Anomaly Detection Method Based on a Hybrid Expert Domain Adaptive Industrial Large Model

Abstract

The present invention discloses an anomaly detection method based on a hybrid expert domain - adaptive industrial large - model. By introducing a hybrid expert model and domain - adaptive technology, this method provides an efficient industrial anomaly detection solution for the problems of training efficiency and adaptability of industrial large - models in multi - domain tasks. It includes data pre - processing, domain division and adaptation, industrial knowledge and data fusion, constructing a hybrid expert model, model training and inference adaptation, etc. The hybrid expert model can automatically select a suitable expert for inference according to the input data, while the domain - adaptive module effectively reduces the distribution differences of data in different domains, thereby improving the model's transfer ability and generalization performance. This method is data - driven, can adapt to multi - modal data in different industrial scenarios, and has strong universality. Compared with the prior art, this method has higher training efficiency and better model adaptability in industrial applications, and both its theoretical and practical aspects are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of industrial large model anomaly detection, and particularly to an anomaly detection method based on a hybrid expert domain adaptive industrial large model. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, the industrial field's demand for artificial intelligence (AI) and large-scale models is continuously growing, especially in aspects such as industrial data analysis, fault prediction, quality control, and optimization decision-making. However, traditional industrial models face many challenges in practical applications, including data heterogeneity, knowledge differences between domains, and the cost of large-scale computing. In particular, industrial data usually has the characteristics of a high knowledge proportion and a low total data volume, which is significantly different from the massive data training method relied on by general large models, resulting in problems such as insufficient generalization ability and low computational efficiency in tasks such as industrial large model anomaly detection.

[0003] Existing industrial large models mainly rely on large-scale historical data for training. However, due to the scarcity of industrial data and the difficulty of obtaining high-quality annotations, pure data-driven methods are difficult to achieve ideal results in complex industrial environments. At the same time, industrial knowledge is often explicit and highly specialized, and traditional data-driven models are difficult to fully mine and utilize this knowledge, easily leading to model overfitting or insufficient adaptability. In addition, when industrial large models process multi-modal data (such as images, text, sensor data, etc.), they also face challenges in multi-source data fusion and cross-modal feature alignment. Summary of the Invention

[0004] The purpose of the present invention is to propose an anomaly detection method based on a hybrid expert domain adaptive industrial large model in view of the deficiencies of the prior art.

[0005] The purpose of the present invention is achieved through the following technical solutions: An anomaly detection method based on a hybrid expert domain adaptive industrial large model, comprising the following steps:

[0006] S1. Collect multi-modal data from different domains in the industrial environment and perform preprocessing according to the data type;

[0007] S2. Project the preprocessed data from different domains into a unified feature space, use a pre-trained domain discriminator for feature alignment, and use a pre-trained domain autoencoder for feature transformation to make the distributions of data from different domains closer;

[0008] S3. Integrate industrial knowledge with the processed multi-modal industrial data through explicit industrial knowledge;

[0009] S4. Construct a mixture-of-experts model and train it. The mixture-of-experts model consists of several mixture-of-experts modules. Each mixture-of-experts module includes a linear projection layer, a self-attention layer, a gating network, a mixture-of-experts layer, and an output layer. The mixture-of-experts layer is an expert layer for different data processing methods.

[0010] S5. Optimize the intelligent computing platform and make it hardware-compatible, and use the optimized mixture-of-experts model to perform anomaly detection inference on the data after multi-modal industrial data fusion.

[0011] Further, the multi-modal data includes one or more of text records, industrial images, video streams, and sensor data.

[0012] Further, the preprocessing according to the data type is specifically as follows:

[0013] For time series data, use the sliding window method to slice it, and use Fourier transform or wavelet transform to extract time-frequency features.

[0014] For text data, perform word segmentation, part-of-speech tagging, and use the word vector model for vectorization.

[0015] For image data, use a convolutional neural network to extract visual features, and use data augmentation technology to improve data diversity.

