An interpretable fault diagnosis method for axial piston pump based on outlet pressure fluctuation order-phase evidence mapping

Abstract

The present application relates to a kind of axial plunger pump explainable fault diagnosis method based on export pressure pulsation order-phase evidence mapping, belong to hydraulic equipment state monitoring and explainable artificial intelligence technical field.The method constructs a rotation pressure sample with shaft rotation period, generates frequency-phase diagram with order-phase coordinate after training set parameter standardization, continuous wavelet transform, combined with block level disturbance attribution and health benchmark energy ratio analysis, the model diagnosis basis is mapped to the physical evidence unit containing shaft order, pumping order, rotation phase and contribution value, can simultaneously output fault category and traceable diagnostic evidence, improve the engineering usability of diagnostic result.The present application can convert the image domain diagnostic basis of deep model into the order-phase physical evidence related to the kinematics of axial plunger pump, improve the physical traceability and engineering explainability of fault diagnosis result.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic equipment condition monitoring and interpretable artificial intelligence technology, specifically involving an interpretable fault diagnosis method for axial piston pumps based on outlet pressure pulsation order-phase evidence mapping. Background Technology

[0002] Axial piston pumps are core power conversion components in hydraulic systems, widely used in aerospace, marine engineering, construction machinery, and high-end manufacturing equipment. Key friction pairs such as the piston assembly, slipper assembly, and flow distribution assembly are prone to wear, leakage, and abnormal flow distribution under high-speed, high-pressure, and high-shear operating environments, leading to problems such as outlet pressure pulsation, flow pulsation, vibration and noise, and efficiency reduction. Outlet pressure pulsation is formed by the combined effects of piston periodic discharge, piston chamber pressure build-up, flow distribution transition, oil compressibility, internal leakage, and external hydraulic line impedance. Compared to external vibration or noise signals, outlet pressure pulsation more directly reflects the coupling state of fluid dynamics and mechanical kinematics within the pump, and pressure sensors are easier to deploy and acquire online in actual hydraulic systems. Therefore, axial piston pump condition monitoring and fault diagnosis based on outlet pressure pulsation signals has high engineering application value.

[0003] The existing fault diagnosis methods for axial piston pumps can be explained as follows:

[0004] (a) Condition assessment of plunger pumps based on pressure pulsation, frequency domain energy, statistical characteristics, or health factors:

[0005] CN117869288A is a scheme that acquires the pressure signal of an axial piston pump, calculates a health factor based on the pulsation characteristics of the pressure signal, and then determines the health status of the axial piston pump based on the health factor; CN115898850A is a scheme that collects signals such as pressure, flow, vibration, and temperature of an axial piston pump through sensors, and extracts time-domain and frequency-domain related feature information at the edge end to achieve efficient transmission of operating status data.

[0006] The aforementioned methods can construct state evaluation indicators using pressure pulsations or multi-source operating signals, offering advantages such as direct implementation, relatively low computational load, and ease of online monitoring. However, for axial piston pumps, pressure pulsations exhibit significant rotational periodicity and piston passage periodicity. If only global frequency domain energy or health factors are used, the correspondence between local changes in pressure pulsations and the pump's kinematic phase can be weakened, leading to insufficient physical traceability of diagnostic conclusions.

[0007] (II) Fault diagnosis of rotating machinery or piston pumps based on S-transform, short-time Fourier transform, continuous wavelet transform, time-frequency graph and deep neural network:

[0008] CN108106830B is a scheme that obtains the normalized time spectrum through multi-resolution generalized S-transform and performs binarization, fusion, and segmentation of the time spectrum to achieve fault diagnosis; CN108830127A is a scheme that performs short-time Fourier time-frequency transform on the vibration signal of rotating machinery to obtain a time-frequency representation and uses it as input to a convolutional neural network; CN109596326A is a scheme that converts the working signal of rotating machinery into a grayscale image and performs fault classification through a convolutional neural network.

[0009] The aforementioned techniques improve the feature representation capability of non-stationary fault signals and reduce the reliance on manually designed features. However, most of them use time-frequency maps obtained by S-transform, short-time Fourier transform, or continuous wavelet transform as ordinary two-dimensional images input into deep networks, focusing on improving classification accuracy or enhancing noise resistance, without establishing complete physical coordinate constraints for the rotational periodicity of the axial piston pump outlet pressure pulsation.

[0010] (III) Interpretability studies of deep diagnostic models:

[0011] Existing research often uses heatmaps, class activation mapping, gradient-weighted class activation mapping, or mask perturbation to display the regions of interest in the model. CN119202905A presents a scheme that uses continuous wavelet transform to obtain a two-dimensional wavelet time-frequency image of vibration signals and introduces gradient-weighted class activation heatmap (Grad-CAM) to visualize the features extracted by the model during the diagnostic process. CN113409213B presents a scheme that uses the gradient magnitude of the feature map of a specific network layer in a convolutional neural network to find key regions in the time-frequency image and removes non-key regions to enhance the time-frequency image.

