A fault diagnosis method and apparatus that integrates alarm and trend event data

Abstract

This invention discloses a fault diagnosis method and apparatus that integrates alarm and trend event data. The method includes the following steps: preprocessing the raw time-series data of an industrial system to obtain preprocessed data; extracting multi-value alarm events and multi-class trend events, and fusing the two types of feature events to form a hybrid sequence; using the Word2Vec core model to vectorize and encode the events in the hybrid sequence; constructing an LSTM model; and employing an instance transfer learning strategy to update the model trained using source domain data to meet the needs of target domain diagnosis, thus achieving cross-domain fault diagnosis. This invention, by simultaneously extracting alarm and trend events, can more comprehensively capture the evolution of faults; through instance selection and weight adjustment, it can effectively solve the problem of scarce target domain data and improve diagnostic accuracy; the method can be extended to complex systems such as shipbuilding, petroleum, and chemical industries, and is compatible with different industrial scenarios. It can be adapted to new systems by adjusting the event definitions, making it widely applicable.

Description

Technical Field

[0001] This invention belongs to the field of industrial process monitoring and fault diagnosis technology, and in particular relates to a fault diagnosis method and device that integrates alarm and trend event data. Background Technology

[0002] With the increasing scale and complexity of industrial systems, fault diagnosis technology has become a crucial component in ensuring the safe, reliable, and stable operation of industrial production. In complex industrial processes and large equipment, due to the tight coupling between systems, a fault in one area can propagate widely along the system's connection path, leading to malfunctions or anomalies in other critical equipment or systems. For example, in a ship's system, a main engine failure can affect generators, transmission systems, and even navigation control systems, potentially causing the ship to lose power or control, and in severe cases, even triggering a maritime accident. Failure to diagnose and address these faults in a timely manner can result in serious consequences such as ship downtime, safety accidents, or environmental pollution. Therefore, timely and accurate fault diagnosis is of paramount importance for ensuring the safe operation of systems and reducing the risk of downtime.

[0003] Existing fault diagnosis methods typically focus only on changes in a single process variable, neglecting the interconnected effects of multidimensional features within the system. For example, relying solely on main engine temperature or vibration signals in a ship's system to determine the presence of a fault may overlook changes in other indicators such as fuel consumption and sluggish steering response, thus affecting the accuracy of the diagnostic results. Furthermore, different indicator parameters exhibit significantly different indicative characteristics, and existing methods fail to integrate transient events with slowly changing trends, making it difficult for current diagnostic schemes to meet the high reliability requirements of complex equipment (such as marine electrical or propulsion systems). Simultaneously, labeled fault samples are usually scarce, and fault events are infrequent and complex, making the collection of large numbers of labeled samples both expensive and time-consuming. Moreover, due to differences in equipment type, operating environment, and operating habits, there are often significant discrepancies between diagnostic data and training data, making it difficult for traditional fault diagnosis models to generalize across different systems or operating environments, impacting diagnostic accuracy and stability. Summary of the Invention

[0004] The problem this invention aims to solve is to provide a fault diagnosis method and device that integrates alarm and trend event data. This method overcomes the aforementioned limitations by combining alarm information, trend event characteristics, and multi-dimensional sensor data from industrial systems to construct a fault diagnosis model based on comprehensive features. This is of great significance for improving the accuracy of diagnostic results for complex equipment. At the same time, by combining techniques such as instance transfer learning, the model can perform well in the environment where the training data is located, while also adapting to environmental changes during real-time diagnosis, which is particularly important for practical applications in industrial processes.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a fault diagnosis method that integrates alarm and trend event data, comprising the following steps: S1: Preprocess the raw time series data of the industrial system to obtain preprocessed data; S2: Extract multi-value alarm events and multi-type trend events, and fuse the two types of feature events to form a mixed sequence; S3: Vectorize the events in the mixed sequence using the Word2Vec core model; S4: Based on the encoded data, construct a fault diagnosis LSTM model based on a long short-term memory network; S5: Employ an instance transfer learning strategy to update the model trained using source domain data to meet the diagnostic requirements of the target domain, thereby achieving cross-domain fault diagnosis.

