A double-constraint incremental axial plunger pump fault diagnosis method based on sample selection playback
By improving the sample selection replay and the dual-constraint incremental learning method with dual-constraint loss function, the problem that the plunger pump fault diagnosis model cannot take into account both new and old fault modes in the online environment is solved, and high accuracy and stable fault diagnosis performance are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-03-19
- Publication Date
- 2026-07-10
Smart Images
Figure CN122365040A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fluid pressure and electrical digital data processing, and specifically relates to a fault diagnosis method for dual-constraint incremental axial piston pumps based on sample selection playback. Technical Background The piston pump has a complex internal structure and diverse failure modes, making it difficult to pre-build a complete dataset covering all potential failure modes throughout its entire lifespan. When a piston pump exhibits new failure modes during actual service, a fault diagnosis model with fixed categories will fail to generalize effectively.
[0002] In theory, the model can be retrained by collecting sufficient new fault data and combining it with all historical data, but this incurs huge storage and computational overhead and a long training cycle. If only new fault data is used to fine-tune the model, it will lead to a catastrophic forgetting of previously learned fault knowledge. This contradiction between the stability of old fault diagnosis capabilities and the plasticity of learning new faults limits the continuous learning ability of the plunger pump online fault diagnosis model. Incremental learning-like approaches, as a technical framework that can accommodate both old and new task modes, provide a solution to improve the continuous learning ability of fault diagnosis models.
[0003] Existing incremental learning algorithms can be divided into four main categories: network architecture-based methods, sample replay-based methods, parameter regularization-based methods, and mapping constraint-based methods. Among these, sample replay-based methods rely on the data itself to mitigate the forgetting of old knowledge. These methods typically preserve a portion of real samples from old tasks or generate representative pseudo-samples, which are then combined with new data for model training. This is the most effective and widely used incremental learning method; however, its performance is limited by the scale and representativeness of the stored old samples. Mapping constraint-based methods, on the other hand, design specific optimization objectives to align the output logic distribution between the new and old models. With almost no reliance on old samples, mapping constraint-based methods exhibit superior performance on old tasks compared to parameter regularization-based methods. However, while these methods alleviate catastrophic forgetting to some extent, they largely ignore the intermediate features encoding key discriminative knowledge.
[0004] Meanwhile, existing methods typically set fixed weights for various optimization objectives, making it impossible to adaptively balance the importance of new and old knowledge during training, thus making it difficult for the model to achieve optimal performance on old tasks.
[0005] Therefore, there is an urgent need to propose a new method for fault diagnosis of axial piston pumps. Summary of the Invention
[0006] To address the existing technical problems, this invention proposes a fault diagnosis method for dual-constraint incremental axial piston pumps based on sample selection and playback. The specific technical solution is as follows: S1. Collect and label multiple fault signals of the axial piston pump at the current moment to obtain fault samples containing fault types; and construct a fault diagnosis model containing a feature extractor and a classifier. S2, merge the faulty samples with the old faulty samples into the current time-incremental dataset; The old fault samples are specifically: Based on the previous incremental dataset and combined with the fault diagnosis model at the current time, the improved sample profile coefficient of each fault sample in the previous incremental dataset is calculated. Only the fault samples with improved sample profile coefficients higher than the threshold in each fault type of the previous incremental dataset are retained as old fault samples. S3, input the current incremental dataset into the fault diagnosis model at the current time and the fault diagnosis model trained at the previous time, respectively, and calculate the total incremental loss function based on the output results; S4. Update all parameters in the fault diagnosis model at the current time according to the total incremental loss function until the total incremental loss function converges, and obtain the fault diagnosis model trained at the current time. S5. Input the newly acquired fault signal into the fault diagnosis model that has been trained at the current moment to obtain the corresponding fault type.
[0007] Furthermore, in S2, the improved profile coefficient of the fault sample is specifically as follows: ; ; ; in, Indicates the first j Class, No. k Each fault sample and its corresponding proxy vector distance, Indicates the first j Class, number k Individual samples and non-similar agent vectors Minimum distance, Indicates the first j Class, No. k Improved sample profile coefficients for each fault sample. , This represents the Euclidean norm, also known as the L2 norm.
[0008] Furthermore, the calculation formula for the proxy vector of the specified class is as follows: ; in, Indicates the current time. j The proxy vector of the fault sample. Indicates the first j The total number of fault samples in a class of fault samples. Indicates the first j Class 1 k One fault sample, This indicates that the fault diagnosis model at the current moment is in the th... j Class, No. k The intermediate feature vector on each fault sample This represents the feature extractor of the fault diagnosis model at the current moment. This represents the total number of fault categories at the current moment. The calculation method for the non-class proxy vectors is the same as that for the class-specific proxy vectors.
[0009] Furthermore, in S3, the total incremental loss function consists of cross-entropy loss, cosine contrast loss, and surrogate loss, specifically: ; in, The total incremental loss function, , , They represent the first i The cross-entropy loss, cosine contrast loss, and surrogate loss calculated in the round loop are: , , They represent the first time. i Normalized weights for cross-entropy loss, cosine contrast loss, and surrogate loss in round loops.
