Online Fault Diagnosis Method for Cranes under Coupled Time-Varying Operating Conditions Based on Size Model

Abstract

This invention discloses an online fault diagnosis method for cranes under coupled time-varying operating conditions based on a large-scale model. The method includes: offline pre-training a general large-scale model using a multi-source fault dataset to enable it to learn extensive general fault diagnosis knowledge, thereby extracting deep feature vectors of various faults from vibration signals; transferring the general fault diagnosis knowledge of the general large-scale model to a lightweight small-scale model through knowledge distillation, and generating an initial fault knowledge base, including a fault category prototype set and an operating condition prior matrix; using the lightweight small-scale model to extract fault feature vectors from real-time acquired crane vibration signals, and combining this with a Bayesian self-correcting online classifier that requires no backpropagation, dynamically updating the fault category prototype set and the operating condition prior matrix in the fault knowledge base, thus completing online fault diagnosis and fault knowledge base updates. This invention achieves lightweight deployment, online dynamic adjustment, and improved diagnostic accuracy of crane diagnostic models under coupled time-varying operating conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for key mechanical components, specifically to an online fault diagnosis method for cranes under coupled time-varying operating conditions based on a size model. Background Technology

[0002] Cranes are key equipment in modern industrial production, and their safe and stable operation is crucial. However, cranes operate under complex conditions for extended periods, frequently subjected to low-frequency cyclic impacts, resulting in significant and variable overall vibration characteristics. Specifically, key operating parameters such as lifting load, lifting speed, and trolley travel speed exhibit nonlinear, synchronously coupled dynamic changes, forming coupled time-varying operating conditions under strong noise. This dynamic condition significantly increases the difficulty of extracting domain-invariant fault features, leading to a decline in the performance of traditional fault diagnosis methods. Existing advanced diagnostic methods, such as domain-adaptive deep learning methods, typically assume that the data distributions of the source domain (training conditions) and the target domain (test conditions) are fixed but differ. These methods employ offline training, which has several fundamental drawbacks:

[0003] (1) Static adaptability: The adaptability of an offline domain adaptive model is fixed after training. For example, a model trained for the "500rpm to 1000rpm" working condition cannot be directly applied to the new "2000rpm" working condition. When the target working condition changes, a large amount of data under the new working condition must be collected again and the model must be retrained.

[0004] (2) Data dependency: Traditional domain adaptation methods require a large amount of labeled or unlabeled target domain data to be acquired in advance during the training phase, which is often difficult to achieve in actual industrial scenarios, especially for sudden or rare working conditions.

[0005] (3) Inability to handle streaming data: The coupled time-varying operating conditions of cranes mean that the distribution of their monitoring data changes continuously and dynamically over time. Monitoring data is usually transmitted in real time in small batches and streaming form. Traditional offline training modes cannot use this real-time data to adjust the model online, and therefore cannot cope with the dynamic drift of data distribution.

[0006] In summary, existing diagnostic methods based on the steady-state operating condition domain adaptation assumption, due to their offline and static nature, cannot meet the online, real-time, and adaptive diagnostic requirements of cranes under coupled time-varying operating conditions, and have obvious application failure problems. Summary of the Invention

[0007] To address the issue of traditional diagnostic methods for cranes failing online under coupled time-varying operating conditions due to static models and reliance on offline training, this invention provides an online fault diagnosis method for cranes under coupled time-varying operating conditions based on a large-scale model. First, in the offline phase, this method leverages the powerful generalization ability of a large model to learn general fault knowledge from multi-source industrial data, and then condenses this knowledge into a lightweight small model suitable for edge deployment through knowledge distillation. Crucially, in the online phase, this invention proposes a Bayesian self-correcting online classifier that requires no backpropagation. This classifier utilizes real-time acquired streaming data and performs dual dynamic self-correction on the fault prototype (likelihood) and operating condition prior (prior) at the top level of the small model through efficient statistical computation. Through this method, the diagnostic model can track data distribution drift caused by changes in operating conditions in real time at resource-constrained edge environments, achieving efficient, accurate, and truly adaptive online fault diagnosis under the challenges of coupled time-varying and streaming data.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The first aspect of the present invention provides an online fault diagnosis method for crane coupled time-varying operating conditions based on a size model, the method comprising:

[0010] Offline pre-training of a general fault diagnosis model is performed using a multi-source fault dataset, enabling it to learn a wide range of general fault diagnosis knowledge and extract deep feature vectors of various faults from vibration signals.

[0011] Knowledge distillation is used to transfer the general fault diagnosis knowledge of the general fault diagnosis model to a lightweight small model, and an initial fault knowledge base is generated according to the fault category. The fault knowledge base includes a set of fault category prototypes and a working condition prior matrix. The initial fault category prototypes are determined by the deep feature vectors of typical samples of each fault category, and the initial working condition prior matrix is ​​an identity matrix of the fault category size.

[0012] A lightweight small model is used to extract fault feature vectors from real-time acquired crane vibration signals. Combined with a Bayesian self-correcting online classifier that requires no backpropagation, the fault category prototype set and the prior matrix of operating conditions in the fault knowledge base are dynamically updated, thereby completing online fault diagnosis and fault knowledge base updates.

[0013] Calculate the prototype matching probability distribution of the extracted fault feature vector with each fault category prototype in the current fault category prototype set. Multiply the prototype matching probability distribution with the prior matrix of the current working condition to obtain the posterior probability distribution. The fault category with the highest probability value in the posterior probability distribution is the fault diagnosis result.

[0014] Update the fault category prototype set according to the prototype matching probability distribution, and update the working condition prior matrix according to the posterior probability distribution.

[0015] Furthermore, the general fault diagnosis model is a general model based on the Transformer architecture, including:

[0016] Instance normalization: For each vibration signal sample, calculate the mean and standard deviation of each channel, and use these statistics to normalize all time points of each channel within the sample, so that the normalized vibration signal sample has zero mean and unit variance within the sample; where each channel refers to the signal at different measurement points or in different directions of the sensor.

[0017] Segmented embedding: The normalized vibration signal sample is divided into several continuous segments of equal length. Each segment is mapped to a high-dimensional feature sequence through a learnable linear embedding layer.

