Alarm flooding data enhancement and root diagnosis method for unbalanced samples

Abstract

The invention discloses an unbalanced sample-oriented alarm flooding data enhancement and root diagnosis method, which comprises the following steps of: acquiring alarm event data generated in an industrial system; the improved SMOTENC algorithm performs data enhancement on minority class samples, and a synthetic alarm sample having highly similar sequential logic with an original fault sample is generated in a feature space to balance a data set; constructing a group of hidden Markov model variant classifiers with different structures; evaluating the performance of each variant classifier by adopting different indexes; and taking different performance indexes as references, fusing diagnosis results of a plurality of classifiers by adopting an integrated learning framework of a weighted voting method, and finally outputting an alarm source diagnosis result with high confidence in real time. By improving the data enhancement method, the problem of sample imbalance in the industrial alarm data can be effectively solved; the advantages of different HMM variants are integrated through an innovative model integration framework, and the accuracy of alarm flooding root diagnosis can be remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of alarm flooding root cause diagnosis technology in complex industrial systems, and in particular relates to an alarm flooding data augmentation and root cause diagnosis method for imbalanced samples. Background Technology

[0002] The primary function of a monitoring and alarm system is to assess the current production status using real-time and historical data, and generate alarm signals when abnormal conditions are detected, alerting operators to take timely measures to prevent further deterioration. However, in today's large-scale, complex industrial systems, these alarm systems commonly suffer from problems such as numerous and frequent interfering alarms, leading to excessive workload for operators and impacting emergency response efficiency. This is particularly true in the marine industry, where the normal operation of electrical and propulsion systems directly affects the safe navigation and overall operational stability of the vessel. The electrical system provides reliable power to various equipment, while the propulsion system is the core for propulsion and critical operations. Due to the complex structure and harsh operating environment of marine systems, alarm overload is especially prominent, severely affecting operators' ability to identify and handle critical faults, and potentially leading to serious safety accidents.

[0003] The causes of alarm overload in marine electrical and propulsion systems are diverse and vary in complexity. For example, short-circuit faults and overload faults in marine electrical systems, as well as engine and propulsion system faults in propulsion systems, can all trigger a large number of alarm signals. In particular, the diagnosis of the root cause of alarm overload triggered by faults is extremely difficult, directly impacting the operators' immediate response and decision-making capabilities. Therefore, researching efficient methods for diagnosing the root causes of alarm overload is of significant scientific theoretical importance and industrial practical value in order to improve the performance of monitoring and alarm systems in marine electrical and propulsion systems.

[0004] Currently, alarm flooding root cause analysis mainly falls into two categories: pattern matching-based methods and classification model-based methods. Pattern matching identifies known alarm flooding types by analyzing historical data, and then uses techniques such as similarity evaluation and cluster analysis to match and identify the root causes of alarm flooding. Classification model-based methods, on the other hand, pre-train diagnostic models to classify the root causes of alarm flooding. In recent years, machine learning has been widely used in alarm flooding root cause diagnosis. Among these methods, Factor Hidden Markov Model (FHMM), Hidden Semi-Markov Model (HSMM), and Discrete Hidden Markov Model (DHMM) have shown promising application prospects in the field of alarm flooding root cause analysis due to their strong interpretability and other characteristics.

[0005] While these HMM variant classifiers meet the interpretability requirements, the difficulty in acquiring high-quality alarm flood data, coupled with the extreme scarcity of data for specific fault types leading to data imbalance, means that using a single model for alarm flood root cause diagnosis may significantly reduce the model's diagnostic ability for minority faults. This increases the risk of overfitting and results in insufficient generalization. This problem is particularly pronounced in marine electrical and propulsion systems, where the complex operating environment and frequent environmental changes lead to various over-limit faults, while aging-related failures are less common. This imbalance in data across different fault types, coupled with high data acquisition and storage costs, further exacerbates the problem. Therefore, designing an effective alarm flood root cause diagnosis method that can handle imbalanced samples is crucial for improving the performance of ship monitoring and alarm systems. Summary of the Invention

