Methods, devices, electronic equipment and storage media for determining the root cause of equipment fault alarms

Abstract

This application relates to the field of data processing technology, and in particular to a method, apparatus, electronic device, and storage medium for determining the root cause of equipment fault alarms. Upon receiving a fault alarm event, the method acquires its textual and structured modal information, extracts features from these two modalities to obtain textual feature vectors and structured feature vectors, fuses them to generate a multimodal feature vector, and inputs this multimodal feature vector into a preset target prediction model trained based on dynamic curriculum learning. This model then outputs the predicted root cause of the fault alarm event. By fusing multimodal information and utilizing a dynamic curriculum learning model, this application effectively solves the problems of traditional methods relying on a single data source and having low fault diagnosis accuracy, enabling more precise location of the root cause of equipment faults.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, electronic device and storage medium for determining the root cause of equipment fault alarms. Background Technology

[0002] With the continuous development of intelligent manufacturing technology, intelligent packaging equipment is increasingly widely used in production lines, and its operational stability directly affects product packaging quality and production efficiency. In complex production environments, root cause analysis (RCA) of intelligent packaging equipment failures is a crucial step in ensuring continuous production line operation and reducing maintenance costs. Supported by the Industrial Internet of Things (IIoT) and big data technologies, the data generated during the operation of intelligent packaging equipment exhibits diverse and heterogeneous characteristics, including multiple modalities of information such as equipment operation logs, fault alarm texts, and sensor monitoring indicators. Traditional root cause analysis methods often rely on a single data source or the experience-based judgment of technical personnel. This makes it impossible to fully capture and reconstruct the true root cause of equipment failures when facing complex failure scenarios, leading to low accuracy in fault diagnosis and slow response times. Especially in large-scale production workshops, this limitation not only affects equipment operation and maintenance efficiency but may also cause production line shutdowns, severely impacting the company's economic benefits.

[0003] Meanwhile, with the increasing automation of intelligent packaging equipment, the interconnectivity between equipment components is enhanced. Minor anomalies in a single component can trigger cascading failures, further increasing the complexity of fault localization. Although existing equipment is equipped with basic fault detection modules, how to effectively integrate this detection data within the intelligent alarm root cause analysis framework to achieve collaborative analysis of multi-dimensional fault information still requires further exploration. To improve the reliability and adaptability of RCA models, existing solutions introduce highly reliable technologies such as adversarial training, federated learning, and interpretability analysis, and combine them with transfer learning and meta-learning methods to enhance the model's adaptability to complex production environments. These methods alleviate the robustness requirements of models in open environments to some extent, but how to effectively integrate multimodal information and adaptively adjust the model learning strategy in intelligent packaging equipment fault alarm scenarios remains an unsolved problem. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method, apparatus, electronic device, and storage medium for determining the root cause of equipment fault alarms. By acquiring multimodal information, extracting and fusing multimodal features, and utilizing a target prediction model based on dynamic course learning training, it effectively solves the problems of traditional methods relying on a single data source, difficulties in collaborative analysis of multidimensional fault information, and the challenges of multimodal information fusion and adaptive model adjustment in the technical background.

[0005] A first aspect of this application provides a method for determining the root cause of equipment fault alarms, the method comprising: When a fault alarm event is received, the text modal information and structured modal information of the fault alarm event are obtained; The text modal information is subjected to feature extraction to obtain a text feature vector; and the structured modal information is subjected to feature extraction to obtain a structured feature vector. The text feature vector and the structured feature vector are fused to generate a multimodal feature vector; The multimodal feature vectors are input into a preset target prediction model so that the target prediction model outputs the predicted alarm root cause corresponding to the fault alarm event; the prediction model is trained based on dynamic course learning.

[0006] In an optional implementation, the method further includes: Obtain a training dataset, which includes alarm multimodal features and corresponding alarm root cause labels; An initial prediction model is constructed based on a multilayer perceptron. The training dataset is input into the initial prediction model. In each iteration, each sample in the training dataset is traversed, and the text length and anomaly index of each sample are calculated. The difficulty of each sample is calculated based on the text length and the anomaly index. Iterate through each sample in the training dataset and calculate the sampling probability of each sample based on the difficulty. Construct mini-batch B based on the sampling probability; The mini-batch B is forward-propagated, and the predicted output is calculated based on the current model parameters; Calculate the loss between the predicted output and the true label in the alarm root cause label, and use backpropagation to update the model parameters based on the loss until the maximum number of iterations is met; The target prediction model is determined by the initial prediction model that satisfies the maximum number of iterations.