[0010] Furthermore, the normalized weights for cross-entropy loss, cosine contrast loss, and surrogate loss are specifically as follows: ; ; ; ; in, , , These represent the cross-entropy loss, cosine contrast loss, and surrogate loss in the initial stage, respectively. , , They represent the first i Cross-entropy loss, cosine contrast loss, and surrogate loss in round loops. , , They represent the first time. i The weights of cross-entropy loss, cosine contrast loss, and surrogate loss in the round-robin iteration. This represents the weight decay coefficient.
[0011] Furthermore, the cross-entropy loss is specifically as follows: ; in, This represents the number of fault categories from the initial time to the current time. This indicates the fault diagnosis model at the current moment for the first... j Class, No. k The output of the fault sample m One logits value, Represents an exponential function. Indicates the first t Stage fault diagnosis model predicts the first j Class, No. k The sample is the first i The logits value of the failure mode; a This represents the fault category label predicted by the fault diagnosis model. This indicates the true fault category label of the fault sample. Used to implement the mapping from real labels to one-hot vectors, when The value is 1 when the condition is met, and 0 otherwise. This indicates the number of fault samples in the current incremental dataset. This indicates the fault diagnosis model's prediction of the current moment. j Class, No. k The fault category to which each fault sample belongs. Indicates the first j Class, No. k One fault sample; The cosine contrast loss is specifically as follows: ; in, This represents the intermediate layer feature vector obtained by the fault diagnosis model at the current moment when the fault sample is input. This represents the intermediate layer feature vector obtained from the fault diagnosis model trained at the previous time step, which is input as a fault sample. Indicates cosine contrast loss. Represents the L2 norm; The agency loss is specifically as follows: ; in, Indicates agency loss, This indicates that the old fault is currently acting as a proxy for losses. This indicates the negative agent loss due to the old fault. This indicates the negative proxy loss due to the new fault.
[0012] Furthermore, the old fault positive proxy loss, the old fault negative proxy loss, and the new fault negative proxy loss are specifically as follows: ; ; ; in, Indicates the scaling factor. This represents a boundary parameter used to control the degree of intra-class similarity and inter-class separation. Indicates the current time. j The number of fault samples of the new fault type. This indicates the storage limit for each type of old fault sample. U This indicates the total storage limit for old fault samples. This represents the total number of fault types corresponding to the old fault samples at the current moment. This represents the dataset of old fault samples at the current moment. This represents the dataset of fault samples collected at the current moment.
[0013] Furthermore, the feature is that the initial incremental dataset is the fault sample collected at the initial time; the fault diagnosis model at the initial time is trained based on the cross-entropy loss function, specifically: ; in, This is the fault diagnosis model at the initial moment.
[0014] Furthermore, in S3, the fault diagnosis model at the current moment is constructed by adding the output node of the new fault category at the current moment to the output layer of its classifier, based on the model trained at the previous moment.
[0015] The beneficial effects of this invention are: (1) The present invention adopts a sample selection and playback dual-constraint incremental fault diagnosis algorithm, which overcomes the catastrophic forgetting problem of learned fault knowledge in the fault diagnosis model in the prior art and improves the continuous learning ability of the plunger pump online fault diagnosis model.
[0016] (2) The present invention adopts an improved sample profile coefficient and a dual-constraint loss function. Based on the improved sample profile coefficient, representative old fault data are selected, which overcomes the problem of large performance fluctuation of old faults in the prior art, thereby improving the accuracy of fault diagnosis.
[0017] (3) The present invention adopts a new loss weight, which includes cosine contrast loss, proxy loss and adaptive weight, overcoming the problems of inter-class separation and intra-class similarity in the prior art, thereby improving the performance of model fault diagnosis. Attached Figure Description
[0018] Figure 1 This is an overall framework diagram of the method of the present invention.
[0019] Figure 2 This is a schematic diagram of filtering old fault samples based on improved sample profile coefficients.
[0020] Figure 3 A detailed training process diagram for each stage of incremental learning.
[0021] Figure 4 The loss function curve and adaptive weight curve are shown when the storage limit is 160.
[0022] Figure 5 Improve incremental fault diagnosis performance for different storage caps.
[0023] Figure 6 This is the confusion matrix for incremental fault diagnosis in stage 4 when the storage limit is 200.
[0024] Figure 7 This is the confusion matrix for incremental fault diagnosis in stage 4 when the storage limit is 160.
[0025] Figure 8 This is the confusion matrix for incremental fault diagnosis in stage 4 when the storage limit is 120.
[0026] Figure 9 Incremental fault diagnosis performance of different methods when the storage limit is 160.
[0027] Figure 10 The results of the t-SNE comparison experiment in stage 4 with a storage limit of 160 are visualized.