[0018] Position encoding: A position encoding vector is superimposed on each high-dimensional feature vector to form a high-dimensional feature sequence; the position encoding vector includes the position information of each segment in the vibration signal;

[0019] Encoder Layers: High-dimensional feature sequences are fed into stacked Transformer encoder layers for deep modeling. Each encoder layer captures individual high-dimensional feature sequences through a multi-head self-attention mechanism and a feedforward neural network. It identifies internal long-range dependencies and outputs a deep feature vector. .

[0020] Furthermore, a multi-task learning head is constructed for the general fault diagnosis model, including:

[0021] Two independent learning heads are connected in parallel at the output of the general fault diagnosis model to perform different pre-training tasks:

[0022] (1) Classification head: A fully connected layer followed by a Softmax activation function, used to classify the deep feature vector Probability distribution mapped to fault categories ;

[0023] (2) Projection head: a fully connected layer used to project the depth feature vector Mapping to low-dimensional features used for contrastive learning ;

[0024] Offline pre-training employs a multi-task hybrid learning framework, which trains a general large-scale fault diagnosis model by jointly optimizing supervised classification tasks and self-supervised contrastive learning tasks.

[0025] Furthermore, multi-task hybrid learning frameworks include:

[0026] Supervised classification task: Train a model using supervised cross-entropy loss to distinguish different fault categories.

[0027] ;

[0028] In the formula, This represents the loss value for supervised learning; N Represents the number of samples in the training batch; K This represents the total number of fault categories in the multi-source fault dataset. Indicates the fault category label; Indicates the first i Each sample belongs to the fault category. k The probability of;

[0029] Self-supervised contrastive learning task: Data augmentation of vibration signals, and training of a model using contrastive loss to learn robust feature representations to changes in operating conditions.

[0030] ;

[0031] In the formula, This represents the average contrast loss value for the entire batch. is the feature vector of the anchor sample, and is the low-dimensional feature of any sample in a certain batch; It is the feature vector of the positive sample, which belongs to the same vibration signal as the anchor sample, and represents the low-dimensional features of the sample after data augmentation. The feature vector of the negative sample is the low-dimensional feature of all other vibration signals in the batch. This is the temperature coefficient, a positive scalar. sim For similarity functions;

[0032] Joint training: By weighted combination of the losses from supervised classification tasks and self-supervised contrastive learning tasks, the pre-training of a general large-scale fault diagnosis model is completed.

[0033] Furthermore, the similarity function uses cosine similarity, as follows:

[0034] ;

[0035] In the formula, A and B This represents two eigenvectors.

[0036] Furthermore, the process of generating the initial fault category prototype set includes:

[0037] Determine the diagnosis of crane faults M Types of faults;

[0038] For each type of fault, select several typical samples;

[0039] Typical samples are input into a general fault diagnosis model to extract... d 3D deep feature vector;

[0040] Calculate the arithmetic mean of the deep feature vectors of all typical samples belonging to the same fault category to obtain the fault category prototype.

[0041] Will M The fault category prototypes of each fault category are combined column by column to ultimately form Initial fault category prototype set of dimensional U ;

[0042] The process of generating the prior matrix of operating conditions includes:

[0043] Build a identity matrix V The first of the matrix i The row represents the signal feature matching the first i Each fault category prototype When the actual fault category is equal to the prior probability distribution of each fault category, the true fault category is equal to the prior probability distribution of each fault category.

[0044] Furthermore, online fault diagnosis specifically includes:

[0045] Extracting real-time crane vibration signals using a lightweight small model Fault feature vector ;

[0046] Calculate the fault feature vector With the current fault category prototype set The cosine similarity of each fault category prototype is calculated and normalized using the Softmax function to obtain a... M Dimensional prototype matching probability distribution :

[0047] ;

[0048] In the formula, M The total number of crane fault categories; This is the prototype of the fault category for the j-th fault category; Represents the fault feature vector With fault category prototype cosine similarity, m =1,2,3… M This allows for the iteration through all fault category prototypes;

[0049] Prototype matching probability distribution Prior matrix of current operating condition Matrix multiplication is used to obtain the posterior probability distribution. :

[0050] ;

[0051] The fault category with the highest probability value in the posterior probability distribution is the one corresponding to the current crane vibration signal. The fault diagnosis results.

[0052] Further updates to the fault knowledge base include:

[0053] (1) Update the fault category prototype set according to the prototype matching probability distribution, including:

[0054] From the prototype matching probability distribution Find the prototype of the fault category with the highest probability. and its index s ;

[0055] Only when the highest prototype matches the probability distribution The fault category prototype set is only updated when the score is greater than or equal to a preset first confidence threshold.

[0056] ;

[0057] In the formula, Norm Indicates vector normalization; This is the first counter, used to record the prototype of the fault category. The number of updates;

[0058] (2) Update the prior matrix of the working conditions according to the posterior probability distribution, including:

[0059] At the same time, according to the posterior probability distribution Update the prior matrix of operating conditions The s OK:

[0060] ;

[0061] In the formula, For the prior matrix of working conditions The s OK; This is the second counter, used to record the prior information in the operating condition matrix. The number of times it is updated.

[0062] According to a second aspect of the present invention, a computer device is provided, comprising: a processor and a memory, the memory storing a program or instructions executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the online fault diagnosis method for crane coupling time-varying operating conditions based on a size model as described in any one of the first aspects.

[0063] According to a third aspect of the present invention, a computer-readable storage medium is provided having a program or instructions stored thereon, which, when executed by a processor, implement the steps of the online fault diagnosis method for crane coupling time-varying operating conditions based on a size model as described in any one of the first aspects.

[0064] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0065] (1) This invention achieves a unification of model capabilities and edge deployment, resolving the core contradiction between high performance and lightweight design. Through offline large-scale model pre-training and knowledge distillation, this invention "condenses" and transfers general and robust fault diagnosis knowledge learned from massive industrial data into a lightweight small model. This enables the model deployed at the edge of the crane to possess both powerful feature extraction and generalization capabilities approaching those of a large model, while also meeting the stringent requirements of limited edge computing resources, providing a novel and feasible technical path for deploying high-performance intelligent diagnostic models in industrial settings.

[0066] (2) An online adaptive mechanism without backpropagation was constructed, realizing real-time tracking and adaptation to coupled time-varying operating conditions. The Bayesian self-correcting online classifier designed in this invention eliminates the high computational overhead required by traditional online fine-tuning. It performs dual dynamic updates on the fault category prototype (likelihood) and operating condition prior (prior) through efficient statistical operations. This enables the diagnostic model to learn and adapt to the continuous drift of data distribution caused by coupled changes in load, speed, and other factors in real time and online, solving the fundamental problem of performance degradation in online diagnosis of traditional static models, and truly realizing intelligent diagnosis that "changes with operating conditions and learns with data".