[0007] In an optional implementation, calculating the difficulty of each sample based on the text length and the anomaly index includes: The difficulty level is determined using the following formula; ; in, The difficulty of the current sample i, The length of the text in the current sample i. The length of the longest text in the training dataset; Used to calculate the average feature complexity of the current sample i across k key stages. Let be the weighting coefficient at stage t; Used to evaluate the information uncertainty of the current sample i after multimodal fusion. This represents the frequency of being correctly identified in stage t, where n is the total number of samples; , and This is a hyperparameter.

[0008] In an optional implementation, calculating the sampling probability of each sample based on the difficulty includes: The sampling probability is determined using the following formula: ; in, Let i be the sampling probability of the current sample i. Let M be a function that increases with the number of iterations t, where M is the total number of samples. This is a hyperparameter.

[0009] In an optional implementation, the step of extracting features from the text modality information to obtain a text feature vector includes: The text modal information is input into a pre-trained language model to perform word segmentation and word embedding processing on the text modal information through the pre-trained language model, thereby obtaining a set of word embedding vectors; The contextual semantic features of each word embedding vector in the word embedding vector set are extracted using the pre-trained language model and used as the text feature vector.

[0010] In an optional implementation, the step of extracting features from the structured modal information to obtain a structured feature vector includes: For the categorical data in the structured modal information, the categorical data is transformed by One-hot encoding; the categorical data after One-hot encoding is input into the dimension alignment layer; the dimension-aligned feature vector is sequentially input into the Embedding layer and the Dense layer, so that the Embedding layer extracts low-dimensional semantic features and the Dense layer extracts high-level feature representations as the structured feature vector of the categorical data; For the numerical data and / or relational data in the structured modal information, preliminary feature processing is completed through standardization, and the structured feature vectors of the numerical data and / or relational data are output.

[0011] In an optional implementation, fusing the text feature vector and the structured feature vector to generate a multimodal feature vector includes: The text feature vector and the structured feature vector are concatenated to obtain a joint feature vector; The joint feature vector is input into at least two consecutive linear mapping layers to output the multimodal feature vector; after processing by each linear mapping layer, a nonlinear transformation is performed using the ReLU activation function, and a Dropout strategy is applied; or The joint feature representation is input to the fusion linear mapping layer to fuse and reduce the feature dimensions, and output the multimodal feature vector. The fusion linear mapping layer uses a non-linear activation function to enhance the feature representation capability after fusion.

[0012] A second aspect of this application provides a device for determining the root cause of equipment fault alarms, the device comprising: The acquisition module is used to acquire the text modal information and structured modal information of the fault alarm event when a fault alarm event is received; The feature extraction module is used to extract features from the text modal information to obtain a text feature vector; and to extract features from the structured modal information to obtain a structured feature vector; The feature fusion module is used to fuse the text feature vector and the structured feature vector to generate a multimodal feature vector; The prediction module is used to input the multimodal feature vector into a preset target prediction model, so that the target prediction model outputs the predicted alarm root cause corresponding to the fault alarm event; the prediction model is trained based on dynamic course learning.

[0013] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the device fault alarm root cause determination method.

[0014] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described device fault alarm root cause determination method.

[0015] In summary, the equipment fault alarm root cause determination method, apparatus, electronic device, and storage medium provided in this application have at least one of the following beneficial effects: 1. Upon receiving a fault alarm event, this application acquires the text modal information and structured modal information of the fault alarm event. By acquiring multiple modal information, it overcomes the limitation of traditional methods relying on a single data source, and can obtain equipment fault-related information from different perspectives, providing a rich data foundation for comprehensive analysis of the root cause of the fault; 2. Text feature vectors are obtained by extracting features from textual modal information, and structured feature vectors are obtained by extracting features from structured modal information. Then, the textual and structured feature vectors are fused to generate a multimodal feature vector. This multimodal feature extraction and fusion method integrates different types of data information, which can more comprehensively reflect the characteristics of equipment faults, helping to more accurately capture and reconstruct the true root cause of equipment faults, thereby improving the accuracy and response speed of fault diagnosis. 3. Input the multimodal feature vectors into the preset target prediction model, which is trained based on dynamic curriculum learning. Dynamic curriculum learning can adaptively adjust the model's learning strategy according to the characteristics of the data and the learning process, enabling the model to better adapt to the complex and ever-changing situations in intelligent packaging equipment fault alarm scenarios, effectively integrate multimodal information, improve the model's adaptability to complex production environments, and thus enhance the model's reliability and adaptability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the architecture of a device fault alarm root cause determination system shown in an embodiment of this application; Figure 2 This is a flowchart illustrating a method for determining the root cause of equipment fault alarms, as shown in an embodiment of this application. Figure 3 This is a general architecture diagram of a method for determining the root cause of equipment fault alarms, as shown in an embodiment of this application. Figure 4 This is a schematic diagram illustrating a multimodal feature extraction and fusion process according to an embodiment of this application; Figure 5 This is a functional block diagram of a device for determining the root cause of equipment fault alarms, as shown in an embodiment of this application. Figure 6 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this invention can be combined interactively without contradicting each other.

