A bridge machine cross-machine fault diagnosis method

Abstract

The application provides a bridge machine cross-machine type fault diagnosis method, relates to the technical field of equipment fault diagnosis, and comprises four stages of source domain high discriminative diagnosis model construction, passive domain self-adaptive migration and high-quality pseudo label generation, incremental learning anti-forgetting and prototype comparison optimization, cross-machine type online diagnosis reasoning and closed-loop evolution. In view of the low reliability of traditional pseudo label generation in the case of no labeled data of a new machine type in a passive domain, a memory library neighborhood aggregation pseudo label generation mechanism is proposed, a local manifold graph of a target domain sample feature space is constructed, pseudo labels are generated by using neighborhood prediction probability weighted aggregation, and low-quality labels are filtered by combining an adaptive confidence threshold screening mechanism, so that the cross-machine type fault diagnosis precision is significantly improved, and incremental learning anti-forgetting is realized through prototype comparison optimization and optimal transport drift calibration.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a method for diagnosing faults across different bridge erecting machines. Background Technology

[0002] Failures in key components of bridge erecting machines, including drive motors, reducers, and bearings, can directly lead to construction interruptions and even major safety accidents. Therefore, accurate and efficient fault diagnosis is crucial for ensuring construction safety and project progress. However, different types of bridge erecting machines exhibit significant differences in structural parameters, load characteristics, and operating conditions, resulting in fundamentally different fault characteristics in key components. Traditional fault diagnosis models are only compatible with specific models, and their accuracy drops sharply after model switching due to domain shifts in data distribution. Furthermore, new models often lack sufficient fault labeling data in the initial stages of deployment, and historical operating and fault data from the source domain are frequently inaccessible due to corporate technical confidentiality and data storage isolation, creating a typical passive domain application scenario. In this scenario, traditional domain adaptation methods fail completely because they cannot simultaneously acquire data from both the source and target domains. The only solution is to generate pseudo-labels to utilize unlabeled data from the target domain for model transfer. The reliability of pseudo-label generation directly determines the accuracy and deployment efficiency of cross-model fault diagnosis models.

[0003] Existing pseudo-label generation methods for new bridge erecting machine models in passive domain scenarios have fundamental flaws, becoming a core bottleneck restricting the implementation of cross-model fault diagnosis technology. In passive domain scenarios, the source domain model's feature extraction of new model data exhibits bias, often resulting in predictions containing numerous low-confidence or erroneous category classifications. Directly introducing such unreliable predictions as supervisory signals (pseudo-labels) into the training process leads to an "error accumulation" effect, where the model continuously reinforces incorrect feature mappings under the guidance of erroneous labels, resulting in target domain diagnostic accuracy far below expectations. Traditional pseudo-label generation mechanisms rely solely on isolated predictions of individual samples after forward propagation, neglecting the local manifold structure of target domain sample data in the feature space. Even in unlabeled data, samples under similar operating conditions should have similar distribution characteristics in the feature space. However, existing methods ignore mechanisms for correcting and verifying prediction results by mining contextual information from neighboring samples, resulting in an inability to effectively filter out erroneous label interference from abnormally fluctuating data.

[0004] Therefore, there is an urgent need for a cross-model fault diagnosis technology that can be quickly adapted in passive domain scenarios using only unlabeled data from new models, in order to meet the actual needs of rapid deployment and intelligent operation and maintenance throughout the entire lifecycle of different models. Summary of the Invention

[0005] The main objective of this invention is to provide a cross-model fault diagnosis method for bridge erecting machines, which solves the problem of low reliability of traditional label generation methods when new models lack labeled data in passive domain scenarios. The invention uses a memory library neighborhood aggregation pseudo-label to improve the accuracy of generating high-quality pseudo-labels, thereby achieving cross-model fault diagnosis.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for diagnosing cross-model faults of bridge erecting machines, comprising the following steps:

[0007] S1. Based on the source domain bridge erecting machine with labeled data, collect the operating data of key components, construct a fault diagnosis model containing feature encoders and classifiers, and simultaneously establish a feature prototype library of source domain fault categories. S2. In the scenario of a passive domain where the original source domain data cannot be accessed, the unlabeled operation data of the bridge erecting machine in the target domain is input into the pre-trained model. Pseudo-labels are generated through neighborhood aggregation and filtered by adaptive threshold to complete the adaptive transfer of the model. S3. Incremental training of the model is completed through prototype constraints and drift calibration, so that the model can learn new fault types while maintaining its ability to identify existing faults. S4. Deploy the optimized model to the target domain bridge erecting machine to achieve online fault diagnosis, and complete model iteration based on on-site feedback data to form a closed-loop evolution.

