Combine harvester fault diagnosis high-precision entity name identification method

Abstract

The combine harvester has a wide application range. In order to quickly find out related fault entities and solve the mechanical faults, the invention provides a high-precision entity name identification method RP-TEBC (RoBERTa-wwm-ext + PGD + Transform-Encoder + BiGRU + CRF) for fault diagnosis of a combine harvester, which is used for solving the problem that mechanical faults inevitably occur as mechanical equipment, and aims to quickly find out related fault entities and solve the mechanical faults. In the RP-TEBC, a dynamically coded RoBERTa-wwm-ext pre-training model is used as a word embedding layer, an adaptive Transform encoder layer is fused with a bidirectional gating unit (BiGRU) as a context encoder, a conditional random field (CRF) is used as a decoding layer, and an optimal path output is found by using a Viterbi algorithm. Meanwhile, the RP-TEBC model generates an adversarial sample by adding some disturbances in a word embedding layer, and the overall robustness and generalization performance of the model can be improved by continuously training and optimizing the model. The result shows that compared with a traditional BERT + BiLSTM + CRF model, the accuracy rate, the recall rate and the F1 value of the model are improved by 1.79%, 1.01% and 1.46% respectively on a constructed combine harvester fault domain named entity recognition data set.

Description

Technical Field

[0001] The present invention belongs to the field of named entity recognition algorithms, and in particular relates to a high-precision entity naming recognition method for combine harvester fault diagnosis. Background Art

[0002] In recent years, with the continuous development of agricultural mechanization in my country, combine harvesters have been mass-produced and widely used as an effective crop harvesting tool. The amount of unstructured text data related to combine harvester maintenance is also growing. The challenge is to efficiently and accurately retrieve the required information from this complex data.

[0003] Named entity recognition (NER), a fundamental task in natural language processing (NLP) information extraction, essentially involves finding the start and end of an entity in a sentence and assigning it a category. With the continuous innovation and development of deep learning technology, the task of named entity recognition has evolved from early rule-based and machine learning methods to deep learning methods, achieving remarkable results.

[0004] In the field of Chinese named entity recognition, Chinese sentences are more difficult to identify because words are connected together, unlike English, which has natural space delimiters. Therefore, the primary task of Chinese named entity recognition is to first segment words and then apply the word-level named entity recognition model to the segmented sentences. However, even the best word segmentation tools can make mistakes, which can lead to errors in entity boundary detection and category prediction during named entity recognition model training, thus affecting the overall effectiveness of the model.

[0005] The meaning of Chinese words often depends on context. The same word can represent different named entity types in different contexts, or it can be both a common word and a proper noun, increasing the complexity of recognition. Some long entities may contain nested entities. For example, in "Peking University Third Hospital," "Peking University" itself is also a named entity. Correctly parsing these nested structures requires more complex models and algorithms.

[0006] Currently, most named entity models are designed for general fields. There are few studies on named entity recognition models for specific fields. There is no report on named entity recognition in the field of combine harvester failures. Summary of the Invention

[0007] Based on the status of the above existing named entity recognition methods, the present invention provides an adaptive Transformer-based named entity recognition model for combine harvester fault diagnosis to overcome at least one technical problem existing in the prior art.

[0008] To achieve the above objectives, the present invention constructs a named entity recognition dataset for combine harvester fault diagnosis, including: This paper collects about 200,000 words of unstructured text in the field of combine harvester faults through crawlers, PDF text conversion and other technical means; and uses the YEDDA annotation tool to annotate the original corpus. The marked entities are divided into four categories, namely fault name (Fault), fault cause (Cause), fault location (Position) and fault repair (Repaies). These four types of entities are annotated in the form of BIO, "B" represents the starting part of the entity, "I" represents the middle and end part of the entity, and "O" represents not an entity;