[0028] Figure 11 The results of the ablation experiment are shown when the storage limit is 160.
[0029] Figure 12 The results of t-SNE visualization for the fourth stage ablation experiment with a storage limit of 160. Detailed Implementation
[0030] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.
[0031] Axial piston pumps mainly consist of key components such as the main shaft, cylinder block, distributor plate, return plate, and slipper-piston assembly. During actual service, they are prone to various failure modes. A fault diagnosis model for piston pumps is used to automatically classify faults based on collected fault data. However, the initial dataset of the fault diagnosis model typically contains only a limited number of fault modes. As the piston pump continues to operate, new fault modes not covered in the initial dataset gradually emerge. To enable the fault diagnosis model to continuously identify these new fault modes, this invention employs incremental learning to continuously train the fault diagnosis model.
[0032] The overall process of this invention is as follows: Figure 1 As shown, during the training phase, an eddy current displacement sensor is first used to collect fault displacement signal data of the cylinder block as fault samples, i.e., fault modes; each fault mode is labeled with a fault category to obtain the real label corresponding to the fault sample.
[0033] The entire training process is divided into ( +1) stages, where stage 0 represents the basic training model, i.e., the initial fault diagnosis model, and stage 1 to... Each stage performs incremental training on new faults that occur in each stage.
[0034] With the first t Taking the first stage as an example, the second stage... t The fault diagnosis model for the stage is denoted as The model consists of a feature extractor. With classifier In the incremental process, the dimension of the intermediate layer feature vector extracted by the fault diagnosis model remains unchanged, while the output dimension of the classifier increases accordingly with the increase of the number of fault modes.
[0035] Considering storage resource constraints and data privacy concerns, old fault samples are often not fully preserved. Therefore, this invention proposes a fault sample replay method based on improved sample profile coefficients, which reduces storage overhead while preserving the distribution characteristics of old fault samples to the greatest extent possible.
[0036] Sample silhouette coefficient By evaluating the intra-class similarity and inter-class discriminative power of each sample, high-quality old fault samples that are located at the cluster centers and best represent the distribution characteristics of that class are selected. In the above formula, a ( i ) indicates the first f The average distance from one sample to all samples of the same type a ( f The smaller the value, the better. f The more similar a sample is to other samples of the same type, the better. b ( f ) indicates the first f The average distance from a sample to all the nearest samples outside the same class. b ( f The larger the value, the more likely it is to be the first. f The more distinct a sample is from samples of different classes, the better.
[0037] Profile coefficient s ( f The closer the value is to 1, the higher the clustering quality of the sample and the stronger its representativeness.
[0038] However, since the formula for calculating the silhouette coefficient requires calculating the distance between all pairs of samples, the computational complexity increases quadratically with the number of samples as the training phase progresses. In order to reduce the computational complexity as much as possible, this invention calculates the average feature vector of each old fault mode in the previous stage as the surrogate vector of that class, and converts the sample silhouette coefficient into the distance between the evaluation sample and the surrogate vector to filter old fault samples.
[0039] No. t During the phase, the first j The formula for calculating the proxy vector of a fault sample class (including old fault samples and newly input fault samples) is as follows: in, Indicates the first j Class of fault samples in the first t The proxy vector of the stage, Indicates the first j The total number of fault samples in the fault sample class. Indicates the first j Class 1 k One fault sample, Indicates the first t The fault diagnosis model in the stage is in the first stage. j Class, No. k The intermediate feature vector on each fault sample Indicates the first t The number of failure modes in a stage.
[0040] like Figure 2 As shown, based on the calculation principle of the profile coefficient, the first... k The distance between the faulty sample and its class's proxy vector is used as a metric for intra-class similarity. k The minimum distance between a faulty sample and a non-class proxy vector is used as a metric for inter-class discriminability. The improved sample silhouette coefficient design is as follows: in, Indicates the first j Class, No. k Each fault sample and its corresponding proxy vector distance, Indicates the first j Class, number k Individual samples and non-similar agent vectors Minimum distance, Indicates the first j Class, No. k Improved sample profile coefficients for each fault sample. , This represents the Euclidean norm, also known as the L2 norm.
[0041] From the above formula, it can be seen that when the sample silhouette coefficient is improved... When it is close to 1, that is, the intra-class distance Extremely small, and the inter-class distance Maximum value indicates faulty sample Samples that are close to the center of their class and far from the centers of other classes are representative of that class.
[0042] Based on the improved sample contour coefficient mentioned above, the first t The incremental dataset for each stage is constructed as follows: (1) Considering class balance, an average allocation strategy is adopted, with the storage limit for each class of old fault samples being [value missing]. ,in U This indicates the total storage limit for old fault samples. Indicates the first t The number of old failure modes in a phase, i.e. t - The number of all failure modes that occur in stage 1; (2) t -1 Stage Real-Class Incremental Dataset Calculate the improved sample profile coefficient for all faulty samples. ; (3) According to Sort the fault samples of the same type in descending order and retain the top ones in each category. The fault sample is used as a representative old fault sample to construct the first fault sample in the formula. j Subset of Fault Samples : in, This indicates the number of elements in the subset.