[0067] (3) The technical solution has high versatility, scalability and robustness. The top-level online adaptation module of this invention is completely decoupled from the bottom-level feature extraction model architecture. This online adaptation capability can be seamlessly integrated into small models based on CNN, Transformer or other architectures.

[0068] (4) Significantly reduced operation and maintenance costs and improved the automation and intelligence level of the diagnostic system. The online adaptive capability proposed in this invention greatly reduces the operation and maintenance burden and cost of frequently and manually re-collecting data, labeling samples, and retraining models due to changes in operating conditions. The diagnostic system can achieve "one-time deployment, long-term adaptation", reducing the deep dependence on operation and maintenance engineers.

[0069] In summary, compared with existing methods, the present invention has significantly improved diagnostic accuracy and real-time performance, and has broad engineering application prospects and great economic value. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the overall process of an online fault diagnosis method for crane coupled time-varying working conditions based on a size model, provided by an embodiment of the present invention.

[0071] Figure 2 This invention provides a general large-scale fault diagnosis model architecture diagram.

[0072] Figure 3 A flowchart of knowledge distillation for a large-scale model is provided as an embodiment of the present invention;

[0073] Figure 4 A flowchart illustrating the generation process of a crane fault knowledge base, as provided in an embodiment of the present invention;

[0074] Figure 5 This is a structural diagram of a Bayesian self-correcting online classifier provided in an embodiment of the present invention;

[0075] Figure 6 This is a schematic diagram illustrating a dynamically updated fault knowledge base provided in an embodiment of the present invention;

[0076] Figure 7 This is a schematic diagram illustrating the working process of an online fault diagnosis model for a crane under coupled time-varying conditions, provided in an embodiment of the present invention.

[0077] Figure 8 This is a schematic diagram of the hardware structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0079] Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, modifications to design, manufacturing, or production based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient disclosure of the present invention.

[0080] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention may be combined with other embodiments without conflict.

[0081] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "a," "an," "an," "the," and similar words used in this invention do not indicate quantity limitation and may indicate singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms "connected," "linked," "coupled," and similar words used in this invention are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "A plurality" used in this invention refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships may exist; for example, "A and / or B" can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects have an "or" relationship. The terms "first," "second," and "third" used in this invention are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0082] Cranes are critical equipment in industrial production, and their safe operation is paramount. However, cranes operate under complex conditions for extended periods, with parameters such as lifting load and operating speed exhibiting nonlinear coupled changes, resulting in coupled time-varying operating conditions. This leads to continuous dynamic changes in the data distribution of vibration signals, posing a severe challenge to the accuracy of fault diagnosis. To address these changing operating conditions, domain adaptation-based deep learning methods have been extensively studied. However, existing domain adaptation methods are mostly offline training methods, designed to solve static operating condition transition problems. When applied to the dynamic diagnosis of cranes, they suffer from three core contradictions:

[0083] (1) The contradiction between static models and dynamic working conditions: The parameters of offline trained models are fixed and cannot be generalized to new working conditions not anticipated during training, resulting in limited adaptability;

[0084] (2) The contradiction between data pre-collection and data streaming: Traditional methods require the collection of a large amount of target domain data in advance, but this is contrary to the reality that crane monitoring data is generated in real time in small batches and streaming form;

[0085] (3) The contradiction between offline training and online adjustment: the model cannot be updated online using real-time data streams, its knowledge will gradually become "outdated" and it will be unable to track the dynamic drift of data distribution.

[0086] In summary, existing diagnostic methods, due to their offline and static nature and deployment challenges, cannot meet the online, adaptive, and lightweight diagnostic needs of cranes under coupled time-varying operating conditions, and technological breakthroughs are urgently needed.

[0087] To address this, the present invention provides an online fault diagnosis method for cranes under coupled time-varying operating conditions based on a size model, aiming to solve the problem that existing fault diagnosis methods are difficult to adapt to the coupled time-varying operating conditions of cranes. Specifically, the present invention aims to solve the following technical problems:

[0088] (1) The problem of poor adaptability of traditional diagnostic methods under dynamic working conditions: Most existing methods are based on static working condition assumptions and cannot adapt in real time to the dynamic drift of monitoring data distribution caused by the coupling changes of parameters such as load and speed;

[0089] (2) The problem that traditional offline domain adaptation models cannot be adjusted online: Existing domain adaptation methods require a large amount of target domain data to be acquired in advance for offline training, and cannot use small batches of streaming data to update the model online in order to cope with unknown or sudden working conditions.

[0090] In other words, how to utilize small-batch real-time monitoring data streams under coupled time-varying operating conditions to dynamically adjust the fault diagnosis model to adapt to the current data distribution in real time, and realize lightweight online high-precision diagnosis of the crane edge-end diagnosis model, is an engineering bottleneck problem that urgently needs to be solved.

[0091] The online fault diagnosis method for cranes under coupled time-varying operating conditions provided by this invention includes the following steps: First, a general large model is pre-trained on a multi-source industrial dataset in the cloud to learn a wide range of fault features. Knowledge distillation is then used to transfer general knowledge to a lightweight small model, generating an initial fault knowledge base. Next, the small model and fault knowledge base are deployed at the edge of the crane, combined with a Bayesian self-correcting online classifier that requires no backpropagation, to process real-time streaming data. Based on statistical methods, the fault category prototypes and operating condition prior matrices in the fault knowledge base are dynamically updated, allowing for real-time and efficient adaptation to the current data distribution. Finally, based on the stable feature extraction capabilities of the small model and the Bayesian self-correcting online classifier, lightweight online fault diagnosis is performed. This invention achieves lightweight deployment, online dynamic adjustment, and improved diagnostic accuracy of crane diagnostic models under coupled time-varying operating conditions.

[0092] like Figure 1 As shown, the online fault diagnosis method for cranes under coupled time-varying conditions based on a size model according to the present invention includes the following steps:

[0093] 1) Deploy a general-purpose large model on a cloud or server with sufficient computing resources, and use multi-source public fault datasets to carry out offline pre-training of the large model to learn a wide range of general fault diagnosis knowledge;

[0094] 2) Using knowledge distillation technology, the general diagnostic knowledge of the pre-trained large model is transferred to a lightweight small model to obtain a lightweight small model with an aligned feature space with the large model that can be efficiently reasoned. The general large model is then used to generate a crane fault knowledge base containing initial fault category prototypes and working condition priors.