[0019] Reference Figure 1The diagram shown illustrates the architecture of a device fault alarm root cause determination system according to an embodiment of this application. The system comprises three layers: the first layer is the raw alarm data generated by the cloud platform, including alarm name, alarm description, potential root cause label, and resource type; the second layer is a BERT-Based-Chinese information filtering module, used to accurately extract textual and source address information from the raw alarms, converting them into deep semantic features, and then performing multimodal feature fusion; the third layer uses dynamic curriculum learning and prediction models to accurately predict the network fault root cause chain. Specifically, in the first layer, features are extracted from the textual modality (such as alarm description) and structured modality (such as alarm type) of the alarm event. First, for textual information, the BERT-Based-Chinese model is used to perform tokenization and embedding operations to generate semantically rich textual feature vectors. Secondly, for structured data (such as alarm types), one-hot encoding of categorical features is performed first, followed by processing through an embedding layer, linear mapping, and ReLU activation function. Finally, a feedforward network generates structured feature vectors. In the second layer, the textual and structured features extracted in the previous stage are concatenated, then passed through two consecutive fusion linear mapping layers. Each linear layer employs ReLU activation function and Dropout strategy to improve the generalization performance of the fused features and avoid overfitting. After this fusion process, a unified and rich multimodal feature vector is generated, providing a foundation for subsequent dynamic curriculum learning strategies. In the third layer, a prediction model is trained based on dynamic curriculum learning, enabling the model to predict the root causes of complex alarms. The fused features are input into the trained model, and the final prediction result is generated through the Softmax or Sigmoid function. Finally, the system visualizes the predicted root cause results and corresponding confidence scores, helping operations personnel quickly and accurately locate the root cause of alarms and optimize the decision-making process.

[0020] In existing intelligent alarm systems, the diverse and unevenly distributed data sources, information silos, and feature redundancy make it difficult to accurately pinpoint the cause of faults. Specifically, there are three main problems: (1) The heterogeneity of multimodal data and the problem of information silos.

[0021] In practical applications, different data modalities (such as system logs, alarm texts, and monitoring metrics) have their own distinct data structures and representations. Direct fusion often results in feature redundancy or information loss. To describe this problem, let the data features of each modality be denoted as... A common fusion method is weighted summation, and its mathematical expression is: ,and ; in Indicates the first i Weights of each modality.

[0022] Determining a reasonable weight allocation and reducing information silos while ensuring effective complementarity of features across modalities are key challenges that urgently need to be addressed.

[0023] (2) Problems of uneven sample distribution and high training difficulty.

[0024] In real-world systems, the amount of data under normal conditions is enormous, while fault samples are relatively scarce, resulting in an imbalanced distribution of training data. This imbalance not only affects the model's convergence speed but may also impair its ability to identify minority class faults. Dynamic Curriculum Learning was introduced to address this problem; its basic idea is to start with simple samples and gradually introduce more challenging ones. Let... The weight update function for samples during training can be represented in the following form:

[0025] in t To determine the number of training iterations, and These are the parameters for adjustment.

[0026] This strategy enables the model to gradually adapt to complex data and effectively alleviates the problem of imbalanced samples. (3) How to achieve efficient and accurate root cause localization in complex fault scenarios.

[0027] For intelligent alarm scenarios, in addition to data fusion and training strategy issues, it is also necessary to consider the extraction and efficient identification of fault features. Since system faults often involve the interaction of multiple factors, a single model is insufficient to capture all key information. The method proposed in this paper combines dynamic curriculum learning and multimodal data fusion to form an end-to-end diagnostic framework, thereby improving the accuracy and response speed of fault root cause localization. To this end, this paper constructs a multimodal fusion function. and loss function Its objective is to minimize the following objective function: ; in, Indicates fault label, These are the model parameters.

[0028] Existing research has adopted multi-sensor data fusion methods to achieve high-precision fault detection. Meanwhile, residual convolutional neural networks and lightweight fusion models have also shown good performance in improving the accuracy of fault diagnosis.

[0029] It is evident that the root cause of existing equipment fault alarms can be clearly defined as three key issues: heterogeneity and information silos during multimodal data fusion, training difficulties caused by uneven sample distribution, and how to efficiently and accurately locate the root cause in complex fault scenarios.

[0030] To facilitate understanding of the inventive concept of this application, the method for determining the root cause of equipment fault alarms will be explained using the example of the root cause of fault alarms in intelligent packaging equipment.

[0031] Reference Figure 2 The diagram shown is a flowchart illustrating a method for determining the root cause of a device fault alarm according to an embodiment of this application. The method for determining the root cause of a device fault alarm includes the following steps.