[0043] (4) Find the union of the fault sample subsets in the formula to obtain the previous set. t -1 set of screened fault samples from stage 1 Together, they constitute the real class increment dataset used for class increment training. : (5) front t - Old fault samples selected in one stage Constituting the first t Real historical datasets of the stage and with the t Newly collected datasets in this phase Together, they constitute the real class increment dataset used for class increment training. In stage 0, it is not necessary to filter old faults, that is... , It should be noted that in step (2) above, the first... t -1 Stage Real-Class Incremental Dataset Similarly, it follows the above process, in the first... t -1 stage truly completes the derivation and construction.
[0044] In the t In the phased incremental learning, the fault diagnosis model It is necessary to learn new failure modes while retaining the ability to diagnose old failure modes.
[0045] In obtaining the first t Phase-based incremental dataset Then, based on the output of the fault diagnosis model, the cross-entropy loss function is calculated. : in, Indicates from stage 0 to stage 1 The total number of fault categories that occurred during the phase. Indicates the first t The fault diagnosis model for the first stage j Class, No.k The first fault sample m One logits value, Represents an exponential function. Indicates the first t Stage fault diagnosis model predicts the first j Class, No. k The sample is the first i The logits value of the failure mode; a The predicted fault category label for the fault sample. This indicates the true fault category label of the fault sample. Used to implement the mapping from real labels to one-hot vectors, when When the model is valid, that is, when the fault category predicted by the model is consistent with the actual labeling of the sample, the value is 1; otherwise, it is 0. express The total number of fault samples included.
[0046] The first summation symbol in the above formula indicates that the summation is performed on the summation of ... The summation is performed on each sample, and the second summation symbol indicates that the output logical value of each sample is summed.
[0047] If relying solely on a small number of old fault samples obtained through screening, the fault diagnosis model will inevitably be affected by the learned old fault features when learning new fault features, leading to catastrophic forgetting. Therefore, while optimization alone can enable the fault diagnosis model to maintain good diagnostic performance on new fault modes, it cannot guarantee that it will have comparable diagnostic performance on old fault modes. To alleviate this phenomenon, this invention introduces cosine contrast loss and surrogate loss into the loss function to constrain the diagnostic performance of old faults.
[0048] Specifically, such as Figure 3 As shown, this invention follows the principle of knowledge distillation, converting the first... t -1 stage training completed the fault diagnosis model As a teacher model, and through the teacher model classifier Add a new output node corresponding to the new fault category to the output layer and initialize the fault diagnosis model. As a student model.
[0049] To preserve the fault diagnosis model's ability to extract features from old faults, a cosine contrast loss was designed to constrain the intermediate layer features of the old and new models. As shown in the following formula in, express Input of fault samples The obtained intermediate layer feature vectors, express Input of fault samples The obtained intermediate layer feature vectors.
[0050] To avoid the influence of new fault features on old fault features, a surrogate loss is designed to constrain the relationship between intermediate layer feature vectors and surrogate vectors. Including losses from old faults and positive agent losses Old fault negative agency loss Negative agent loss due to new faults : in, This represents the scaling factor, used to control the constraint strength of the proxy vector. This represents a boundary parameter used to control the degree of intra-class similarity and inter-class separation. This represents the number of samples of the j-th type of new fault in stage t. By increasing the similarity between old fault features and the proxy vector of their respective class, intra-class consistency of old fault patterns is ensured. By reducing the similarity between old fault features and vectors of dissimilar proxies, the inter-class separability of old fault patterns is enhanced. - By reducing the similarity between new fault features and old fault proxy vectors, the influence of new fault features on old fault features can be avoided.
[0051] During training, the importance of new and old fault features dynamically changes with each training epoch. To adaptively balance the importance of new and old fault features during training, a cross-entropy loss mechanism is used. Cosine contrast loss With agency losses An adaptive weighting system was designed, which includes a loss function normalization term and a weight decay term: in, , , These represent the initial cross-entropy loss, cosine contrast loss, and surrogate loss, respectively. , , They represent the firsti Cross-entropy loss, cosine contrast loss, and surrogate loss in round training. , , They represent the first time. i The weights of cross-entropy loss, cosine contrast loss, and surrogate loss in each round of training. , , These represent the weights after normalization. This represents the weight decay coefficient, which controls the relative contribution of the normalization term and the weight decay term in the loss function.
[0052] Therefore, the total incremental loss function As shown in the following formula The incremental fault diagnosis method proposed in this invention includes initial training in phase 0 and phase 1-phase 2. Phase-based incremental training.