[0095] 3) Deploy lightweight mini-models and crane fault knowledge bases to crane edge computing devices;

[0096] 4) Utilize a lightweight small model to extract high-dimensional features from real-time streaming monitoring data. Combine this with a Bayesian self-correcting online classifier that does not require backpropagation to perform dual dynamic updates: update the fault category prototype in the fault knowledge base using the current signal features to track signal morphology drift; and update the working condition prior matrix using the diagnostic posterior probability to learn the dynamic correlation between working conditions and faults. Adapt to the current data distribution in real time and establish a dynamic and reliable mapping relationship between the high-dimensional features extracted by the small model and the fault category prediction to enhance the accuracy of fault diagnosis.

[0097] Through the above steps, an online fault diagnosis model for coupled time-varying operating conditions of cranes was constructed at the resource-constrained edge, which efficiently and accurately solved the problem of fault diagnosis under coupled time-varying operating conditions.

[0098] In some embodiments, step 1) specifically involves the following steps:

[0099] ① Data Preparation and Integration: Collect and integrate multiple publicly available, multi-source industrial machinery fault datasets, such as the Case Western Reserve University (CWRU) bearing dataset, the Xi'an Jiaotong University XJTU-SY bearing dataset, and the Southeast University gearbox dataset. Perform unified preprocessing on these data from different sources and under different operating conditions, such as signal slicing and normalization, to construct a large-scale, diverse pre-training dataset.

[0100] Multi-source public fault datasets are constructed by collecting and integrating multiple public mechanical fault datasets covering different equipment (such as bearings, gearboxes, motors, etc.) and operating conditions to build a large-scale, diverse pre-trained dataset.

[0101] ② Construct a large-parameter signal encoder based on deep learning as a general large model (teacher model). The model can adopt a Transformer-based architecture, which uses its core multi-head self-attention mechanism to capture long-range temporal dependencies in vibration signals.

[0102] The core components of the general large model include a segmented embedding layer for converting the input vibration signal into a high-dimensional feature sequence, and a stacked multi-head self-attention module and a feedforward network module for context-aware modeling of the feature sequence, so as to achieve effective extraction of non-local, long-term dynamic features in the signal.

[0103] ③ To ensure that the features extracted by the large model possess both discriminative and generalizable qualities, a multi-task hybrid learning framework is employed for pre-training. A joint loss function simultaneously optimizes the following two tasks:

[0104] Supervised classification task: Using the fault category labels in the dataset, train the model to distinguish different fault categories through the cross-entropy loss function;

[0105] Self-supervised contrastive learning task: Random data augmentation (such as adding noise or time shifting) is applied to the signal, and the model is trained to learn a feature representation that is robust to changes in operating conditions by using a contrastive loss function.

[0106] This step yields a teacher model capable of extracting robust, general features from complex industrial signals.

[0107] In some embodiments, step 2) specifically involves the following steps:

[0108] ① Lightweight small model design: Design a lightweight signal encoder with a small number of parameters and low computational complexity as a student model. Its network structure can be a variant of lightweight CNN such as MobileNet or ShuffleNet to meet the deployment requirements of the crane edge.

[0109] The lightweight small model is a lightweight vibration signal encoder with a small number of parameters and low computational complexity. It can achieve real-time inference on mainstream industrial edge computing devices (such as NVIDIA Jetson series and embedded ARM platforms) and meet the deployment requirements of crane edge devices.

[0110] ② Knowledge Distillation: A feature-aligned knowledge distillation method is employed. Signals from the pre-training dataset are simultaneously input into both the teacher and student models. The student model is trained with the goal of minimizing the mean squared error between their output feature vectors. After distillation, the student model obtains a feature space aligned with the teacher model, achieving efficient and high-quality feature extraction capabilities.

[0111] Knowledge distillation is a technique that uses a large model as the teacher model and a lightweight small model as the student model, transferring knowledge from the teacher model to the student model. Specifically, signals from the pre-training dataset are simultaneously input into both the teacher and student models. The student model is trained to minimize the mean squared error between their output feature vectors. This process yields a lightweight small model with an aligned feature space with the teacher model, enabling efficient reasoning.

[0112] ③ Generation of fault knowledge base:

[0113] Generate an initial set of fault category prototypes Typical fault samples covering predefined fault categories of cranes are input into the teacher model to extract their deep features. The initial feature prototype for each fault category is calculated by clustering or averaging features belonging to the same category. Combined into an initial set of fault category prototypes ;

[0114] Generate the initial working condition prior matrix : The prior matrix of working conditions Initialize it as an identity matrix to establish an initial, unbiased one-to-one correspondence between fault prototypes and categories.

[0115] Among them, the fault category prototype set It is A matrix of dimension, where, M The total number of predefined crane fault categories (including normal status). d The dimension of the eigenvectors. Each row of the matrix. Representing the m The initial, standardized feature representation of each fault category, i.e., the fault category prototype;

[0116] Operating condition prior matrix It is A dimensional matrix is ​​used to represent the initial, unbiased prior beliefs about the working conditions; This represents the probability that a fault prototype corresponds to the final category, used to perceive the distribution of new categories under the current operating conditions / environment. Dimensions A one-hot encoded vector, only the index is m The value at index 1 is 0, while the values ​​at all other indices are 0.

[0117] Initial Fault Category Prototype Set and the initial working condition prior matrix The entire generation process is completed offline in one go; after generation and The matrix itself is a lightweight data structure that occupies little storage space and can be loaded and updated without a lot of computing resources, making it suitable for deployment together with lightweight small models on resource-constrained crane edge computing devices.

[0118] In some embodiments, step 3) specifically involves the following steps:

[0119] ① The lightweight student model trained by knowledge distillation and the generated initial set of fault category prototypes. and the initial working condition prior matrix Package it into a deployment unit;

[0120] ② Deploy the deployment unit to an edge computing device installed on a crane, such as an embedded GPU module (like the NVIDIA Jetson series) or a high-performance industrial PC.

[0121] In some embodiments, step 4) specifically involves the following steps:

[0122] ① Streaming monitoring data refers to real-time monitoring data from sensors on key crane components. This data is processed using a pre-deployed lightweight student model to analyze the real-time monitoring signals. Perform fast forward computation to extract its high-dimensional feature vector. .