[0009] To address the above issues, this paper proposes a PR-TEBC named entity recognition model. The model structure is mainly divided into four parts: using the RoBERTa-wwm-ext pre-trained model as the word embedding layer; adding adversarial training to each embedded word vector to generate adversarial samples to improve the robustness and generalization of the model; using an adaptive Transformer encoder fused with a bidirectional gated recurrent unit (BiGRU) as a context encoder; and finally using a conditional random field (CRF) as the output layer to predict labels, thereby obtaining entities in the input text: Word embedding layer, BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer architecture, which has achieved remarkable results in the field of natural language processing (NLP). BERT's word embeddings are one of its core components, which are used to convert words or subwords in a text into high-dimensional vector representations. These vectors can capture the semantic and contextual information of the words. BERT's word embeddings have the following characteristics: (1) Bidirectionality: Unlike traditional unidirectional language models from left to right or right to left, BERT is bidirectional. This means that when generating the embedding of a word, BERT considers both the left and right contexts of the word. This enables BERT to capture the semantic information of the word more accurately. (2) Word-based representation: BERT uses the WordPiece algorithm for word segmentation, breaking down words into smaller subword units. This subword-based representation method can handle out-of-vocabulary words and spelling errors because even if a complete word does not appear in the training set, its subword units may have appeared in the training set. (3) Context-sensitive: BERT's word embeddings are context-sensitive, that is, the same word will have different embedding representations in different contexts. This is achieved through BERT's pre-training process, in which the model is trained on a large amount of text data to learn how to generate appropriate embedding representations based on the context. (4) High-dimensional vector representation: BERT's word embeddings are usually high-dimensional vectors (such as 768 dimensions, 1024 dimensions, etc.), which can capture the complex semantics and contextual information of words. These vectors can be compared using metrics such as cosine similarity to measure the semantic similarity between words; The RoBERTa pre-training model is an improvement on the BERT model. The RoBERTa pre-training model uses a dynamic masking strategy and more diverse training data, and eliminates the NSP during the training process. Through these improvements, the performance of the RoBERTa pre-training model is better than that of the BERT pre-training model. RoBERTa-wwm-ext is a new pre-training model obtained by retraining RoBERTa using a Chinese dataset based on the whole word masking (WWM) strategy proposed by Cui et al.

[14] . The whole word masking strategy is to perform masking language training on words. Because Chinese words are composed of multiple characters, directly masking a single character may cause the loss of semantic information. By using the whole word masking strategy, the complete semantics of Chinese words can be better captured, thereby improving the performance of the model in Chinese natural language processing tasks. The context encoder layer is mainly composed of an adaptive Transformer encoder and a bidirectional gated recurrent unit (BiGRU). The Transformer encoder is different from the traditional recurrent neural network (RNN) type encoder. It uses a multi-head attention mechanism to better learn the context information of the input text, solving the problems of gradient vanishing and gradient explosion that exist in traditional recurrent neural networks when learning longer input texts. However, the performance of the traditional Transformer encoder in the named entity recognition task is far inferior to its performance in other natural language processing tasks. Therefore, this paper uses the adaptive Transformer encoder proposed by Yan et al.

[15] to encode the character-level features and word-level features of the input text for the named entity recognition task. The sinusoidal position embedding in the original Transformer encoder is more sensitive to distance perception, but not very sensitive to directionality perception. To address this problem, the adaptive Transformer encoder adopts relative position encoding and improves the relative position encoding. The improved relative position encoding not only uses fewer parameters but also has better performance; Decoder layer, CRF (Conditional Random Field) is often used as a decoder in Named Entity Recognition (NER). In named entity recognition, CRF models are often used to treat entity recognition in text as a sequence labeling problem. Specifically, given an input sequence (i.e., the text to be processed), the CRF model outputs a sequence of tags, where each tag corresponds to a word or subword in the input sequence and indicates whether the word or subword belongs to a certain named entity category; its main advantage is that it can use the contextual information of the input sequence to make predictions. Specifically, the model considers the entire input sequence and previously annotated entity information to predict the tag at the current position. This enables the model to more accurately identify named entities in the text, especially for those entities that have specific meanings or structures in the context. Adversarial training, PGD (Projected Gradient Descent) adversarial training is a training method to improve the robustness and generalization ability of the model. Its core idea is to introduce adversarial samples during the training process, that is, to generate new samples by applying small perturbations to the input samples. These new samples are intended to increase the loss of the model. Then, these adversarial samples are used together with the original samples to train the model to improve the model's defense against adversarial samples and its generalization ability to the original samples. The specific steps of PGD adversarial training include: (1) Generate adversarial samples: Generally, the PGD algorithm is used to generate adversarial samples. Specifically, by applying small perturbations to the input samples and iteratively updating in the direction of increasing the model loss, the adversarial samples are finally obtained. (2) Train the model: Use the generated adversarial samples and the original samples to train the model. During the training process, the model needs to learn how to distinguish between real samples and adversarial samples, thereby improving its defense against adversarial samples (3) Repeat the iteration: Repeat the above two steps to generate new adversarial samples and train the model to continuously improve the robustness and generalization ability of the model;