[0053] In stage 0, the fault diagnosis model Train using the cross-entropy loss function until the loss function converges: In Phase 1 - In the first phase, a real-class incremental dataset is constructed based on the improved sample silhouette coefficients. Then, calculate the cross-entropy loss, cosine contrast loss, and surrogate loss. Finally, combine the adaptive weights and train the fault diagnosis model according to the following formula. : When the St stage loss function Upon convergence, the fault diagnosis model is considered to have completed training. Based on this, if it is necessary to add fault categories, the displacement data corresponding to the new fault category is input into the trained fault diagnosis model as the... St +1 phase, repeat the above training steps.
[0054] During the inference phase, the detected displacement signal is input into the trained fault diagnosis model to obtain the corresponding fault category.
[0055] To verify the incremental fault diagnosis method proposed in this invention, this experiment collected a total of 13 fault modes and simulated the fault evolution trend from early weak faults to later severe faults to sudden faults. In stage 0, the normal state and the distribution disk wear state 1, plunger wear state 1, and slipper wear state 1 with the smallest average relative wear were injected to train the initial model. In stages 1 and 2, wear states with progressively increasing average relative wear were injected to simulate the early weak evolution. In stages 3 and 4, return disk fracture, ball joint jamming, spindle spline peeling, and distribution disk burn were injected to simulate sudden faults.
[0056] In Phase 1 and Phase 4, faults that existed before the current phase's fault injection are considered old fault modes, while faults injected in the current phase are considered new fault modes, as shown in Table 1.
[0057] Table 1 Storage limit U This determines the number of old fault samples that participate in incremental training. This experiment sets up a total of... U =200、 U =160 and U The storage limit is 120. The initial model structure is shown in Table 2. The feature extractor consists of 9 convolutional layers, and the output intermediate layer feature vector dimension is... D The number of neurons is 256; the classifier consists of one fully connected layer, and the number of output neurons in stage 0 is 4. In stages 0-4, the structure of the feature extractor remains unchanged, and the number of output neurons of the classifier increases sequentially according to the number of fault modes defined in Table 2.
[0058] Table 2 In the initial training phase 0, the optimizer is Adam, the batch size is 16, the loss function is cross-entropy, the maximum number of training epochs is 100, and the initial learning rate is 4 × 10⁻⁶. -4 In the incremental learning stages 1-4, the feature extractor parameters for the current stage are initialized using the feature extractor parameters from the previous stage. The classifier for the current stage is initialized using a random initialization method. The optimizer is Adam, the batch size is 16, and the loss function includes cross-entropy loss, cosine contrast loss, and surrogate loss. The maximum number of training epochs is 200, and the initial learning rate is 4×10⁻⁶. -4 Based on experience, the scaling factor for agent loss... Set to 8, and boundary parameters to 0.1. Since incremental learning retains some old fault samples, the initial weights [ , , Setting the value to [1,0,0] allows the model to first learn the cross-entropy generated by the new and old fault samples and the true labels, quickly establishing the ability to distinguish between new and old fault modes; the attenuation coefficient... t Set to 10 -4 The weights of the cosine contrast loss and the surrogate loss are adaptively adjusted to gradually enhance the retention of performance for old faults. To eliminate random errors in model initialization and optimization direction, all training processes are repeated 5 times and trained until convergence. The average test accuracy is used as the evaluation metric for fault diagnosis performance.
[0059] like Figure 4 The loss function curve and adaptive weight curve of the proposed method are shown. From the figure, it can be observed that the total loss in stage 0 is the cross-entropy loss. L entropy The number of training rounds decreased rapidly in the first 25 rounds; the total loss for stages 1-4 was... L cL Cross-entropy loss L entropy Cosine contrast loss L cos and agency losses L proxy With adaptive weights , , Obtained by weighted summation. In the first 20 training epochs, L entropy exist L CL China holds a dominant position. L CL and L entropy Synchronous rapid decline, L proxy Slowly descending, and L cos The trend is upward. This is because the feature extractor parameters in the current stage are initialized by the feature extractor parameters in the previous stage. In the first training epoch, the intermediate feature vector of the current stage... Intermediate feature vectors from the previous stage Almost identical, L cos Relatively small; in training rounds 1-20, due to the presence of new fault samples and the cosine contrast loss weights The difference is relatively small; the current stage of feature extractor parameter distribution shifts, leading to... L cos Increase. In training rounds 21-200, L cos and L proxy Starting with total lossesL cL China holds a dominant position. L entropy , L cos and L proxy All showed a downward trend. L cL Approaching stability.
[0060] Further analysis Figure 4 From (b) to (e), it can be seen that during the entire training process, due to L entropy The largest decrease was observed in the cross-entropy loss weights. It shows a continuous downward trend, dropping to around 0.7 by the 200th training round; cosine contrast loss weights Weighted with agency loss An upward trend, and due to L cos A trend of first rising and then falling. Always slightly larger Both values rose to the 0.1-0.2 range by the 200th training round. To avoid... L cos Excessive descent leads to overfitting of new fault modes, so training is stopped after the 200th training epoch.
[0061] Figure 5 The results of incremental fault diagnosis for different storage caps are shown.