[0006] Furthermore, in S1, wavelet analysis is used to remove noise from the original time series and normalize the data to the same scale.

[0007] Furthermore, in S2, the multi-value alarm events include alarm events that extract individual variables based on statistical distributions and integrate them into an alarm sequence in chronological order.

[0008] Furthermore, in S2, the trend event extraction includes segmenting the non-stationary time series into a linear representation, fitting the slope and intercept of each segment through linear regression, characterizing the trend change into multiple types of events, and generating a trend sequence.

[0009] Furthermore, in S3, the vectorized encoding adopts the Skip-Gram model, which maps mixed event labels to word vectors in a high-dimensional space through matrix operations; during the training process of the vectorized encoding, the loss function of the vectorized encoding is constructed by maximizing the log probability of observed positive samples and minimizing the log probability of randomly selected negative samples.

[0010] Furthermore, in S4, the gating structure of the LSTM model includes a forget gate, an input gate, and an output gate, which are used to capture long-term dependencies in the event sequence.

[0011] Furthermore, S5 includes the following steps: S51: Based on the improved Smith-Waterman algorithm, calculate the similarity between different source domain instances and the target domain, select instances from the source domain that are suitable for the target domain by setting a similarity threshold, and retrain the model after adjusting the sample weights. S52: The difference in distribution between the source domain and the target domain is measured by calculating the maximum mean difference index (MMD) between the distributions of instances in the source domain and instances in the target domain. S53: The loss during the transfer learning training process is evaluated by weighting the MMD loss and the target domain instance classification loss; S54: Model transfer is achieved by minimizing the weighted loss.

[0012] Furthermore, the present invention provides a fault diagnosis system that integrates alarm and trend event data, and implements the above-mentioned fault diagnosis method that integrates alarm and trend event data, including: The data acquisition module is used to acquire sensor data from industrial equipment in real time. The event extraction module is used to extract and merge alarm events and trend events; The vector encoding module is used to implement word embedding representation of event data; The model training module is used to build and train LSTM diagnostic models; The migration adaptation module is used to enable instance migration of cross-domain models.

[0013] Furthermore, the present invention provides an apparatus including a memory, a processor, and an algorithm stored in the memory and executable on the processor, wherein the processor implements the above-described data processing method when executing a computer program.

[0014] Furthermore, the present invention provides a computer-readable storage medium storing a computer algorithm, which, when executed by a processor, performs the aforementioned data processing.

[0015] The advantages and positive effects of this invention are: This invention can capture the evolution of faults more comprehensively by simultaneously extracting alarm events (such as exceeding limits and sudden changes) and trend events (such as slow rise and step drop). Through instance selection and weight adjustment, it can effectively solve the problem of scarce target domain data and improve diagnostic accuracy. This method can be extended to complex systems such as ships, petroleum, and chemical industries, and is compatible with different industrial scenarios. It can be adapted to new systems by adjusting the event definition (such as adding equipment log text), and has a wide range of applications. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention.

[0017] Figure 2 This is a flowchart illustrating the overall steps of model construction in an embodiment of the present invention.

[0018] Figure 3 This is a confusion matrix diagram of four sets of experiments in an embodiment of the present invention.

[0019] Figure 4 This is an overall structural diagram of an embodiment of the device of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The embodiments of the present invention will be further described below with reference to the accompanying drawings: like Figure 1 , Figure 2 As shown, a fault diagnosis method that integrates alarm and trend event data includes the following steps.

[0022] S1: Preprocess the raw time series data of the industrial system to obtain preprocessed data. This time series data preprocessing includes noise removal and normalization of the multivariate time series. Specifically, this embodiment uses wavelet analysis to remove noise from the raw time series and then normalizes the data to the same scale.