[0010] To solve the above problems and enable users to use the named entity recognition model for combine harvester fault diagnosis, a user-oriented named entity recognition system for combine harvester fault diagnosis was invented, which includes: The data acquisition module is responsible for collecting the working status data of each combine harvester component in real time or periodically, including but not limited to sensor data (such as vibration, temperature, pressure, flow, etc.), control unit signals, operation logs, etc. This module may need to integrate the CAN bus communication protocol to obtain data; The preprocessing module cleans, formats, and standardizes the received raw data to remove noise and fill in missing values for subsequent analysis. This step is crucial for improving the accuracy of fault identification. The feature extraction module extracts key features that are helpful for fault diagnosis from the preprocessed data. These features may include vibration amplitude at a specific frequency, temperature change rate, operating current anomalies, etc., which help distinguish different types of fault modes; Named Entity Recognition (NER) module: This module is specifically used to identify key entities in fault descriptions, such as the fault component name, fault type, and fault severity. It is usually implemented using natural language processing techniques such as part-of-speech tagging, sequence tagging, or deep learning models (such as BERT and LSTM-CRF); The fault diagnosis module uses machine learning algorithms (such as support vector machines (SVMs), random forests, neural networks), fuzzy logic, rule engines, or hybrid methods to classify and diagnose faults based on the extracted features and identified named entities. This module may include a model-based knowledge base for storing and querying known fault cases and their solutions. The decision support and recommendation module provides operators or maintenance personnel with intuitive fault reporting and diagnostic results, including graphical interfaces, alarm prompts, and mobile application notifications, making it easier for them to understand and take action. The data storage and analysis module stores fault diagnosis data over a long period of time for subsequent fault mode learning, performance evaluation, and system optimization. This also helps build fault prediction models and provide early warning of potential problems. The remote communication module allows remote monitoring and diagnosis, and transmits field data to a central server or expert system via a wireless network, enabling remote technical support and software updates.

[0011] In order to solve the above problem, the present invention further provides an electronic device, the electronic device comprising: At least one processor, a memory, and a computer program stored in the memory and executable on the processor, so that when the processor executes the program, the steps of the user-oriented named entity recognition system for combined harvester fault diagnosis as described above can be implemented.

[0012] In order to solve the above problems, the present invention also provides a computer-readable storage medium, a computer-readable storage medium, and at least one or more programs in the computer-readable storage medium. When the program is executed by the at least one processor, it implements a user-oriented combine harvester fault diagnosis named entity recognition system as described above.

[0013] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages:

[0014] This invention provides a user-friendly named entity recognition system for combine harvester fault diagnosis. The system comprises the following core components: First, the system uses a data acquisition module to collect a large amount of combine harvester operating sensor data, maintenance records, and fault reports from actual operations. A preprocessing module then cleans and standardizes this raw data. Next, a feature extraction module filters and extracts key features useful for fault diagnosis. Based on these features, a model training module employs advanced machine learning or deep learning algorithms to construct a named entity recognition model. This model accurately identifies key entities in fault descriptions, such as the faulty component, fault type, and possible cause. Finally, through a user-friendly diagnostic interface, the system presents complex fault diagnosis information to operators or maintenance technicians in an intuitive and understandable manner. This significantly improves the efficiency and accuracy of fault diagnosis, reduces downtime caused by misdiagnosis, and ensures the smooth progress of agricultural production activities. Results show that, on a constructed named entity recognition dataset for combine harvester faults, the proposed model achieves 1.79% higher precision, 1.01% higher recall, and 1.46% higher F1-score than the traditional BERT+BiLSTM+CRF model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more intuitively illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required in the embodiments. The drawings are only used to intuitively explain the present invention and are not considered to limit the present invention. Figure 1 The overall architecture diagram of the RP-TECB model is shown; Figure 2 A schematic diagram of a named entity recognition system for combine harvester fault diagnosis in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0016] Exemplary embodiments of the present disclosure will be described below in more detail with reference to the accompanying drawings.

[0017] Based on the above problems in the prior art, the present invention mainly provides a named entity recognition system for combine harvester fault diagnosis, which aims to solve the problem in the prior art of how to efficiently and accurately retrieve the required information from these complex data.

[0018] This study collected about 130,000 words of unstructured text in the field of combine harvester failures through crawlers, PDF text conversion and other technical means. The original corpus was annotated using the YEDDA[8] annotation tool. The marked entities were divided into four categories: fault name (Fault), fault cause (Cause), fault location (Position) and fault repair (Repaies). These four types of entities were annotated using the BIO annotation format, where "B" represents the beginning of the entity, "I" represents the middle and end of the entity, and "O" represents not an entity. After randomly shuffling the constructed data set, the training set and test set were divided into a ratio of 8:2.