[0062] Depend on Figure 5 From (a), we can see that in U =200、 U =160 and U When the storage limit is 120, the accuracy rate of new fault diagnosis in stages 0-4 is above 99%, indicating that the storage limit has little impact on the accuracy rate of new fault diagnosis. The accuracy rate of new fault diagnosis in stage 1 is relatively low. U =200、 U =160 and UWhen the accuracy is 120, the rates are 99.13±1.13%, 99.47±0.91%, and 99.53±0.58%, respectively. This is likely because in the first stage, the number of old fault modes and the number of new fault modes are 4 and 3, respectively, and the number of training samples for old faults is the same as that for new faults. Therefore, the old and new fault categories are relatively balanced, and the model needs to directly weigh between the two during training, resulting in a relatively low accuracy rate for new fault diagnosis in the first stage. In stages 2-4, as the number of old fault modes increases while the total number of old fault training samples remains unchanged, the old and new fault categories gradually become unbalanced, and the accuracy rate for new fault diagnosis rebounds, reaching 100% in most cases.
[0063] Depend on Figure 5 As shown in (b), the accuracy of old fault diagnosis shows a continuous downward trend as the incremental stage progresses. In stage 1, the accuracy of old fault diagnosis for each storage limit remains at around 99%, a decrease of about 1% compared to the fault diagnosis accuracy in stage 0; in stages 2-4, the decrease in the accuracy of old fault diagnosis becomes relatively larger. U When the value was 200, the percentages decreased by 0.42%, 3.37%, and 1.44% respectively in each stage. U When the value was 160, the percentages decreased by 2.39%, 4.42%, and 2.67% respectively in each stage. U When the value was 160, the percentages decreased by 2.17%, 4.18%, and 7.72% respectively in each stage. Among them, U = The overall accuracy of fault diagnosis is highest when I is 200; U =160 and U The old fault diagnosis accuracy at =120 was basically the same in stages 2 and 3, but in stage 4, the former was 4.99% higher than the latter. After incremental learning, the average fault diagnosis accuracy for all fault modes was 94.83±2.46%. U =200), 91.23±2.15% ( U =160) and 86.96±1.69% ( U =120). In general, the larger the storage limit, the more old fault samples participate in incremental learning, and the better the old fault diagnosis performance.
[0064] Figures 6-8The following are the confusion matrices for different storage limits in stage 4. It can be observed that in incremental learning, although the performance of old fault diagnosis generally degrades, the degree of degradation varies among different old fault modes, and fault modes learned earlier do not necessarily exhibit more severe performance degradation. Overall, the degradation levels of plunger wear state 1, slipper wear state 1, and slipper wear state 2 are relatively low, maintaining a class accuracy of over 90% across different storage limits. In contrast, distributor plate wear state 2 and ball joint jamming tend to be misdiagnosed as normal states; plunger wear state 2 tends to be misdiagnosed as plunger wear state 3; and distributor plate wear state 3 tends to be misdiagnosed as distributor plate wear state 1.
[0065] This invention further compares the proposed method with mainstream incremental learning methods LwF and iCaRL. LwF is a knowledge distillation-based incremental learning method that mitigates catastrophic forgetting by aligning the output logistic distribution between the new and old models. iCaRL is a nearest-neighbor replay-based incremental learning method that combines a prototype model with knowledge distillation to preserve the performance of older tasks. Both LwF and iCaRL use Adam as the optimizer, with a batch size of 16, a maximum training epochs of 100 per stage, and an initial learning rate of 4 × 10⁻⁶. -4 To ensure fairness in the comparison, storage limits are used. U The performance of the proposed method was compared at a coefficient of 160. The proposed method uses an improved sample silhouette coefficient to filter old fault samples, iCaRL uses the nearest neighbor criterion to filter old fault samples, and LwF retains old fault samples by random sampling. All training processes were repeated 5 times and trained until convergence. The average test accuracy was used as the evaluation index of fault diagnosis performance.
[0066] Performance of incremental fault diagnosis methods as follows: Figure 9 As shown.
[0067] Depend on Figure 9 As shown in (a), in stages 0-4, the accuracy of the proposed method, random sampling + LwF, and iCaRL in diagnosing new faults all showed a trend of first decreasing and then increasing: in stage 1, the accuracy of the proposed method and random sampling + LwF in diagnosing new faults was the lowest, at 99.47±0.91% and 98.45±2.89%, respectively; the accuracy of iCaRL in diagnosing new faults was the lowest in stage 2, at 99.81±0.22%; in stages 3 and 4, the accuracy of all methods in diagnosing new faults reached 100%.