[0023] S2: Extract multi-valued alarm events and multi-class trend events, and fuse the two types of feature events to form a hybrid sequence. Specifically, multi-valued alarm events include alarm events extracted based on single variables according to statistical distribution, and integrated into an alarm sequence in chronological order. Trend event extraction includes: performing piecewise linear representation on non-stationary time series, fitting the slope and intercept of each segment through linear regression; qualitatively characterizing trend changes as multi-class events, and generating a trend sequence.

[0024] Specifically, for multi-valued alarm event extraction, the alarm events for each individual process variable are first extracted, and then all alarm events are arranged and integrated in chronological order to form an alarm sequence. Assuming the alarm system monitors m process variables, and an alarm event is defined as a variable triggering a certain alarm category at a specific time, then... To express.

[0025] in, The time when the alarm occurs. Alarm category ( ), This represents the number of alarm events for a single process variable during a fault.

[0026] The total number of alarm events generated by all variables is .

[0027] An alarm sequence consisting of all alarm events is defined as follows: .

[0028] Extracting multiple types of trend events, preprocessing the data, and determining the stationarity of data for different variables; converting time series data... If the data is divided into M contiguous data segments, then the m-th data segment will be represented as: ,in, and Let these be the first and last data points of this segment. Assume a linear regression model is used for each data segment, approximating the result as follows: .

[0029] in, , and The slope and intercept are respectively obtained by least squares estimation; the trend sequence composed of all extracted trend events is defined as... .

[0030] The above steps yield alarm events and trend events. These two types of characteristic events are then fused to obtain a mixed sequence. A characteristic event is defined as a variable undergoing a certain characteristic change at a specific moment, including alarm and trend changes. To express.

[0031] in, The time when the characteristic change occurs. For feature categories, , The total number of feature events contained in the mixed sequence is denoted as . .

[0032] In the formula, n is the total number of alarm events that occurred; Let be the i-th alarm event in the alarm flooding sequence X.

[0033] S3: Vectorization encoding method for event data. It mainly considers that alarm and trend event data are semi-structured data, which are usually not suitable for direct use in building data-driven fault diagnosis models. Therefore, the Word2Vec core model Skip-Gram model is used to map mixed event labels into word vectors in high-dimensional space through matrix operations.

[0034] During the training process of vectorized encoding, the loss function of vectorized encoding is constructed by maximizing the log probability of observed positive samples while minimizing the log probability of randomly selected negative samples.

[0035] Specifically, for a mixed sequence containing V unique alarm and trend event labels (hereinafter collectively referred to as characteristic events) Each event in the sequence is represented using one-hot encoding. The position corresponding to the i-th variable is "1", and the other positions are "0". At this point, the entire sequence is initially encoded as a V*V dimensional sparse matrix. The a-th feature event in the mixed sequence is... This means mapping it to a high-dimensional N-dimensional embedding space. This mapping process involves the weight matrix between the input layer and the hidden layer. By calculating the inner product of the weight matrix W and the one-hot vector corresponding to each feature event, we obtain the N-dimensional encoding vector corresponding to each row in W. ,in Hidden layer weights are used This indicates that the matrix between the hidden layer and the output layer is... ,in In this way, the representation of each variable can be obtained through corresponding matrix operations, thereby realizing the transformation from the original event label to the vectorized representation, providing a foundation for further data analysis and model training.

[0036] The formula for calculating each variable is as follows: .

[0037] in, Representation matrix The j-th line.

[0038] The model's loss function aims to maximize the log probability of observed positive samples while minimizing the log probability of randomly selected negative samples. The formula for the loss function E is defined as follows: in, For output variables (i.e., positive samples); It is the corresponding encoded vector; The Softmax function is used to compute the output of a given hidden layer. When, the output probability of a positive sample and a set of negative samples; T represents the transpose operation. Includes positive samples Negative samples were obtained by sampling from the distribution. To represent characteristic events The encoded vector, then the mixed sequence Will be updated to .