[0006] The problem this invention aims to solve is to provide a data augmentation and root cause diagnosis method for alarm flooding based on imbalanced samples. This method addresses the low accuracy problem caused by sample imbalance and the limitations of a single model in alarm flooding diagnosis, and achieves highly robust root cause fault identification.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for alarm flooding data augmentation and root cause diagnosis for imbalanced samples, comprising the following steps: S1: Acquire alarm event data generated in the industrial system; S2: Improve the SMOTENC algorithm to perform data augmentation on minority class samples, and generate synthetic alarm samples with highly similar temporal logic to the original fault samples in the feature space to balance the dataset; S3: Construct a set of Hidden Markov Model (HMM) variant classifiers with different structures, including Factor Hidden Markov Model (FHMM), Hidden Semi-Markov Model (HSMM), and Discrete Hidden Markov Model (DHMM). S4: Evaluate the performance of each variant classifier using different metrics; S5: Using different performance metrics as benchmarks, an ensemble learning framework employing weighted voting integrates the diagnostic results of multiple classifiers, ultimately outputting alarm root cause diagnostic results with high confidence in real time.

[0008] Furthermore, step S2 includes the following steps: S21: For each minority class sample, based on the preset feature similarity measurement criteria, identify the K nearest neighbor samples in its class to form its local neighborhood. S22: Randomly select one of the K nearest neighbor samples as the synthetic benchmark for generating new samples together with the current minority class samples; S23: For each discrete feature to be generated, a direct selection strategy or a frequency sampling strategy is used to determine its feature value; S24: Repeat steps S21-S23 until the number of synthesized samples reaches the preset balance ratio, and integrate all synthesized samples with the original dataset.

[0009] Furthermore, S3 includes the following steps: S31: FHMM classifier construction: By combining multiple independent Markov chains in parallel, the multi-dimensional processing capability of the model is enhanced, the forward and backward probabilities of different alarm sequences are calculated, and the fault state of the system is analyzed and determined by combining the state transition characteristics of the system and the generation probability of alarms. S32: HSMM Classifier Construction: Adding the concept of state duration allows for consideration of situations where the state remains unchanged over a period of time. By distinguishing between macroscopic and microscopic states, it provides an effective analytical tool for handling scenarios where the state duration changes significantly. S33: DHMM classifier construction: Assuming that the state follows an unobservable Markov process, determined by the probability vector of the initial state, the state transition probability matrix, and the observation probability matrix, the model parameters are adjusted through training samples to maximize the probability of alarm flooding sequences occurring.

[0010] Furthermore, the parameters of the FHMM classifier are defined as follows: Wherein, the state transition matrix A is the state transition matrix of each independent Markov chain. The transition matrix has N layers, and the transition probability of the i-th layer is... B represents the observation probability matrix, and the probability distribution of independently generated observations at each layer is as follows: π represents the initial state distribution, and the initial probability vector for each layer is... N represents the hidden state dimension; The probability of generating observation sequences in FHMM is: in, For all layers, O represents the hidden state sequence; T represents the time length of the sequence; and N represents the hidden state dimension. Let be the state transition probability of the i-th layer; Let be the observation probability of the i-th layer.

[0011] Furthermore, the parameter set of the HSMM classifier is as follows: in, Represents the state duration matrix, Let i be the probability that state i lasts for d time units. The probability of generating the observed sequence in HSMM is: in, It is a macroscopic state sequence; It is a duration sequence; Initial state The probability of; Indicates from state Transition to state The transition probability; In the state Generate observations The probability of observation.

[0012] Furthermore, the parameter set of the DHMM classifier is as follows: The probability of generating the observed sequence in DHMM is: in, For a moment Forward probability ( For the current moment, (Total length of the observation sequence), recursive formula: in, State transition probability matrix From the state Transition to state The probability of; The observation probability matrix In the state Generate observations The probability of.

