Storage cluster fault diagnosis method and system based on spiking neural membrane calculation model

Abstract

The invention provides a storage cluster fault diagnosis method and system based on a spiking neural P calculation model, and the method comprises the steps: collecting operation indexes containing static data and dynamic data from a monitoring system of a storage cluster, defining a neuron for each operation index after preprocessing through the construction and training of the spiking neural P calculation model, and carrying out the fault diagnosis of the storage cluster. Initializing an initial state value of each neuron according to the collected data; and performing state updating and reasoning by using pulse signal propagation and fuzzy reasoning through a trained spiking neural membrane calculation model, judging a fault according to a state value of a neuron, and when the state value of a certain neuron exceeds a set fault threshold value, considering that an index has a fault or a potential fault, and identifying a specific fault type in the storage cluster through state combination of the plurality of neurons. According to the invention, efficient fault diagnosis of the distributed storage system is realized.

Description

Technical Field

[0001] The present invention relates to the field of fault diagnosis, and in particular to a storage cluster fault diagnosis method and system based on a pulse neural membrane computing model. Background Art

[0002] With the rapid development of cloud computing, big data, and artificial intelligence, distributed storage clusters have become core infrastructure in data centers. Their performance and reliability directly impact the stability and service quality of the entire system. However, due to the large number of nodes, complex topology, and highly dynamic operational state of distributed storage systems, they face a high risk of failure. These failures may include disk corruption, network latency, and node downtime, which in severe cases can impact system availability, data integrity, and access performance. Therefore, how to quickly identify and accurately diagnose various types of failures in distributed storage systems has become a key technical challenge to ensure stable system operation. Existing fault diagnosis methods mainly include log analysis, rule matching, and traditional machine learning models. However, these methods suffer from the following shortcomings in practical applications: First, they suffer from poor real-time performance. Most rely on offline data analysis, making them difficult to meet the high-timeliness requirements of the system. Second, they have limited ability to handle uncertainty and fuzziness, making it difficult to accurately determine complex nonlinear fault patterns and monitoring data with noise and incompleteness. Third, they suffer from poor scalability. As clusters expand, traditional centralized diagnosis algorithms struggle to adapt to the diagnostic efficiency requirements of high-concurrency scenarios. Fourth, they lack the ability to model dynamic features. Existing methods are mostly based on static rules or fixed models, making them unable to adapt to the continuous changes in system operating conditions. Fifth, they suffer from insufficient distributed adaptability. Most solutions fail to fully consider the parallel computing and collaborative diagnosis mechanisms in distributed architectures, resulting in low diagnostic efficiency. Currently, the most promising technical solutions include rule-based matching, classification models based on traditional machine learning, and uncertainty reasoning methods based on fuzzy logic. While these approaches have improved fault detection capabilities to some extent, they still have significant limitations: rule-based matching methods lack flexibility and high maintenance costs; machine learning methods rely on large amounts of labeled data, have long training cycles, and lack real-time response capabilities. While fuzzy logic methods have some noise immunity, they are often based on static rule sets and lack the ability to adapt to dynamic system changes.

[0003] Although existing technologies have solved the fault diagnosis problem in distributed storage systems to a certain extent, they still have obvious deficiencies in terms of real-time performance, dynamic adaptability, and distributed concurrent processing capabilities. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] The present invention proposes a storage cluster fault diagnosis method based on the pulse neural membrane computing model. Combining the pulse neural membrane computing model and fuzzy reasoning, it can accurately analyze and predict faults in the storage cluster, especially when facing complex and dynamic storage environments, and can significantly reduce false positives and missed positives.

[0006] Another object of the present invention is to provide a storage cluster fault diagnosis system based on the pulse neural membrane computing model.

[0007] To achieve the above objectives, the present invention provides a storage cluster fault diagnosis method based on a pulsed neural membrane computing model, comprising:

[0008] Collect operational indicators, including static and dynamic data, from the storage cluster's monitoring system. The static data includes storage topology information, disk configuration, and hardware information; the dynamic data includes disk I / O performance, network latency, CPU usage, memory utilization, hardware sensor information, and data synchronization status.

[0009] Preprocess the collected data, including data normalization, missing value processing and noise removal;

[0010] By building and training a pulse neural membrane computing model, a neuron is defined for each running indicator after preprocessing, and the initial state value of each neuron is initialized according to the collected data;

[0011] Through the trained pulse neural membrane computing model, pulse signal propagation and fuzzy reasoning are used to update and reason the state. Each neuron calculates its state based on the current input data and updates the state according to the rule set. When the state value of a neuron exceeds the preset threshold, a pulse signal is triggered to propagate to other neurons, and fuzzy reasoning methods are used to deal with the ambiguity and uncertainty of the indicators.

[0012] Faults are judged based on the status values ​​of neurons. When the status value of a neuron exceeds the set fault threshold, the indicator is considered to have a fault or potential fault. The specific fault type in the storage cluster is identified by combining the status of multiple neurons.

[0013] The storage cluster fault diagnosis method based on the pulse neural membrane computing model according to the embodiment of the present invention may also have the following additional technical features:

[0014] In one embodiment of the present invention, training a spiking neural membrane computational model includes:

[0015] Obtain multi-dimensional operating indicators from the storage cluster's monitoring system and use them as training data. The training data includes static data, dynamic data, and labeled data.

[0016] Construct a pulse neural membrane computing model, define a neuron for each indicator, and initialize the neuron weight; design a rule set for the indicator neurons, each rule defining the neuron's state change, signal propagation, and fault judgment logic; design the connection relationship between indicator neurons to define the causal relationship and information propagation path between indicators;

[0017] The training data is input into the pulse neural membrane computing model for model training. The neuron state is calculated according to the input data, the pulse output is updated, and the connection weights of the neurons are adjusted through the backpropagation algorithm to minimize the prediction error.