[0032] S21, when a fault alarm event is received, obtain the text modal information and structured modal information of the fault alarm event.

[0033] In some embodiments, a dedicated data receiving module is deployed in the device monitoring system. This module has the ability to listen to network ports or message queues in real time. For example, using Socket programming, while running on the server side, message middleware (such as RabbitMQ, Kafka, etc.) can also be used to receive fault alarm events. Taking RabbitMQ as an example, a specific queue (such as fault_alarm_queue) is created. The device or monitoring system, as the producer, sends fault alarm events to this queue, and the data receiving module, as the consumer, retrieves messages from the queue. A specific port (such as 8080) is opened to continuously listen for fault alarm messages from the device or monitoring equipment.

[0034] In intelligent alarm scenarios, alarm text (such as alarm descriptions and root cause hints) and structured fields (such as alarm type and resource type) carry key information in different dimensions. When the data receiving module receives a fault alarm event, the electronic device can parse it to extract text modal information and structured modal information. For text modal information, the fault alarm event may contain natural language descriptions of the fault phenomenon, device location, operator feedback, etc. For structured modal information, the fault alarm event usually includes some structured data, such as device operating parameters (voltage, current, temperature, etc.), the timestamp of the fault occurrence, and the device's unique identifier.

[0035] S22, extract features from the text modal information to obtain a text feature vector; and extract features from the structured modal information to obtain a structured feature vector.

[0036] Refer to together Figure 4 For textual modal information in fault alarm events, a pre-trained language model (such as BERT) is used for semantic encoding. In this embodiment, the pre-trained language model uses the existing open-source BERT-Base-Chinese model. Tokenization and embedding operations are performed through the BERT-Base-Chinese model to generate semantically rich text feature vectors containing contextual semantic information.

[0037] In an optional implementation, the step of extracting features from the text modality information to obtain a text feature vector includes: The text modal information is input into a pre-trained language model to perform word segmentation and word embedding processing on the text modal information through the pre-trained language model, thereby obtaining a set of word embedding vectors; The contextual semantic features of each word embedding vector in the word embedding vector set are extracted using the pre-trained language model and used as the text feature vector.

[0038] In some embodiments, the electronic device first performs word segmentation on the text modal information, splitting the text into independent lexical units. Then, it performs word embedding processing, converting each lexical unit into a fixed-dimensional vector representation, outputting a segmented word sequence and a corresponding set of word embedding vectors. Next, the set of word embedding vectors is input into the BERT-Base-Chinese model. Through the model's multi-layered structure, the contextual semantic features of each word embedding vector in the set are extracted, thereby generating a high-dimensional text feature representation. Specifically, the BERT-Base-Chinese model first employs a character- and subword-based word segmentation strategy. For the input Chinese text, it splits the text into independent character or subword units. Two special markers, such as [CLS] and [SEP], are added to the beginning and end of the segmented sequence, respectively. [CLS] indicates the start of the entire sequence, and [SEP] separates different sentences or text segments or marks the end of the text. Next, the segmented and specially tagged character or sub-word sequence is converted into a corresponding index value based on the predefined vocabulary of the BERT-Base-Chinese model, resulting in a sequence of index values. Then, based on the vocabulary index sequence, the corresponding word vector is retrieved from the word embedding matrix of the BERT-Base-Chinese model to obtain the word embedding sequence. The word embedding matrix is ​​learned during pre-training and maps each character or sub-word in the vocabulary to a fixed-dimensional vector space. Positional embedding vectors are then generated and added to the previous two to obtain the input embedding sequence, i.e., the set of word embedding vectors. This set of word embedding vectors is then input into a multi-layer Transformer encoder. Its self-attention mechanism allows the model to pay attention to other words while processing each word, capturing long-distance dependencies and context. After multi-layer processing, each word incorporates full-text context information, outputting a text feature vector.

[0039] Compared with the traditional bag-of-words model, the feature extraction method based on BERT can capture the deep semantic information of alarm text more comprehensively and provide richer clues for fault diagnosis.

[0040] Meanwhile, for the structured information in fault alarm events (such as alarm type, resource type, root cause label, time series and topology information, etc.), appropriate data preprocessing and feature extraction are performed.

[0041] In an optional implementation, the step of extracting features from the structured modal information to obtain a structured feature vector includes: For the categorical data in the structured modal information, the categorical data is transformed by One-hot encoding; the categorical data after One-hot encoding is input into the dimension alignment layer; the dimension-aligned feature vector is sequentially input into the Embedding layer and the Dense layer, so that the Embedding layer extracts low-dimensional semantic features and the Dense layer extracts high-level feature representations as the structured feature vector of the categorical data; For the numerical data and / or relational data in the structured modal information, preliminary feature processing is completed through standardization, and the structured feature vectors of the numerical data and / or relational data are output.