[0013] Furthermore, in S4, the performance of each variant classifier is evaluated by calculating macro-precision, macro-recall, and macro-F1 score.

[0014] Furthermore, in S5, a weighted voting method is used to integrate the diagnostic results of three different HMM variant classifiers. The weight coefficients of each classifier are not fixed values, but are dynamically calculated based on their macro F1 scores on the cross-validation set. Higher weights are assigned to classifiers that perform better in specific test scenarios, thereby achieving complementary advantages and decision optimization.

[0015] Furthermore, the present invention also provides an alarm flood data augmentation and root cause diagnosis device for imbalanced samples, which, when running the above-described alarm flood data augmentation and root cause diagnosis method for imbalanced samples, includes: The data acquisition module is used to acquire alarm event data generated in industrial systems, either offline or in real time. The data augmentation module is used to augment minority class samples using an improved SMOTENC algorithm, generating new synthetic samples to balance the dataset. The classifier building module is used to build FHMM, HSMM, and DHMM classifiers; The ensemble learning module is used to integrate the diagnostic results of multiple classifiers using a weighted voting ensemble learning framework, thereby improving the model's generalization ability and robustness.

[0016] Furthermore, the present invention also provides a computer-readable storage medium storing a computer algorithm, which, when executed by a processor, performs the above-described data processing.

[0017] The advantages and positive effects of this invention are: This invention effectively solves the problem of sample imbalance in industrial alarm data by improving data augmentation methods; by combining the advantages of different HMM variants through an innovative model integration framework, it can significantly improve the accuracy of diagnosing the root causes of alarm proliferation. The technical solution of this invention is particularly suitable for complex operating conditions such as marine electrical systems and power systems, and has important scientific theoretical significance and industrial practical value. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of an embodiment of the method of the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of an embodiment of the device of the present invention.

[0020] Figure 3 This is a schematic diagram of the data sample size under various faults after the enhancement of the embodiments of the present invention.

[0021] Figure 4 This is a schematic diagram of the original data sample size under various faults in the embodiments of the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the principle of FHMM in an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram illustrating the principle of HSMM in an embodiment of the present invention.

[0024] Figure 7 This is a confusion matrix diagram obtained by FSDHMM after data enhancement in an embodiment of the present invention.

[0025] Figure 8 This is a confusion matrix diagram obtained from the original sample using FSDHMM in an embodiment of the present invention. Detailed Implementation

[0026] 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.

[0027] The embodiments of the present invention will be further described below with reference to the accompanying drawings: like Figure 1 As shown, a method for alarm flooding data augmentation and root cause diagnosis for imbalanced samples includes the following steps: S1: Acquire alarm event data generated in the industrial system.

[0028] S2: The improved SMOTENC algorithm performs data augmentation on minority class samples, generating synthetic alarm samples with highly similar temporal logic to the original fault samples in the feature space to balance the dataset. Specifically, S2 includes the following steps: S21: Neighborhood determination step: For each minority class sample, based on the preset feature similarity measurement criteria, identify K nearest neighbor samples in its class to form its local neighborhood.

[0029] S22: Synthesis benchmark selection step: Randomly select one from the K nearest neighbor samples as the synthesis benchmark for generating new samples together with the current minority class samples.

[0030] S23: New sample feature construction steps: For each discrete feature to be generated, use a direct selection strategy or a frequency sampling strategy to determine its feature value.

[0031] S24: Iteration completion step: Repeat steps S21-S23 until the number of synthesized samples reaches the preset balance ratio, and integrate all synthesized samples with the original dataset.