[0039] S4: Based on the encoded data, construct a fault diagnosis model based on a Long Short-Term Memory (LSTM) network.

[0040] The method for constructing a fault diagnosis model by integrating multi-feature event data mainly considers that faults may be caused by the accumulation of complex events, which may be far apart in time. Therefore, a Long Short-Term Memory (LSTM) network is used to construct the fault diagnosis model. After the mixed events are represented by data vectorization, they are mapped to the interval (-1, 1) through the Tanh activation function. For an LSTM neuron at time t, its three inputs are: the latent memory state of the neuron at time t-1. Hidden state output Vectorized encoding of alarm event data at time t The corresponding output is the memory state of the neural unit at the current moment. and hidden state output The operation of an LSTM cell involves several key steps, each controlled by a different gating structure, including forget gates, input gates, and output gates, as well as cell state updates to capture long-term dependencies in a sequence of events.

[0041] S5: Employing an instance transfer learning strategy, the model trained using source domain data is updated to meet the diagnostic requirements of the target domain, achieving cross-domain fault diagnosis. Specifically, the instance transfer learning strategy includes: The model is retrained after calculating the similarity between different source domain instances and the target domain based on the improved Smith-Waterman algorithm, selecting instances from the source domain that fit the target domain by setting a similarity threshold, adjusting the sample weights, and then retraining the model.

[0042] The difference in distribution between the source and target domains is measured by calculating the maximum mean difference (MMD) metric between the distributions of instances in the source and target domains.

[0043] The loss during transfer learning training is evaluated by weighting the MMD loss and the target domain instance classification loss.

[0044] Model transfer is achieved by minimizing the weighted loss.

[0045] Specifically, given the high difficulty and cost of obtaining large-scale and accurately labeled datasets, an instance migration strategy for fault diagnosis models is proposed. Before starting the migration task, it is necessary to measure the distributional differences between the source and target domains. Domain difference measurement is the core concept of instance migration, and the Maximum Mean Discrepancy (MMD) is adopted, with the following formula: in, It is a domain difference measure, defined as the difference between all mapped functions in a given function class. The upper bound (maximum value) of the absolute value of the difference between the expected values ​​in the source domain and the target domain. S and T represent the data distributions in the source and target domains, respectively, used to map the data to a higher-dimensional space, and E represents the expectation. and These are samples drawn from the source domain and the target domain, respectively. The objective function of transfer learning aims to minimize the loss in the target domain and the difference between the source and target domains. The general form of the objective function is: in, In the hypothesis space The optimal function in ; It is a loss function; It is a regularization parameter; The actual label or target value.

[0046] Instance selection strategies aim to select the most useful data instances from the source domain for learning the target domain. These strategies primarily include two types: class weight adjustment and sample relabeling. Currently, the most commonly used method is weight-based instance selection, which involves finding a weight vector... To minimize the empirical risk in the target domain, the formula is expressed as: in, It is the optimal weight vector; It is the number of samples in the source domain; and These are the i-th mixed sequence in the source domain and its label, respectively; It is a loss function; It is a learning model; It is a regularization term; It is a regularization parameter; The weight assigned to the i-th mixed sequence in the source domain indicates its importance to the target task. For instance selection, an improved Smith-Waterman algorithm is used, which determines suitable instances for transfer learning by setting a similarity threshold.

[0047] The diagnostic performance of the model was evaluated based on accuracy, macro-F1 score, and confusion matrix.

[0048] The following section uses the design and simulation experiment of the 3W fault diagnosis dataset from Petrobras, Brazil's state-owned oil company, as an example to illustrate the invention in detail: The dataset includes eight types of faults, as shown in Table 1.