[0012] The methods described above can, to some extent, show which regions in the time-frequency map or input image the deep model focuses on. However, the interpretation results are mostly expressed as heatmaps, salient regions, class activation maps, or gradient response regions, with the interpretation mainly remaining in the image or pixel domain. Simply using heatmaps, salient maps, CAM, Grad-CAM, or similar mask visualization methods is insufficient to generate engineering-readable and traceable order-phase physical evidence.

[0013] In summary, while existing technologies have made some progress in pressure pulsation health assessment, time-frequency plot-based deep diagnosis, and deep model interpretability, they still have the following shortcomings: First, state assessment methods based on pressure pulsation health factors or frequency domain energy typically focus on global indicators, making it difficult to express local order-phase diagnostic evidence within one revolution cycle; Second, diagnostic methods based on S-transform, short-time Fourier transform, CWT time-frequency plots, and deep neural networks typically treat time-frequency plots as ordinary image processing, lacking explicit mappings between them and shaft rotation frequency, plunger passage frequency, pumping order, and one-revolution phase; Third, interpretable methods based on heatmaps, saliency maps, CAM, Grad-CAM, or mask perturbations can mostly only point out the image regions of interest to the model, and cannot yet form physical evidence units that include frequency range, shaft order range, pumping order range, one-revolution phase range, and contribution value; Fourth, existing interpretation methods typically do not jointly analyze the model contribution with the health baseline wavelet energy ratio, making it difficult to distinguish between energy enhancement evidence, energy decay evidence, energy redistribution evidence, and non-dominant energy change regions.

[0014] Therefore, it is necessary to propose an interpretable fault diagnosis method for axial piston pump outlet pressure pulsation signals. This method establishes a complete information chain between one-revolution sample construction, CWT frequency-phase characterization, diagnostic model identification, local contribution extraction, and physical coordinate mapping. This allows the image domain diagnostic basis of the deep model to be converted into order-phase physical evidence related to the kinematics of the axial piston pump. Furthermore, the energy change type corresponding to this evidence can be interpreted by combining the health baseline energy ratio, thereby improving the physical traceability, engineering interpretability, and practical application credibility of the fault diagnosis results. Summary of the Invention

[0015] The purpose of this invention is to solve the problems in existing axial piston pump fault diagnosis methods based on pressure pulsation time-frequency maps, such as the lack of shaft order, pumping order, and one-revolution phase constraints in the time-frequency map pixels, the difficulty in converting local evidence from the deep model into engineering-readable physical evidence, and the difficulty in distinguishing model contributions from the type of pressure pulsation energy change. This invention provides an interpretable fault diagnosis method for axial piston pumps based on outlet pressure pulsation order-phase evidence mapping.

[0016] To achieve the above objectives, the technical solution of the present invention is: an interpretable fault diagnosis method for axial piston pumps based on outlet pressure pulsation order-phase evidence mapping, comprising:

[0017] Collect the outlet pressure pulsation signal of the axial piston pump, obtain the number of pistons, shaft speed and sampling frequency, calculate the shaft rotation frequency and piston passing frequency, and establish the coordinate mapping relationship between frequency components and shaft order and pumping order.

[0018] Using the shaft rotation period as a constraint, the continuous pressure pulsation signal is divided into one-rotation pressure samples, and one-rotation phase coordinates are established for each one-rotation pressure sample.

[0019] The pressure samples of the first rotation are standardized and then subjected to continuous wavelet transform to obtain a frequency-phase map synchronized with the phase coordinates of the first rotation, while retaining the original wavelet energy map corresponding to the pixel position of the frequency-phase map.

[0020] Input the frequency-phase diagram into the diagnostic model to output the health status or fault category of the axial piston pump;

[0021] The frequency-phase map is divided into multiple image blocks according to the frequency or order direction and the phase direction. The contribution value of each image block to the target category is calculated based on the block-level perturbation attribution.

[0022] Based on the coordinate mapping relationship, the pixel index range of each image block is converted into the axis order range, pumping order range and one-turn phase range to form a physical evidence unit;

[0023] Based on the original wavelet energy map, the health baseline energy ratio of the target category relative to the health status is calculated.

[0024] By jointly determining the contribution value and contribution ratio of image patches with the energy ratio of the health baseline, evidence of energy enhancement, energy decay, energy redistribution, or non-dominant energy change areas can be obtained.

[0025] The output includes fault category, physical evidence unit, contribution value, health baseline energy ratio, and order-phase interpretable diagnostic results for the evidence type.

[0026] Furthermore, the construction and standardization of the single-rotation pressure sample includes:

[0027] For the original pressure record p i = [p i (1), p i (2), ..., p i (L i )] T The number of sampling points per revolution is calculated using the following formula:

[0028] N r = round(f s / f r )

[0029] Where L i f is the length of the original pressure record. s f is the sampling frequency. r The shaft rotation frequency;

[0030] The number of one-revolution pressure samples Q is determined according to the following formula.i And define the qth first-rotation pressure sample x i,q :

[0031] Q i = [L i / N r ]

[0032] x i,q = [p i ((q-1)N r +1), ..., p i (qN r )] T , q=1,2,…, Q i

[0033] Each pressure sample is resampled to a uniform length K, and a deterministic phase coordinate is established as follows:

[0034] θ k = 2π(k-1) / (K-1),k=1,2,…,K

[0035] The standardization is based on the global mean of the training set. and global standard deviation of the training set The standardized expression is:

[0036]

[0037] Where ε is a positive number to prevent division by zero; the and It is calculated only from the training set to avoid leakage of information from the validation and test sets.