[0008] Preferably, in step S1, the multi-source heterogeneous sensing data includes: The triaxial vibration acceleration signal was acquired by a triaxial MEMS accelerometer, the bearing and gearbox surface temperature signal was acquired by a PT100 temperature sensor, the three-phase current signal was acquired by a Hall-effect current sensor, and the line voltage signal was acquired by an isolation voltage sensor. Preprocessing includes wavelet thresholding for noise reduction, moving average filtering, and normalization.

[0009] Preferably, in step S1, the fault diagnosis model is trained using an adaptive label smoothing strategy, which includes: using the temporary model of the current iteration state to perform forward inference on the samples during the training process, obtaining the predicted probability distribution and calculating the information entropy, dynamically assigning a smoothing coefficient to each sample based on the information entropy, replacing the traditional one-hot label with an adaptive smoothing label, calculating the adaptive label smoothing cross-entropy loss and optimizing the model parameters.

[0010] Preferably, in step S2, the dynamic memory library adopts a first-in-first-out storage strategy. Each record stores a feature vector, a prediction probability vector, and an uncertainty score triplet. When the library is full, the sample with the highest uncertainty score is eliminated first. The neighborhood prediction probability weighted aggregation is specifically as follows: calculate the Gaussian kernel similarity between samples in the memory library to construct a local feature manifold graph, and perform weighted aggregation of the prediction probabilities of neighborhood samples based on the similarity to generate sample pseudo-labels.

[0011] Preferably, in step S2, the adaptive confidence threshold screening mechanism includes: defining the pseudo-label confidence of a sample as the maximum class probability value after neighborhood weighted aggregation, adopting an adaptive threshold based on statistical distribution, dynamically adjusting the threshold according to the mean and standard deviation of the confidence of the current batch of samples, and retaining only samples with confidence greater than or equal to the adaptive threshold and their pseudo-labels to form a high-quality target domain pseudo-label dataset.

[0012] Preferably, in step S2, during the model adaptive transfer process, all parameters of the classifier are frozen, and only the parameters of the feature encoder are fine-tuned and optimized.

[0013] Preferably, in step S3, the prototype exclusion constraint specifically means: when the distance between any two different categories of prototypes is less than a preset safety threshold, the prototype exclusion regularization term is activated to increase the prototype distance between categories in order to ensure the class distinguishability of the feature space.

[0014] Preferably, in step S3, the prototype drift calibration specifically involves: solving for the optimal mapping matrix by minimizing the Wasserstein distance between the old and new feature spaces, and mapping the old prototype to the updated feature space to complete the calibration; the slicing Wasserstein distance is used to approximate the solution, and the calibration operation is performed after each incremental training iteration.

[0015] Preferably, in step S3, the loss function used for incremental training includes classification loss for new fault categories, knowledge distillation loss for constraining feature distribution consistency, and hybrid contrastive learning loss anchored by prototypes.

[0016] Preferably, in step S4, the closed-loop evolution specifically involves: periodically evaluating the accuracy of the pseudo-label generation strategy based on real fault labels from on-site feedback, and automatically optimizing hyperparameters when the accuracy is lower than the preset target; during non-operation periods of the bridge erecting machine, incrementally fine-tuning the model using accumulated real label data and high-confidence pseudo-label data to complete model iteration. When adapting to a new bridge erecting machine model, repeat steps S2 to S4 without accessing the original annotation data of the source domain.

