A device health assessment method based on node embedding

Abstract

The present invention discloses a device health assessment method based on node embedding, which relates to a device health assessment method. At present, the accuracy of device health assessment cannot be guaranteed while reducing the technical requirements for personnel; the present invention includes the steps of: constructing a device association graph model, and defining the node features in the device association graph through feature extraction; performing device health assessment based on the device association graph; the technical solution adopts a device association graph to characterize the complex associations between devices, effectively capturing all the associations between devices, and can more accurately reflect the effects of factors affecting device health, and obtain more accurate health values. The device features are defined through the most intuitive basic information and operation information of the device, getting rid of the dependence of the existing method on expert knowledge; for different types of equipment, it can also operate normally without the exclusive features defined by expert knowledge, accurately realize equipment health assessment, and reduce the requirements for personnel.

Description

Technical Field

[0001] The present invention relates to a device health evaluation method, and in particular to a device health evaluation method based on node embedding. Background Art

[0002] With the improvement of automation and integration level of information equipment, the classic equipment asset management method can no longer meet the current equipment management needs. The basic information and operation status data of a large number of routers, switches and professional production equipment exceed the analysis capabilities of traditional expert experience, and it is urgent to migrate and apply intelligent analysis methods.

[0003] Most of the existing equipment health assessment methods are based on the reliability-centered maintenance (RCM) concept. They describe historical failure data through quantitative modeling and combine expert evaluation to determine the life and reliability of equipment, make preventive maintenance decisions, and reduce potential production downtime losses.

[0004] Among them, traditional methods are mostly based on statistical definition and regression analysis technology, and typical ones are relative health models and their improved models for different fields. Based on expert knowledge, statistics in application scenarios are defined, such as equipment operation indicators, equipment temperature, and related product technical indicators, and then statistical weights are calculated through multivariate regression, entropy correction and other technologies. Finally, the relative health model obtained is used to predict the health of future equipment.

[0005] The development of machine learning technology has introduced new ideas for equipment health assessment. Existing machine learning-based methods also define key equipment features based on expert knowledge, such as basic equipment information and operating indicators, and then use machine learning algorithms such as XGBoost and clustering to perform feature modeling, and obtain good classification or prediction models through training with large amounts of data.

[0006] Deep learning technology has also been applied, but due to the limited amount of data and labels required by deep learning models, the migration of related technologies is still in its early stages. Some researchers have used sequence models to predict the future health of a single device, but the accuracy needs to be improved. In particular, the method based on graph node embedding has not yet been migrated and applied to the field of equipment health assessment.

[0007] Defects and shortcomings:

[0008] Equipment health prediction methods based on different technologies have different defects, including:

[0009] 1. Classic statistical methods cannot effectively extract high-level features from data, and their classification or prediction capabilities are insufficient, resulting in low accuracy of health assessment and limited guidance for equipment maintenance in production environments.

[0010] 2. The machine learning-based method has improved the evaluation accuracy to a certain extent, but it also relies on the introduction of expert knowledge and has insufficient migration capabilities for different application scenarios and different device states.

[0011] 3. The application of deep learning methods in the field of equipment health assessment needs further research. Methods such as sequence prediction are only applicable to a single device and consume a lot of computing resources, making it difficult to apply them in real scenarios with a large number of devices. Summary of the invention

[0012] The technical problem to be solved and the technical task proposed by the present invention are to improve and perfect the existing technical solutions and provide a device health assessment method based on node embedding, so as to achieve the purpose of reducing the technical requirements for personnel while ensuring the accuracy of the assessment. To this end, the present invention adopts the following technical solutions.

[0013] A device health assessment method based on node embedding includes the following steps:

[0014] 1) Build a device association graph model and define node features in the device association graph through feature extraction;

[0015] 11) Obtain device information;

[0016] Equipment information includes basic equipment information, equipment usage information and information for specific equipment types; basic equipment information includes manufacturer, production time and equipment type; equipment usage information includes physical location of equipment, average daily working hours, average daily number of failures and average daily temperature; information for specific equipment types includes working information based on equipment type;

[0017] 12) Determine the device association relationship;

[0018] Device association refers to the association between device information. If a certain piece of information of two devices is the same, it is considered that there is an association between the two devices.

[0019] 13) Construct a device association diagram based on the device association relationship;

[0020] First, the original data is processed. For each device, its device information vector is generated as the device feature. Then, a node is constructed for each device, and the device feature vector is added as an attribute. If the health status is known, it is added as a label. Then, the weights between nodes are calculated pairwise, and edges with corresponding weights greater than 0 are generated.

[0021] 2) Evaluate equipment health based on equipment association diagram;

[0022] Based on the edge weights of the device association graph, all associations between devices are captured and the health of all unknown devices is predicted.