[0038] Furthermore, the continuous wavelet transform employs a wavelet function with time-frequency localization capability, including the complex Morlet wavelet, and constructs the frequency-phase diagram and the original wavelet energy diagram using the following formula:

[0039]

[0040]

[0041] in, This is a frequency-phase diagram. For wavelet energy maps, For frequency, For phase.

[0042] Furthermore, the diagnostic model employs a convolutional neural network model, comprising several convolutional blocks, a global average pooling layer, a Dropout layer, and a Softmax classification layer; each convolutional block includes a two-dimensional convolutional layer, a batch normalization layer, a nonlinear activation layer, and a pooling layer, wherein the global average pooling layer is used to obtain a feature vector that can characterize global order-phase features.

[0043] Furthermore, the frequency-phase map is divided into multiple image blocks according to the frequency or order direction and the phase direction, specifically including:

[0044] Divide the frequency-phase diagram into N f ×N θ N non-overlapping image patches, where N f N is the number of divisions in the direction of frequency or order. θ The number of image patches is M = N, representing the number of divisions in the phase direction. f N θ For each image patch, establish a binary mask vector z=[ , ,…, ] T , =1 indicates that the m-th image patch is retained. =0 indicates that the m-th image patch is occluded; the occluded image patch is not zero-filled, but is replaced by a region at the same position in the average frequency-phase map of the training set, in order to reduce the impact of non-physical perturbations on the attribution results.

[0045] Furthermore, the block-level perturbation attribution employs the Kernel SHAP method. Based on the block-level perturbation attribution, the contribution value of each image block to the target category is calculated. Specifically, for target category c, the perturbation input under different mask combinations is fed into the diagnostic model to obtain the model output. The Shapley contribution value of the m-th image block is estimated through a weighted linear substitution model, where the calculation of the weighted linear substitution model is as follows:

[0046]

[0047] in, Let m be the contribution value of the m-th image patch to the target category output. This is the baseline output of the model for category c.

[0048] Furthermore, the physical evidence unit is represented as follows:

[0049]

[0050] in, and For the corresponding axis order range, and To correspond to the pumping order range, and To correspond to the phase range of one revolution, This represents the contribution value of the corresponding image patch to category c.

[0051] Furthermore, the physical evidence units are aggregated and statistically analyzed according to the low-order modulation region, the plunger passing through the fundamental frequency region, the second pumping harmonic region, the third pumping harmonic region, the adjacent order modulation region, and the high-order broadband response region.

[0052] Furthermore, the health baseline energy ratio and positive contribution percentage relative to the health state H are calculated for category c and the m-th image patch using the following formulas:

[0053]

[0054]

[0055] in, Let CWT be the average CWT energy of category c on image patch m. The average CWT energy of the healthy state on image patch m. ε is a positive number to prevent division by zero. For the first The average contribution value of each image patch, where M is the total number of image patches.

[0056] Furthermore, the joint determination specifically includes:

[0057] when and ρ m (c) When > 1 + δ, it is considered evidence supporting energy enhancement; when and ρ m (c) When < 1 - δ, it is considered evidence supporting energy decay; when and 1 - δ ≤ ρ m (c) When ≤ 1 + δ, it is considered evidence supporting energy redistribution; when and |ρ m (c) When -1| > δ, it is determined to be a non-dominant energy change region, where The contribution threshold is δ, and the energy tolerance is δ.

[0058] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0059] The present invention also provides a computer-readable medium having processor-executable non-volatile program code that causes the processor to perform the above-described method.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. This invention uses the axis rotation period as a sample construction constraint to avoid the problems of incomplete rotation period and phase inconsistency caused by arbitrary sliding windows.

[0062] 2. In the CWT characterization, the present invention explicitly retains the frequency, shaft order, pumping order and first-turn phase, so that each local region in the time-frequency diagram has a traceable physical meaning.

[0063] 3. This invention obtains local diagnostic evidence through block-level Kernel SHAP and maps image domain evidence into order-phase evidence units that engineers can understand.

[0064] 4. This invention interprets the model contribution in conjunction with the health baseline energy ratio, which can distinguish between energy enhancement, energy decay and energy redistribution, and avoids simply equating positive contribution with energy increase.

[0065] 5. This invention can improve the physical interpretability, engineering reliability, and online monitoring applicability of diagnostic conclusions while maintaining a high fault identification capability. Attached Figure Description

[0066] Figure 1 This is a flowchart of the axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping as described in this invention.

[0067] Figure 2 This is a schematic diagram illustrating the file-level test performance evaluation of the fault diagnosis model in this invention.

[0068] Figure 3 This is a schematic diagram comparing the CWT frequency-phase diagram, SHAP contribution, and health baseline energy ratio of different types of samples in this invention.