[0032] S3: Construct a set of Hidden Markov Model (HMM) variant classifiers with different structures, including Factor Hidden Markov Model (FHMM), Hidden Semi-Markov Model (HSMM), and Discrete Hidden Markov Model (DHMM), and integrate the diagnostic results of these classifiers through ensemble learning methods to improve the accuracy and robustness of the model. Specifically, S3 includes the following steps: S31: FHMM Classifier Construction: By combining multiple independent Markov chains in parallel to enhance the multi-dimensional processing capability of the model, the forward and backward probabilities of different alarm sequences are calculated. Combining the system's state transition characteristics and alarm generation probabilities, the system's fault state is analyzed and determined. Its parameters are defined as follows: Wherein, the state transition matrix A is the state transition matrix of each independent Markov chain. The transition matrix has N layers, and the transition probability of the i-th layer is... B represents the observation probability matrix, and the probability distribution of independently generated observations at each layer is as follows: π represents the initial state distribution, and the initial probability vector for each layer is... N represents the hidden state dimension.

[0033] The probability of generating observation sequences in FHMM is: in, For all layers, O represents the hidden state sequence; O represents the observed alarm sequence; T represents the time length of the sequence; and N represents the hidden state dimension. Let be the state transition probability of the i-th layer; Let be the observation probability of the i-th layer.

[0034] S32: HSMM Classifier Construction: The concept of state duration is added, allowing consideration of situations where states remain unchanged over a period of time. By distinguishing between macroscopic and microscopic states, it provides an effective analytical tool for scenarios with significantly varying state durations, overcoming the limitations of the Markov chain assumption in HMMs. HSMM introduces the state duration characteristic, with the following parameter set: in, Represents the state duration matrix, Let be the probability that state i lasts for d time units.

[0035] The probability of generating the observed sequence in HSMM is: in, It is a macroscopic state sequence; It is a duration sequence; Initial state The probability of; Indicates from state Transition to state The transition probability; In the state Generate observations The probability of observation.

[0036] S33: DHMM Classifier Construction: Assuming the state follows an unobservable Markov process, determined by the initial state probability vector, state transition probability matrix, and observation probability matrix, the model parameters are adjusted using training samples to maximize the probability of alarm flooding sequences occurring. DHMM is a discrete observation version of the classic HMM, with the following parameter set: The probability of generating the observed sequence in DHMM is: in, For a moment Forward probability ( For the current moment, (Total length of the observation sequence), recursive formula: in, State transition probability matrix From the state Transition to state The probability of; The observation probability matrix In the state Generate observations The probability of.

[0037] S4: Evaluate the performance of each variant classifier using different metrics. Specifically, evaluate the performance of each variant classifier by calculating macro-precision, macro-recall, and macro-F1 score.

[0038] S5: Using different performance metrics as benchmarks, an ensemble learning framework employing weighted voting integrates the diagnostic results of multiple classifiers, ultimately outputting high-confidence alarm root cause diagnostic results in real time. Specifically, a weighted voting method is used to integrate the diagnostic results of three HMM variant classifiers with different structures; where the weight coefficients of each classifier are not fixed values, but are dynamically calculated based on their macro F1 scores on the cross-validation set, allocating higher weights to classifiers that perform better in specific test scenarios, thereby achieving complementary advantages and decision optimization.

[0039] Specifically, this invention constructs an ensemble learning framework based on weighted voting, which performs weighted voting on the classification results of FHMM, HSMM, and DHMM. By assigning higher weights to more accurate models, the judgments of each classifier are aggregated to optimize the decision-making process.

[0040] Diagnostic probability output: for each classifier ( The model parameters are estimated using either the EM algorithm or the Baum-Welch algorithm. ,maximize For a given observation sequence O, calculate its relationship to the fault category. The posterior probability is: in, This represents the total number of fault categories. Let be the prior probability of the category.

[0041] Weighting: Weight With classifier The validation set performance is positively correlated, and the specific calculation is as follows: Among them, Macro-F1 i For classifier The Macro-F1 score on the validation set is calculated using the following formula: Macro-Precision and Macro-Recall represent macro precision and macro recall, calculated using the following formulas: in, , , , respectively, represent the number of true positives, false positives, and false negatives in class j.