[0016] Further, the pre-training process of the domain discriminator specifically includes: making the domain discriminator unable to distinguish the data distributions of the source domain and the target domain through adversarial training, and using a gradient reversal layer to optimize the loss function.

[0017] Further, the domain autoencoder performs joint training in multiple domains during the pre-training process.

[0018] Further, the fusion of industrial knowledge and the processed multi-modal industrial data through explicit industrial knowledge is specifically as follows:

[0019] S3.1. Extract explicit industrial knowledge from the knowledge base in the industrial field, transform it into structured data, rule sets, or construct an industrial knowledge graph.

[0020] S3.2. Use multi-modal encoders to extract features from different data types respectively, and map the data of multiple modalities to a unified feature space.

[0021] S3.3. By constructing a context learning architecture based on the domain knowledge graph, perform context learning in combination with the domain knowledge graph.

[0022] Further, in the mixture-of-experts model, the expert layers for different data processing methods specifically include the following experts:

[0023] Fully-connected expert: includes one or more fully-connected layers

[0024] Convolutional expert: includes one or more convolutional layers

[0025] Transformer expert: includes a Transformer architecture with an attention mechanism

[0026] Zero expert: directly discards the input;

[0027] Copy expert: skips the current mixture-of-experts layer;

[0028] Constant expert: performs constant correction and replacement on the input;

[0029] The expert layers in each mixture-of-experts module include four of them.

[0030] Furthermore, the gated neural network scores the input features, calculates the activation probabilities of each expert, and selects the top-2 activated experts. The activation probability is specifically:

[0031]

[0032] On the other hand, the specification also provides an anomaly detection device based on a mixture-of-experts domain-adaptive industrial large model, including a memory and one or more processors. An executable code is stored in the memory. When the processor executes the executable code, it implements the above-mentioned anomaly detection method based on a mixture-of-experts domain-adaptive industrial large model.

[0033] On the other hand, the specification also provides a computer-readable storage medium with a program stored thereon. When the program is executed by a processor, it implements the above-mentioned anomaly detection method based on a mixture-of-experts domain-adaptive industrial large model.

[0034] Advantages of the present invention:

[0035] In the first aspect, the present invention first fuses explicit industrial knowledge and multi-modal industrial data through a context learning architecture. This architecture is based on consecutive training stages of pre-training and context learning inference (i.e., model warm-up). On the basis of improving the context learning ability of the large model, it further optimizes the knowledge expression of the model. By combining explicit industrial knowledge (such as industry norms, operation manuals, etc.) with multi-modal data in the industrial field (such as images, texts, sensor data, etc.), the overfitting phenomenon in data-driven models is significantly reduced.

[0036] In a second aspect, the present invention proposes an innovative Mixture of Experts (MoE) model. By introducing multiple expert networks and a dynamic selection mechanism, the model can dynamically activate appropriate expert networks for reasoning when facing various industrial tasks, thereby improving computational efficiency and task adaptability. In the MoE architecture, by designing multiple types of neural network experts, the operation efficiency is improved, the computational overhead is reduced, and the reasoning efficiency is enhanced. In a third aspect, in order to achieve efficient training and reasoning on different intelligent computing platforms, the present invention optimizes the training and reasoning processes of the large model on domestic intelligent computing platforms. On the one hand, the present invention can be adapted to domestic chips (such as Ascend, Kunlun) and deep learning frameworks (such as MindSpore, PaddlePaddle), and optimize compilation options and computing libraries. For example, on Ascend chips, the CANN operator library is used to optimize matrix calculations to reduce computational latency. On the other hand, the present invention adjusts the underlying operators and adopts means such as operator fusion (fusing BatchNorm and Conv to reduce memory access and accelerate convolutional calculations in image anomaly detection tasks), precision optimization (using BF16 / INT8 to accelerate reasoning and adopting low-precision calculations to improve reasoning speed in time series anomaly detection tasks), and computational parallelization (adopting data parallel and model parallel strategies to improve computational efficiency in a multi-card training environment). The industrial large model of the present invention can perform efficient multi-task parallel training and reasoning on domestic platforms with high-performance computing clusters and large-scale data storage.