[0123] ② Computational characteristics With fault category prototype set The matching probability (i.e., normalized likelihood) of each prototype is then compared with the current working condition prior matrix. Perform matrix multiplication to obtain the final posterior probability distribution, i.e., for The fault diagnosis results.

[0124] ③ Determine whether the highest confidence level of the diagnostic result obtained in step ② exceeds the preset threshold. If the limit is exceeded, the following update will be performed:

[0125] Fault category prototype update: using current signal characteristics Update the fault prototype that best matches it. To track feature drift caused by changes in operating conditions;

[0126] Update the prior matrix of operating conditions: using diagnostic posterior probabilities This is used to update the corresponding rows in the prior matrix V of operating conditions in order to learn the dynamic relationship between operating conditions and failure probabilities.

[0127] A Bayesian self-correcting online classifier is used to transform high-dimensional features extracted by a small model into fault category predictions. Its working mechanism is based on the fault feature vector extracted by the small model. Fault Category Prototype Set and working condition prior matrix The dynamic updating of the fault category prototypes and operating condition prior matrices in the fault knowledge base is based on real-time acquired streaming data. This lightweight, backpropagation-free online adaptation method ensures extremely high efficiency and minimal computational resource consumption, perfectly suited for edge computing scenarios. Compared to traditional methods of adjusting normalization layer parameters, its core advantage lies in the fact that traditional methods remain at a shallow statistical level adapting to data distribution, while this invention directly adjusts at a high-level semantic level by updating fault prototypes and operating condition priors, without needing to update the parameters of the smaller model. Furthermore, the update mechanism of this invention operates at the top level of the model, exhibiting natural robustness to noisy data through confidence filtering; and it is independent of the specific architecture of the feature extractor (small model), allowing seamless integration with both CNNs and Transformers, making it far more versatile than methods that rely solely on a specific normalization layer.

[0128] Through the collaborative efforts of the above steps, a complete online fault diagnosis model for cranes under coupled time-varying conditions was finally constructed and implemented at the resource-constrained edge of the crane system. This model learns general knowledge offline and performs gradient-free, lightweight dynamic adaptation online, thus efficiently and accurately addressing the continuous data distribution drift caused by coupled changes in factors such as load and speed, solving the challenge of fault diagnosis under coupled time-varying conditions.

[0129] Specifically, the online fault diagnosis method for crane coupled time-varying operating conditions based on a size model according to embodiments of the present invention, such as... Figure 1 As shown, it includes the following steps:

[0130] 1) On a cloud or server with sufficient computing resources, conduct pre-training of a general fault diagnosis model to learn a teacher model capable of extracting general and robust features from various industrial vibration signals. The specific steps are as follows:

[0131] ① Construction of a general fault diagnosis model: Refer to Figure 2 This embodiment constructs a general-purpose large model (teacher model) based on the Transformer architecture. Its input processing flow is as follows:

[0132] Instance normalization: For each vibration signal sample, the mean and standard deviation of each channel are calculated, and these statistics are used to normalize all time points of each channel within the sample, so that the normalized vibration signal sample has zero mean and unit variance within the sample. Here, each channel refers to the signal from different measurement points or different directions of the sensor. This step can eliminate the influence of the distribution differences of vibration signal samples in the input multi-source dataset caused by different working conditions and equipment, allowing the general large model to focus on the morphological pattern of the vibration signal sample rather than the physical amplitude.

[0133] Segmented embedding: The normalized vibration signal sample is divided into several continuous segments of equal length. Each segment is mapped to a high-dimensional feature sequence through a learnable linear embedding layer. This step transforms the continuous signal into a set of abstract feature representations, that is, maps the mechanical signal to a token space that the Transformer can understand, in order to adapt to the input requirements of the Transformer architecture.

[0134] Positional encoding: To compensate for the temporal insensitivity of the Transformer's self-attention mechanism, a unique positional encoding vector is superimposed on each vector of the aforementioned feature sequence. This positional encoding vector contains the absolute positional information of each signal segment in the original time series, enabling the model to understand and utilize the sequential relationship of the signals.

[0135] After the above processing, the original signal is transformed into a high-dimensional feature sequence that contains both rich feature information and key temporal information. This sequence is then fed into... W Depth modeling is performed using stacked Transformer encoder layers.

[0136] Each encoder layer, through its core multi-head self-attention mechanism and feedforward neural network, captures long-range dependencies within the signal sequence and outputs a deep feature vector. ,in, It is a general-purpose large model used for feature extraction.

[0137] ② Construction of the multi-task learning head: in the teacher model The output of the device is connected in parallel with two independent learning heads to perform different pre-training tasks:

[0138] Classification Header: A fully connected layer followed by a Softmax activation function, used to classify the deep feature vector. Probability distribution mapped to fault categories .

[0139] Projection head: a fully connected layer that will... Output deep feature vector Mapping to low-dimensional features used for contrastive learning .

[0140] ③ Multi-task hybrid learning: A multi-task hybrid learning framework is adopted, which trains the model by jointly optimizing the following two tasks. The specific steps are as follows:

[0141] Supervised classification task: Utilizing the labels in the dataset, the supervised cross-entropy loss is expressed as shown in the following formula. The trained model distinguishes between different fault categories:

[0142] ;

[0143] in, This represents the loss value for supervised learning, and the goal of model training is to minimize this loss. N Represents the number of samples in the training batch; K This represents the total number of fault categories in a multi-source public fault dataset. Indicates the true category label; This represents the predicted probability of the model, i.e., the probability that the model predicts the nth... i Each sample belongs to category k The probability is obtained by processing the features extracted from the large model through a classification head.

[0144] Self-supervised contrastive learning task: By augmenting the signal with data, the model is trained using the contrastive loss shown in the following formula to learn intrinsic features that are insensitive to changes in operating conditions, noise, and other disturbances.