[0069] Figure 4 This is a schematic diagram illustrating the aggregated statistics of physical evidence units of different categories of samples in this invention, and the energy ratio of relatively healthy samples. Detailed Implementation

[0070] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0071] This invention provides an interpretable fault diagnosis method for axial piston pumps based on outlet pressure pulsation order-phase evidence mapping, including:

[0072] Collect the outlet pressure pulsation signal of the axial piston pump, obtain the number of pistons, shaft speed and sampling frequency, calculate the shaft rotation frequency and piston passing frequency, and establish the coordinate mapping relationship between frequency components and shaft order and pumping order.

[0073] Using the shaft rotation period as a constraint, the continuous pressure pulsation signal is divided into one-rotation pressure samples, and one-rotation phase coordinates are established for each one-rotation pressure sample.

[0074] The pressure samples of the first rotation are standardized and then subjected to continuous wavelet transform to obtain a frequency-phase map synchronized with the phase coordinates of the first rotation, while retaining the original wavelet energy map corresponding to the pixel position of the frequency-phase map.

[0075] Input the frequency-phase diagram into the diagnostic model to output the health status or fault category of the axial piston pump;

[0076] The frequency-phase map is divided into multiple image blocks according to the frequency or order direction and the phase direction. The contribution value of each image block to the target category is calculated based on the block-level perturbation attribution.

[0077] Based on the coordinate mapping relationship, the pixel index range of each image block is converted into frequency range, axis order range, pumping order range and one-turn phase range to form a physical evidence unit.

[0078] Based on the original wavelet energy map, the health baseline energy ratio of the target category relative to the health status is calculated.

[0079] By jointly determining the contribution value and contribution ratio of image patches with the energy ratio of the health baseline, evidence of energy enhancement, energy decay, energy redistribution, or non-dominant energy change areas can be obtained.

[0080] The output includes fault category, physical evidence unit, contribution value, health baseline energy ratio, and order-phase interpretable diagnostic results for the evidence type.

[0081] The following is a detailed implementation process of the present invention.

[0082] This example provides an interpretable fault diagnosis method for axial piston pumps based on outlet pressure pulsation order-phase evidence mapping, which will be implemented according to... Figure 1 The process shown is used to construct an interpretable diagnostic model for the physical coordinates of the pressure pulsation of an axial piston pump.

[0083] S1. Collect the outlet pressure p(t) of the axial piston pump. For an axial piston pump with z pistons and a shaft speed of n r / min, the shaft rotation frequency f r and the plunger through frequency f p They are respectively:

[0084] f r =n / 60, fp =zf r (1)

[0085] For any frequency component f in the pressure pulsation, its shaft order and pumping order are defined as follows:

[0086] O s = f / f r O p = f / f p (2)

[0087] Among them, O s This indicates the order of the frequency relative to the shaft rotation frequency, O p This indicates the order of the frequency relative to the frequency at which the plunger passes. Using this definition, the ordinary frequency axis can be converted into a physical coordinate axis directly related to the pump's kinematics.

[0088] The collected continuous pressure is expressed as p i = [p i (1), p i (2), ..., p i (L i )] T Calculate the number of sampling points per revolution:

[0089] N r = round(f s / f r (3)

[0090] Among them, L i To record the length, f s The sampling frequency.

[0091] The continuous pressure record is divided into Q. i = ⌊L i / N r The ⌋th one-turn sample, the qth one-turn sample is:

[0092] x i,q = [p i ((q-1)N r +1), ..., p i (qN r )] T (4)

[0093] Each sample covers one complete shaft rotation cycle. To accommodate differences in sample length under different sampling frequencies or rotation speeds, each sample is resampled to a uniform length K, and phase coordinates are established.

[0094] θ k= 2π(k-1) / (K-1) , k = 1, 2, ..., K (5)

[0095] Through this step, the horizontal axis of the sample is no longer an arbitrary time index, but a phase coordinate with a defined physical meaning.

[0096] To avoid data leakage, the mean is calculated only from the training set. and standard deviation The samples in the training, validation, and test sets are standardized using the same parameters:

[0097] (6)

[0098] in, This is a resampled sample. To prevent division by zero for positive numbers, this step ensures that model training, validation, and testing are performed on the same statistical scale.

[0099] S2. Perform continuous wavelet transform on the standardized one-transform sample to obtain the wavelet coefficients:

[0100] (7)

[0101] in, 'a' is the scale parameter, 'b' is the translation parameter, and 'ψ(t)' is the mother wavelet function. Scale 'a' is mapped to a pseudo-frequency. And further mapped to axis order. :

[0102] (8)

[0103] in, The center frequency of the mother wavelet. This is the time interval between adjacent sampling points in a resampled sample. Therefore, each CWT pixel corresponds to a frequency, an axis order, a pumping order, and a phase position.

[0104] To reduce the amplitude difference between strong pumping harmonics and weak fault disturbances, the logarithmic amplitude diagram will be used. As input to a convolutional neural network:

[0105] (9)

[0106] At the same time, the original CWT energy map is retained. :

[0107] (10)

[0108] The logarithmic magnitude plot is used for model learning, and the original energy plot is used for subsequent calculation of the health baseline energy ratio.

[0109] The CWT frequency-phase map is adjusted to a uniform size and then input into a compact convolutional neural network. The network consists of three convolutional blocks, each including two-dimensional convolution, batch normalization, ReLU activation, and max pooling; after the convolutional blocks, a global average pooling layer, a Dropout layer, and a Softmax classification layer are connected.