[0049] Table 1. Eight Fault Types in the 3W Dataset The known data (source domain) is defined as subsets 1, 11, and 12, and the unknown data (target domain) is defined as subsets 2, 3, 5, 9, 10, and 13. A fault diagnosis model is constructed from the source domain, and instance migration is used to effectively distinguish faults in the target domain. The fault selection is shown in Table 2.

[0050] Table 2. Fault selection of source and target domains and corresponding dataset numbers Specifically, it includes the following steps: Event extraction and encoding: Extract fault data alarms and trend events from the source and target domains, and use the Skip-gram model to embed words into the events, with the word vector dimension set to 50.

[0051] Diagnostic Model Construction: The overall steps for constructing a multi-feature event encoding and fault diagnosis model are as follows: Figure 1 As shown, based on the vector form of event data, decision tree, SVM, RNN and LSTM fault diagnosis models are constructed, their performance is compared, and the best model is selected as the transfer learning basis.

[0052] Instance migration: Apply the optimal model to diagnose faults in the target domain, including instance migration of no migration, single fault, and double fault types, and evaluate performance to verify the effectiveness of the method.

[0053] Five-fold cross-validation was used to compare the performance of four different machine learning and deep learning models (decision tree, SVM, RNN, and LSTM) on fault diagnosis tasks. The experimental results are shown in Table 3. Compared with decision tree, SVM, and RNN, LSTM showed higher diagnostic accuracy, all above 0.9. This result may be due to the correlation between event data, and the limited ability of decision tree and SVM, due to their relatively simple model structure, to capture the complex features of time series data.

[0054] Instance transfer was performed using an LSTM-based model to diagnose faults in the target domain. To better compare the instance transfer performance, four control experiments were conducted: Experiment 1: Migrate without selecting any instances from the source domain; Experiment 2: Select F2 instances in the source domain and target domain that meet the similarity threshold for migration; Experiment 3: Select F4 instances from the source domain and the target domain that meet the similarity threshold for migration; Experiment 4: Select the F2 and F4 instances that have met the similarity threshold above for migration.

[0055] Table 3. Experimental comparison of different algorithms on the source domain dataset. For Experiment 1, the pre-trained LSTM model was directly tested on 7 F2 and 15 F4 samples in the target domain. For Experiments 2, 3, and 4, similarity matrices were obtained through sequence alignment, and a similarity threshold of 0.7 was set. Sample instances meeting the similarity threshold were selected from the source domain, along with 1 / 3 of the F2 and F4 samples (2 F2 and 5 F4) from the target domain, as training data. The sample weights for the source and target domains were set to 0.7 and 1, respectively. The pre-trained LSTM model was retrained and adjusted, and the remaining 5 F2 and 10 F4 samples in the target domain were used for testing. Specific experimental results are shown in Table 4, and the corresponding confusion matrix results are as follows: Figure 3 As shown.

[0056] Table 4. Experimental comparison of different migration methods The experimental results above show that the Word2Vec+LSTM modeling method achieved the highest accuracy and macro-F1 score. After introducing source domain instances for transfer learning, the overall accuracy of the model increases with the increase of source domain fault types. Furthermore, the model can effectively identify and learn the features of specific fault types and, by integrating multi-class feature knowledge from the source domain, achieve simultaneous adaptation to multiple fault types, demonstrating good adaptability.

[0057] Therefore, it can be concluded that the proposed multi-feature event encoding and fault diagnosis model construction method is very effective in improving the model's adaptability, significantly improving the model's ability to diagnose specific fault types and achieving a high diagnostic accuracy.

[0058] As an optional implementation, this embodiment provides a fault diagnosis device that integrates alarm and trend event data, referencing... Figure 4 The device includes the following modules: The data acquisition module is used to acquire sensor data from industrial equipment in real time. The event extraction module is used to extract and merge alarm events and trend events; The vector encoding module is used to implement word embedding representation of event data; The model training module is used to build and train LSTM diagnostic models; The migration adaptation module is used to enable instance migration of cross-domain models.