[0042] Final Diagnostic Category Decided by weighted voting: Ultimately, the fault category that maximizes the sum of weighted probabilities is selected. If the results are equal, the output of FHMM will be selected first, mainly because it has stronger multidimensional modeling capabilities.

[0043] This invention selects accuracy, as well as Macro-Precision, Macro-Recall, and Macro-F1, as indicators to evaluate the diagnostic performance of the integrated model for root cause faults of alarm proliferation.

[0044] To implement the alarm flood data augmentation and root cause diagnosis method for imbalanced samples proposed in this invention, this embodiment also designs an alarm flood data augmentation and root cause diagnosis device for imbalanced samples, such as... Figure 2 As shown, its structure and function are described below: The data acquisition module is used to acquire alarm event data generated in industrial systems, either offline or in real time.

[0045] Data Augmentation Module: Performs data augmentation on minority class samples using an improved SMOTENC algorithm.

[0046] Classifier building module: Builds HMM variant classifiers based on FHMM, HSMM, and DHMM.

[0047] Ensemble learning module: An ensemble learning framework using weighted voting integrates the diagnostic results of multiple classifiers and outputs the final fault diagnosis result.

[0048] The present invention will now be described in detail with reference to specific embodiments: Using a publicly available VAM industrial simulation model, multiple fault types were actively triggered to obtain the required alarm flooding dataset. This embodiment selected 13 fault types, collecting 20 historical alarm event logs for 4 of these fault types, 10 for 5 of them, and 3 for the last 4. The resulting alarm flooding sequences revealed a significant sample imbalance in the original dataset.

[0049] S1: For the collected alarm event data, firstly, extract the alarm flood sequence, and ensure the alarm flood length is consistent by truncation. Then, perform data augmentation and integrated classification. The overall steps are as follows: Figure 1 As shown.

[0050] S2: To address the issue of imbalanced alarm data samples, an improved SMOTENC algorithm is used for data augmentation. The specific steps are as follows: Synthetic sample generation: For discrete alarm event features, sampling is performed directly based on the discrete feature values ​​of existing minority class samples.

[0051] Find the nearest neighbor sample: For each selected minority class sample, find K nearest neighbor samples in its feature space (K=5 in this embodiment), and evaluate them based on the matching degree or similarity of feature values.

[0052] Generating new samples: For a selected minority class sample and its randomly selected nearest neighbor sample, the sample is determined by directly selecting a value from the selected sample or its nearest neighbor sample, or by probability sampling based on the co-occurrence frequency of the feature value.

[0053] Repeat the above process until a sufficient number of synthetic samples are generated for the minority class to achieve the predetermined oversampling ratio. The resulting augmented data has better balance than the original data, such as... Figure 3 , Figure 4 As shown.

[0054] S3: Then construct three variant classifiers of HMM and perform ensemble classification.

[0055] FHMM classifier: The principle of constructing an FHMM classifier is as follows: Figure 5 As shown, the multi-dimensional processing capability of the model is enhanced by combining multiple independent Markov chains in parallel. Assuming the dataset contains 3 layers of Hidden Markov Chains, each with 5 states, then the entire system contains 5... 3 A state transition matrix with three dimensions is used. The FHMM classifier is trained using the training dataset. By calculating the forward and backward probabilities of different alarm sequences, and combining the system's state transition characteristics with the alarm generation probability, the system's fault state is analyzed and determined.

[0056] HSMM Classifier: The principle of constructing an HSMM classifier is as follows: Figure 6 As shown, the concept of state duration is added, allowing for the consideration of situations where the state remains unchanged over a period of time. The maximum state duration is set to 10 time units, and the state duration probability matrix is ​​calculated based on actual data. The HSMM classifier is trained using the training dataset to improve the diagnostic capability for scenarios with significant changes in state duration.