[0018] In one embodiment of the present invention, tuning the pulse neural membrane computational model includes:

[0019] Tuning hyperparameters in a spiking neural membrane computational model, including learning rate, number of neurons, and connection weight range;

[0020] Use early stopping and regularization techniques to optimize training strategies to avoid model overfitting;

[0021] The spiking neural membrane computational model was evaluated using the cross-validation method, with evaluation metrics including accuracy, precision, recall, and F1 score.

[0022] In actual operation, the model is dynamically adjusted through incremental learning or online learning methods, and the model is evaluated based on the comparison between the predicted results and the actual fault conditions, and the model is feedback-tuned based on the evaluation results.

[0023] In one embodiment of the present invention, the tuning of the learning rate is selected by cross-validation; the tuning of the number of neurons selects an appropriate number of neurons by balancing the model complexity and prediction performance; the connection weight initialization adopts random initialization or initialization based on existing data or rules; the early stopping method monitors the error of the validation set during training; the regularization technology includes L2 regularization or Dropout technology; the incremental learning adapts to new data by regularly retraining the model; the online learning obtains new data in real time during the operation of the system, dynamically updates the model, and ensures that it is always in the best state; model evaluation also includes adjustment of the model structure, design of the rule set or adjustment of the neuron connection relationship; the feedback tuning includes adjustment of hyperparameters, neuron initialization, rule set design and other aspects.

[0024] In one embodiment of the present invention, each neuron n i is represented as a seven-tuple:

[0025] n i =(o i ,σi ,R i ,P i ,E i ,syn i ,out)

[0026] Among them, i represents neuron n i The internal identifier of a single set of symbol symbols; i Represents the state of the neuron, which is a real value representing the running data of a certain dimension of the storage cluster; R i Represents the rule set of neurons, used to process and update the neuron state, in the form of: R i =(c i ,p i ,f i ); where c i : Conditions for updating the current state; p i : The weight change function of the state, used to describe how the input signal affects the neuron state; f i : state transition function, which defines how neurons transition from one state to the next; P i represents the initial input pulse set of the neuron; E i Represents input signals from external environmental signals or interactions with other neurons; i It represents the connection relationship between neurons and is used to define the signal flow between neurons. Out represents the output pulse of the neuron, which sends the result to other neurons or is ultimately used for fault prediction output.

[0027] In one embodiment of the present invention, the initial state of each neuron is determined by the operating indicator data in the monitoring system:

[0028] Disk I / O usage: mapped to σ io ∈[0,1];

[0029] Node CPU usage: mapped to σ io ∈[0,1];

[0030] Hardware CE error rate: mapped to σ io ∈[0,1];

[0031] Network delay: mapped to σ io ∈[0,1];

[0032] Mapping method: Assume that the actual value of a certain operating indicator is X, and its standard range is [X min ,X max ], then the normalization formula is:

[0033]

[0034] Among them, X represents the actual indicator value collected by the monitoring system; X min , X max Indicates the minimum and maximum values ​​of the indicator.

[0035] In one embodiment of the present invention, the rule set R i Contains the following sections:

[0036] R i ={r1,r2,…,r k}

[0037] The formal representation of each rule is as follows:

[0038]

[0039] Condition c i :Define the triggering conditions for the current state update; when the indicator σ i Exceeding the threshold T threshold When the state changes:

[0040] c j :σ i >T threshold

[0041] State transition function σ ′ i : Calculation based on the current input signal and rule weights:

[0042] σ ′ i =f i (σ i ,P i ,w ij )

[0043] Where: P i : represents the external input pulse received by the neuron; w ij : represents the weight of the connection with other neurons; f i : It is defined as follows:

[0044]

[0045] When σ ′ i When the set fault threshold is exceeded, it is considered a potential fault.

[0046] In one embodiment of the present invention, for disk I / O performance neurons n io , the rules are defined as follows:

[0047] Rule r1:

[0048]

[0049] When the disk I / O usage exceeds 80%, the indicator is judged to have a potential failure risk and the status changes to the failure state;

[0050] Rule r2:

[0051]

[0052] If the disk I / O usage is less than or equal to 80%, the current status remains unchanged.

[0053] In one embodiment of the present invention, the mathematical form of the connection relationship between indicator neurons is:

[0054] The connection of each neuron is represented by a directed graph G = (V, E); V: a set of neurons representing the indicators of each dimension of monitoring; E: a directed connection relationship between neurons;

[0055] Connection weight w ij Describes the correlation between two indicators; the weights are initialized and updated as follows:

[0056] Initialization weight: According to the correlation Corr(X i ,X j )calculate:

[0057] w ij =Corr(X i ,X j )

[0058] Where Corr(X i ,X j ) is the index X i and X j Pearson correlation coefficient;

[0059] Dynamic weight update: When changes in correlation are detected during operation, the weight is adjusted in real time:

[0060] w ij =w ij +Δw

[0061] Where Δw is calculated and adjusted based on real-time correlation.