[0037] In a fourth aspect, the model of the present invention can effectively migrate from one industrial field to another through domain adaptation technology. This method jointly trains data from multiple domains and utilizes shared knowledge and cross-domain learning to enhance the adaptability and generalization ability of the model in unknown domains. Specifically, when processing different industrial scenarios, the model can dynamically adjust the activation mode of its expert network, thereby improving the processing ability for different tasks and data. In a fifth aspect, the method of the present invention adopts a strategy combining large-scale pre-training and fine-tuning. First, the large model is pre-trained through self-supervised learning or generative models to learn the general features of the data. Then, using domain adaptation technology, the pre-trained model is applied to specific industrial tasks, such as equipment fault diagnosis, production line optimization, etc., to further improve the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of the method for anomaly detection of the domain-adaptive industrial large model based on the mixture of experts implemented by the present invention;

[0039] Figure 2 is the functional architecture specifically included in each module of the present invention;

[0040] Figure 3It is the structure diagram of the mixture-of-experts model for industrial large models implemented by the present invention;

[0041] Figure 4 It is the schematic diagram of the structure of the mixture-of-experts layer (MoE Layer) in the mixture-of-experts model of the present invention;

[0042] Figure 5 It is the flowchart of the mixture-of-experts model of the present invention for processing data;

[0043] Figure 6 They are various experts involved in the mixture-of-experts model of the present invention;

[0044] Figure 7 It is the schematic diagram of an anomaly detection device based on a mixture-of-experts domain adaptive industrial large model provided by an embodiment of the present invention. Specific Embodiments

[0045] The following further elaborates on the specific embodiments of the present invention with reference to the accompanying drawings.

[0046] As Figure 1 shown, the present invention provides an anomaly detection method based on a mixture-of-experts domain adaptive industrial large model, and the specific implementation method is as follows:

[0047] Step 1: Collect multimodal data in an industrial environment, including but not limited to text records (text data such as production logs, maintenance reports, fault descriptions, etc.), industrial images (image data such as equipment operation status, product appearance inspection, etc.), video streams (image frame data such as production line monitoring videos, equipment operation processes, etc.), and sensor data (time series data such as temperature, pressure, vibration, etc.).

[0048] Step 2: Clean and standardize the collected data, including denoising, filling in missing values, anomaly detection, and removing redundant data. For data of different modalities, this method has different processing methods.

[0049] For time series data, the present invention uses a sliding window method for slicing and extracts time-frequency features using Fourier transform or wavelet transform;

[0050] For text data, the present invention performs word segmentation, part-of-speech tagging, and vectorization using a word vector model (such as Word2Vec or BERT);

[0051] For image data, the present invention uses a convolutional neural network (CNN) to extract visual features and uses data augmentation techniques to enhance data diversity.

[0052] Step 3: Project the preprocessed data into a unified feature space and use domain adaptation techniques to reduce the differences in data distributions between different domains. The domain discriminator network structure can adopt a fully connected neural network (FCNN), and adversarial training is carried out through the input features and domain labels; the domain autoencoder can use a convolutional neural network (CNN) or a multi-layer perceptron (MLP), and the domain features of the input data are effectively extracted by minimizing the reconstruction error. Taking the semiconductor manufacturing and petrochemical industries as two examples, it can specifically include the following sub-steps:

[0053] (1) Map the industrial data in different domains to a unified feature space to make it easier to align in the high-dimensional space. Use self-supervised learning methods, such as contrastive learning or variational autoencoder (VAE), to enhance the feature expression ability.

[0054] For example, during training, pair the data of semiconductor factories (source domain) and petrochemical factories (target domain), compare the feature distances between normal samples and abnormal samples, and ensure that normal samples under normal process conditions have similar feature vectors.

[0055] (2) Introduce a domain discriminator and adopt the adversarial training method to make the domain discriminator unable to distinguish the data distributions of the source domain and the target domain, thereby achieving feature alignment. Specifically, use a gradient reversal layer (GRL) to optimize the loss function, so that the domain discriminator can simultaneously minimize the task loss and maximize the domain discrimination loss.