[0017] Beneficial effects: 1. Solves the core problem of cross-model diagnosis in passive domain scenarios. It can quickly adapt the model without accessing the original data of the source domain, and can achieve rapid model adaptation only through the unlabeled data of the target domain. It breaks through the strong dependence of traditional domain adaptation methods on source domain data and perfectly adapts to the actual engineering scenarios of enterprise technology confidentiality and cross-enterprise data isolation. 2. The proposed dynamic memory library neighborhood aggregation pseudo-label generation mechanism fully explores the feature manifold structure of unlabeled data in the target domain, corrects single-sample prediction bias through neighborhood weighted aggregation, and improves pseudo-label accuracy by combining adaptive confidence threshold screening. 3. A complete incremental learning anti-forgetting framework has been constructed. By constraining feature space drift through knowledge distillation, calibrating prototype offset through optimal transfer theory, and enhancing inter-class discrimination through hybrid contrastive learning, it can achieve rapid learning of new fault types and adapt to the full life cycle operation and maintenance needs of bridge erecting machines during long-term operation. 4. A closed-loop evolution mechanism driven by operation and maintenance feedback was designed, which can continuously optimize model parameters and pseudo-label generation strategies based on real-world fault data. The model diagnostic performance can be continuously improved with runtime. At the same time, it can be quickly adapted to more bridge erecting machine models such as the third and fourth types, and has strong generalization ability and engineering implementation value. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the overall technical process of this invention; Figure 2 This is a flowchart of the passive domain adaptive migration process of the present invention; Figure 3 This is the flowchart of the incremental learning anti-forgetting process of this invention; Figure 4 This is a flowchart of the cross-model fault diagnosis reasoning process of the present invention. Detailed Implementation

[0019] Example 1 like Figure 1 As shown, a method for diagnosing cross-model faults in bridge erecting machines specifically includes: S1. Construction of a source domain high-discrimination diagnostic model: Using a source domain bridge-building machine with a large amount of existing labeled data, a fault diagnosis model with strong discriminative ability and generalization potential is trained. S2, Passive Domain Adaptive Migration and High-Quality Pseudo-Label Generation: In passive domain scenarios where source domain data is inaccessible due to confidentiality or storage limitations, high-quality pseudo-labels are generated and adaptive migration of the model is achieved using only the unlabeled running data of the new target domain model. S3. Incremental learning anti-forgetting and prototype comparison optimization: When new types of faults continuously appear during the operation of the bridge erecting machine, the model maintains its ability to identify existing fault types without decreasing while learning new category knowledge. S4. Cross-model online diagnostic reasoning and closed-loop evolution: Deploy the optimized model obtained from the training above to the target model to achieve online real-time diagnosis, and form a closed-loop iteration through on-site feedback data.

[0020] According to an embodiment of the present invention, in step S1, a highway bridge erecting machine is selected as the source domain model, and a multi-source heterogeneous sensor array is deployed in key components such as the drive motor, reducer, lifting mechanism, and traveling mechanism, including: The vibration sensor array uses a triaxial MEMS accelerometer to acquire triaxial vibration acceleration signals. , and The vibration sensor should be installed no more than 10cm away from the fault source, and the sampling frequency should be set to more than 5 times the fault characteristic frequency. The PT100 temperature sensor collects the surface temperature of bearings and gearboxes. The temperature sensor uses a thermocouple for direct contact measurement; The current sensor uses the Hall effect replacement current sensor HIT100 to collect three-phase current. , and ; The voltage sensor uses the LEM MBT19 isolated voltage sensor to collect line voltage. , and The current and voltage sensors achieve non-contact measurement using the Hall effect principle, with an isolation voltage ≥2500V to ensure the accuracy and safety of data acquisition. The acquired raw data undergoes preprocessing, including wavelet threshold denoising, moving average filtering, and normalization, and is divided into training and validation sets in a 7:3 ratio. The data acquisition system employs a time synchronization mechanism, with all sensor data stamped with a unified GPS timestamp.

[0021] The collected data covers 15 classic fault types, including: gear pitting, broken gear teeth, bearing inner ring wear, bearing outer ring wear, bearing rolling element damage, shaft misalignment, shaft bending, loose coupling, reducer oil shortage, broken motor rotor bars, short circuit in motor stator windings, motor bearing overheating, brake failure, encoder failure, and limit switch failure. Each fault is categorized into three severity levels: mild, moderate, and severe. A three-level labeling system is used: the first level label indicates the fault type, the second level label indicates the location of the fault, and the third level label indicates the degree of verification.