[0042] In some embodiments, for categorical data (such as alarm types and resource types), the electronic device first performs feature transformation using one-hot encoding. Then, the categorical data transformed by one-hot encoding is input into a dimension alignment layer, ensuring that the structured features are dimensionally consistent with the text features. Next, the dimension-aligned feature vector is input into an embedding layer, which transforms the high-dimensional sparse vector representation into a low-dimensional dense vector representation, learning the latent semantic information in the data to obtain a more representative feature vector (i.e., low-dimensional semantic features). Finally, the low-dimensional semantic features are input to a Dense layer, a fully connected layer that performs a non-linear transformation on the input vector. By setting multiple neurons and appropriate activation functions (such as ReLU, Sigmoid, etc.) in the Dense layer, the model can learn more complex and higher-level feature representations. The Dense layer performs a non-linear transformation on the input feature vector to further extract higher-level feature representations, thereby obtaining the structured feature vector of the processed categorical data.

[0043] For numerical data and / or relational data such as time series or topological structures, electronic devices employ normalization or standardization methods for preliminary feature processing to ensure the comparability and stability of feature scales in subsequent fusion stages. Specifically, assuming the normalization method used is Min-Max normalization, the time series data is linearly transformed to the [0, 1] interval. Assuming the standardization method used is Z-score standardization, the time series data is converted into a distribution with a mean of 0 and a standard deviation of 1. Next, a sliding window is used to slide across the time series, calculating the statistical characteristics of the data within the window, such as mean, variance, maximum, and minimum values. Then, a Fourier transform is used to convert the time series from the time domain to the frequency domain, extracting frequency domain features. The Fourier transform provides the amplitude and phase information of different frequency components in the time series, reflecting its periodicity and trend. Finally, the normalized or standardized original data, along with the extracted statistical or frequency domain features, are concatenated to form a fixed-dimensional numerical feature vector. For example, normalized time series data can be concatenated with its mean and variance to obtain a new feature vector, which is a structured feature vector of numerical data and / or relational data.

[0044] Furthermore, the processed categorical structured feature vectors and numerical and relational structured feature vectors are integrated. The integration method can be chosen according to specific needs; for example, a simple concatenation operation can be used to join two feature vectors together to form a longer feature vector. This longer feature vector is the final structured feature vector, containing rich features extracted from different types of structured information.

[0045] S23, the text feature vector and the structured feature vector are fused to generate a multimodal feature vector.

[0046] In some embodiments, after the independent extraction of text feature vectors and structured feature vectors is completed, in order to fully integrate the complementary information of text feature vectors and structured feature vectors, it is necessary to fuse text feature vectors and structured feature vectors.

[0047] In an optional implementation, fusing the text feature vector and the structured feature vector to generate a multimodal feature vector includes: The text feature vector and the structured feature vector are concatenated to obtain a joint feature vector; The joint feature vector is input into at least two consecutive linear mapping layers to output the multimodal feature vector; after processing by each linear mapping layer, a nonlinear transformation is performed using the ReLU activation function, and a Dropout strategy is applied; or The joint feature representation is input to the fusion linear mapping layer to fuse and reduce the feature dimensions, and output the multimodal feature vector. The fusion linear mapping layer uses a non-linear activation function to enhance the feature representation capability after fusion.

[0048] Refer to together Figure 4 In some embodiments, electronic devices can concatenate text feature vectors and structured feature vectors to form a joint feature vector. Concatenation refers to directly merging text feature vectors and structured feature vectors to form a new feature vector. This is simple and results in minimal information loss. Assuming the text feature vector has a dimension of m and the structured feature vector has a dimension of n, concatenating these two vectors results in a joint feature vector with a dimension of m+n. For example, if the text feature vector X = [x1, x2, ..., xm] and the structured feature vector Y = [y1, y2, ..., yn], then the joint feature vector is [x1, x2, ..., xm, y1, y2, ..., yn].

[0049] Further, the joint feature vector after concatenation is fused and its dimensionality reduced using Fusion Linear, and the expressive power of the fused features is further enhanced by using appropriate nonlinear activation functions (such as GELU or ReLU), including: Option 1: The joint feature vector is sequentially input into at least two consecutive linear mapping layers (FusionLinear layers). After processing by each linear mapping layer, a ReLU activation function is applied for non-linear transformation, followed by a Dropout strategy, where the dropout rate of the Dropout strategy is within a preset range. Specifically, the joint feature vector is input into the first linear mapping layer for linear transformation, then the ReLU activation function is applied to the output of the first linear mapping layer for non-linear transformation, followed by the application of the Dropout strategy. The vector after Dropout processing is input into the second linear mapping layer for linear transformation, again the ReLU activation function is applied to the output of the second linear mapping layer for non-linear transformation, and finally the Dropout strategy is applied to obtain the final multimodal feature vector.