[0145] ;

[0146] in, This represents the average contrast loss value for the entire batch. N This also represents the number of original samples in a training batch; The feature vector of the anchor sample is a low-dimensional feature from any sample in a certain batch; The feature vector of the positive sample comes from the same original signal as the anchor point, and represents the low-dimensional feature vector of the sample after different data augmentation. The feature vectors of negative samples are the low-dimensional feature vectors of all other samples in the batch that are different from the original signal. The low-dimensional feature vectors mentioned above are all obtained by outputting the features extracted by the large model after passing through the projection head. The anchor point and the positive sample form a unique "same class pair", and the anchor point and the negative sample form a "different class pair". sim It is a similarity function, and in this embodiment, cosine similarity is preferably used to calculate the similarity between two feature vectors (e.g., A and B The directional consistency in the feature space is calculated using the following formula: ; It is a temperature coefficient, a positive scalar, used to adjust the distribution of similarity scores; it is relatively small. This will make the distribution sharper, increasing the difficulty for the model to distinguish negative samples, and thus leading to better learning results.

[0147] Joint training: By using a joint loss function, the two tasks mentioned above are optimized simultaneously to conduct offline pre-training of a general-purpose large model. This allows the large model to simultaneously learn to identify known faults and understand signal structures, thereby obtaining more robust and generalizable feature representations and mastering a wide range of general fault diagnosis knowledge.

[0148] ;

[0149] in, The total loss from pre-training the large model is used to update the parameters of the large model; These are weighting coefficients used to balance the importance of supervised loss and self-supervised loss.

[0150] 2) Utilize knowledge distillation to transfer the extensive knowledge from the pre-trained general fault diagnosis model to a lightweight small model, generating a fault knowledge base to prepare for deployment at the edge of the crane. The specific steps are as follows:

[0151] ① Knowledge distillation: Figure 3 This is a flowchart of the knowledge distillation process for the large and small models of this invention. To achieve lightweight deployment, this embodiment uses the teacher model... Knowledge transfer to a lightweight small model Specifically, the signals in the pre-training dataset Simultaneously inputting a fixed teacher model and a student model to be trained, the student model is trained by minimizing the mean squared error (i.e., knowledge distillation loss) between the output features of the two backbone networks:

[0152] ;

[0153] in, For the teacher model backbone network (excluding the multi-task learning head) Extracted features; Student model pair Extracted features; This represents the square of the L2 norm, i.e., the mean squared error. This process aligns the feature space of the student model with that of the teacher model.

[0154] ② The crane fault knowledge base generation consists of two parts: initial fault category prototypes. U and working condition prior matrix V ,like Figure 4 As shown.

[0155] a. Initial Fault Category Prototype U The specific steps involved in generating the product include:

[0156] First, clearly define the identification required for crane diagnostic tasks. M Fault categories, for example: {Category 1: Normal operation, Category 2: Bearing inner ring wear, Category 3: Gear tooth breakage, ..., Category ...} M : Motor winding short circuit}

[0157] For each of the above fault categories, select a set (e.g., 10-50) of the most typical, clearest, and most representative signal samples from multi-source publicly available fault datasets or historical crane operation and maintenance data. Merge the typical samples from all categories to form a typical fault sample set.

[0158] Each sample in the typical fault sample set is input into the pre-trained general fault diagnosis model. Utilizing the model's powerful and well-generalized feature extraction capabilities, a feature is generated for each typical sample. d 3D deep feature vectors.

[0159] The deep feature vectors of all typical samples belonging to the same fault category are aggregated to compute a unique, centered fault prototype for that category. The preferred aggregation method is to calculate the arithmetic mean of these deep feature vectors:

[0160] ;

[0161] in, It is a category m Typical sample size, It belongs to the category m The sample.

[0162] Finally, all M A calculated fault prototype By combining columns, the final result is... Initial fault category prototype set of dimensional U .

[0163] b. Initial operating condition prior matrix V The specific steps involved in generating the product include:

[0164] First, build a square array V The first matrix i The row represents the signal feature matching the first row. i Each fault category prototype When the true category is the prior probability distribution of each fault, the true category is the true category of each fault.

[0165] The matrix is ​​calculated according to the following formula. V Initialize to an identity matrix:

[0166] .

[0167] Initialized identity matrix, initial working condition prior matrix V It is generated, representing a most direct and unbiased initial belief: if the characteristics of a signal look like the first... iIf it's the prototype of this type of fault, then we are 100% certain that it is the first one. i This initial assumption, which assumes a strict one-to-one correspondence between the prototype and the category before seeing online data for any specific operating condition, avoids introducing any subjective bias. This initial belief will be updated during subsequent Bayesian self-correcting online classification.

[0168] 3) Lightweight small model and fault knowledge base ( U , V Deployed at the edge of the crane, it utilizes a Bayesian self-correcting online classifier to perform online diagnostics and fault knowledge base updates, such as... Figure 5 As shown, the specific steps are as follows:

[0169] ① Online diagnosis based on Bayesian self-correcting online classifier:

[0170] a. Feature extraction: Real-time acquisition of each frame or small batch of signal data by the sensor. It is input into the deployed lightweight small model and its... 3D Fault Feature Vector .

[0171] b. Prototype matching probability calculation: Calculate the fault feature vector according to the following formula. With fault category prototype Each prototype The cosine similarity is calculated and normalized using the Softmax function to obtain a... M Dimensional prototype matching probability distribution :

[0172] ;

[0173] In the formula, M The total number of crane fault categories; For the first j A prototype of a fault category; Represents the fault feature vector With fault category prototype cosine similarity, m =1,2,3… M This allows us to iterate through all fault category prototypes.

[0174] c. Diagnostic decision: Matching the prototype probability distribution With the current working condition prior matrix Performing matrix multiplication, as shown in the equation below, yields the final posterior probability distribution. The class with the highest probability value in this distribution represents the class of the current data. Fault diagnosis results:

[0175] .

[0176] ②See the diagram illustrating the dynamic updating of the fault knowledge base. Figure 6 The specific steps are as follows:

[0177] a. Select the update target: the prototype matching probability distribution calculated from diagnostic steps a and b of the Bayesian self-correcting online classifier. In the process, find the fault prototype with the highest probability. and its index s .

[0178] b. Conditional judgment: Only when the highest matching probability is met. Greater than the preset confidence threshold Only when necessary should the fault knowledge base be updated to avoid interference from low-quality or noisy data.

[0179] c. Fault Category Prototype Set Update: Using the current signal fault feature vector Update the most suitable fault prototype The update method is shown in the following formula:

[0180] ;

[0181] in, Norm Indicates vector normalization, This is the first counter used to record fault prototypes. The number of times the fault category prototype is updated. Updating the fault category prototype allows its representation to shift towards the feature distribution of the new test data, thereby matching the new distribution of the current operating condition in real time.