[0022] In this embodiment, for the purpose of explaining the inventive concept, both manual feature extraction and deep learning approaches are analyzed. After screening, sensitive features with a Pearson correlation coefficient greater than 0.7 include four time-domain features (mean, variance, kurtosis, waveform factor), four frequency-domain features (centroid frequency, mean square frequency, frequency variance, frequency standard deviation), and the energy of the first four sub-bands after wavelet packet decomposition, totaling 12 fault-sensitive features. However, to preserve complete signal information and achieve end-to-end learning, the deep learning model constructed in this invention directly uses the preprocessed multi-channel original time series signal as input, rather than manually generated feature vectors. The model will automatically learn the optimal deep feature representation.

[0023] The feature encoder consists of three one-dimensional convolutional layers with kernel sizes of 32×1, 64×1, and 128×1, a stride of 1, and ReLU activation. The classifier consists of two fully connected layers with 256 and 15 neurons respectively. The output layer uses Softmax activation. An adaptive label smoothing strategy is employed to optimize the training process. During training, a temporary model in the current iteration state is first used to perform forward inference on the samples to obtain the predicted probability distribution. The information entropy of this distribution is calculated as shown in equation (1): (1); Among them, information entropy The range of values When the model predicts a class probability of a sample that is much less than 1, the larger the entropy value, the closer the model predicts a class probability to 1, and the smaller the entropy value.

[0024] Based on information entropy, a smoothing coefficient is dynamically assigned to each sample, as shown in equation (2) below: (2); in Based on the smoothing coefficient, This is the operating condition complexity adjustment factor. For the first The final smoothing coefficient for each sample.

[0025] Traditional one-hot labels concentrate all probability quality on the true class, which can lead to model overconfidence and push logits to a maximum value in the later stages of training, reducing the model's generalization ability. Therefore, the traditional one-hot labels are replaced with adaptive smooth labels, as shown in equation (3) below: (3); Label smoothing, by uniformly distributing a portion of the probability mass across all categories, serves as a regularization mechanism, preventing the model from overfitting the training data. The adaptive label smoothing cross-entropy loss is calculated as shown in equation (4) below: (4); The loss is optimized through backpropagation. During training, the model automatically adapts to the labeling confidence of different samples. The training parameters are set as follows: batch size 32, learning rate 0.001, Adam optimizer, and 500 iterations. After training, feature prototypes for each category in the source domain are immediately generated and stored as the knowledge carrier of the model.

[0026] According to an embodiment of the present invention, in step S2, to address the technical problem of low reliability of traditional pseudo-label generation when new models lack labeled data in passive domain scenarios, a memory-based neighborhood aggregation pseudo-label generation mechanism is proposed. A railway bridge-building machine is selected as the target domain model, and a sensor array of the same model and specifications as the source domain is deployed to ensure consistent sensor characteristics and collect label-free operational data. .

[0027] The target domain cannot access the original data in the source domain; it can only use the parameters of the pre-trained model in the source domain. and After the target domain data undergoes preprocessing consistent with that of the source domain, it is input into the source domain pre-trained model, as shown in equations (6) and (7) below: (6); (7); like Figure 2 As shown, this invention constructs a dynamic memory library. The in-memory database, with a capacity of 1024 entries, stores the deep feature representations and prediction statistics of samples in the target domain. It employs a first-in, first-out (FIFO) update strategy. Each record stores a triplet including a feature vector, a prediction probability vector, and an uncertainty score. When the database is full, the oldest sample with the highest uncertainty score is prioritized for deletion to ensure the prediction reliability of the samples within the database.

[0028] For each target domain sample feature in the memory library Calculate its features compared to all samples in the database. Gaussian kernel similarity is used to construct a local feature manifold graph. Similarity The calculation formula is shown in equation (8) below: (8); in, The Gaussian kernel width parameter is set to the median distance between all sample pairs.

[0029] Based on the graph diffusion mechanism, target domain samples The pseudo-labels are obtained by weighted aggregation of neighborhood prediction probabilities, as shown in equation (9) below: (9); This method utilizes the global structural information of the feature manifold to effectively suppress the interference of isolated noise samples on the generation of pseudo-labels, and significantly improves the accuracy of pseudo-labels.