[0056] For example, construct a domain discriminator , and judge its source (semiconductor factory or petrochemical factory) after inputting the features. Through gradient reversal (GRL), make the feature distributions as close as possible during the training process, weaken the domain differences, and its loss function formula can be as follows:

[0057]

[0058] Among them, is the feature of the input data, is the predicted probability of the domain classifier.

[0059] (3) Adopt a domain autoencoder for feature transformation to make the data distributions in different domains closer. In the pre-training stage, use multi-source domain data for joint training to enhance the cross-domain generalization ability of the domain autoencoder.

[0060] For example, when training a Domain Autoencoder, the data from semiconductor factories and petrochemical factories are reconstructed in the same feature space. During the pre-training phase, joint training is carried out to ensure cross-domain generalization ability. In the inference phase, the model can automatically adjust the features of different factories to make them closer to the same distribution.

[0061] Simply put, this step realizes the alignment of feature distributions by constructing the adversarial loss between the source domain and the target domain. At the same time, the domain transfer learning method is adopted. Based on retaining domain characteristics and enhancing feature expression ability, through cross-domain joint training and knowledge sharing mechanisms, the generalization ability of the large model for data in different domains is improved.

[0062] Step 4: Extract explicit industrial knowledge (such as rules, specifications, fault modes, etc.) from the knowledge base in the industrial field, and integrate the industrial knowledge with multi-modal industrial data (text, images, etc.) to improve the model's understanding and application ability for industrial scenarios. This process includes multiple sub-steps to ensure the seamless integration of knowledge and data, and the specific implementation methods are as follows:

[0063] (1) First of all, explicit industrial knowledge (such as rules, specifications, fault modes, etc.) needs to be extracted from the knowledge base in the industrial field. Common sources of industrial knowledge include industrial standards and specifications (such as equipment operation standards, fault diagnosis specifications, maintenance manuals), failure mode and effect analysis (this method can help identify and analyze potential equipment failure modes and their possible impacts), expert experience (qualitatively analyze the operation of equipment through experts' knowledge and intuition to identify early warning signals of equipment failures), knowledge graphs (structuring industrial knowledge and representing it in the form of graphs, enabling the system to obtain knowledge through querying, reasoning, etc.), etc.

[0064] When converting this knowledge into a form that can be processed by the model, rules such as fault diagnosis and maintenance processes can be converted into structured data or rule sets, enabling the system to be directly applied to anomaly detection or prediction tasks. Or, by constructing an industrial knowledge graph, elements such as equipment, faults, maintenance, and operations are presented in the form of graph nodes, and reasoning and querying are carried out through the connection relationships of the graph.

[0065] (2) In this stage, the goal is to map data of multiple modalities such as text, images, and sensor data into a unified feature space, enabling different modalities of data to interact with each other and extract cross-modal information. This process includes using multi-modal encoders (such as variants of Transformer, BERT, etc.) to extract features from different data types respectively. For example, using CNN to extract features of image data, using RNN or Transformer to process time series data, and using BERT to process text data. These encoders convert their respective features into a unified vector representation. To enhance the spatio-temporal correlation of multi-modal data, relative position encoding is used to adjust the relationship between different modalities. For example, in the fusion of image and text data, one-dimensional or two-dimensional position encoding can be used to optimize the representation of the feature sequence, ensuring effective spatial alignment between the image and the text.

[0066] (3) In the process of fusing multi-modal data, it is crucial to combine domain knowledge graphs for context learning of data. In the data fusion process, by constructing a context learning architecture based on the domain knowledge graph, a Transformer with a cross-attention mechanism is used to interpret and infer abnormal patterns in the data according to known industrial standards and knowledge backgrounds.

[0067] For example, in the scenario of fault prediction and diagnosis of CNC (Computer Numerical Control) equipment in a factory in the intelligent manufacturing environment. Through learning explicit knowledge such as industrial standards (GB / T 20759-2020 "Reliability Evaluation of Computer Numerical Control Machine Tools") and expert experience, the model can infer abnormal types. For example, when the spindle vibration frequency exceeds 5 kHz and the temperature rises by more than 10 °C, it may be that the spindle bearing is worn. When the abnormal increase in motor power consumption (exceeds 20% of the normal value), it may be that the tool is stuck or the spindle is overloaded.