[0059] The advantages and positive effects of this invention are: This invention can capture the evolution of faults more comprehensively by simultaneously extracting alarm events (such as exceeding limits and sudden changes) and trend events (such as slow rise and step drop). Through instance selection and weight adjustment, it can effectively solve the problem of scarce target domain data and improve diagnostic accuracy. This method can be extended to complex systems such as ships, petroleum, and chemical industries, and is compatible with different industrial scenarios. It can be adapted to new systems by adjusting the event definition (such as adding equipment log text), and has a wide range of applications.

[0060] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A fault diagnosis method that integrates alarm and trend event data, characterized in that: Includes the following steps, S1: Preprocess the raw time series data of the industrial system to obtain preprocessed data; S2: Extract multi-value alarm events and multi-type trend events, and fuse the two types of feature events to form a mixed sequence; S3: Vectorize the events in the mixed sequence using the Word2Vec core model; S4: Based on the encoded data, construct a fault diagnosis LSTM model based on a long short-term memory network; S5: Employ an instance transfer learning strategy to update the model trained using source domain data to meet the diagnostic requirements of the target domain, thereby achieving cross-domain fault diagnosis.

2. The fault diagnosis method that integrates alarm and trend event data according to claim 1, characterized in that: In S1, wavelet analysis is used to remove noise from the original time series and normalize the data to the same scale.

3. A fault diagnosis method that integrates alarm and trend event data according to claim 1 or 2, characterized in that: In S2, the multi-value alarm events include alarm events that extract individual variables based on statistical distributions and integrate them into an alarm sequence in chronological order.

4. A fault diagnosis method that integrates alarm and trend event data according to claim 1 or 2, characterized in that: In S2, the trend event extraction includes segmenting the non-stationary time series into a linear representation, fitting the slope and intercept of each segment through linear regression, characterizing the trend change into multiple types of events, and generating a trend sequence.

5. A fault diagnosis method that integrates alarm and trend event data according to claim 1 or 2, characterized in that: In S3, the vectorized encoding adopts the Skip-Gram model, which maps mixed event labels to word vectors in a high-dimensional space through matrix operations. During the training process of the vectorized encoding, the loss function of the vectorized encoding is constructed by maximizing the log probability of observed positive samples and minimizing the log probability of randomly selected negative samples.

6. A fault diagnosis method that integrates alarm and trend event data according to claim 1 or 2, characterized in that: In S4, the gating structure of the LSTM model includes a forget gate, an input gate, and an output gate, which are used to capture long-term dependencies in the event sequence.

7. A fault diagnosis method that integrates alarm and trend event data according to claim 1 or 2, characterized in that: S5 includes the following steps: S51: Based on the improved Smith-Waterman algorithm, calculate the similarity between different source domain instances and the target domain, select instances from the source domain that are suitable for the target domain by setting a similarity threshold, and retrain the model after adjusting the sample weights. S52: The difference in distribution between the source domain and the target domain is measured by calculating the maximum mean difference index (MMD) between the distributions of instances in the source domain and instances in the target domain. S53: The loss during the transfer learning training process is evaluated by weighting the MMD loss and the target domain instance classification loss; S54: Model transfer is achieved by minimizing the weighted loss.

8. A fault diagnosis system that integrates alarm and trend event data, characterized in that: The fault diagnosis method for integrating alarm and trend event data as described in any one of claims 1 to 7 includes, The data acquisition module is used to acquire sensor data from industrial equipment in real time. The event extraction module is used to extract and merge alarm events and trend events; The vector encoding module is used to implement word embedding representation of event data; The model training module is used to build and train LSTM diagnostic models; The migration adaptation module is used to enable instance migration of cross-domain models.

9. An apparatus comprising a memory, a processor, and an algorithm stored in the memory and executable on the processor, characterized in that: When the processor executes a computer program, it implements the data processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer algorithm, characterized in that, When the computer algorithm is executed by the processor, it performs the data processing as described in any one of claims 1 to 7.

Images