[0062] To achieve the above objectives, the present invention further provides a storage cluster fault diagnosis system based on a pulsed neural membrane computing model, comprising:

[0063] A data collection module is used to collect operational indicators including static data and dynamic data from the monitoring system of the storage cluster. The static data includes storage topology information, disk configuration and hardware information; the dynamic data includes disk I / O performance, network latency, CPU usage, memory utilization, hardware sensor information and data synchronization status;

[0064] Data preprocessing module, used to preprocess the collected data, including data normalization, missing value processing and noise removal;

[0065] The model training module is used to build and train the pulse neural membrane computing model, define a neuron for each running indicator after preprocessing, and initialize the initial state value of each neuron according to the collected data;

[0066] The state update and reasoning module is used to perform state update and reasoning using a trained spiking neural membrane computational model, pulse signal propagation, and fuzzy reasoning. Each neuron calculates its state based on the current input data and updates its state according to a set of rules. When the state value of a neuron exceeds a preset threshold, a pulse signal is triggered to propagate to other neurons, and fuzzy reasoning methods are used to handle the ambiguity and uncertainty of the indicator.

[0067] The fault diagnosis module is used to judge faults based on the status values ​​of neurons. When the status value of a neuron exceeds the set fault threshold, the indicator is considered to have a fault or potential fault. The specific fault type in the storage cluster is identified by combining the status of multiple neurons.

[0068] The storage cluster fault diagnosis method and system based on the Spiking Neural Membrane Computing Model (SNPS) in this embodiment of the present invention combines the Spiking Neural Membrane Computing Model (SNPS) with fuzzy reasoning technology, leveraging the strengths of both to overcome the shortcomings of traditional fault diagnosis methods when processing large-scale, dynamic, and multi-dimensional data. The SNPS can simulate complex dynamic processes, while fuzzy reasoning can handle uncertainty and ambiguity, providing higher fault tolerance and accuracy.

[0069] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0071] Figure 1 This is a flow chart of a storage cluster fault diagnosis method based on a pulse neural membrane computing model according to an embodiment of the present invention;

[0072] Figure 2 is a data collection flow chart according to an embodiment of the present invention;

[0073] Figure 3 is a flowchart of model training and tuning according to an embodiment of the present invention;

[0074] Figure 4 is another flow chart of a storage cluster fault diagnosis method based on a pulsed neural membrane computing model according to an embodiment of the present invention;

[0075] Figure 5 4 is a structural diagram of a storage cluster fault diagnosis system based on a pulse neural membrane computing model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0077] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0078] The following describes a storage cluster fault diagnosis method and system based on a pulse neural membrane computing model according to an embodiment of the present invention with reference to the accompanying drawings.

[0079] The technical terms appearing in the present invention are defined below:

[0080] The Spiking Neural P System (SNPS) is a bioinspired model based on distributed parallel computing. It mimics the workings of neurons in the biological brain, particularly the process of pulse emission. Its key features include: 1. Discreteness and dynamics: The SNPS system is discrete in time and space, making it suitable for handling dynamic problems. 2. Nonlinearity: It can handle the nonlinear characteristics of power grid faults, especially the dynamic evolution of complex faults. 3. Modularity and parallel processing: Different membranes can independently handle specific tasks while interacting with other membranes, providing powerful parallel processing capabilities.

[0081] Fuzzy reasoning real number spiking neural membrane system (frSNPS) is an extension of spiking neural membrane system (SNPS), which introduces fuzzy logic and real number computing to meet the computational needs of dynamic and complex problems.

[0082] Figure 1 FIG. 1 is a flow chart of a storage cluster fault diagnosis method based on a pulse neural membrane computing model according to an embodiment of the present invention. Figure 1 Shown, including:

[0083] S1, collects operational indicators including static data and dynamic data from the monitoring system of the storage cluster. The static data includes storage topology information, disk configuration and hardware information; the dynamic data includes disk I / O performance, network latency, CPU usage, memory utilization, hardware sensor information and data synchronization status;

[0084] S2, preprocessing the collected data, including data normalization, missing value processing and noise removal;

[0085] S3, using the constructed and trained pulse neural membrane computing model, defines a neuron for each running indicator after preprocessing, and initializes the initial state value of each neuron according to the collected data;

[0086] S4, through the trained pulse neural membrane computing model, uses pulse signal propagation and fuzzy reasoning to perform state update and reasoning. In this process, each neuron calculates its state based on the current input data and updates its state according to the rule set. When the state value of a neuron exceeds the preset threshold, a pulse signal is triggered to propagate to other neurons, and fuzzy reasoning methods are used to deal with the ambiguity and uncertainty of the indicator.

[0087] S5, judge the fault based on the state value of the neuron. When the state value of a neuron exceeds the set fault threshold, the indicator is considered to have a fault or potential fault. The specific fault type in the storage cluster is identified by combining the states of multiple neurons.

[0088] Specifically, first, the present invention constructs and trains a pulse neural membrane computational model, and the process is as follows:

[0089] Obtain multi-dimensional operating indicators from the storage cluster's monitoring system, including disk I / O performance, network latency, node CPU usage, memory utilization, hardware CE errors, node internal sensor temperatures, power consumption, data synchronization status, etc.

[0090] Using fuzzy membership functions, continuous indicators (such as disk I / O performance) are mapped to discrete fuzzy values ​​(such as "normal", "high load", and "abnormal"). Fuzzy membership functions can be defined using trigonometric functions, Gaussian functions, or trapezoidal functions. The data collection process for multi-dimensional operating indicators is as follows: Figure 2 shown.