[0050] Option 2: Input the joint feature vector into a fusion linear mapping layer, i.e., a Fusion Linear fusion layer. This layer achieves feature dimension fusion and dimensionality reduction, and uses a non-linear activation function (selected from GELU or ReLU) to enhance the expressive power of the fused features. Specifically, the joint feature vector is input into the Fusion Linear fusion layer for linear transformation, and then the output of the Fusion Linear fusion layer is non-linearly transformed using the GELU activation function to enhance the expressive power of the fused features. After processing by the GELU activation function, a multimodal feature vector is obtained. Alternatively, the ReLU activation function can be used, and the multimodal feature vector is obtained after processing by the ReLU activation function.

[0051] By using any of the above methods, the fusion of text feature vectors and structured feature vectors can be completed to generate high-dimensional multimodal feature vectors. These multimodal feature vectors simultaneously cover the comprehensive feature representation of text semantics and structured data, and serve as important inputs for subsequent dynamic course learning and model prediction stages, thereby helping the system to accurately locate and analyze the root cause in complex alarm scenarios.

[0052] By acquiring multimodal data, extracting features, and fusing features, we not only effectively solved the problem of information silos between text and structured data, but also provided a solid and reliable data foundation for subsequent model training and prediction.

[0053] S24, the multimodal feature vector is input into a preset target prediction model so that the target prediction model outputs the predicted alarm root cause corresponding to the fault alarm event.

[0054] After fusing to obtain multimodal feature vectors, the multimodal feature vectors are input into the trained target prediction model. The target prediction model generates the final prediction result, i.e., the prediction of the root cause of the alarm, through the Softmax or Sigmoid function.

[0055] In addition, electronic devices can visualize the predicted root cause results and corresponding confidence levels, helping maintenance personnel to quickly and accurately locate the root cause of alarms and optimize the decision-making process.

[0056] In some embodiments, the target prediction model is constructed and trained based on a multilayer perceptron (MLP). During the training of the prediction model, a Dynamic Curriculum Learning (DCL) strategy is introduced to adaptively adjust the sampling order and weights of samples during training. The DCL strategy mimics the human learning process from easy to difficult, allowing the model to prioritize learning easily understood alarm samples in the early stages. As training progresses, more complex and challenging samples are gradually introduced, effectively alleviating the problem of uneven sample distribution and improving the model's generalization ability. Starting from model initialization, the difficulty of each sample is quantified, the sampling probability is calculated and updated, and then a mini-batch is constructed using dynamic sampling. This is followed by a cyclical process of forward propagation, loss calculation, and backpropagation to update the model parameters. Finally, the sample difficulty index is updated before entering the next iteration.

[0057] Specifically, firstly, an initial prediction model is built based on MLP, and a training dataset D is obtained, where... , To provide alarms based on multimodal features, Set the corresponding root cause label for the alarm. Also, set the maximum number of training epochs T and initialize the model parameters. .

[0058] Set the iteration counter t to 1, input the training dataset D into the initial prediction model, and enter the outer loop, which executes sequentially from t=1 to the maximum training round number T.

[0059] In each iteration, the inner layer first iterates through each sample i in the training dataset D. For each sample i, the text length of each sample i is calculated. and abnormal indicators Among them, text length and abnormal indicators It is typically used to measure the basic properties of a sample and the degree of potential anomalies.

[0060] Next, based on text length and abnormal indicators The difficulty level of each sample i is calculated to provide a basis for dynamic curriculum learning, thereby distinguishing between easy and difficult samples. To quantify the difficulty of each sample, the following formula is used to calculate the sample difficulty. Perform comprehensive measurement: ; in, This indicates the text length of the current sample i; The length of the longest text in the training dataset D is used to normalize the text complexity. Used to calculate the mean feature complexity of the current sample i within k key "windows" (or "stages"). This indicates the number of times the threshold is exceeded or an anomaly indicator in multimodal information related to the current sample i. Let be the weighting coefficient at stage t; Then, from the perspective of feature structuring, the information uncertainty of the current sample i after multimodal fusion is evaluated. This represents the frequency at which the model correctly identifies a sample in stage t, where n is the total number of samples in the training dataset D. , and This is a hyperparameter used to balance the importance of different factors in difficulty calculation.

[0061] Iterate through all samples in the training set, and in the t-th round of training, calculate the sampling probability for each sample i. This reflects the importance or learning difficulty of the sample in the current iteration, facilitating dynamic sampling in subsequent steps. Sampling probability It can take the following form: ; in, To control the hyperparameters that affect the degree of difficulty, A function that increases with the number of iterations t (e.g. As training progresses, more challenging samples will gradually occupy a larger proportion of the sampling in the later stages, enabling the model to steadily improve its ability to identify the root causes of complex alarms during the training process, which starts with easier samples and gradually becomes more difficult.