[0019] Figure 1 This is the overall architecture of the RP-TEBC model. The model structure consists of four main parts: using the RoBERTa-wwm-ext pre-trained model as the word embedding layer; generating adversarial examples by incorporating adversarial training into each embedded word vector to improve the model's robustness and generalization; utilizing an adaptive Transformer encoder fused with a bidirectional gated recurrent unit (BiGRU) as the context encoder; and finally, using a conditional random field (CRF) as the output layer to predict labels and thus obtain entities in the input text.

[0020] The RoBERTa-wwm-ext pre-trained model serves as the word embedding layer. RoBERTa (Robustly Optimized BERTPretraining Approach) is an improved version of the original BERT model. It further improves model performance by removing the next sentence prediction task, increasing the amount of training data, and increasing the number of training steps. "wwm-ext" indicates the use of whole word masking and additional Chinese pre-training data. This is particularly important for Chinese processing because it can better understand and represent the overall meaning of Chinese words rather than the meaning of individual characters, thereby improving the model's semantic understanding ability. As a word embedding layer, RoBERTa-wwm-ext provides a high-dimensional vector representation for each input word. This vector contains the semantic information of the word and its contextual associations in a large-scale corpus. After encoding the input text through the RoBERTa-wwm-ext pre-trained model, a vector encoding with better linguistic information is obtained.

[0021] Adaptive Transformer Encoder Fusion with BiGRU. This part combines two powerful sequence modeling techniques: the Transformer encoder and the Bidirectional Gated Recurrent Unit (BiGRU). The Transformer captures long-range dependencies through a self-attention mechanism, while the BiGRU excels at capturing sequential information in a sequence. "Adaptive" here may refer to the model's ability to dynamically adjust its behavior based on the input, such as by learning weight distribution to balance the contributions of the Transformer and BiGRU. The fusion of the two aims to combine the Transformer's global perspective and the BiGRU's sensitivity to sequence order to form a more comprehensive contextual understanding capability. Unlike traditional recurrent neural network (RNN)-type encoders, the Transformer encoder uses a multi-head attention mechanism to better learn the contextual information of the input text, solving the problems of vanishing and exploding gradients that traditional RNNs have when learning longer input text.

[0022] The performance of traditional Transformer encoders in named entity recognition tasks is far inferior to its performance in other natural language processing tasks. Therefore, this paper uses the adaptive Transformer encoder proposed by Yan et al. to encode the character-level and word-level features of the input text for named entity recognition tasks. The sinusoidal position embedding in the original Transformer encoder is sensitive to distance perception but not to direction perception. To address this problem, the adaptive Transformer encoder adopts relative position encoding and improves the relative position encoding. The improved relative position encoding not only uses fewer parameters but also has better performance.

[0023] Considering that the attention distribution of the traditional Transformer encoder is scaled and smooth, but not all words need to be noticed in the task of named entity recognition, a sparse attention mechanism is obviously more suitable for the named entity recognition task. For this purpose, the adaptive Transformer encoder uses a non-scaled and sharp attention mechanism. In the formula Here, t represents the index of the target word, j represents the index of the context word, Q, K, V are the query vector, key vector and value vector respectively, and W q and W v is the weight parameter, is the result of a matrix operation. In, R t-j is a relative position encoding, and Where i represents the index variable, the range is In the formula middle is the attention score between two tokens; Here t stands for mark, which indicates the deviation from a specific relative distance; u T K j Here, j also represents the mark, which represents the deviation from the mark; v T R t-j is the deviation term for a specific distance and direction. It can be seen that the position information is multiplied by the query vector (Qt) of the word, and the relative position encoding (R t-j ) can also reflect the position information well. By adding 'multi-head', the performance of the attention mechanism is further improved. Based on the formula Get the formula Because sin(-x)=-sin(x), cos(x)=cos(-x). Where t represents the offset, which means that for offset t, the forward and backward relative position codes are relative to cos(c i t) term is the same, but with respect to sin(c i t) is the opposite. Therefore, by using R t-j , the attention score can distinguish different directions and distances.