[0091] Furthermore, each neuron in the spiking neural membrane system is a basic computing unit used to simulate and store the operating data of different dimensions of the cluster. Its mathematical formal definition is as follows:

[0092] Formal definition of neurons: Each neuron n i The form can be expressed as a seven-tuple:

[0093] n i =(o i ,σ i ,R i ,P i ,E i ,syn i ,out)

[0094] Among them, i represents neuron n i The internal identifier of a single set of symbol symbols; i Indicates the state of the neuron, which is a real value representing the operating data of a certain dimension of the storage cluster, such as disk I / O, CPU usage, hardware errors, etc., and σ i ∈[0,1]; R i Represents the rule set of neurons, used to process and update the neuron state, in the form of: R i =(c i ,p i ,f i ). Among them, c i : The condition for updating the current state (e.g. σ i >T threshold , where T threshold is the set threshold); p i : The weight change function of the state, used to describe how the input signal affects the neuron state; f i : state transition function, which defines how neurons transition from one state to the next; P i Represents the initial input pulse set of the neuron (the pulse is quantified into a numerical value, such as the amplitude of a change in an operating indicator); E i Represents input signals from external environmental signals or interactions with other neurons; i It represents the connection relationship between neurons (i.e., topology), which is used to define the signal flow between neurons; out represents the output pulse of the neuron, which sends the result to other neurons or is ultimately used for fault prediction output.

[0095] The initial state of each neuron is determined by the operating indicator data in the monitoring system. They are listed as follows:

[0096] Disk I / O usage: mapped to σ io∈[0,1];

[0097] Node CPU usage: mapped to σ io ∈[0,1];

[0098] Hardware CE error rate: mapped to σ io ∈[0,1];

[0099] Network delay: mapped to σ io ∈[0,1];

[0100] Mapping method: Assume that the actual value of a certain operating indicator is X, and its standard range is [X min ,X max ], then the normalization formula is:

[0101]

[0102] Among them, X represents the actual indicator value collected by the monitoring system; X min , X max Indicates the minimum and maximum values ​​of the indicator.

[0103] Furthermore, the rule set is the core of the spiking neural membrane system to achieve fault diagnosis. Each rule defines the state change of neurons, signal propagation and fault judgment logic.

[0104] Rule Set R i Contains the following sections:

[0105] R i ={r1,r2,…,r k}

[0106] The formal representation of each rule is as follows:

[0107]

[0108] Condition c i :Define the triggering conditions for the current state update. For example, when the indicator σ i Exceeding the threshold T threshold When the state changes:

[0109] c j :σ i >T threshold

[0110] State transition function σ ′ i : Calculation based on the current input signal and rule weights:

[0111] σ′ i =f i (σ i ,Pi ,w ij )

[0112] Where: P i : represents the external input pulse received by the neuron; w ij : represents the weight of the connection with other neurons; f i : can be defined as follows:

[0113]

[0114] When σ′ i When the set fault threshold is exceeded, it is considered a potential fault.

[0115] For example, for disk I / O performance neuron n io , the rules are defined as follows:

[0116] Rule r1:

[0117]

[0118] When the disk I / O usage exceeds 80%, the indicator is considered to have a potential failure risk and the status changes to "Failed".

[0119] Rule r2:

[0120]

[0121] If the disk I / O usage is less than or equal to 80%, the current status remains unchanged.

[0122] Furthermore, the connection relationship between neurons (syn i ) is the topological structure of the system, which defines the causal relationship and information propagation path between indicators.

[0123] Mathematical form of connection relationship: The connection of each neuron can be represented by a directed graph G = (V, E);

[0124] V: A set of neurons representing the indicators of each dimension to be monitored; E: The directed connection relationship between neurons.

[0125] Connection weight w ij Describes the correlation between two indicators. The weights are initialized and updated as follows:

[0126] Initialization weight: According to the correlation Corr(X i ,X j )calculate:

[0127] w ij =Corr(X i ,Xj )

[0128] Where Corr(X i ,X j ) is the index X i and X j Pearson correlation coefficient.

[0129] Dynamic weight update: When changes in correlation are detected during operation, the weight is adjusted in real time:

[0130] w ij =w ij +Δw

[0131] Where Δw is calculated and adjusted based on real-time correlation.

[0132] Disk I / O performance io and network delay n latency There is a direct relationship between them (e.g., delays in data transmission can lead to decreased I / O performance). The connection weights are initialized as:

[0133] w io,latency =Corr(σ io ,σ latency )

[0134] Node CPU usage n cpu and memory utilization n memory The connection weight between is:

[0135] w cpu,memory =Corr(σ cpu ,σ memory )

[0136] Furthermore, the goal of model training is to enable the neural network to learn the associations between the failure modes and indicators of the storage cluster from historical data. The training process includes data preprocessing, model initialization, rule learning, and network structure optimization. The training process is as follows:

[0137] Step 1: Data Preparation: Before building a model, you must first collect and prepare the data for training. This data includes three categories: static data (such as storage topology, disk configuration, and hardware information), dynamic data (such as real-time disk I / O performance, network latency, CPU usage, memory utilization, sensor temperature, and power consumption), and labeled data (fault annotation information corresponding to each point in time, such as whether a node has failed). This data forms the foundation for subsequent model training.

[0138] Step 2: Data Preprocessing: To improve the efficiency and accuracy of model training, preprocessing is required before using the raw data. This primarily involves normalizing the data, mapping all input metrics to a uniform range (typically [0, 1]), thereby eliminating the effects of differences between different dimensions. This step helps improve model convergence speed and prediction accuracy.

[0139] Step 3: Neuron Initialization: Next, define a corresponding neuron for each data metric and set its initial weights. At the same time, you also need to initialize a set of rules. These rules can be manually set based on expert experience or automatically generated using machine learning methods (such as decision trees) to guide the behavior and response mechanisms of the neurons.