[0068] Step 5: Design a mixture of experts model to complete the core ability training of the large model through an efficient mixture of experts architecture.

[0069] In this step, referring to Figure 3 the overall architecture diagram of the mixture of experts model and Figure 4 the schematic diagram of the MoE layer structure in Figure 5 , the mixture of experts model consists of N MoE Blocks. The training and inference processes of this model can refer to

[0070] (1) Input data , where is the input sequence length, is the batch size, is the feature dimension. For the input feature Perform a query (Query, ), key (Key, ), and value (Value, ) for linear projection to generate the basic input for the attention mechanism: , , , where , , is the projection weight matrix.

[0071] (2) The features are passed to the attention mechanism in parallel, where the attention mechanism is used to calculate the global attention features, and the calculation formula is , where is the attention space dimension. This step uses multiple attention heads to calculate in parallel, and the output feature dimension remains unchanged, and the global features are extracted to support cross-domain shared knowledge.

[0072] (3) The output result of the attention mechanism is passed into the routing mechanism, and the routing mechanism dynamically selects the top-2 activated experts, that is, uses the gating network (Gating Network) to score the input features and calculate the activation probability of each expert . Where is the full-text feature representation obtained by the attention mechanism in (2), is the learnable parameter matrix of the gating network, is the bias term of the gating network.

[0073] (4) The expert head processes the domain-specific features, and activates the most relevant experts according to , avoiding all experts from participating in the calculation to improve efficiency. The activated experts receive input and perform feature extraction respectively.

[0074] (5) Perform weighted fusion on the outputs of the shared head and the expert head, and fuse the attention mechanism output and the expert head output by weight , where , is the fusion weight parameter.

[0075] (6) Generate the final output through a feed-forward neural network.

[0076] The expert-based neural network structure included in this step can refer to Figure 6 , specifically as follows:

[0077] Fully Connected Expert: It contains one or more fully connected layers and is suitable for processing low-dimensional, unstructured feature inputs such as sequence embeddings or flattened image features;

[0078] Convolutional Experts: It contains one or more convolutional layers and is applicable to processing data with local correlations such as images, time series, speech signals, etc.;

[0079] Transformer Experts: It contains the Transformer architecture with attention mechanism, emphasizing the modeling of long-range dependencies and global semantic information;

[0080] Zero Expert: It directly discards the input;

[0081] Copy Expert: It skips the current mixture-of-experts layer;

[0082] Constant Expert: It performs constant correction and replacement on the input.

[0083] (7) Loss function: Cross-Entropy Loss for the anomaly detection task is used as the main task loss; to prevent some experts from being overused, Load Balancing Loss (negative entropy loss) is adopted as the expert regularization loss to encourage the balanced use of different experts:

[0084]

[0085] For the routing mechanism, to ensure the sparse activation of experts, L1 regularization loss is usually added as the gating loss:

[0086]

[0087] where are the weights of the router.

[0088] (8) Backpropagation and optimization: Calculate the gradients of the loss function and use optimization algorithms (such as Adam or SGD) for parameter updates.

[0089] The multiple Experts designed in the present invention expand the combined mode of the model to process inputs, expanding from the original single mode to multiple modes, thereby enhancing the model's expressive ability and task processing flexibility. All inputs will be processed by two experts together, so different data processing methods can be applied to different types of data, thus enhancing the model's expressive ability and task processing flexibility. In addition, three experts that do not require computational cost, namely the Constant Expert, Zero Experts, and Copy Experts, can significantly reduce the computational overhead, achieving the purpose of reducing the consumption of computing resources.

[0090] Step 6: For the intelligent computing platform, optimize the training architecture and underlying operators. After achieving efficient model training and inference, use the model with optimized operators for industrial anomaly detection to obtain the final anomaly detection results.