[0110] The model training uses the cross-entropy loss function with a regularization term:

[0111] (11)

[0112] Where N is the number of training samples and C is the number of classes. For real labels, Let W be the model's predicted probability, W be the set of trainable weights, and λ be the regularization coefficient. To reduce the risk of overfitting with small samples, file-level partitioning, early stopping strategies, and a mechanism for preserving the weights of the model with the minimum validation loss can be employed.

[0113] S3, The frequency-phase diagram S i,q ∈ ℝ H × W Divided into N f N θ There are n non-overlapping image patches, and the total number of image patches is M = N. f N θ For the m-th image patch Define binary mask variables .like =1, then retain the image patch; if =0, then use the training set average frequency-phase map. The corresponding region in the image is replaced with the image patch. The perturbation input is represented as:

[0114] (12)

[0115] For target class c, the perturbation input is fed into the trained model. A linear substitution model is fitted using Kernel SHAP as follows:

[0116] g(z) = φ0 + (13)

[0117] in, Let m be the contribution value of the m-th image patch to the target category output. This is the baseline output of the model for category c. A value greater than zero indicates that the image patch supports category c. A value less than zero indicates that the image patch has a suppressive effect on category c.

[0118] S4. For the m-th image patch, determine its pixel index range in the frequency direction. ] and the pixel index range in the phase direction [ Based on the CWT coordinate vector, convert it to a frequency range. and phase range ]; combined with f r and f p This yields the axis order range and the pumping order range. Therefore, the m-th image block can be represented as the following physical evidence unit:

[0119] (14)

[0120] This physical evidence unit enables local image evidence of the deep model to be translated into engineering-readable statements, such as "fault support evidence was obtained in the region near the seventh axis order and in a certain phase interval of the model".

[0121] To facilitate engineering interpretation, the physical evidence units are further aggregated into regions of different orders. For a z-plunger axial plunger pump, the k-th pumping harmonic region can be defined as:

[0122] (15)

[0123] in, Let m be the order of the center axis of the m-th image patch. The half-bandwidth is determined based on the CWT frequency resolution or image block bandwidth. The remaining image blocks can be assigned to low-order modulation regions, pump harmonic adjacent modulation regions, and high-order broadband response regions. This grouping is not directly equivalent to fault mechanism determination, but rather provides an organizational scale related to pump kinematics for statistical evidence.

[0124] S5. For category c and image patch m, calculate the energy ratio relative to the healthy state H:

[0125] (16)

[0126] in, Let CWT be the average CWT energy of category c on image patch m. Let the average CWT energy be the health status energy on image patch m. Further calculation of the positive contribution percentage is then performed.

[0127] (17)

[0128] Based on contribution threshold And the energy tolerance δ, establish the following joint interpretation rule: when and ρ m (c)When the value is greater than 1 + δ, this region is interpreted as evidence supporting energy enhancement; when... and ρ m (c) When < 1 - δ, this region is interpreted as evidence supporting energy decay; when and 1 - δ ≤ ρ m (c) When ≤ 1 + δ, this region is interpreted as supporting evidence for energy redistribution; when and |ρ m (c) When -1| > δ, this region is interpreted as a region of non-dominant energy change.

[0129] S6. In one specific embodiment, the method of the present invention is verified using the outlet pressure pulsation data of a seven-plunger axial piston pump. The axial piston pump has a plunger number z=7, a shaft speed n=2000 r / min, and a working pressure of 10 MPa. The shaft rotation frequency f can be obtained from equation (1). r =33.33 Hz, the plunger passes through at a frequency f p =233.33 Hz. Therefore, the 7th axis order corresponds to the fundamental frequency of the plunger, the 14th axis order corresponds to the second pumping harmonic, and the 21st axis order corresponds to the third pumping harmonic.

[0130] A. The original outlet pressure pulsation is segmented according to the shaft rotation cycle, so that each sample corresponds to a complete shaft rotation cycle. According to formula (3), the sampling points of the one-rotation pressure sample are calculated to be 1536 sampling points, and the one-rotation phase coordinates are established according to formula (5), where K=1536. The embodiment obtains four types of state samples, including healthy state, slipper pair wear, plunger pair wear, and distribution pair wear. Among them, slipper pair wear, plunger pair wear, and distribution pair wear are represented by fault 1, fault 2, and fault 3, respectively, with 62 one-rotation samples in each category, for a total of 248 one-rotation samples. In order to prevent adjacent one-rotation samples in the same original file from entering the training set and the test set at the same time and causing data leakage, a file-level partitioning method is adopted to divide the 248 samples into 160 training samples, 32 verification samples, and 56 test samples.

[0131] B. Calculate the global mean using the training set. and global standard deviation The training set, validation set, and test set are processed with the same standardized parameters according to equation (6). A continuous wavelet transform is performed on the standardized one-turn pressure samples, using a complex Morlet wavelet as the mother wavelet with a center frequency f. c=1 Hz. The retained CWT frequency range is 90.68-7657.90 Hz, corresponding to an axis order range of 2.72-229.74. This range covers low-order frequency-related components, the piston through fundamental frequency, the second and third pumping harmonics, and higher-order broadband pressure pulsation responses. The logarithmic amplitude is calculated according to Equation (9) and adjusted to a size of 128×128 before being used as input to the convolutional neural network. At the same time, the CWT energy map is calculated according to 2.10 as the basis for subsequent health baseline energy ratio calculation.