[0140] Step 4: Training process: After initialization, the model training phase begins. During this process, neurons calculate their states based on the input preprocessed data and generate corresponding pulse outputs. By using backpropagation algorithms or other optimization algorithms (such as simulated annealing or genetic algorithms), the connection weights between neurons are continuously adjusted to minimize prediction errors. The goal of training is to enable the model to accurately identify whether the system is in a fault state, thereby achieving efficient fault judgment and early warning.

[0141] Step 5: Model Tuning: The purpose of model tuning is to optimize model performance based on training results so that it can stably and accurately predict faults under different scenarios and data conditions. Model tuning steps may include:

[0142] 1) In spiking neural membrane systems, multiple hyperparameters such as learning rate, number of neurons, and connection weight range have a key impact on model performance. First, the learning rate is an important parameter that controls the speed at which neuron states are updated. If the learning rate is set too high, it may cause instability during model training; conversely, if it is too low, it may slow down the convergence rate. Therefore, it is particularly important to select a suitable learning rate through cross-validation methods. Secondly, the number of neurons directly determines the complexity of the model. A small number of neurons may not be enough to capture all the features in the data, while too many may lead to a waste of resources. The appropriate number of neurons needs to be determined based on the balance between model complexity and predictive performance. Finally, the initialization of connection weights is also extremely important. It can be performed through random initialization or based on existing data and rules to optimize the model training effect.

[0143] 2) To improve the model's generalization and avoid overfitting, adopting an effective training strategy is crucial. Early stopping is a common technique. During training, the validation set error is monitored and training is stopped as soon as the error begins to rise, effectively preventing the model from overfitting the training data. Regularization techniques, such as L2 regularization or dropout, are also effective means of improving model generalization. These techniques can reduce the model's reliance on specific training examples and enhance its performance when exposed to unseen data.

[0144] 3) Comprehensively evaluating the model using cross-validation is an important step in ensuring its reliability and effectiveness. Specifically, the entire dataset is divided into several subsets, each selected as a validation set, while the remaining subset is used for training. By repeating this process multiple times, the model's performance can be comprehensively evaluated from different perspectives. Common evaluation metrics include accuracy, precision, recall, and F1 score, which can reflect the actual performance of the model from multiple perspectives.

[0145] 4) In practical applications, as new data patterns or fault types emerge, the model needs to be able to self-update and optimize. Incremental learning allows the model to update the existing model with new data without retraining the entire network, primarily by adjusting the connection weights between neurons. Online learning goes a step further, acquiring new data in real time while the system is running and dynamically adjusting model parameters to ensure that the model is always in optimal condition. In addition, the accuracy and reliability of the model are evaluated by comparing the model's predictions with the actual situation, and hyperparameters, neuron initialization, and rule set design are tuned based on the feedback results to ensure that the model always maintains efficient fault diagnosis capabilities.

[0146] Furthermore, the model training and tuning process of the present invention is as follows Figure 3 shown.

[0147] In summary, after the above continuous training, a trained pulse neural membrane computing model is obtained, and the trained pulse neural membrane computing model is applied to the actual data processing scenario. The overall processing flow is as follows Figure 4 shown.

[0148] S11 comprehensively collects multi-dimensional operating indicators from the storage cluster's monitoring system, including static and dynamic data. Static data mainly involves storage topology information, disk configuration, and hardware information. Specifically, this covers details such as the cluster's network architecture, node configuration, disk layout, and the specifications, models, and historical status of each disk. Dynamic data focuses on real-time performance indicators such as disk I / O performance (read and write speeds, response time), network latency, CPU usage, memory utilization, hardware sensor information (temperature, power consumption, fault reports), and data synchronization status. These data provide the basis for subsequent data analysis.

[0149] S12, after obtaining the raw data, the next step is to preprocess it to improve data quality and model training efficiency. First, data normalization is performed to map all data to the same range (such as [0,1]) to eliminate the influence of different dimensions. Common methods include minimum-maximum normalization or Z-score normalization. Next, missing values ​​are processed to ensure that the data is complete through interpolation or mean filling. In addition, filtering techniques such as Gaussian filtering are used to remove noise in the data, further enhancing the reliability of the signal and laying a solid foundation for accurate modeling.

[0150] In S13, a neuron is defined for each of the collected operational metrics, and its state value is initialized based on the specific data of that metric. For example, neurons are set for key performance indicators such as disk I / O performance and CPU utilization. This process not only helps simulate the operating mechanisms of the actual system but also, through appropriate initial settings, effectively improves the initial expressiveness of the model, laying a good foundation for further state updates and reasoning.

[0151] S14, in the pulse neural membrane system, each neuron calculates its own state based on the input data and updates it according to a pre-set set of rules. When the state value of a neuron exceeds a specific threshold, it triggers a pulse signal to propagate to other associated neurons. In this process, the intensity and propagation path of the pulse are determined by the rule set and the connection relationship. At the same time, fuzzy reasoning methods are used to deal with the ambiguity and uncertainty of indicators. For example, the input signal is converted into a fuzzy set through a fuzzy membership function, and the fuzzy rules are applied to generate the final reasoning result. Once the state value of a neuron is found to exceed the fault threshold, it can be judged that the corresponding indicator is abnormal, and the specific fault type can be comprehensively identified by combining the states of multiple neurons.

[0152] S15, through comprehensive analysis of all neuron states, can precisely determine the location and type of failures within the storage cluster. Decision rules or clustering algorithms are used to categorize and identify different failure modes, such as disk failures and network latency. Furthermore, historical data and neuron outputs can be used to predict the probability or timing of future failures using regression models or time series analysis. Upon detecting any failure or potential risk, the system immediately triggers an alarm to notify operations and maintenance personnel and automatically implements protective measures based on the diagnostic results, such as disconnecting the faulty node and activating backup systems, thereby minimizing losses.