[0091] This step mainly optimizes the training and inference processes of the large model, including operator tuning, distributed training, and hardware compatibility design, to improve the performance of the model on the local intelligent computing platform. In the training stage, efficient processing of multi-modal data is achieved by optimizing the underlying operators, supporting multi-task parallel training and effectively improving the model performance. In the inference stage, combined with model optimization techniques such as pruning and quantization, the computational complexity is reduced and the inference efficiency is improved.

[0092] According to the above method embodiments, the specification of the present invention also provides an embodiment of an anomaly detection system based on a hybrid expert domain adaptive industrial large model, as Figure 2 shown. The system includes a data preprocessing module, a domain adaptation module, an industrial knowledge and data fusion module, and a hybrid expert model;

[0093] The data preprocessing module preprocesses the multi-modal data collected from different domains according to the data type;

[0094] The domain adaptation module projects the preprocessed data from different domains into a unified feature space, uses a pre-trained domain discriminator for feature alignment, and uses a pre-trained domain autoencoder for feature transformation to make the distributions of data from different domains closer;

[0095] The industrial knowledge and data fusion module fuses industrial knowledge with the processed multi-modal industrial data through explicit industrial knowledge;

[0096] The hybrid expert model is composed of several hybrid expert modules. Each hybrid expert module includes a linear projection layer, a self-attention layer, a gating network, a hybrid expert layer, and an output layer. The hybrid expert layer is an expert layer for different data processing methods.

[0097] The specific implementation process of the present invention is as follows: Build and deploy the above system. After the data preprocessing module completes the formatting and feature extraction of the input data, the domain adaptation module calibrates the feature distribution, the fusion module conducts joint modeling of data and knowledge, and finally completes the training of the large model in the hybrid attention expert module. In the model inference stage, a real-time and efficient inference process is achieved through the optimized intelligent computing platform to meet the application requirements of anomaly detection in industrial sites. For the remaining detailed steps, please refer to the implementation process of the corresponding steps in the above method, which will not be elaborated here.

[0098] Corresponding to the embodiment of the foregoing anomaly detection method based on a domain-adaptive industrial large model with hybrid experts, the present invention also provides an embodiment of an anomaly detection device based on a domain-adaptive industrial large model with hybrid experts.

[0099] See Figure 7 , an anomaly detection device based on a domain-adaptive industrial large model with hybrid experts provided by an embodiment of the present invention includes a memory and one or more processors. An executable code is stored in the memory. When the processor executes the executable code, it is used to implement the anomaly detection method based on a domain-adaptive industrial large model with hybrid experts in the above embodiment.

[0100] The embodiment of the anomaly detection device based on a domain-adaptive industrial large model with hybrid experts provided by the present invention can be applied to any device with data processing capabilities. Such a device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by the processor of any device with data processing capabilities reading the corresponding computer program instructions in the non-volatile memory into the memory for operation. From the hardware level, as Figure 7 shown, it is a hardware structure diagram of any device with data processing capabilities where the anomaly detection device based on a domain-adaptive industrial large model with hybrid experts provided by the present invention is located. In addition to Figure 7 the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities where the device in the embodiment is located usually also includes other hardware according to the actual functions of the device with data processing capabilities, which will not be elaborated here.

[0101] The implementation process of the functions and roles of each unit in the above device is specifically detailed in the implementation process of the corresponding steps in the above method, which will not be elaborated here.

[0102] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0103] An embodiment of the present invention also provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements an anomaly detection method based on a hybrid expert domain adaptive industrial large model in the above embodiments.

[0104] The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium may also be an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Further, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store the data that has been output or will be output.

[0105] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the anomaly detection method based on a hybrid expert domain adaptive industrial large model described above.

[0106] After considering the specification and practicing the content disclosed herein, those skilled in the art will readily think of other embodiments of the present application. The present application aims to cover any variations, uses, or adaptive changes of the present application, which follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of the present application are pointed out by the claims.