[0057] DHMM Classifier: Construct a DHMM classifier, assuming the states follow an unobservable Markov process. Train the DHMM classifier using a training dataset, adjusting the model parameters to maximize the probability of alarm flooding sequences occurring. Assuming the dataset contains 5 possible fault states and 10 possible alarm types, obtain the optimal state transition matrix and observation probability matrix through training.

[0058] An ensemble learning framework based on weighted voting was constructed to integrate the diagnostic results of FHMM, HSMM, and DHMM classifiers. The performance of each classifier was tested using 10-fold cross-validation, and Table 1 shows the calculated Macro-F1 scores for the 10-fold cross-validation. Weights were assigned based on accuracy: FHMM weight was 40%, HSMM weight was 35%, and DHMM weight was 25%.

[0059] Table 1. Accuracy and weight allocation of each classifier The ensemble model FSDHMM, obtained by integrating the three classifiers FHMM, HSMM, and DHMM, was tested using augmented samples. The results show that FSDHMM has an accuracy of 93.8%, a macro-precision of 94.7%, a macro-recall of 93.8%, and a macro-F1 score of 93.4%. The confusion matrix is ​​attached. Figure 7 As shown, the samples on the diagonal are correctly classified, indicating that the model has high classification accuracy and robustness. It can be seen that the technical solution proposed in this invention has excellent diagnostic performance and can effectively solve the problem of diagnosing the root causes of imbalanced sample alarms.

[0060] The diagnostic performance of three models—FHMM, HSMM, and DHMM—was also tested separately in the experiment and compared with the FSDHMM proposed in this invention. Table 2 shows the evaluation metrics for root cause diagnosis using each method on the data-augmented sample set.

[0061] Table 2 Evaluation metrics of different methods on the data-augmented sample set To verify the effectiveness of the data augmentation method, the performance of FSDHMM was tested using the original imbalanced sample set. The test results show that FSDHMM achieves an accuracy of 76.6%, a macro-precision of 75.3%, a macro-recall of 76.6%, and a macro-F1 score of 73.4% on the original imbalanced sample set. The confusion matrix is ​​attached. Figure 8 As shown, the model has a weak ability to classify minority class samples.

[0062] The above experiments confirm that the alarm flooding data enhancement and root cause diagnosis method proposed in this invention can effectively solve the problems of difficult root cause diagnosis of alarm flooding and data sample imbalance in existing monitoring and alarm systems, and significantly improve diagnostic accuracy and generalization ability.

[0063] 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 method for alarm flooding data augmentation and root cause diagnosis of imbalanced samples, characterized in that: Includes the following steps, S1: Acquire alarm event data generated in the industrial system; S2: Improve the SMOTENC algorithm to perform data augmentation on minority class samples, and generate synthetic alarm samples with highly similar temporal logic to the original fault samples in the feature space to balance the dataset; S3: Construct a set of Hidden Markov Model (HMM) variant classifiers with different structures, including Factor Hidden Markov Model (FHMM), Hidden Semi-Markov Model (HSMM), and Discrete Hidden Markov Model (DHMM). S4: Evaluate the performance of each variant classifier using different metrics; S5: Using different performance metrics as benchmarks, an ensemble learning framework employing weighted voting integrates the diagnostic results of multiple classifiers, ultimately outputting alarm root cause diagnostic results with high confidence in real time.

2. The method for alarm flooding data augmentation and root cause diagnosis for imbalanced samples according to claim 1, characterized in that: S2 includes the following steps: S21: For each minority class sample, based on the preset feature similarity measurement criteria, identify the K nearest neighbor samples in its class to form its local neighborhood. S22: Randomly select one of the K nearest neighbor samples as the synthetic benchmark for generating new samples together with the current minority class samples; S23: For each discrete feature to be generated, a direct selection strategy or a frequency sampling strategy is used to determine its feature value; S24: Repeat steps S21-S23 until the number of synthesized samples reaches the preset balance ratio, and integrate all synthesized samples with the original dataset.