[0023] As a preferred technical means: in step 13), the device association graph is a type of graph structure, including nodes and edges;

[0024] node:

[0025] Each node represents a device, and the attribute of the node is the feature vector composed of the information of the device, where the continuous value is normalized to the interval [0,1] by the following formula, and the discrete value is processed as a one-hot code;

[0026] The standardization formula is as follows:

[0027]

[0028] where y i Refers to the normalized result of field i, x i Refers to the value of this device in field i, max(x) and min(x) refer to the maximum and minimum values ​​of all devices in field i;

[0029] The node label is the health of the device, and its value is assigned to the corresponding node as a label;

[0030] side:

[0031] Each edge in the device association graph connects two nodes. The edge has no direction but has a weight. The construction of the edge follows the following process:

[0032] a) The weight of the edge between any two nodes is 0;

[0033] b) For any two nodes, the information of their corresponding devices is examined; for every identical field in the device information, the weight of the edge between the two nodes increases by 1;

[0034] c) Generate all edges between nodes with weight greater than 0.

[0035] As a preferred technical means: in step 13), constructing a device association diagram includes the following steps:

[0036] 131) Obtain device information and generate device characteristics;

[0037] 132) Generate nodes according to device characteristics;

[0038] 133) Determine whether all devices have generated nodes, if not, return to step 131), if yes, proceed to the next step;

[0039] 134) Calculate the weight between nodes; Calculate the weight between nodes pairwise;

[0040] 135) Generate edges with corresponding weights greater than 0.

[0041] As a preferred technical means: step 2) comprises the following steps:

[0042] 21) Based on the device association graph, a random walk is performed on the graph to obtain the node sequence;

[0043] 22) Then the node embedding vector is calculated based on the node sequence using the Woed2Vec algorithm;

[0044] 23) Based on the three-layer perceptron, the node embedding vector and the known label are input to the perceptron model. The perceptron model predicts the node labels of all label positions and evaluates the health of the equipment.

[0045] As a preferred technical means: In step 21), since the device association graph is a static isomorphic graph, the transition probability of the random walk needs to be adjusted: the probability of the current node v being transferred to node t in the next step is adjusted to:

[0046]

[0047]

[0048] Where weight(t,v) represents the weight of the edge between node v and node t, N w (v) represents the sum of the weights of the edges between node v and all its neighboring nodes; t i Refers to every node adjacent to v, That is, sum the edge weights between each node adjacent to v and v.

[0049] As a preferred technical means: the number of nodes in the path obtained by the selected walk is 10. Otherwise refers to "others". In the formula, when (t,v)∈E, there is In other cases, P(t|v)=0

[0050] As a preferred technical means: in step 22), Word2vec is a graph node embedding algorithm that embeds nodes into vectors and makes the embedded vectors obtained by nodes with similar attributes as close as possible;

[0051] For any n of the paths obtained by random walk in step 21) v-c ,…,n v+c For the node v in word2vec, the objective function to be maximized is:

[0052]

[0053] The probability in the above formula is converted into the product of a series of probabilities, and the final objective function can be converted to:

[0054]

[0055] in Representative node n i The input vector of , that is, its attributes, Representative node n i The output vector of is the embedding vector; V represents the number of all nodes; in the calculation process of i growing to V, the above formula calculates the embedding vectors of all nodes. T transpose, e natural constant, Σ sum.

[0056] As a preferred technical means: when training the perceptron model in step 23), the embedding vectors of the corresponding nodes of all devices with known health are used as the input of the three-layer perceptron, and their health values ​​are used as the labels to be fitted; after the perceptron model training is completed, the embedding vectors of the nodes with unknown labels are input into the perceptron model, and the output obtained is the predicted health of the corresponding node, that is, the corresponding device;

[0057] Each layer in the three-layer perceptron is a fully connected layer, and each neuron follows the following calculation formula;

[0058] output = f(net-θ)

[0059] net=∑z i ·v i

[0060] Among them, z i is the output value of the i-th neuron in the previous layer, v i is the weight of the ith neuron in the previous layer linking this neuron; θ is the bias value of this neuron, which is set to 0; f(x) is the activation function, which is the sigmoid function. net is the variable name, i.e., net in output = f(net-θ).

[0061] Beneficial effects:

[0062] 1. The device association graph proposed in the present invention can effectively characterize the complex associations between devices. Traditional methods often focus on the historical operation data of a single device, and the analysis of the associations between devices is limited to the weight similarity in the regression equation brought about by the association of basic information such as the same manufacturer. The influence of associations such as the physical location of the device is not obtained. For example, devices that share a common power supply may be affected by the same power supply status and have similar health. The edge weight calculation method of the association graph proposed in the present invention effectively captures all the associations between devices, and can more accurately reflect the effects of factors affecting device health, thereby obtaining more accurate health values.

[0063] 2. This invention introduces node embedding based on random walk and Word2Vec into the field of equipment health assessment, which introduces new ideas for research and development in this field. Compared with the existing methods based on statistics and machine learning, the node embedding method reduces the dimension of the feature vector and reduces the complexity of subsequent calculations; on the other hand, the vector value of a device after embedding is affected by the devices with which it has a strong association, and can extract higher-level features to achieve more accurate health assessment.