[0062] Based on the calculated sampling probability Construct a mini-batch B, which is a small batch of samples drawn in each round of training. Here, mini-batch B is used in accordance with the meaning of batch size. The sampling probability is then used... This determines the probability of sample i being selected into the mini-batch, ensuring that the training samples in each round can reflect the current learning needs of the model.

[0063] Subsequently, forward propagation is performed on mini-batch B, utilizing the current model parameters. Obtain the predicted output.

[0064] The loss function is calculated by comparing the difference between the predicted output and the true label. This metric directly reflects the current prediction error of the model. In the embodiments of this application, the loss function is the ReLU function.

[0065] Using the backpropagation algorithm based on the loss Calculate the gradient and update the model parameters. To continuously optimize model performance.

[0066] Furthermore, the sample feedback metrics are updated based on the current predicted performance. This indicator will provide a reference for the dynamic adjustment of sample difficulty and sampling probability in subsequent iterations.

[0067] After the entire loop is complete (until t > T), the trained model parameters are returned on line 14. This algorithm effectively implements a progressive learning strategy from simple to difficult samples by dynamically adjusting the sampling probability according to the sample difficulty in each iteration, thereby helping to improve the overall learning efficiency and prediction performance of the model.

[0068] To facilitate understanding of the inventive concept of this application, the following table 1 shows the overall training process of a prediction model based on dynamic course learning.

[0069] Table 1:

[0070] Reference Figure 5 The diagram shown is a functional block diagram of the device for determining the root cause of equipment fault alarms according to an embodiment of this application.

[0071] In some embodiments, the device fault alarm root cause determination device 50 may include a plurality of functional modules composed of computer program segments. The computer programs for each program segment of the device fault alarm root cause determination device 50 may be stored in the memory of the electronic device and executed by at least one processor to perform (see details). Figure 2 (Description) This describes the function of determining the root cause of equipment fault alarms. Based on the functions it performs, it can be divided into multiple functional modules. These functional modules may include: an acquisition module 501, a feature extraction module 502, a feature fusion module 503, and a prediction module 504. The term "module" in this application refers to a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0072] The acquisition module 501 is used to acquire the text modal information and structured modal information of the fault alarm event when a fault alarm event is received.

[0073] The feature extraction module 502 is used to extract features from the text modal information to obtain a text feature vector; and to extract features from the structured modal information to obtain a structured feature vector.

[0074] The feature fusion module 503 is used to fuse the text feature vector and the structured feature vector to generate a multimodal feature vector.

[0075] The prediction module 504 is used to input the multimodal feature vector into a preset target prediction model, so that the target prediction model outputs the predicted alarm root cause corresponding to the fault alarm event; the prediction model is trained based on dynamic course learning.

[0076] It should be understood that the various variations and specific embodiments of the equipment fault alarm root cause determination method provided in the above embodiments are also applicable to the equipment fault alarm root cause determination device of this embodiment. Through the foregoing detailed description of the equipment fault alarm root cause determination method, those skilled in the art can clearly understand the implementation method of the equipment fault alarm root cause determination device of this embodiment. For the sake of brevity, it will not be described in detail here.

[0077] See Figure 6 The diagram shown is a schematic representation of the structure of an electronic device according to an embodiment of this application. In a preferred embodiment of this application, the electronic device 6 includes a memory 61, at least one processor 62, and at least one communication bus 63.

[0078] Those skilled in the art should understand that Figure 6 The structure of the electronic device shown does not constitute a limitation of the embodiments of this application. It can be a bus structure or a star structure. The electronic device 6 may also include more or fewer other hardware or software than shown, or different component arrangements.

[0079] In some embodiments, the electronic device 6 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), digital processors, and embedded devices. The electronic device 6 may also include user equipment, which includes, but is not limited to, any electronic product capable of human-computer interaction with a user via a keyboard, mouse, remote control, touchpad, or voice control device, such as a personal computer, tablet computer, smartphone, or digital camera.

[0080] In the embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, computer-readable storage media, and electronic devices can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple components or modules may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices, components, or modules may be electrical, mechanical, or other forms.

[0081] The components described as separate parts may or may not be physically separate. The components shown as components may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the components can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each component can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0083] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drive, portable hard drive, read-only memory (ROM). Various media that can store program code, such as only memory, random access memory (RAM), magnetic disks or optical disks.

