component feature decoupler with embedded prior decomposition factors to extract key component features; An intrinsic task-guided strategy is proposed, and the ability to extract key fault features is enhanced by a multivariate operating condition distribution removal method.

2. The method for fault diagnosis of a lightweight hydraulic pump applicable to all working conditions as described in claim 1, characterized in that, The method includes the following steps: Step S1: Use a data acquisition card to acquire the sound pressure or vibration signal of the hydraulic pump equipment, as well as the pressure and speed signals under the current operating conditions; Step S2: Based on the rotation speed, the collected sound pressure or vibration signal is resampled using dual characteristics, retaining the high-frequency and low-frequency signals respectively; Step S3: Design an angle domain order calibration module to remove resampling errors and high-frequency artifacts; Step S4: Construct a multi-cycle, multi-scale component feature decoupler driven by prior decomposition factors, which captures the features of the dominant order and serves as a stable cross-condition feature extractor. Step S5: Based on the cause and location of the fault, the main task is decomposed into two sub-tasks as guiding items to improve the ability to extract key fault features of the main task. Step S6: Design a DANN-based domain discriminator branch to suppress the representation of working condition information and encourage the feature extractor to retain the purest fault features. Step S7: Construct a multi-task loss function, which includes a main fault diagnosis task, a fault factor task, a fault location subtask, a speed discriminator, and a pressure discriminator.

3. The lightweight hydraulic pump fault diagnosis method applicable to all working conditions according to claim 2, characterized in that, The implementation method of step S1 is as follows: according to the sensor arrangement method suitable for the hydraulic pump being measured, select a sound pressure or vibration sensor as the model signal input, and synchronously collect or interact with peripheral equipment to obtain real-time speed and pressure transient values.

4. The lightweight hydraulic pump fault diagnosis method applicable to all working conditions according to claim 2, characterized in that, The implementation method of step S2 is as follows: perform nearest neighbor interpolation and cubic spline interpolation on the signal obtained in step S1 according to the instantaneous rotation speed, and uniformly resample to the highest rotation speed supported by the device; merge the two interpolation results into a dual-channel signal, which is used as the input timing signal of the model.

5. The method for fault diagnosis of a lightweight hydraulic pump applicable to all working conditions as described in claim 4, characterized in that, Step S3 specifically includes the following steps: S31. Perform FFT decomposition on the two channel signals respectively to obtain the real and imaginary parts; S32. Convolution feature extraction is performed on the real and imaginary parts respectively. Through multiple iterations, pseudo-features, including resampling errors and artifacts, will be gradually removed during the iteration process. S33. After convolution learning, the real and imaginary parts are merged back into a complex structure, and then iFFT inverse transformation is performed to convert it back into a time-series signal. S34. Perform residual connection between the signal obtained in step S33 and the resampled signal of the original input, retaining local features and learned global features.

6. The lightweight hydraulic pump fault diagnosis method applicable to all working conditions according to claim 2, characterized in that, Step S4 specifically includes the following steps: S41. Perform FFT analysis on the dual-channel resampled signal obtained in step S2, normalize the maximum and minimum values ​​of the angle domain amplitude, and obtain the first K dominant orders as multi-period decomposition factors for subsequent signals. S42. Transform the signal according to the dominant order, convert the time series into a two-dimensional tensor with multiple periods, and embed the intra-period and inter-period variations into columns and rows respectively; S43. Design a shared feature extractor to automatically extract features of the dominant order, suppress other non-order features, and focus on retaining the essential features of the order that have generalizability. S44. After extracting the corresponding order features, restore the two-dimensional order feature map to a one-dimensional sequence and remove the padding introduced by the previous splicing operation. S45. Use average pooling to extract features that are translation invariant. S46. Design a SelfNorm module suitable for one-dimensional feature sequences to alleviate the distribution difference between the training and testing phases; the SelfNorm module uses the mean and standard deviation within the channel as the statistical basis, and uses lightweight convolution to adaptively generate scaling and offset factors to dynamically adjust the feature distribution. S47. After the feature decoupling module of each component, a set of shared one-dimensional convolutional layers is used to realize feature convergence. In the feature convergence process, average pooling operation is introduced to compress the feature length, reduce redundant information, and contribute to the feature convergence layer so that the domain-independent gradient can more fully act on the shallow feature extractor during backpropagation and suppress the expression of working condition features. S48. Design a residual connection method to align the length of the feature output of the previous layer through convolution and pooling, and connect it to the next layer as a residual term. This alleviates the gradient vanishing problem and promotes the efficient transfer of deep features, thereby improving the training stability and expressive power of the model in a multi-layer structure.

7. A fault diagnosis method for a lightweight hydraulic pump applicable to all working conditions as described in claim 2, characterized in that, The implementation method of step S5 is as follows: design a fault state master classifier driven by the fault type and fault location sub-tasks, and decouple feature type and fault location information through a fully connected layer to improve the attention to key fault features.

8. A fault diagnosis method for a lightweight hydraulic pump applicable to all working conditions as described in claim 2, characterized in that, The implementation method of step S6 is as follows: Since the signal is affected by rotational speed and pressure, a gradient inversion layer is introduced to invert the gradient during back propagation, which prompts the feature extractor to extract shared features that are independent of the operating conditions, thereby improving the generalization ability of fault diagnosis.

9. A fault diagnosis method for a lightweight hydraulic pump applicable to all working conditions as described in claim 2, characterized in that, Step S7 is implemented by simultaneously optimizing two objectives: minimizing the class loss function L. y and maximizing the neighborhood classification loss function L d This guides the model to extract general features that are independent of operating conditions.

10. A computer-readable storage medium having stored thereon computer program instructions executable by a processor, wherein when the processor executes the computer program instructions, it is able to implement the steps of the method as described in any one of claims 1-9.

Images