[0030] To filter out low-quality pseudo-labels that may still exist after aggregation, this invention introduces an adaptive confidence threshold screening mechanism. (Sample definition follows.) Confidence of pseudo-labels The maximum class probability value after neighborhood weighted aggregation is shown in equation (10) below: (10); Unlike traditional fixed thresholds, this invention uses an adaptive threshold based on statistical distribution. The confidence level of the current batch of samples is dynamically adjusted based on the mean and standard deviation, as shown in equation (11) below: (11); in, and The mean and standard deviation of the confidence scores for the current batch of samples, with an adjustment factor. The value is 0.5. Only keep this value. The samples and their pseudo-labels form a high-quality target domain pseudo-label dataset. .

[0031] Under passive domain constraints, a domain adaptation strategy of classifier freezing and feature encoder fine-tuning is adopted. The classifier is frozen. Update all parameters, only the feature encoder. The joint loss function is shown in equation (12) below: (12); Among them, classification loss For pseudo-label cross-entropy loss and consistency loss Constrain the consistency of predictions before and after sample augmentation.

[0032] According to an embodiment of the present invention, in step S3, incremental learning anti-forgetting and prototype comparison optimization are performed to ensure that the model maintains its ability to identify existing fault types without degradation when new fault types are added.

[0033] The mean value of the sample features for each fault type in the source domain is calculated and stored as prototype features in the prototype library, as shown in equation (13) below: (13); in, For the source domain belonging to the first A sample collection of classes. Store all prototype vectors in the prototype library. .

[0034] When a new fault type appears in the target domain, after accumulating no fewer than 20 reliable samples, the mean of the characteristics of this type of sample is calculated as the new prototype. Insert the prototype into the prototype library. Simultaneously, the Euclidean distance between the new prototype and all old prototypes is checked; if the distance between any two prototypes of different categories is less than a safety threshold... Then, a prototype exclusion regularization term is added to the loss function, as shown in equation (14): (14); This mechanism ensures that different categories maintain sufficient distinguishability in the feature space, preventing new categories from eroding the boundaries of old categories.

[0035] like Figure 3 As shown, due to parameter updates in the feature encoder when learning new categories, representational drift occurs in the feature space, and the old prototype will no longer accurately reflect the position of the corresponding category in the new feature space. This invention introduces optimal transfer theory to calibrate the prototype to correct this drift.

[0036] The optimal mapping matrix is ​​obtained by minimizing the Wasserstein distance between the old and new feature spaces. The old prototype Mapping to a new space to obtain a calibrated prototype This embodiment uses the slice Wasserstein distance approximation, which significantly reduces computational complexity and allows prototype calibration to be completed quickly after each gradient update.

[0037] The training loss function in the incremental learning phase consists of three parts: the new category classification loss, the knowledge distillation loss, and the mixed contrastive learning loss. The new category classification loss... Calculate the cross-entropy loss and knowledge distillation loss for new category samples. With the previous feature encoder For the teacher network, the updated encoder For the student network, the Kullback-Leibler divergence is calculated for the old category samples to constrain the output distribution of the old and new encoders to be consistent, as shown in Equation (15): (15); Hybrid contrastive learning loss By using prototype vectors from the prototype library as anchor points for contrastive learning, the distance between sample features and their respective class prototypes is narrowed, while the distance between them and prototypes from other classes is increased. The total loss function is... ,in , By optimizing this loss, the model's knowledge forgetting rate can be controlled to below 5%.

[0038] According to an embodiment of the present invention, in step S4, the optimized model obtained from the aforementioned training is deployed on the target model to achieve online real-time diagnosis, and a closed-loop iteration is formed through on-site feedback data to continuously improve diagnostic performance.

[0039] like Figure 4 As shown, the online monitoring data of the target domain model, after preprocessing and feature extraction, is input into the optimized diagnostic model. Model quantization is used for acceleration. The feature encoder maps the input features to the domain-invariant feature space. The classifier calculates the Softmax probability of the feature vector belonging to each known fault category, and outputs the fault type with the highest probability, the corresponding fault location in the preset knowledge base, and the severity level based on the feature amplitude mapping, thus meeting the real-time monitoring needs of industrial sites.