[0182] d. Prior condition matrix Update: The posterior probability of diagnosis obtained in step c of the diagnosis using a Bayesian self-correcting online classifier. Update the prior matrix of operating conditions The s Okay, the update method is as follows:

[0183] ;

[0184] in, This is the second counter, used to record the prior information in the operating condition matrix. The number of updates; The fault prototype representation is selected from step a in the diagnostics process. This step enables the model to learn the dynamic relationship between specific operating conditions and the probability of fault occurrence, which is crucial for adapting to coupled time-varying operating conditions.

[0185] 4) Through the collaborative work of the above steps, a complete online fault diagnosis model for cranes under coupled time-varying operating conditions was constructed and implemented at the resource-constrained edge of the crane system. The workflow of the model is described in [link to model]. Figure 7 The specific steps are as follows:

[0186] a. First, a lightweight small model deployed at the edge is used to quickly extract features from real-time streaming vibration signals;

[0187] b. Then, the Bayesian self-correcting online classifier takes over the diagnostic decision-making. It requires no backpropagation at all, but instead performs real-time dual dynamic updates through highly efficient statistical computation:

[0188] ① Update fault prototypes: Continuously fine-tune the fault prototypes of "standard faults" with new data to track signal morphology drift caused by changes in operating conditions (such as load and speed changes) in real time;

[0189] ② Update prior operating conditions: Dynamically learn the empirical patterns of "what kind of failure is more likely to occur under what operating conditions", so that diagnostic decisions are integrated with the judgment of current operating conditions risks;

[0190] c. Finally, an online loop of "feature extraction → diagnostic decision → fault knowledge base update" is continuously performed on the real-time data stream. Through this continuous and lightweight self-correction, the model achieves dynamic adaptation to the coupled time-varying operating conditions of the crane, completes high-precision online fault diagnosis, and achieves truly online and lightweight deployment with extremely low computational overhead.

[0191] In summary, this invention designs a large-to-small model knowledge transfer framework. By fully pre-training a general-purpose large model on multi-source data in the cloud and employing knowledge distillation technology, it efficiently transfers and solidifies the general diagnostic knowledge into a lightweight small model. This method effectively solves the core contradiction that high-performance diagnostic models, due to their massive parameter count, are difficult to deploy at resource-constrained crane edge environments. The deployed small model possesses both the powerful generalization capabilities of the large model and meets the practical engineering requirements of lightweight design and high efficiency.

[0192] Furthermore, this invention proposes a Bayesian self-correcting online classifier that requires no backpropagation. Operating on the top-level decision module of the model, it utilizes real-time acquired streaming monitoring data and employs efficient statistical computation to dynamically update both the fault category prototype (likelihood) and the operating condition prior matrix (prior). This mechanism enables the diagnostic model to track the continuous drift in data distribution caused by coupled time-varying operating conditions in real time, fundamentally solving the problem of performance degradation in online diagnostics of traditional static models. It achieves true "deploy once, adapt long-term," significantly improving the accuracy and robustness of the diagnosis.

[0193] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0194] In addition, combined Figure 1 The online fault diagnosis method for crane coupled time-varying operating conditions based on a size model, as described in this embodiment of the invention, can be implemented by a computer device. Figure 8 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Figure 8 As shown, the device may include a processor 301 and a memory 302 storing computer program instructions.

[0195] Specifically, the processor 301 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of the present invention.

[0196] Memory 302 may include a large-capacity memory for data or instructions. For example, and not limitingly, memory 302 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 302 may include removable or non-removable (or fixed) media. Where appropriate, memory 302 may be internal or external to a data processing device. In a particular embodiment, memory 302 is non-volatile memory. In a particular embodiment, memory 302 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0197] The memory 302 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 301.

[0198] The processor 301 reads and executes computer program instructions stored in the memory 302 to implement any of the online fault diagnosis methods for crane coupled time-varying working conditions based on size models in the above embodiments.

[0199] In some embodiments, the computer device may further include a communication interface 303 and a bus 300. For example, Figure 8 As shown, the processor 301, memory 302, and communication interface 303 are connected through bus 300 and complete communication with each other.

[0200] The communication interface 303 is used to enable communication between the various modules, devices, units, and / or equipment in the embodiments of the present invention. The communication interface 303 can also enable data communication with other components such as external devices, image / data acquisition devices, databases, external storage, and image / data processing workstations.

[0201] Bus 300 includes hardware, software, or both, that couples components of a computer device together. Bus 300 includes, but is not limited to, at least one of the following: data bus, address bus, control bus, expansion bus, and local bus. For example, and not as a limitation, bus 300 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 300 may include one or more buses. Although specific buses are described and illustrated in embodiments of the invention, the invention contemplates any suitable bus or interconnect.

[0202] The computer device can execute the online fault diagnosis method for crane coupled time-varying operating conditions based on a size model, as described in this embodiment of the invention, thereby achieving a combination of... Figure 1 This paper describes an online fault diagnosis method for cranes under coupled time-varying operating conditions based on a size model.

[0203] Furthermore, in conjunction with the crane coupled time-varying operating condition online fault diagnosis method based on a size model in the above embodiments, this invention can be implemented using a computer-readable storage medium. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the crane coupled time-varying operating condition online fault diagnosis methods based on a size model in the above embodiments.