[0024] Since the traditional recurrent neural network (RNN) will flow through the tanh node and the matrix product node when backpropagating to update the gradient parameters in the time direction, the derivative of y=tanh(x) is The derivative shows that when the value of the derivative is less than 1, it approaches 0 as the value of x increases in the positive direction. This means that if the gradient passes through too many tanh nodes, the derivative will gradually approach 0, resulting in a vanishing gradient phenomenon. Once the gradient vanishes, the weight parameters cannot be updated. This is one of the main reasons why traditional recurrent neural networks cannot learn long-term dependencies. When the gradient passes through a matrix multiplication node, the gradient grows exponentially with the increase in time steps. When the gradient is too large, it becomes non-numeric, preventing the neural network from learning and causing gradient explosion. The Long Short-Term Memory (LSTM) network alleviates these problems of traditional recurrent neural networks to some extent by introducing input gates, forget gates, and output gates. During backpropagation, the LSTM uses element-wise product operations. Each element is multiplied according to different gate values, thus alleviating the problems of vanishing and exploding gradients. The Gated Recurrent Unit (GRU) is an upgrade and improvement to the LSTM. Since the GRU only has a reset gate and an update gate, its computational cost and parameters are less than those of the LSTM, and its performance is comparable to that of the LSTM. The calculation formula is as follows:

[0025] Adversarial training enhances model robustness and generalization. Adversarial training is a technique that improves model robustness and generalization by adding carefully crafted small perturbations (adversarial examples) to the input data. In this model, adversarial examples are generated directly within each word embedding vector, requiring the model to not only learn conventional text representations but also learn to correctly interpret slightly distorted inputs. This mechanism forces the model to learn smoother, more invariant feature representations, reducing overfitting and improving performance on unseen data.

[0026] Focusing on the optimization of the adversarial and robustness of the model, the PGD (Projected Gradient Descent) adversarial training algorithm proposed by Madry et al. provides a good help in solving this problem. Assume that the data distribution of a standard classification task is D, and the data x∈R d , label y∈[k], loss function is L(θ,x,y), θ is a fixed parameter, and the empirical risk minimization (ERM) is used to find the optimal parameters of the model, namely: However, ERM does not produce a model that is robust to adversarial data, so it is necessary to expand the ERM paradigm. By introducing a set of perturbations of size S to each data point x, using the original sample to generate adversarial samples, and then using the adversarial samples to obtain the expectation, this is the famous Min-Max formula, as shown in the formula As shown. Min-Max mainly consists of internal loss maximization and external experience minimization. The goal of the internal maximization problem is to find the adversarial example for a given data point x to achieve the highest loss; the external experience risk minimization is to find the optimal parameters of the model to minimize the adversarial loss. PGD uses the strategy of "taking small steps and approaching little by little" to ensure that the perturbation is not too large. If it goes beyond the perturbation radius, it will be remapped to the "sphere", calculated as follows: t+1 =Π x+S (x t +∝sgn(▽ x L(θ,x,y)))

[0027] Conditional Random Fields (CRFs) are used as input layers. CRFs are a probabilistic graphical model for sequence labeling and segmentation, particularly well-suited for tasks with structured outputs, such as named entity recognition (NER). The CRF layer takes into account the dependencies between labels in the labeled sequence, ensuring that the output label sequence is optimal overall, not just the optimal selection of individual labels. Adding a CRF layer to the end of the model allows the use of previously predicted labels to guide the prediction of the current label, thereby improving the accuracy and consistency of the final labeling.

[0028] Conditional random fields are proposed based on the hidden Markov model (HMM) and the maximum entropy model (EM). They break the hidden Markov assumption and make the label prediction more reasonable. At the same time, they also correct the label bias problem of the EM model and make it possible to achieve global normalization. Given a sequence s = [s1, s2, ..., s T ] Its corresponding label sequence is y=[y1,y2,…,y T ],Y (s) Represents the sequence of all valid labels, the probability of y can be expressed by Calculation. In the formula f(y t-1 ,y t ,s) is the calculation of y t-1 to y t The conversion score is used to maximize P(y|s) and the Viterbi algorithm is used to find the optimal label path output.

[0029] The BERT+BiLSTM+CRF model architecture is currently a common one in named entity recognition. This model uses the original BERT pre-trained model as the word embedding layer, employs a bidirectional long short-term memory network (BiLSTM) to contextually encode the input text, and finally decodes it using a conditional random field (CRF). Due to its excellent portability, it has been widely used in both academia and industry. To verify the effectiveness of the RP-TEBC model for named entity recognition in the combine harvester fault domain, this paper conducted a comparative experiment under the same experimental environment. The experimental results are shown in Table 3. The experiment used the currently mainstream named entity recognition model BERT+BiLSTM+CRF as the baseline model. As shown in the table, the RP-TEBC model achieved improvements of 1.79%, 1.01%, and 1.46% in precision, recall, and F1 score, respectively, compared to the baseline model. This demonstrates that the RP-TEBC model is more efficient than traditional models in named entity recognition in the combine harvester fault domain.