[0068] Depend on Figure 9As shown in (b), the accuracy of old fault diagnosis for the proposed method, random sampling + LwF, and iCaRL decreased in stages 0-4. In stage 1, the old fault diagnosis performance of all methods was similar, with the accuracy remaining above 98%. However, in stage 4, the old fault diagnosis accuracy of the proposed method, random sampling + LwF, and iCaRL were 89.77±2.51%, 80.76±6.28%, and 84.98±4.95%, respectively. Among them, random sampling + LwF showed relatively poor old fault diagnosis performance and relatively large performance fluctuations. On the one hand, random sampling resulted in inconsistent old fault sample quality, which could not fully represent the original data distribution, leading to large fluctuations in old fault diagnosis performance and poor overall performance. On the other hand, LwF's constraint strength based solely on KL divergence for knowledge distillation was relatively limited, making it difficult to fully maintain the model's diagnostic ability for old faults. Although iCaRL improves the performance of old fault diagnosis and stabilizes performance fluctuations to some extent by screening old fault samples through nearest neighbor playback, the sample screening only considers intra-class relationships and ignores inter-class relationships, so its ability to maintain the performance of old fault diagnosis in the long term is still not ideal.
[0069] Figure 10 The t-SNE visualization results of the intermediate layer feature vectors and output are presented. It can be observed that although both the proposed method and iCaRL exhibit some confusion in distinguishing between the normal state and the distribution disk wear state 2, the proposed method and iCaRL show better inter-class discrimination and intra-class similarity than random sampling + LwF in terms of the separability of the intermediate feature vectors and output results. This indicates that filtering and retaining old fault samples that can characterize the original data distribution can effectively improve the performance of old fault diagnosis.
[0070] This invention revolves around cosine contrast loss L cos Agency losses L proxy With adaptive weights, further ablation experiments were conducted. All training processes were repeated 5 times until convergence, and the average test accuracy was used as the evaluation metric for fault diagnosis performance.
[0071] Ablation test results as follows Figure 11 As shown.
[0072] Depend on Figure 11As shown in (a), in stages 0-4, the accuracy of new fault diagnosis using different ablation methods is above 99%, and all show a trend of decreasing in stage 1 and increasing in stages 2-4. In stage 1, the difference in accuracy of new fault diagnosis among different ablation methods is greatest: compared to the result using only cross-entropy loss, introducing cosine contrast loss decreases the accuracy of new fault diagnosis by 0.07%; further, introducing surrogate loss further decreases the accuracy by 0.53%; while introducing adaptive weights increases the accuracy by 0.27%. This is because the old and new fault categories are relatively balanced in stage 1, and cosine contrast loss and surrogate loss strengthen the model's retention of old fault diagnosis performance, while adaptive weights help the model dynamically balance the performance of new and old fault diagnosis.
[0073] Depend on Figure 11 As shown in (b), cosine contrast loss, surrogate loss, and adaptive weights all improve the performance of the old fault diagnosis. In the first - In stage 4, compared to the results using only cross-entropy loss, cosine contrast loss improved the accuracy of the old fault diagnosis by 2.50%, 5.56%, 5.98%, and 7.14%, respectively. Based on this, after introducing surrogate loss, in stage 2... - The performance of the old fault diagnosis in stages 4 further increased by 0.40%, 0.97%, and 1.62%, respectively, while the performance of the old fault diagnosis in stage 1 decreased by 0.55%, which may be a suboptimal result due to the fixed weights. Under the effect of adaptive weights, the performance of the old fault diagnosis in stage 1 rebounded by 0.65%, and the performance in stage 2... - The accuracy of old fault diagnosis in stage 4 was further improved by 0.29%, 0.41%, and 0.84%. Overall, the cosine contrast loss had the most significant effect on improving the performance of old fault diagnosis.
[0074] Visualization results of t-SNE in ablation experiments are as follows Figure 12 As shown. It can be seen that relying solely on L entropy The intermediate layer features are relatively dispersed, with significant overlap between categories. (Introduction) L cos Afterwards, intra-class similarity and inter-class separation significantly improved, but the inter-class separation between distribution disk wear state 1 and distribution disk wear state 3, as well as the inter-class separation between slipper wear state 2, ball joint jamming, and distribution disk burning, remained insufficient. (Introduction) L proxy Afterwards, the samples within the same class became more closely packed, but some confusion still existed between the aforementioned fault categories. By further introducing adaptive weights, although there was overlap between the normal state and the distribution plate wear state 2, the degree of confusion between similar fault categories was significantly reduced. Overall, L cos ,L proxy Both adaptive weights effectively improve intra-class similarity and inter-class separation in the feature distribution space.
[0075] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.
Claims
1. A fault diagnosis method for dual-constraint incremental axial piston pumps based on sample selection and playback, characterized in that, include: S1, collect and label multiple fault signals of the axial piston pump at the current moment to obtain a fault sample containing fault types; And construct a fault diagnosis model that includes a feature extractor and a classifier; S2, merge the faulty samples with the old faulty samples into the current time-incremental dataset; The old fault samples are specifically: Based on the previous incremental dataset and combined with the fault diagnosis model at the current time, the improved sample profile coefficient of each fault sample in the previous incremental dataset is calculated. Only the fault samples with improved sample profile coefficients higher than the threshold in each fault type of the previous incremental dataset are retained as old fault samples. S3, input the current incremental dataset into the fault diagnosis model at the current time and the fault diagnosis model trained at the previous time, respectively, and calculate the total incremental loss function based on the output results; S4. Update all parameters in the fault diagnosis model at the current time according to the total incremental loss function until the total incremental loss function converges, and obtain the fault diagnosis model trained at the current time. S5. Input the newly acquired fault signal into the fault diagnosis model that has been trained at the current moment to obtain the corresponding fault type.
2. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 1, characterized in that, In S2, the improved profile coefficient of the fault sample is specifically as follows: ； ； ； in, Indicates the first j Class, No. k Each fault sample and its corresponding proxy vector distance, Indicates the first j Class, number k Individual samples and non-similar agent vectors Minimum distance, Indicates the first j Class, No. k Improved sample profile coefficients for each fault sample. , This represents the Euclidean norm, also known as the L2 norm.
3. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 2, characterized in that, The formula for calculating the proxy vector of the specified class is as follows: ； in, Indicates the current time. j The proxy vector of the fault sample. Indicates the first j The total number of fault samples in a class of fault samples. Indicates the first j Class 1 k One fault sample, This indicates that the fault diagnosis model at the current moment is in the th... j Class, No. k The intermediate feature vector on each fault sample This represents the feature extractor of the fault diagnosis model at the current moment. This represents the total number of fault categories at the current moment. The calculation method for the non-class proxy vectors is the same as that for the class-specific proxy vectors.
4. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 1, characterized in that, In S3, the total incremental loss function consists of cross-entropy loss, cosine contrast loss, and surrogate loss; specifically: ； in, The total incremental loss function, , , They represent the first i The cross-entropy loss, cosine contrast loss, and surrogate loss calculated in the round loop are: , , They represent the first time. i Normalized weights for cross-entropy loss, cosine contrast loss, and surrogate loss in round loops.
5. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 4, characterized in that, The normalized weights for cross-entropy loss, cosine contrast loss, and surrogate loss are specifically: ； ； ； ； in, , , These represent the cross-entropy loss, cosine contrast loss, and surrogate loss in the initial stage, respectively. , , They represent the first i Cross-entropy loss, cosine contrast loss, and surrogate loss in round loops. , , They represent the first time. i The weights of cross-entropy loss, cosine contrast loss, and surrogate loss in the round-robin iteration. This represents the weight decay coefficient.
6. A fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 1 or 4, characterized in that, The cross-entropy loss is specifically as follows: ； in, This represents the number of fault categories from the initial time to the current time. This indicates the fault diagnosis model at the current moment for the first... j Class, No. k The output of the fault sample m One logits value, Represents an exponential function. Indicates the first t Stage fault diagnosis model predicts the first j Class, No. k The sample is the first i The logits value of the fault mode; a This represents the fault category label predicted by the fault diagnosis model. This indicates the true fault category label of the fault sample. Used to implement the mapping from real labels to one-hot vectors, when The value is 1 when the condition is met, and 0 otherwise. This indicates the number of fault samples in the current incremental dataset. This indicates the fault diagnosis model's prediction of the current moment. j Class, No. k The fault category to which each fault sample belongs. Indicates the first j Class, No. k One fault sample; The cosine contrast loss is specifically as follows: ； in, This represents the intermediate layer feature vector obtained by the fault diagnosis model at the current moment when the fault sample is input. This represents the intermediate layer feature vector obtained from the fault diagnosis model trained at the previous time step, which is input as a fault sample. Indicates cosine contrast loss. Represents the L2 norm; The agency loss is specifically as follows: ； in, Indicates agency loss, This indicates that the old fault is currently acting as a proxy for losses. This indicates the negative agent loss due to the old fault. This indicates the negative proxy loss due to the new fault.
7. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 6, characterized in that, The old fault positive proxy loss, the old fault negative proxy loss, and the new fault negative proxy loss are specifically as follows: ； ； ； in, Indicates the scaling factor. This represents a boundary parameter used to control the degree of intra-class similarity and inter-class separation. Indicates the current time. j The number of fault samples of the new fault type. This indicates the storage limit for each type of old fault sample. U This indicates the total storage limit for old fault samples. This represents the total number of fault types corresponding to the old fault samples at the current moment. This represents the dataset of old fault samples at the current moment. This represents the dataset of fault samples collected at the current moment.
8. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 1, characterized in that, The initial incremental dataset consists of fault samples collected at the initial time; the initial fault diagnosis model is trained using the cross-entropy loss function, specifically: ； in, This is the fault diagnosis model at the initial moment.
9. The fault diagnosis method for a dual-constraint incremental axial piston pump based on sample selection and playback according to claim 1, characterized in that, In S3, the fault diagnosis model at the current moment is constructed by adding the output node of the new fault category at the current moment to the output layer of its classifier, based on the model trained at the previous moment.