[0132] A 128×128 frequency-phase map was input into a compact convolutional neural network. The model was trained using the Adam optimizer with an initial learning rate of 1×10^-3 and a batch size of 16. The loss function was cross-entropy loss with an added L2 regularization term, where the L2 regularization coefficient λ = 1×10^-4. Training was stopped when the validation set loss did not decrease for 20 consecutive training epochs, and the model weights with the minimum validation loss were saved for testing. The trained CWT-CNN model was used on a file-level test set, which contained 14 one-turn samples per class, for a total of 56 one-turn samples across the four classes. The test results are as follows: Figure 2 As shown, the test samples for healthy status, fault 1, fault 2, and fault 3 were all correctly identified, with an overall test accuracy of 100%. The precision, recall, and F1-score for each category were all 1.000. Furthermore, the feature vectors output by the global average pooling layer were projected using two-dimensional PCA, and the four classes of samples formed discriminative clusters in the projection space. The first and second principal components explained 41.2% and 32.2% of the feature variance, respectively. The maximum softmax probability of the model output was higher than 0.975 for most samples and higher than 0.990, indicating that the CWT frequency-phase representation in this embodiment can provide a stable classification basis for the compact network.

[0133] C. Divide each 128×128 frequency-phase diagram into N sections along the frequency or order direction. f =12 segments, divided into N along the first phase direction. θ =8 segments, resulting in a total of M=96 non-overlapping image patches. For each image patch Establish a binary mask variable z m When z m When = 0, the average frequency-phase graph of the training set is used. Perturbed images with different mask combinations are input into a trained CWT-CNN model, and the contribution value φ of each image patch to the target class c is estimated using the Kernel SHAP linear substitution model. m .Depend on Figure 3It can be seen that although the CWT frequency-phase diagrams of all four states contain the main strip structure formed by the plunger action, the block-level SHAP evidence diagrams show obvious class differences. In the healthy state, positive evidence is mainly distributed in the high-order broadband region above the first three pumping harmonics, and a secondary support region appears near the 14th axis order, while the contribution of the low-order region below the 7th axis order is weak or slightly suppressed, indicating that the model's judgment of the healthy state does not rely on a single harmonic peak, but on regular multi-order texture and the absence of abnormal low-order modulation. The positive evidence for fault 1 is more phase-selective near the 14th axis order and in the high-order broadband region, with local contributions concentrated in several corner intervals, indicating that the discrimination criterion for this state mainly comes from local order-phase perturbations, rather than uniform harmonic enhancement around the entire circumference. Positive evidence for fault 2 is concentrated in the high-order broadband band from approximately the 50th to the 80th axis, spanning a relatively wide phase range. The contributions of the main pumping harmonics of the 7th, 14th, and 21st axes are relatively low, indicating that the model primarily utilizes broadband textures above the main pumping harmonics to identify this state. Positive evidence for fault 3 is distributed across a relatively wide high-order range from approximately the 50th to the 150th axis, with a larger evidence bandwidth than fault 2, indicating that this state corresponds to a wider range of high-order pressure pulsations.

[0134] For the m-th image block, first determine its frequency direction index range and phase direction index range in the 128×128 image, and then convert it to the frequency range using the CWT coordinate vector. and phase range ]. Combined with f r =33.33 Hz and f p =233.33 Hz, further obtaining the axis order range [ and the range of pumping stages [ This forms the physical evidence unit of equation (14). In this embodiment, the physical evidence unit is aggregated into the low-order modulation region, the plunger-through fundamental frequency region, the second-pump harmonic region, the third-pump harmonic region, the adjacent-order modulation region, and the high-order broadband response region. Figure 4 (a) It can be seen that the positive SHAP evidence for the health status is distributed across multiple order regions. The low-order modulation region, the second pumping harmonic region, the third pumping harmonic region, and the mid-order broadband response region all contribute to the health status, indicating that the health category corresponds to a relatively stable multi-order rule representation, rather than a single-point feature of an isolated order. The largest positive contribution of fault 1 comes from the second pumping harmonic region, with a contribution ratio of approximately 0.48. The low-order modulation region also provides a certain positive contribution. Figure 4(b) The mid-to-high order broadband response region shows a strong negative contribution to fault 1, indicating that this region plays a greater role in suppressing the judgment of fault 1. The positive and absolute contributions of fault 2 are both significantly concentrated in the mid-to-high order broadband response region, with the positive contribution accounting for approximately 0.63% and the absolute contribution accounting for approximately 0.57%. The contribution of the main pumping harmonic region is lower than that of this broadband region, indicating that the judgment basis of fault 2 is mainly the difference in mid-to-high order broadband texture. The evidence distribution of fault 3 is more dispersed than that of fault 2. The mid-to-high order broadband response region is still the main positive contribution region, while the low-order modulation region and the plunger through the fundamental frequency region also participate in the judgment. Figure 4 (b) The secondary pumping harmonic region has a significant negative contribution to fault 3, indicating that the relatively stable secondary harmonic structure under healthy or other conditions is more reflected as missing or attenuated information in fault 3 discrimination. Figure 4 (c) To further explain, the absolute SHAP contribution not only includes positive evidence supporting the target class, but also negative evidence excluding the target class, thus it can more comprehensively reflect the degree to which regions of each order participate in the model's discrimination.