[0030] In order to verify the effectiveness of adding adaptive Transformer encoder and introducing adversarial training for named entity recognition in the field of combine harvester failure, an ablation test was conducted. (1) The accuracy of the RoBERTa-wwm-ext+BiGRU+CRF model is 74.95%, the recall rate is 86.28%, and the F1 value is 80.22%. (2) The accuracy of the RoBERTa-wwm-ext+BiG RU+Transformer+CRF model is 75.12%, the recall rate is 86.74%, and the F1 value is 80.51%. (3) The accuracy of the RoBERTa-wwm-ext+BiGRU+PGD+CRF model is 75.78%, the recall rate is 86.93%, and the F1 value is 80.97%. (4) The accuracy of the RoBERTa-wwm-ext+BiGRU+Transformer+PGD+CRF model is 76.33%, the recall rate is 87.11%, and the F1 value is 81.36%. Compared to Model 1, Model 2 achieved a 0.17% improvement in accuracy, a 0.46% improvement in recall, and a 0.29% improvement in F1 score. Model 3 introduced adversarial training, which improved its generalization and robustness. Compared to Model 1, Model 3 achieved a 0.83% improvement in accuracy, a 0.65% improvement in recall, and a 0.75% improvement in F1 score. Model 4 (RP-TEBC) incorporated both an adaptive Transformer encoder and adversarial training, enabling it to not only effectively learn the semantic information of the input text but also improve its generalization and robustness. This resulted in Model 4 achieving higher accuracy, recall, and F1 score than Models 1, 2, and 3. This suggests that incorporating an adaptive Transformer encoder and adversarial training can improve named entity recognition in the combine harvester fault domain.

[0031] Model packaging. (1) API interface development: Encapsulate the trained NER model into one or more easily callable API services, which can be easily interacted with other systems or front-end applications. (2) Microservice architecture: If applicable, the model can be deployed as part of a microservice to facilitate horizontal expansion and independent maintenance.

[0032] Environment configuration. (1) Hardware selection: Select a suitable server or cloud service environment based on the model's resource requirements (CPU, GPU, memory, etc.). (2) Software environment: Configure the necessary operating environment, including the operating system, programming language version, and dependency libraries, to ensure consistency or compatibility with the development environment. (3) Containerization: Use container technologies such as Docker to package the application and its dependencies for easy deployment and management across environments.

[0033] Security considerations. (1) Data encryption: Ensure data encryption during transmission to protect sensitive information. (2) Access control: Implement authentication and authorization mechanisms to limit access to the API. (3) Security auditing: Record key operation logs and conduct regular security audits and vulnerability scans.

[0034] Performance optimization. (1) Load balancing: Deploy a load balancer to distribute requests and ensure system stability under high concurrency. (2) Caching strategy: Use cache for frequently queried but infrequently changing data to reduce database access pressure and improve response speed. (3) Resource control: Deploy monitoring tools to track system resource usage in real time and identify and resolve problems promptly.

[0035] Integration testing. (1) Unit testing: Perform unit testing on API services to ensure that each functional point works as expected. (2) Integration testing: Simulate real-world scenarios to test whether the model interacts smoothly with other system components. (3) Stress testing: Simulate high-load conditions to evaluate system performance limits and optimize performance bottlenecks.

[0036] Deployment strategy. (1) Grayscale release: Deploy the new version to a limited number of users or regions first, then gradually expand the scope to facilitate problem detection and rollback. (2) Rolling update: Update service instances in batches without downtime to ensure service continuity.

[0037] Monitoring and maintenance. (1) Log analysis: Regularly analyze system logs to monitor errors and anomalies. (2) Continuous improvement: Based on user feedback and technological advancements, continuously iterate and optimize models and services. (3) Backup and recovery: Based on user feedback and technological advancements, continuously iterate and optimize models and services.

[0038] It should be noted that the present invention also provides an electronic device, which includes at least one processor, a memory, and a computer program stored in the memory and runnable on the processor, so that when the processor executes the program, it can implement the steps in the user-oriented combine harvester fault diagnosis named entity recognition system as described above. The present invention also provides a computer-readable storage medium, a computer-readable storage medium, and the computer-readable storage medium contains at least one or more programs, which, when executed by the at least one processor, implements the user-oriented combine harvester fault diagnosis named entity recognition system as described above.