[0084] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0085] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0086] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for determining the root cause of equipment fault alarms, characterized in that, The method includes: When a fault alarm event is received, the text modal information and structured modal information of the fault alarm event are obtained; The text modal information is subjected to feature extraction to obtain a text feature vector; and the structured modal information is subjected to feature extraction to obtain a structured feature vector. The text feature vector and the structured feature vector are fused to generate a multimodal feature vector; The multimodal feature vectors are input into a preset target prediction model so that the target prediction model outputs the predicted root cause of the fault alarm event; the prediction model is trained based on dynamic course learning.

2. The method for determining the root cause of equipment fault alarms according to claim 1, characterized in that, The method further includes: Obtain a training dataset, which includes alarm multimodal features and corresponding alarm root cause labels; An initial prediction model is constructed based on a multilayer perceptron. The training dataset is input into the initial prediction model. In each iteration, each sample in the training dataset is traversed, and the text length and anomaly index of each sample are calculated. The difficulty of each sample is calculated based on the text length and the anomaly index. Iterate through each sample in the training dataset and calculate the sampling probability of each sample based on the difficulty. Construct mini-batch B based on the sampling probability; The mini-batch B is forward-propagated, and the predicted output is calculated based on the current model parameters; Calculate the loss between the predicted output and the true label in the alarm root cause label, and use backpropagation to update the model parameters based on the loss until the maximum number of iterations is met; The target prediction model is determined by the initial prediction model that satisfies the maximum number of iterations.

3. The method for determining the root cause of equipment fault alarms according to claim 2, characterized in that, The calculation of the difficulty of each sample based on the text length and the anomaly index includes: The difficulty level is determined using the following formula; ； in, Given the difficulty of the current sample i, The length of the text in the current sample i. The length of the longest text in the training dataset; This is used to calculate the average feature complexity of the current sample i across k key stages. Let be the weighting coefficient at stage t; Used to evaluate the information uncertainty of the current sample i after multimodal fusion. This represents the frequency of being correctly identified in stage t, where n is the total number of samples; , and This is a hyperparameter.

4. The method for determining the root cause of equipment fault alarms according to claim 2, characterized in that, The calculation of the sampling probability of each sample based on the difficulty includes: The sampling probability is determined using the following formula: ； in, Let i be the sampling probability of the current sample i. Let M be a function that increases with the number of iterations t, where M is the total number of samples. This is a hyperparameter.

5. The method for determining the root cause of equipment fault alarms according to claim 1, characterized in that, The step of extracting features from the text modality information to obtain the text feature vector includes: The text modal information is input into a pre-trained language model to perform word segmentation and word embedding processing on the text modal information through the pre-trained language model, thereby obtaining a set of word embedding vectors; The contextual semantic features of each word embedding vector in the word embedding vector set are extracted using the pre-trained language model and used as the text feature vector.

6. The method for determining the root cause of equipment fault alarms according to claim 1, characterized in that, The step of extracting features from the structured modal information to obtain a structured feature vector includes: For the categorical data in the structured modal information, the categorical data is transformed by One-hot encoding; the categorical data after One-hot encoding is input into the dimension alignment layer; the dimension-aligned feature vector is sequentially input into the Embedding layer and the Dense layer, so that the Embedding layer extracts low-dimensional semantic features and the Dense layer extracts high-level feature representations as the structured feature vector of the categorical data; For the numerical data and / or relational data in the structured modal information, preliminary feature processing is completed through standardization, and the structured feature vectors of the numerical data and / or relational data are output.

7. The method for determining the root cause of equipment fault alarms according to claim 1, characterized in that, The step of fusing the text feature vector and the structured feature vector to generate a multimodal feature vector includes: The text feature vector and the structured feature vector are concatenated to obtain a joint feature vector; The joint feature vector is input into at least two consecutive linear mapping layers to output the multimodal feature vector; after processing by each linear mapping layer, a nonlinear transformation is performed using the ReLU activation function, and a Dropout strategy is applied; or The joint feature representation is input to the fusion linear mapping layer to fuse and reduce the feature dimensions, and output the multimodal feature vector. The fusion linear mapping layer uses a non-linear activation function to enhance the feature representation capability after fusion.

8. A device for determining the root cause of equipment fault alarms, characterized in that, The device includes: The acquisition module is used to acquire the text modal information and structured modal information of the fault alarm event when a fault alarm event is received; The feature extraction module is used to extract features from the text modal information to obtain a text feature vector; and to extract features from the structured modal information to obtain a structured feature vector; The feature fusion module is used to fuse the text feature vector and the structured feature vector to generate a multimodal feature vector; The prediction module is used to input the multimodal feature vector into a preset target prediction model, so that the target prediction model outputs the predicted alarm root cause corresponding to the fault alarm event; the prediction model is trained based on dynamic course learning.

9. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the device fault alarm root cause determination method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the device fault alarm root cause determination method according to any one of claims 1 to 7.

Images