[0135] Using the CWT energy of the healthy state samples as a benchmark, the healthy benchmark energy ratio ρ for fault 1, fault 2, and fault 3 is calculated in each image patch. m (c) And combined with the proportion of positive contribution η m (c) An interpretation type determination is performed. In this embodiment, the contribution threshold is used. Take 0.05, and the energy tolerance δ is taken as 0.05. Combined with... Figure 3 (e) and Figure 3 (f) shows that the model evidence for fault 1 includes both the positive harmonic correlation contribution near the 14th axis order and the energy attenuation information in the adjacent region. Figure 4 (d) The healthy baseline energy ratio in the second pumping harmonic region is approximately 1.06, while the energy ratio in the third pumping harmonic region and the mid-order broadband response region is as low as 0.88, indicating that the interpretable evidence for fault 1 belongs to a mixed mode of "local energy enhancement + adjacent energy attenuation". Combined with... Figure 3 (h) Figure 3 (i) and Figure 4 (d) It can be seen that the main positive evidence for fault 2 is concentrated in the mid-order broadband response region, but the healthy baseline energy ratio in this region is approximately 1.01, close to the healthy baseline. Although the energy ratios of the plunger through the fundamental frequency region and the second pumping harmonic region are approximately 1.11 and 1.18 respectively, their positive SHAP contribution is lower than that in the mid-order broadband response region. Therefore, fault 2 is not dominated by the simple amplification of the main pumping harmonics, but is more consistent with redistribution-type evidence of changes in the texture or energy distribution pattern in the mid-order broadband region. Figure 3 (k) Figure 3 (l) and Figure 4(d) It can be seen that the positive SHAP evidence for fault 3 occurs simultaneously with a large-scale energy attenuation. The energy ratio in the secondary pumping harmonic region is approximately 0.80, in the plunger passing through the fundamental frequency region it is approximately 0.68, and in the mid-order broadband response region it is approximately 0.61, indicating that the model can use the reduction of the original regular pressure pulsation component under healthy conditions as a classification criterion. Therefore, the positive SHAP contribution in this method does not necessarily indicate a local energy enhancement; its physical meaning needs to be judged in conjunction with the healthy baseline energy ratio.

[0136] For the healthy state, the model's positive evidence mainly manifests as relatively dispersed high-order evidence, with energy levels close to the healthy baseline. For Fault 1, the model evidence exhibits phase-selective harmonic correlation evidence, accompanied by a mixture of local energy enhancement and energy attenuation in adjacent regions, with local energy enhancement near the second pumping harmonic and slight attenuation in some adjacent regions observed. For Fault 2, the model's positive evidence is mainly concentrated in the high-order broadband region of approximately orders 50-80 within the retained CWT axis, spanning a wide phase range. The positive contribution of the main pumping harmonic region is lower than that of this high-order region, indicating that its discrimination criterion is closer to the energy redistribution within the high-order broadband region. For Fault 3, the model's positive evidence is distributed across a wider high-order range, approximately orders 50-150, while the energy ratio plot shows significant energy attenuation in both the main pumping harmonic and high-order broadband regions, indicating that the model can utilize the reduction in the original regular pressure pulsation component under the healthy state as classification evidence.

[0137] The above embodiments demonstrate that the present invention does not merely output health or fault categories, but rather transforms the local discriminative evidence of the CWT-CNN model into physical evidence units containing frequency range, axial order range, pumping order range, one-turn phase range, and contribution values. Furthermore, it combines this with the health baseline energy ratio to explain whether the evidence represents energy enhancement, energy decay, or energy redistribution. Thus, the model's judgment criteria can be traced back from the ordinary image domain to the order-phase physical coordinates of the axial plunger pump pressure pulsations, improving the physical traceability and engineering interpretability of fault diagnosis results.

[0138] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0139] The present invention also provides a computer-readable medium having processor-executable non-volatile program code that causes the processor to perform the above-described method.

[0140] 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 method for interpretable fault diagnosis of axial piston pumps based on outlet pressure pulsation order-phase evidence mapping, characterized in that, include: Collect the outlet pressure pulsation signal of the axial piston pump, obtain the number of pistons, shaft speed and sampling frequency, calculate the shaft rotation frequency and piston passing frequency, and establish the coordinate mapping relationship between frequency components and shaft order and pumping order. Using the shaft rotation period as a constraint, the continuous pressure pulsation signal is divided into one-rotation pressure samples, and one-rotation phase coordinates are established for each one-rotation pressure sample. The pressure samples of the first rotation are standardized and then subjected to continuous wavelet transform to obtain a frequency-phase map synchronized with the phase coordinates of the first rotation, while retaining the original wavelet energy map corresponding to the pixel position of the frequency-phase map. Input the frequency-phase diagram into the diagnostic model to output the health status or fault category of the axial piston pump; The frequency-phase map is divided into multiple image blocks according to the frequency or order direction and the phase direction. The contribution value of each image block to the target category is calculated based on the block-level perturbation attribution. Based on the coordinate mapping relationship, the pixel index range of each image block is converted into the axis order range, pumping order range and one-turn phase range to form a physical evidence unit; Based on the original wavelet energy map, the health baseline energy ratio of the target category relative to the health status is calculated. By jointly determining the contribution value and contribution ratio of image patches with the energy ratio of the health baseline, evidence of energy enhancement, energy decay, energy redistribution, or non-dominant energy change areas can be obtained. The output includes fault category, physical evidence unit, contribution value, health baseline energy ratio, and order-phase interpretable diagnostic results for the evidence type.

2. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping as described in claim 1, characterized in that, The construction and standardization of the single-rotation pressure sample includes: For the original pressure record p i = [p i (1), p i (2), ..., p i (L i )] T The number of sampling points per revolution is calculated using the following formula: N r = round(f s / f r ) Where L i f is the length of the original pressure record. s f is the sampling frequency. r The shaft rotation frequency; The number of one-revolution pressure samples Q is determined according to the following formula. i And define the qth first-rotation pressure sample x i,q : Q i = [L i / N r ] x i,q = [p i ((q-1)N r +1), ..., p i (qN r )] T , q=1,2,…, Q i Each pressure sample is resampled to a uniform length K, and a deterministic phase coordinate is established as follows: θ k = 2π(k-1) / (K-1),k=1,2,…,K The standardization is based on the global mean of the training set. and the global standard deviation of the training set The standardized expression is: Where ε is a positive number to prevent division by zero; the and It is calculated only from the training set to avoid leakage of validation and test set information.

3. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping as described in claim 1, characterized in that, The continuous wavelet transform employs wavelet functions with time-frequency localization capabilities, including complex Morlet wavelets, and constructs the frequency-phase diagram and the original wavelet energy diagram using the following formula: in, This is a frequency-phase diagram. For wavelet energy maps, For frequency, For phase.

4. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 1, characterized in that, The diagnostic model employs a convolutional neural network model, comprising several convolutional blocks, a global average pooling layer, a Dropout layer, and a Softmax classification layer. Each convolutional block includes a two-dimensional convolutional layer, a batch normalization layer, a nonlinear activation layer, and a pooling layer. The global average pooling layer is used to obtain a feature vector that can characterize global order-phase features.

5. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 1, characterized in that, The frequency-phase map is divided into multiple image blocks according to the frequency or order direction and the phase direction, specifically including: Divide the frequency-phase diagram into N f ×N θ N non-overlapping image patches, where N f N is the number of divisions in the direction of frequency or order. θ The number of image patches is M = N, representing the number of divisions in the phase direction. f N θ For each image patch, establish a binary mask vector z=[ , ,…, ] T , =1 indicates that the m-th image patch is retained. =0 indicates that the m-th image patch is occluded; the occluded image patch is not zero-filled, but is replaced by a region at the same position in the average frequency-phase map of the training set, in order to reduce the impact of non-physical perturbations on the attribution results.

6. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 5, characterized in that, The block-level perturbation attribution employs the Kernel SHAP method. Based on the block-level perturbation attribution, the contribution value of each image block to the target category is calculated. Specifically, for target category c, the perturbation input under different mask combinations is fed into the diagnostic model to obtain the model output. The Shapley contribution value of the m-th image block is estimated through a weighted linear substitution model, where the weighted linear substitution model is calculated as follows: in, Let m be the contribution value of the m-th image patch to the target category output. This is the baseline output of the model for category c.

7. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 1, characterized in that, The physical evidence unit is represented as follows: in, and For the corresponding axis order range, and To correspond to the pumping order range, and To correspond to the phase range of one revolution, This represents the contribution value of the corresponding image patch to category c.

8. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 1, characterized in that, The physical evidence units are aggregated and statistically analyzed according to the low-order modulation region, the plunger passing through the fundamental frequency region, the second pumping harmonic region, the third pumping harmonic region, the adjacent order modulation region, and the high-order broadband response region.

9. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 1, characterized in that, The following formulas are used to calculate the health baseline energy ratio and positive contribution percentage relative to the health state H for category c and the m-th image patch, respectively: in, Let CWT be the average CWT energy of category c on image patch m. The average CWT energy of the healthy state on image block m. ε is a positive number to prevent division by zero. For the first The average contribution value of each image patch, where M is the total number of image patches.

10. The axial piston pump interpretable fault diagnosis method based on outlet pressure pulsation order-phase evidence mapping according to claim 9, characterized in that, The joint determination specifically includes: when and ρ m (c) When > 1 + δ, it is considered evidence supporting energy enhancement; when and ρ m (c) When < 1 - δ, it is considered evidence supporting energy decay; when and 1 - δ ≤ ρ m (c) When ≤ 1 + δ, it is considered evidence supporting energy redistribution; when and |ρ m (c) When -1| > δ, it is determined to be a non-dominant energy change region, where The contribution threshold is δ, and the energy tolerance is δ.

Images