Claims

1. A high-precision entity naming and recognition method for combine harvester fault diagnosis, characterized in that: The steps include: In the first step, we collected unstructured text in the field of combine harvester failures through crawlers, PDF text conversion and other technical means; The second step is to use the YEDDA annotation tool to annotate the original corpus; The third step is to classify the marked entities into four categories: fault name (Fault), fault cause (Cause), fault location (Position) and fault repair (Repaies). The fourth step is to mark these four types of entities using the BIO annotation format, where "B" represents the beginning of the entity, "I" represents the middle and end of the entity, and "O" represents not an entity.

2. The high-precision entity naming and recognition method for combine harvester fault diagnosis according to claim 1 is characterized by: This includes proposing the RP-TEBC model, which is mainly divided into four parts. It uses the RoBERTa-wwm-ext pre-trained model as the word embedding layer; generates adversarial samples by adding adversarial training to each embedded word vector to improve the robustness and generalization of the model; uses the adaptive Transformer encoder to fuse the bidirectional gated recurrent unit (BiGRU) as the context encoder; and finally uses the conditional random field (CRF) as the output layer to predict labels, thereby obtaining entities in the input text.

3. The high-precision entity naming and recognition method for combine harvester fault diagnosis according to claim 2, characterized in that: Including the RoBERTa-wwm-ext pre-training model as the word embedding layer: the BERT pre-training model is proposed in the word embedding layer stage, and the multi-head attention mechanism is introduced in the Transformer encoding structure through the BERT pre-training model to obtain the semantic information of the input text, and the masked language model (MLM) and next sentence prediction (NSP) sub-tasks are used to train the word vector representation of the corpus. In this representation method, the word vector is calculated based on the context of the current word, so the same word appears in different places in the same sentence. The word vector representation is different, and the semantics expressed is also different. The input of the BERT model is composed of two text segments. After BERT modeling, the contextual semantic representation of the output text is obtained, and finally the masked language model and next sentence prediction are learned. RoBERTa The pre-training model is improved based on the BERT model. The RoBERTa pre-training model uses a dynamic masking strategy and more diverse training data, and cancels NSP during the training process. Through various improvements, the performance of the RoBERTa pre-training model is better than that of the BERT pre-training model. RoBERTa-wwm-ext is a new pre-training model obtained by retraining RoBERTa using a Chinese dataset based on the whole-word masking (WWM) strategy proposed by Cui et al. The whole-word masking strategy is to perform masked language training on words. Since Chinese words are composed of multiple characters, directly masking a single character may lead to the loss of semantic information. By using the whole-word masking strategy, the complete semantics of Chinese words can be better captured, thereby improving the model's performance in Chinese natural language processing tasks.

4. The high-precision entity naming and recognition method for combine harvester fault diagnosis according to claim 3 is characterized by: It includes a context encoder layer, which is mainly composed of an adaptive Transformer encoder and a bidirectional gated recurrent unit (BiGRU). The adaptive Transformer encoder uses relative position encoding and improves relative position encoding. The improved relative position encoding uses fewer parameters and has better performance. For this adaptive Transformer encoder, a non-scaled and sharp attention mechanism is used. In the formula Q, K, V = HW q , HW v Here, t represents the index of the target word, j represents the index of the context word, Q, K, V are the query vector, key vector and value vector respectively, and W q and W v is the weight parameter, is the result of a matrix operation. In, R t-j is a relative position encoding, and Where i represents the index variable, the range is In the formula middle is the attention score between two tokens; Here t stands for mark, which indicates the deviation from a specific relative distance; u T K j Here, j also represents the mark, which represents the deviation from the mark; v T R t-j is the deviation term for a specific distance and direction. It can be seen that the position information is multiplied by the query vector (Qt) of the word, and the relative position encoding (R t-j ) can also reflect the position information well. By adding 'multi-head', the performance of the attention mechanism is further improved. Based on the formula Get the formula Because sin(-x)=-sin(x), cos(x)=cos(-x). Where t represents the offset, which means that for offset t, the forward and backward relative position codes are relative to cos(c i t) term is the same, but with respect to sin(c i t) term is the opposite, so by using R t-j , the attention score can distinguish different directions and distances. Since the traditional recurrent neural network (RNN) will flow through the tanh node and the matrix product node when backpropagating to update the gradient parameters in the time direction, the derivative of y = tanh (x) is According to its derivative, when the value of the derivative is less than 1, as the value of x increases in the positive direction, the value of the derivative is getting closer and closer to 0. This means that if the gradient passes through too many tanh nodes, the value of the derivative will slowly approach 0, resulting in the phenomenon of gradient disappearance. Once the gradient disappears, the weight parameters will not be able to be updated. This is also one of the main reasons why traditional recurrent neural networks cannot learn long-term dependencies. When the gradient passes through the matrix multiplication node, the gradient will show an exponential growth with the increase of this time step. When the gradient is too large, a non-numeric value will appear, causing the neural network to be unable to learn, thereby causing Gradient explosion. The Long Short-Term Memory (LSTM) network alleviates the problems brought by traditional recurrent neural networks to a certain extent by introducing input gates, forget gates, and output gates. LSTM uses the product operation of corresponding elements during backpropagation. The corresponding elements are multiplied according to different gate values each time, so it alleviates the problems of gradient disappearance and gradient explosion. The Gated Recurrent Unit (GRU) is an upgraded improvement of LSTM. Since GRU only has reset gates and update gates, the computational cost and parameters are less than LSTM, and the performance is comparable to LSTM. The calculation formula is as follows 5. The high-precision entity naming and recognition method for combine harvester fault diagnosis according to claim 4 is characterized in that: Including adversarial training: Formula Min-Max mainly consists of internal loss maximization and external experience minimization. The purpose of the internal maximization problem is to find the adversarial sample of a given data point x to achieve the highest loss; the external experience risk minimization is to find the optimal parameters of the model to minimize the adversarial loss. PGD uses the strategy of "taking small steps and approaching little by little" to ensure that the disturbance is not too large. If it goes out of the disturbance radius, it will be remapped to the "sphere". The calculation is as follows shown.