[0107] It should be understood that the foregoing general description and the following detailed description are merely exemplary and explanatory, and do not limit this application. This application is not limited to the exact structures that have been described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An anomaly detection method based on a hybrid expert domain - adaptive industrial large - model, characterized in that, It includes the following steps: S1. Collect multi-modal data from different fields in the industrial environment and perform preprocessing according to the data type; S2. Project the preprocessed data from different fields into a unified feature space, use a pre-trained domain discriminator for feature alignment, and use a pre-trained domain autoencoder for feature transformation to make the distributions of data from different fields closer; S3. Integrate industrial knowledge with the processed multi-modal industrial data through explicit industrial knowledge; specifically: S3.

1. Extract explicit industrial knowledge from the knowledge base of the industrial field, transform it into structured data, rule sets or construct an industrial knowledge graph; S3.

2. Use a multi-modal encoder to extract features from different data types respectively, and map data of multiple modalities into a unified feature space; S3.

3. Construct a context learning architecture based on the domain knowledge graph and perform context learning in combination with the domain knowledge graph; S4. Construct and train a mixture of experts model, which consists of several mixture of experts modules. Each mixture of experts module includes a linear projection layer, a self-attention layer, a gating network, a mixture of experts layer and an output layer. The mixture of experts layer is an expert layer for different data processing methods; S5. Optimize and make hardware compatible with the intelligent computing platform, and use the optimized mixture of experts model to perform anomaly detection inference on the data after multi-modal industrial data integration.

2. An anomaly detection method based on a hybrid expert domain adaptive industrial large model according to claim 1, characterized in that, The multi-modal data includes one or more of the following: text records, industrial images, video streams, sensor data.

3. An anomaly detection method based on a hybrid expert domain adaptive industrial large model according to claim 1, characterized in that The specific preprocessing according to the data type is as follows: For time series data, use the sliding window method for slicing and extract time-frequency features using Fourier transform or wavelet transform; For text data, perform word segmentation, part-of-speech tagging, and vectorize using a word vector model; For image data, use a convolutional neural network to extract visual features and use data augmentation techniques to improve data diversity.

4. An anomaly detection method based on a hybrid expert domain adaptive industrial large model according to claim 1, characterized in that, The domain discriminator is a fully connected neural network. The specific pre-training process includes: making the domain discriminator unable to distinguish the data distributions of the source domain and the target domain through adversarial training, and using a gradient reversal layer to optimize the loss function.

5. The anomaly detection method based on a hybrid expert domain adaptive industrial large model according to claim 1, wherein, The domain autoencoder uses a convolutional neural network or a multi-layer perceptron, and performs joint training across multiple domains during the pre-training process. By minimizing the reconstruction error, it ensures that the domain features of the input data are effectively extracted.

6. The anomaly detection method based on a hybrid expert domain adaptive industrial large model according to claim 1, characterized in that, In the mixture of experts model, the expert layer for different data processing methods specifically includes the following experts: Fully connected expert: contains one or more fully connected layers; Convolutional expert: contains one or more convolutional layers; Transformer expert: contains a Transformer architecture with an attention mechanism; Zero expert: directly discards the input; Copy expert: skips the current mixture of experts layer; Constant expert: performs constant correction and replacement on the input; The expert layer in each mixture of experts module includes four of them.

7. An anomaly detection method based on a hybrid expert domain adaptive industrial large model according to claim 6, characterized in that, In the training of the mixture-of-experts model, the loss function is specifically as follows: The cross-entropy loss of the anomaly detection task is used as the main task loss, and the negative entropy loss is used as the expert regularization loss to encourage the balanced use of different experts. For the routing mechanism, the L1 regularization loss is added as the gating loss.

8. An anomaly detection device based on a hybrid expert domain adaptive industrial large model, comprising a memory and one or more processors, wherein executable code is stored in the memory, characterized in that, When the processor executes the executable code, it implements an anomaly detection method based on a mixture-of-experts domain adaptive industrial large model as described in any one of claims 1-7.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements an anomaly detection method based on a mixture-of-experts domain adaptive industrial large model as described in any one of claims 1-7.

Images