[0040] After on-site maintenance personnel perform repair operations based on the diagnostic results, they record the confirmed fault information through a mobile terminal or onboard HMI interface. The data is then uploaded to a local server or cloud database via industrial Ethernet. The system periodically evaluates the accuracy of the fake tag generation strategy based on the returned real tags. If the current accuracy is lower than the target value of 90%, the hyperparameter optimization module is automatically triggered to fine-tune the following parameters within a preset range: Confidence threshold adjustment factor Search for the optimal value in the range of 0.3 to 0.7; The memory library capacity can be adjusted within the range of 512 to 2048; Gaussian kernel width parameter Recalculate based on the median distance of the current feature distribution.

[0041] During non-operational periods of the bridge erecting machine (such as nighttime downtime), the deployed model is incrementally fine-tuned using the accumulated real-label dataset and high-confidence pseudo-label dataset. Fine-tuning employs the Adam optimizer with a learning rate of 0.0002 and 100 epochs. The updated model parameters automatically overwrite the original model, completing one round of closed-loop evolution. When adapting this method to a third type of bridge erecting machine, steps S2 to S4 only need to be repeated; there is no need to access the original labeled data of the source domain highway bridge erecting machine again.

[0042] Example 2 This embodiment verifies the continuous adaptation capability of the present invention to a third or more machine models. The core verification is that the technical effect of cross-domain adaptation of multiple machine models can be completed without tracing back the original data of the source domain, using only the initial source domain model parameters.

[0043] This embodiment uses the highway bridge erecting machine in Embodiment 1 as the initial source domain model, the railway bridge erecting machine adapted in Embodiment 1 as the intermediate adapted model, and the segmental assembly bridge erecting machine as the new secondary target domain model. Throughout the process, it does not access the original labeled data of the source domain highway bridge erecting machine, but only calls the pre-trained model parameters of the initial source domain, the prototype library, and the optimized model parameters from Embodiment 1. The complete implementation process is as follows: For target domain data acquisition and preprocessing, considering the structural characteristics of the segmental bridge erecting machine, a multi-source heterogeneous sensor array of the same model and specifications as the source domain is deployed in key components such as the drive motor, reducer, hoisting device, traveling mechanism, and synchronous hydraulic system. Two additional pressure sensors are added to collect the working pressure of the hydraulic system. All sensors are kept in GPS time synchronization, and the sampling frequency is consistent with the source domain. Unlabeled operational data of the segmental bridge erecting machine during the entire process of beam segment hoisting, alignment, and beam placement are collected. A preprocessing workflow identical to that of the source domain, including wavelet threshold denoising, moving average filtering, and min-max normalization, is used to eliminate dimensional differences and environmental noise interference.

[0044] Passive domain adaptive migration only calls the pre-trained model parameters and prototype library of the initial source domain, without accessing the original source domain data, and performs feature extraction on the preprocessed unlabeled data of the segmental assembly bridge erecting machine. Considering the frequent switching of working conditions and large load fluctuations of the segmental assembly bridge erecting machine, the dynamic memory library capacity was adjusted to 2048 records, the nearest neighbor sample K value was set to 30, and the confidence threshold adjustment factor was adjusted. Pseudo-labels are generated through a memory-based neighborhood aggregation mechanism, and a high-quality pseudo-label dataset is obtained through adaptive threshold filtering. Domain adaptive transfer is achieved using a strategy of classifier freezing and feature encoder fine-tuning, with joint loss function weight coefficients... The training parameters were set to batch size 32, initial learning rate 0.0005, Adam optimizer, and maximum number of iterations 200, to obtain the segmental assembly bridge erection machine domain adaptation model.

[0045] Incremental learning is used to adapt to specific fault types. Two new fault types not covered by the source domain are identified: spreader synchronization deviation faults and hydraulic system pressure abnormality faults, both unique to segmental bridge erecting machines. At least 25 high-confidence pseudo-label samples are collected for each fault type. New fault prototypes are calculated and inserted into the prototype library. Old prototype drift calibration is performed using slice Wasserstein distance. Model optimization is achieved using the incremental training total loss function, with weight coefficients... , and After training is complete, update the model parameters and prototype library.

[0046] Online deployment and closed-loop optimization involve quantizing and deploying the optimized model to the onboard edge terminal of the segmental bridge erecting machine, enabling real-time online diagnostics. Based on real fault labels from on-site maintenance feedback, pseudo-labels are optimized monthly to generate hyperparameters, allowing for incremental fine-tuning of the model during non-operational periods of the bridge erecting machine.