[0153] S16: To adapt to changing environmental conditions and new fault types, the model must be self-optimizing. In actual operation, incremental learning and online learning methods are used to adjust the neuron connection weights and rule sets in real time to ensure the model remains highly efficient. Simultaneously, the model's prediction accuracy is regularly evaluated to check for false positives or false negatives. If any issues are identified, the model needs to be re-tuned to the model tuning stage to readjust the hyperparameters, rule sets, or neuron structure to continuously improve the model's reliability and accuracy.

[0154] In summary, this invention proposes an innovative solution for storage cluster fault diagnosis by integrating the spiking neural membrane computation model (SNPS) with fuzzy reasoning technology. Its main technical improvements include: First, the combination of the spiking neural membrane computation model and fuzzy reasoning, leveraging the former's ability to simulate complex dynamic processes and the latter's ability to handle uncertainty and ambiguity, overcomes the shortcomings of traditional methods in processing large-scale, dynamic, and multi-dimensional data; second, the fault diagnosis method based on the real-number spiking neural membrane system (frSNPS) not only accurately models the fault modes in the storage cluster, but also responds to dynamic changes in the system in real time; third, the adaptive model and online learning mechanism enable the model to self-adjust based on newly collected data, enhancing the system's adaptability; finally, efficient data preprocessing and fault diagnosis processes improve the model's accuracy and diagnostic efficiency.

[0155] In terms of technical advantages, the present invention achieves high precision and low false alarm rate, especially in complex and dynamic storage environments; its real-time nature benefits from the characteristics of the pulse neural membrane system, which can complete fault detection and positioning in a very short time; its adaptability and scalability ensure that the model can be dynamically adjusted and expanded as the scale of the storage cluster and the failure mode change; multi-level fault analysis conducts a comprehensive analysis of faults from multiple dimensions to provide comprehensive and accurate diagnostic results.

[0156] The key points of the present invention are the spiking neural membrane computing model (SNPS), which can handle complex nonlinear and dynamic systems; the combination of fuzzy reasoning and spiking neural networks is used to deal with the ambiguity of fault indicators; frSNPS as the core model realizes efficient and accurate fault diagnosis; the data acquisition and preprocessing module ensures efficient and stable processing of complex real-time data; and the online learning and incremental learning mechanisms ensure the model's adaptability to new fault types and accurate monitoring.

[0157] This invention's design and implementation steps for a storage cluster fault diagnosis method based on a spiking neural membrane computational model include: defining the states of neurons in the spiking neural membrane system and designing a rule set; a fault diagnosis module combining fuzzy reasoning and spiking neural networks; dynamic optimization of online learning and incremental learning mechanisms; and multi-dimensional data acquisition and fuzzy processing methods. These form the foundation for fault prediction and diagnosis and demonstrate a high degree of innovation and practicality. This integration of technologies and methods aims to provide a more intelligent, accurate, and efficient fault diagnosis solution for storage clusters.

[0158] Furthermore, the technical solution of the present invention can also be applied to the following scenarios:

[0159] For example, in cloud computing environments, the spiking neural membrane computing model can be used to monitor the operating status of components such as virtual machines, storage nodes, and network links. By analyzing multi-dimensional data, it can predict potential system failures and avoid service interruptions. Furthermore, the system can adjust resource allocation based on real-time computing load and traffic changes, achieving intelligent load balancing across the cloud platform.

[0160] For example, in edge computing devices (such as gateways and sensor nodes), fault diagnosis can be performed by combining sensor data with device operating status. These devices typically have limited computing power, and the use of lightweight spiking neural membrane models can effectively improve the real-time and accuracy of fault diagnosis, ensuring the stable operation of edge devices.

[0161] For example, in a data center, multiple critical devices (such as servers, storage systems, and network switches) must operate continuously. By monitoring device environmental data (such as temperature, power consumption, and load) in real time and using this model to predict potential failures, the risk of downtime caused by equipment failure can be reduced, ensuring the normal operation of the data center.

[0162] For example, it can predict faults in network switches, routers, and data transmission paths, combining real-time network traffic, latency, packet loss, and other metrics to diagnose potential bottlenecks or faults in the network. By predicting network faults in advance, it improves network management efficiency and ensures smooth data transmission.

[0163] For example, in industrial production, sensors are used to monitor equipment status (such as temperature, vibration, and pressure), and the health status of equipment is assessed by combining pulse neural membrane computing models and fuzzy reasoning. This allows for the prediction of potential failure types (such as mechanical failures and electrical failures) before they occur, and preventive maintenance measures can be taken based on the diagnostic results, reducing downtime and maintenance costs.

[0164] For example, by monitoring and predicting faults in multiple systems within the manufacturing process (such as conveyor belts, automated robots, and temperature and humidity control systems), this technology can help achieve real-time adjustments and optimization of the manufacturing system. This technology can also improve production efficiency and product quality by integrating real-time data streams for incremental learning and automatic optimization.

[0165] For example, a pulsed neural membrane computing model can be used to monitor power equipment (such as transformers, power transmission lines, and generators) and predict faults by combining real-time multi-dimensional data such as current, voltage, and power. In smart grids, this system can be used to predict potential faults in power equipment and provide early warning of potential failures.

[0166] For example, by modeling the operating data of each node in the power grid and analyzing the fault propagation path between different devices, intelligent scheduling and optimization of the power grid can be achieved, reducing the impact of faults on the stability of the power system and power supply reliability.