6. The high-precision entity naming and recognition method for combine harvester fault diagnosis according to claim 5, characterized in that: Including decoder layer: Use conditional random field (CRF) as the decoding layer of RP-TEBC model. The conditional random field is proposed based on the hidden Markov model (HMM) and the maximum entropy model (EM). Given a sequence s = [s1, s2, ..., s T ] Its corresponding label sequence is y=[y1,y2,…,y T ],Y (s) Represents the sequence of all valid labels, the probability of y can be expressed by Calculation, in the formula f(y t-1 ,y t ,s) is the calculation of y t-1 to y t The conversion score is used to maximize P(y|s) and the Viterbi algorithm is used to find the optimal label path output.

7. A high-precision entity naming and recognition method for combine harvester fault diagnosis according to any one of claims 1 to 6, characterized in that: A high-precision entity naming and recognition system for combine harvester fault diagnosis is provided, wherein the system comprises: The data acquisition module is responsible for collecting the working status data of each component of the combine harvester in real time or regularly, including sensor data (such as vibration, temperature, pressure, flow, etc.), control unit signals, and operation logs. The preprocessing module cleans, formats and standardizes the received raw data, removes noise, and fills in missing values. The feature extraction module extracts key features that are helpful for fault diagnosis from the preprocessed data, including vibration amplitude at a specific frequency, temperature change rate, abnormal working current, etc. Named Entity Recognition (NER) module, which includes the name of the faulty component, fault type, fault severity, etc. It is usually implemented using natural language processing techniques such as part-of-speech tagging, sequence tagging, or deep learning models (such as BERT and LSTM-CRF); The fault diagnosis module classifies and diagnoses faults based on the extracted features and identified named entities using machine learning algorithms (such as support vector machines (SVMs), random forests, neural networks), fuzzy logic, rule engines, or hybrid methods. It includes a model-based knowledge base for storing and querying known fault cases and their solutions. Decision support and recommendation module, which provides operators or maintenance personnel with intuitive fault reporting and diagnostic results display, including graphical interface, alarm prompts, and mobile application notifications; Data storage and analysis module, which stores fault diagnosis data in the long term for subsequent fault mode learning, performance evaluation and system optimization; The remote communication module allows remote monitoring and diagnosis, and transmits field data to a central server or expert system via a wireless network, enabling remote technical support and software updates.

8. A high-precision entity naming and recognition method for combine harvester fault diagnosis according to any one of claims 1 to 6, characterized in that: The invention comprises an electronic device, which comprises: at least one processor, a memory and a computer program stored in the memory and executable on the processor, so that when the processor executes the program, the steps in the user-oriented named entity recognition system for combined harvester fault diagnosis as described above can be implemented.

9. A high-precision entity naming and recognition method for combine harvester fault diagnosis according to any one of claims 1 to 6, characterized in that: A storage medium is included, comprising: the storage medium stores one or more programs, and when the programs are executed by the at least one processor, the named entity recognition method for combine harvester fault diagnosis according to any one of claims 1 to 6 is implemented.

Images