[0047] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A bridge machine cross-machine fault diagnosis method, characterized in that, Includes the following steps: S1. Based on the source domain bridge erecting machine with labeled data, collect the operating data of key components, construct a fault diagnosis model containing feature encoders and classifiers, and simultaneously establish a feature prototype library of source domain fault categories. S2. In the scenario of a passive domain where the original source domain data cannot be accessed, the unlabeled operation data of the bridge erecting machine in the target domain is input into the pre-trained model. Pseudo-labels are generated through neighborhood aggregation and filtered by adaptive threshold to complete the adaptive transfer of the model. S3. Incremental training of the model is completed through prototype constraints and drift calibration, so that the model can learn new fault types while maintaining its ability to identify existing faults. S4. Deploy the optimized model to the target domain bridge erecting machine to achieve online fault diagnosis, and complete model iteration based on on-site feedback data to form a closed-loop evolution.

2. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S1, the multi-source heterogeneous sensing data includes: The triaxial vibration acceleration signal was acquired by a triaxial MEMS accelerometer, the bearing and gearbox surface temperature signal was acquired by a PT100 temperature sensor, the three-phase current signal was acquired by a Hall-effect current sensor, and the line voltage signal was acquired by an isolation voltage sensor. Preprocessing includes wavelet thresholding for noise reduction, moving average filtering, and normalization.

3. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S1, an adaptive label smoothing strategy is used to train the fault diagnosis model, including: using the temporary model of the current iteration state to perform forward inference on the samples during the training process, obtaining the predicted probability distribution and calculating the information entropy, dynamically assigning a smoothing coefficient to each sample based on the information entropy, replacing the traditional one-hot label with an adaptive smoothing label, calculating the adaptive label smoothing cross-entropy loss and optimizing the model parameters.

4. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S2, the dynamic memory library adopts a first-in-first-out storage strategy. Each record stores a feature vector, a prediction probability vector, and an uncertainty score triplet. When the library is full, the sample with the highest uncertainty score is eliminated first. The neighborhood prediction probability weighted aggregation is specifically as follows: calculate the Gaussian kernel similarity between samples in the memory library to construct a local feature manifold graph, and perform weighted aggregation of the prediction probabilities of neighborhood samples based on the similarity to generate sample pseudo-labels.

5. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S2, the adaptive confidence threshold screening mechanism includes: defining the pseudo-label confidence of a sample as the maximum class probability value after neighborhood weighted aggregation; adopting an adaptive threshold based on statistical distribution; dynamically adjusting the threshold according to the mean and standard deviation of the confidence of the current batch of samples; and retaining only samples with confidence greater than or equal to the adaptive threshold and their pseudo-labels to form a high-quality target domain pseudo-label dataset.

6. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S2, during the adaptive transfer of the model, all parameters of the classifier are frozen, and only the parameters of the feature encoder are fine-tuned and optimized.

7. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S3, the prototype exclusion constraint is specifically implemented as follows: when the distance between any two different categories of prototypes is less than a preset safety threshold, the prototype exclusion regularization term is activated to increase the prototype distance between categories in order to ensure the class distinguishability of the feature space.

8. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S3, the prototype drift calibration is specifically performed by: solving for the optimal mapping matrix by minimizing the Wasserstein distance between the old and new feature spaces, and mapping the old prototype to the updated feature space to complete the calibration; the Waserstein distance of the slice is used to approximate the solution, and the calibration operation is performed after each incremental training iteration.

9. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S3, the loss functions used for incremental training include classification loss for new fault categories, knowledge distillation loss for constraining feature distribution consistency, and hybrid contrastive learning loss anchored by prototypes.

10. The method for diagnosing cross-model faults of a bridge erecting machine according to claim 1, characterized in that, In step S4, the closed-loop evolution specifically involves periodically evaluating the accuracy of the pseudo-label generation strategy based on real fault labels from on-site feedback, and automatically optimizing hyperparameters when the accuracy is lower than the preset target. During the non-operation period of the bridge erecting machine, the model is incrementally fine-tuned using accumulated real label data and high-confidence pseudo label data to complete the model iteration. When adapting to a new bridge erecting machine model, repeat steps S2 to S4 without accessing the original annotation data of the source domain.

Images