3. A method for alarm flooding data augmentation and root cause diagnosis of imbalanced samples according to claim 1 or 2, characterized in that: S3 includes the following steps: S31: FHMM classifier construction: By combining multiple independent Markov chains in parallel, the multi-dimensional processing capability of the model is enhanced, the forward and backward probabilities of different alarm sequences are calculated, and the fault state of the system is analyzed and determined by combining the state transition characteristics of the system and the generation probability of alarms. S32: HSMM Classifier Construction: Adding the concept of state duration allows for consideration of situations where the state remains unchanged over a period of time. By distinguishing between macroscopic and microscopic states, it provides an effective analytical tool for handling scenarios where the state duration changes significantly. S33: DHMM classifier construction: Assuming that the state follows an unobservable Markov process, determined by the probability vector of the initial state, the state transition probability matrix, and the observation probability matrix, the model parameters are adjusted through training samples to maximize the probability of alarm flooding sequences occurring.

4. The method for alarm flooding data augmentation and root cause diagnosis for imbalanced samples according to claim 3, characterized in that: The parameters of the FHMM classifier are defined as follows: Wherein, the state transition matrix A is the state transition matrix of each independent Markov chain. The transition matrix has N layers, and the transition probability of the i-th layer is... B represents the observation probability matrix, and the probability distribution of independently generated observations at each layer is as follows: π represents the initial state distribution, and the initial probability vector for each layer is... N represents the hidden state dimension; The probability of generating observation sequences in FHMM is: in, For all layers, O represents the hidden state sequence; T represents the time length of the sequence; and N represents the hidden state dimension. Let be the state transition probability of the i-th layer; Let be the observation probability of the i-th layer.

5. The method for alarm flooding data augmentation and root cause diagnosis for imbalanced samples according to claim 3, characterized in that: The parameter set of the HSMM classifier is as follows: in, Represents the state duration matrix, Let i be the probability that state i lasts for d time units. The probability of generating the observed sequence in HSMM is: in, It is a macroscopic state sequence; It is a duration sequence; Initial state The probability of; Indicates from state Transition to state The transition probability; In the state Generate observations The probability of observation.

6. The method for alarm flooding data augmentation and root cause diagnosis for imbalanced samples according to claim 3, characterized in that: The parameter set of the DHMM classifier is as follows: The probability of generating the observed sequence in DHMM is: in, For a moment Forward probability ( For the current moment, (Total length of the observation sequence), recursive formula: in, State transition probability matrix From the state Transition to state The probability of; The observation probability matrix In the state Generate observations The probability of.

7. A method for alarm flooding data augmentation and root cause diagnosis of imbalanced samples according to claim 1 or 2, characterized in that: In S4, the performance of each variant classifier is evaluated by calculating macro precision, macro recall, and macro F1 score.

8. A method for alarm flooding data augmentation and root cause diagnosis of imbalanced samples according to claim 1 or 2, characterized in that: In S5, a weighted voting method is used to integrate the diagnostic results of three HMM variant classifiers with different structures. The weight coefficients of each classifier are not fixed values, but are dynamically calculated based on their macro F1 scores on the cross-validation set. Higher weights are assigned to classifiers that perform better in specific test scenarios, thereby achieving complementary advantages and decision optimization.

9. A data augmentation and root cause diagnosis device for alarm flooding of imbalanced samples, characterized in that: The alarm flood data augmentation and root cause diagnosis method for imbalanced samples as described in any one of claims 1 to 8 is implemented. include, The data acquisition module is used to acquire alarm event data generated in industrial systems, either offline or in real time. The data augmentation module is used to augment minority class samples using an improved SMOTENC algorithm, generating new synthetic samples to balance the dataset. The classifier building module is used to build FHMM, HSMM, and DHMM classifiers; The ensemble learning module is used to integrate the diagnostic results of multiple classifiers using a weighted voting ensemble learning framework, thereby improving the model's generalization ability and robustness.

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 8.

Images