[0167] According to an embodiment of the present invention, the storage cluster fault diagnosis method based on the spiking neural membrane computing model combines the spiking neural membrane computing model (SNPS) with fuzzy reasoning technology, leveraging the advantages of both to overcome the shortcomings of traditional fault diagnosis methods when processing large-scale, dynamic, multi-dimensional data. The spiking neural membrane computing model can simulate complex dynamic processes, while fuzzy reasoning can handle uncertainty and ambiguity, providing higher fault tolerance and accuracy. By constructing a fuzzy reasoning real spiking neural membrane system (frSNPS), not only can the fault modes in the storage cluster be accurately modeled, but real-time responses to dynamic changes in the system can also be made, achieving efficient fault diagnosis. Through incremental learning and online learning methods, the model of the present invention can self-adjust based on newly collected data, avoiding the lag of traditional methods when facing new fault types or data patterns, and enhancing the system's adaptability. By fuzzifying and normalizing multi-dimensional data, the system can effectively convert diverse indicators into a unified standard format, improving the model's accuracy and diagnostic efficiency.

[0168] In order to implement the above embodiment, Figure 5 As shown, this embodiment also provides a storage cluster fault diagnosis system 10 based on a pulse neural membrane computing model, including:

[0169] The data collection module 100 is used to collect operational indicators including static data and dynamic data from the monitoring system of the storage cluster. The static data includes storage topology information, disk configuration and hardware information. The dynamic data includes disk I / O performance, network latency, CPU usage, memory utilization, hardware sensor information and data synchronization status.

[0170] The data preprocessing module 200 is used to preprocess the collected data, including data normalization, missing value processing and noise removal;

[0171] The model training module 300 is used to construct and train the pulse neural membrane computing model, define a neuron for each operating indicator after preprocessing, and initialize the initial state value of each neuron according to the collected data;

[0172] A state update and reasoning module 400 is used to perform state update and reasoning using a trained spiking neural membrane computation model, utilizing pulse signal propagation and fuzzy reasoning. Each neuron calculates its state based on current input data and updates its state according to a set of rules. When a neuron's state value exceeds a preset threshold, a pulse signal is triggered to propagate to other neurons. Fuzzy reasoning methods are used to handle the ambiguity and uncertainty of indicators.

[0173] The fault diagnosis module 500 is used to judge faults based on the state values ​​of neurons. When the state value of a neuron exceeds the set fault threshold, the indicator is considered to have a fault or potential fault. The specific fault type in the storage cluster is identified by combining the states of multiple neurons.

[0174] According to an embodiment of the present invention, the storage cluster fault diagnosis system based on the spiking neural membrane computing model combines the spiking neural membrane computing model (SNPS) with fuzzy reasoning technology, leveraging the advantages of both to overcome the shortcomings of traditional fault diagnosis methods when processing large-scale, dynamic, multi-dimensional data. The spiking neural membrane computing model can simulate complex dynamic processes, while fuzzy reasoning can handle uncertainty and ambiguity, providing higher fault tolerance and accuracy. By constructing a fuzzy reasoning real spiking neural membrane system (frSNPS), not only can the fault modes in the storage cluster be accurately modeled, but the system's dynamic changes can also be responded to in real time, achieving efficient fault diagnosis. Through incremental learning and online learning methods, the model of the present invention can self-adjust based on newly collected data, avoiding the lag of traditional methods when facing new fault types or data patterns, and enhancing the system's adaptability. By fuzzifying and normalizing multi-dimensional data, the system can effectively convert diverse indicators into a unified standard format, improving the model's accuracy and diagnostic efficiency.

[0175] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0176] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A storage cluster fault diagnosis method based on a pulse neural membrane computing model, characterized in that: include: Collect operational indicators, including static and dynamic data, from the storage cluster's monitoring system. The static data includes storage topology information, disk configuration, and hardware information; the dynamic data includes disk I / O performance, network latency, CPU usage, memory utilization, hardware sensor information, and data synchronization status. Preprocess the collected data, including data normalization, missing value processing and noise removal; By building and training a pulse neural membrane computing model, a neuron is defined for each running indicator after preprocessing, and the initial state value of each neuron is initialized according to the collected data; Through the trained pulse neural membrane computing model, pulse signal propagation and fuzzy reasoning are used to update and reason the state. Each neuron calculates its state based on the current input data and updates the state according to the rule set. When the state value of a neuron exceeds the preset threshold, a pulse signal is triggered to propagate to other neurons, and fuzzy reasoning methods are used to deal with the ambiguity and uncertainty of the indicators. Faults are judged based on the status values ​​of neurons. When the status value of a neuron exceeds the set fault threshold, the indicator is considered to have a fault or potential fault. The specific fault type in the storage cluster is identified by combining the status of multiple neurons.

2. The method according to claim 1, characterized in that Training spiking neural membrane computational models, including: Obtain multi-dimensional operating indicators from the storage cluster's monitoring system and use them as training data. The training data includes static data, dynamic data, and labeled data. Construct a pulse neural membrane computing model, define a neuron for each indicator, and initialize the neuron weight; design a rule set for the indicator neurons, each rule defining the neuron's state change, signal propagation, and fault judgment logic; design the connection relationship between indicator neurons to define the causal relationship and information propagation path between indicators; The training data is input into the pulse neural membrane computing model for model training. The neuron state is calculated according to the input data, the pulse output is updated, and the connection weights of the neurons are adjusted through the backpropagation algorithm to minimize the prediction error.