[0204] It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. In addition, depending on the implementation needs, the various steps / components described in this invention can be broken down into more steps / components, or two or more steps / components or parts of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0205] It will be readily understood by those skilled in the art that the above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A size model based online fault diagnosis method for time-varying working conditions of crane coupling, characterized in that, The method comprises: The general fault diagnosis large model is pre-trained offline using a multi-source fault data set, so that it learns extensive fault general diagnosis knowledge to extract deep feature vectors of various faults from vibration signals; The fault general diagnosis knowledge of the general fault diagnosis large model is migrated to a lightweight small model through knowledge distillation, and an initial fault knowledge base is generated according to fault categories; the fault knowledge base comprises a fault category prototype set and a working condition prior matrix, and the initial fault category prototype is determined by deep feature vectors of typical samples of each fault category, and the initial working condition prior matrix is a unit matrix of the size of the fault category; The fault feature vector of the crane vibration signal acquired in real time is extracted using the lightweight small model, and a completely back-propagation-free Bayesian self-correction online classifier is combined to dynamically update the fault category prototype set and the working condition prior matrix in the fault knowledge base, so as to complete online fault diagnosis and fault knowledge base updating: The prototype matching probability distribution of the extracted fault feature vector and each fault category prototype in the current fault category prototype set is calculated, the prototype matching probability distribution is multiplied with the current working condition prior matrix to obtain a posterior probability distribution, and the fault category with the highest probability value in the posterior probability distribution is the fault diagnosis result; The fault category prototype set is updated according to the prototype matching probability distribution, and the working condition prior matrix is updated according to the posterior probability distribution; The fault knowledge base updating comprises: (1) updating the fault category prototype set according to the prototype matching probability distribution, comprising: from the prototype matching probability distribution the most probable fault class prototype and its index s ; only when the highest prototype match probability distribution the fault class prototype set is updated only when the highest prototype match probability distribution is greater than or equal to a preset first confidence threshold ； In the formula, Norm represents vector normalization; is a first counter for recording the number of updates of the fault category prototype ; is a fault feature vector of the crane vibration signal ; (2) updating the working condition prior matrix according to the posterior probability distribution, comprising: At the same time, according to the posterior probability distribution updating the working condition prior matrix the first s row: ； In the formula, is the working condition prior matrix is the first row of the working condition prior matrix s ; is the second counter for recording the number of updates of the working condition prior matrix .

2. The size model based online fault diagnosis method for coupled time-varying working conditions of a crane according to claim 1, characterized in that, The general fault diagnosis large model is a general large model based on a Transformer architecture, comprising: Instance normalization: the mean and standard deviation of each channel of each vibration signal sample are calculated respectively, and the mean and standard deviation of each channel are used to normalize all time points in each channel of the sample, so that the normalized vibration signal sample presents zero mean and unit variance within the sample; wherein each channel refers to signals at different measurement points or in different directions of a sensor; Segment embedding: the normalized vibration signal sample is segmented into a plurality of continuous equal-length segments, and each segment is mapped to a high-dimensional feature sequence through a learnable linear embedding layer; Position encoding: a position encoding vector is superimposed on each high-dimensional feature vector to form a high-dimensional feature sequence; the position encoding vector comprises position information of each segment in the vibration signal; Encoder layers: high-dimensional feature sequences are fed into stacked Transformer encoder layers for deep modeling, each of which captures long-range dependencies within each high-dimensional feature sequence through multi-head self-attention mechanisms and feed-forward neural networks and outputs a deep feature vector .

3. The size model based online fault diagnosis method for coupled time-varying working conditions of a crane according to claim 2, characterized in that, A multi-task learning head is constructed for the general fault diagnosis large model, comprising: Two independent learning heads are connected in parallel at the output end of the general fault diagnosis large model to perform different pre-training tasks: (1) Classification head: a fully connected layer followed by a Softmax activation function to map the deep feature vector to a probability distribution over the failure categories ; (2) Projection head: a fully connected layer to map the deep feature vector to a low-dimensional feature for contrastive learning ; The offline pre-training adopts a multi-task mixed learning framework, which trains the general fault diagnosis large model by jointly optimizing a supervised classification task and a self-supervised contrast learning task.

4. The size model based online fault diagnosis method for coupled time-varying working conditions of a crane according to claim 3, characterized in that, The multi-task mixed learning framework comprises: Supervised classification task: the model is trained to distinguish different fault categories through a supervised cross-entropy loss: ； wherein, represents a loss value of supervised learning; N represents a number of samples of a training batch; K represents a total number of failure classes in a multi-source failure data set; denotes a failure class label; denotes a probability that an i th sample belongs to a failure class k . Self-supervised contrast learning task: the vibration signal is data-augmented, and the model is trained to learn a feature representation robust to working condition changes through a contrast loss: ； In the formula, represent the average contrast loss value of the whole batch; is the feature vector of the anchor sample, and is the low-dimensional feature of any sample in a batch; is the feature vector of the positive sample, which belongs to the same vibration signal as the anchor sample and represents the low-dimensional feature of the sample after data enhancement; is the feature vector of the negative sample, which is the low-dimensional feature of all other vibration signal samples in the batch; is the temperature coefficient, which is a positive scalar; sim is a similarity function; Joint training: the pre-training of the general large model for fault diagnosis is completed by weighting combining the loss of the supervised classification task and the self-supervised contrastive learning task.

5. The size model based online fault diagnosis method for coupled time-varying working conditions of a crane according to claim 4, characterized in that, The similarity function adopts cosine similarity, as follows: ； wherein A and B denote two feature vectors.

6. The size model based online fault diagnosis method for coupled time-varying working conditions of a crane according to claim 1, characterized in that, The generation process of the initial fault class prototype set includes: Determining crane fault diagnosis M A fault category; For each fault class, select several typical samples; inputting a typical sample into a fault diagnosis general large model, extracting d a deep feature vector of the dimension Calculate the arithmetic mean of the deep feature vectors of all typical samples belonging to the same fault class to obtain the fault class prototype of the fault class; Will M The fault category prototypes of each fault category are combined column by column to ultimately form Initial fault category prototype set of dimensional U ; The generation process of the working condition prior matrix includes: Construct a identity matrix V , the first i row of the matrix represents the prior probability distribution of various fault categories when the signal feature matches the first i fault category prototype .

7. The size model based online fault diagnosis method for coupled time-varying working conditions of a crane according to claim 1, characterized in that, The fault online diagnosis specifically includes: Extracting real-time acquired crane vibration signals using lightweight small models fault feature vectors ; Computing a fault feature vector Cosine similarity with each of the current set of fault class prototypes M A prototype matching probability distribution of dimension :​ ； wherein M is the total number of crane failure classes; is the failure class prototype of the jth failure class; denotes the failure feature vector is the cosine similarity of the failure feature vector to the failure class prototype m = 1,2,3... M to traverse all failure class prototypes; matching probability distribution to the current operating condition prior matrix matrix multiplication to obtain the posterior probability distribution : ； The fault category with the highest probability value in the posterior probability distribution is the fault diagnosis result of the current crane vibration signal .

8. A computer device, comprising: including: A processor and a memory, the memory storing a program or instructions executable on the processor, the program or instructions being executed by the processor to implement the steps of the crane coupled time-varying working condition online fault diagnosis method based on the size model in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, A processor and a memory, the memory storing a program or instructions executable on the processor, the program or instructions being executed by the processor to implement the steps of the crane coupled time-varying working condition online fault diagnosis method based on the size model in any one of claims 1 to 7.

Images