3. The method according to claim 1, characterized in that Tuning of the spiking neural membrane computational model, including: Tuning hyperparameters in a spiking neural membrane computational model, including learning rate, number of neurons, and connection weight range; Use early stopping and regularization techniques to optimize training strategies to avoid model overfitting; The spiking neural membrane computational model was evaluated using the cross-validation method, with evaluation metrics including accuracy, precision, recall, and F1 score. In actual operation, the model is dynamically adjusted through incremental learning or online learning methods, and the model is evaluated based on the comparison between the predicted results and the actual fault conditions, and the model is feedback-tuned based on the evaluation results.

4. The method according to claim 3, characterized in that The learning rate is tuned by cross-validation; the number of neurons is tuned by selecting the appropriate number of neurons by balancing the model complexity and prediction performance; the connection weight initialization adopts random initialization or initialization based on existing data or rules; the early stopping method monitors the error of the validation set during training; the regularization technology includes L2 regularization or Dropout technology; the incremental learning adapts to new data by regularly retraining the model; the online learning obtains new data in real time during the operation of the system, dynamically updates the model, and ensures that it is always in the best state; model evaluation also includes adjustment of model structure, design of rule set or adjustment of neuron connection relationship; the feedback tuning includes adjustment of hyperparameters, neuron initialization, rule set design and other aspects.

5. The method according to claim 1, wherein Each neuron n i is represented as a seven-tuple: n i =(o i ,σ i ,R i ,P i ,E i ,son i (out) Among them, i represents neuron n i The internal identifier of a single set of symbol symbols; i Represents the state of the neuron, which is a real value representing the running data of a certain dimension of the storage cluster; R i Represents the rule set of neurons, used to process and update the neuron state, in the form of: R i =(c i ,p i ,f i ); where c i : Conditions for updating the current state; p i : The weight change function of the state, used to describe how the input signal affects the neuron state; f i : state transition function, which defines how neurons transition from one state to the next; P i represents the initial input pulse set of the neuron; E i Represents input signals from external environmental signals or interactions with other neurons; i It represents the connection relationship between neurons and is used to define the signal flow between neurons. Out represents the output pulse of the neuron, which sends the result to other neurons or is ultimately used for fault prediction output.

6. The method according to claim 1, characterized in that The initial state of each neuron is determined by the operating indicator data in the monitoring system: Disk I / O usage: mapped to σ io ∈[0,1]; Node CPU usage: mapped to σ io ∈[0,1]; Hardware CE error rate: mapped to σ io ∈[0,1]; Network delay: mapped to σ io ∈[0,1]; Mapping method: Assume that the actual value of a certain operating indicator is X, and its standard range is [X min ,X max ], then the normalization formula is: Among them, X represents the actual indicator value collected by the monitoring system; X min , X max Indicates the minimum and maximum values ​​of the indicator.

7. The method according to claim 1, characterized in that Rule Set R i Contains the following sections: R i ＝{r1,r2,…,r k} The formal representation of each rule is as follows: Condition c i :Define the triggering conditions for the current state update; when the indicator σ i Exceeding the threshold T threshold When the state changes: c j :s i >T threshold State transition function σ ′ i : Calculation based on the current input signal and rule weights: s ′ i =f i (s i ,P i ,w ij ) Where: P i : represents the external input pulse received by the neuron; w ij : represents the weight of the connection with other neurons; f i : It is defined as follows: When σ ′ i When the set fault threshold is exceeded, it is considered a potential fault.

8. The method according to claim 7, characterized in that For disk I / O performance neuron io , the rules are defined as follows: Rule r1: When the disk I / O usage exceeds 80%, the indicator is judged to have a potential failure risk and the status changes to the failure state; Rule r2: If the disk I / O usage is less than or equal to 80%, the current status remains unchanged.

9. The method according to claim 2, characterized in that The mathematical form of the connection relationship between indicator neurons: The connection of each neuron is represented by a directed graph G = (V, E); V: neuron set, representing the indicators of each dimension of monitoring; E: directed connection relationship between neurons; Connection weight w ij Describes the correlation between two indicators; the weights are initialized and updated as follows: Initialization weight: According to the correlation Corr(X i ,X j )calculate: w ij =Corr(X i ,X j ) Where Corr(X i ,X j ) is the index X i and X j Pearson correlation coefficient; Dynamic weight update: When changes in correlation are detected during operation, the weight is adjusted in real time: In ij =in ij +Δw Where Δw is calculated and adjusted based on real-time correlation.

10. A storage cluster fault diagnosis system based on a pulse neural membrane computing model, characterized in that: include: A data collection module is used to collect operational indicators including static data and dynamic data from the monitoring system of the storage cluster. The static data includes storage topology information, disk configuration and hardware information; the dynamic data includes disk I / O performance, network latency, CPU usage, memory utilization, hardware sensor information and data synchronization status; Data preprocessing module, used to preprocess the collected data, including data normalization, missing value processing and noise removal; The model training module is used to build and train the pulse neural membrane computing model, define a neuron for each running indicator after preprocessing, and initialize the initial state value of each neuron according to the collected data; The state update and reasoning module is used to perform state update and reasoning using a trained spiking neural membrane computational model, pulse signal propagation, and fuzzy reasoning. Each neuron calculates its state based on the current input data and updates its state according to a set of rules. When the state value of a neuron exceeds a preset threshold, a pulse signal is triggered to propagate to other neurons, and fuzzy reasoning methods are used to handle the ambiguity and uncertainty of the indicator. The fault diagnosis module is used to judge faults based on the status values ​​of neurons. When the status value of a neuron exceeds the set fault threshold, the indicator is considered to have a fault or potential fault. The specific fault type in the storage cluster is identified by combining the status of multiple neurons.

Images