[0064] 3. The present invention performs equipment health assessment based on a graph structure and can predict the health of all unknown devices through a unified node embedding, thus solving the problem that the existing equipment health assessment method based on deep learning focuses on the prediction of the health of a single device.

[0065] 4. The present invention defines the device features through the most intuitive basic information and operation information of the device, getting rid of the reliance of existing methods on expert knowledge. For different types of equipment, the present invention can also operate normally without the exclusive features defined by expert knowledge, accurately implement the device health assessment, and reduce the threshold for the application of the present invention in terms of personnel and data completeness. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart for generating a device association diagram of the present invention.

[0067] Figure 2 It is a flow chart of equipment health assessment of the present invention. DETAILED DESCRIPTION

[0068] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0069] The present invention comprises the steps of:

[0070] S1: Build a device association graph model and define node features in the device association graph through feature extraction;

[0071] S11: Obtain device information;

[0072] Equipment information includes basic equipment information, equipment usage information, and other information for specific equipment types. Among them, basic equipment information includes manufacturer, production time, equipment type, etc.; equipment usage information includes equipment physical location, average daily working hours, average daily number of failures, average daily temperature, etc.; other information for specific equipment types refers to working information based on equipment types, such as network switches including average daily forwarding volume, average daily fan speed, etc.

[0073] Based on specific scenarios and devices, device information can be added without limit, thus forming more complete device features.

[0074] S12: Determine the device association relationship;

[0075] Device association refers to the association between device information. If two devices have the same information, they are considered to be associated. In actual production scenarios, the health of devices with stronger associations is more similar. For example, devices from the same batch or devices running at the same temperature.

[0076] S13: constructing a device association graph according to the device association relationship;

[0077] In order to effectively describe the information of the device and the relationship between the devices, it is necessary to construct a device association diagram. Device association diagram G =<V,E> It is a type of graph structure.

[0078] node:

[0079] Each node represents a device, and the attribute of the node is the feature vector composed of the information of the device. The continuous value is normalized to the interval [0,1] by the following formula, and the discrete value is processed as a one-hot code.

[0080] The standardization formula is as follows:

[0081]

[0082] where y i Refers to the normalized result of field i, x i Refers to the value of this device in field i, max(x) and min(x) refer to the maximum and minimum values ​​of all devices in field i.

[0083] The label of the node is the health of the device. The health of the data in the training set is known, and its value is assigned to the corresponding node as a label.

[0084] side:

[0085] Each edge in the device association graph connects two nodes. The edge has no direction but has weight. The construction of the edge follows the following process:

[0086] a: The weight of the edge between any two nodes is 0

[0087] b: For any two nodes, examine the corresponding device information. For every identical field in the device information, the weight of the edge between the two nodes increases by 1.

[0088] c: Generate all edges with weight greater than 0 between nodes.

[0089] The process of generating device association diagram is as follows: Figure 1As shown in the figure. First, the original data is processed, and for each device, its device information vector is generated as the device feature; then a node is constructed for each device, and the device feature vector is added as an attribute. If the health is known, it is added as a label; then the weights between nodes are calculated pairwise, and the corresponding edges with weights greater than 0 are generated.

[0090] This technical solution effectively represents the complex associations between devices through the device association graph. The edge weight calculation method of the association graph effectively captures all the associations between devices, and can more accurately reflect the effects of factors affecting device health, thereby obtaining a more accurate health value.

[0091] By defining the device features through the most intuitive basic information and operation information of the device, the existing methods get rid of the reliance on expert knowledge. For different types of equipment, under the premise of no exclusive features defined by expert knowledge, this technical solution can also operate normally and accurately implement the health assessment of the equipment, thus lowering the threshold for the application of the invention in terms of personnel and data completeness.

[0092] S2: Evaluate the health of the equipment based on the equipment association diagram; the algorithm execution process is as shown in the attached Figure 2 shown.

[0093] S21: Based on the device association graph, a random walk is performed on the graph to obtain a node sequence;

[0094] Most of the existing random walk methods can be applied to heterogeneous graphs, heterogeneous information networks and other structures. The device association graph is a static isomorphic graph, so the transition probability of the random walk needs to be adjusted. Specifically, we adjust the probability of the current node v to the next node t as follows:

[0095]

[0096]

[0097] Where weight(t,v) represents the weight of the edge between node v and node t, N w (v) represents the sum of the weights of the edges between node v and all its neighboring nodes. We choose 10 as the number of nodes in the path obtained by the walk.

[0098] S22: Then, the node embedding vector is calculated based on the node sequence using the Woed2Vec algorithm;

[0099] The Word2vec algorithm is a coding method. We transform it into a graph node embedding algorithm, embedding nodes into vectors and making the embedding vectors of nodes with similar attributes as close as possible.

[0100] For any n of the paths obtained by random walk in step S21 v-c ,…,n v+c For the node v in word2vec, the objective function that word2vec needs to maximize is

[0101]

[0102] The probability in the above formula can be converted into the product of a series of probabilities, and the final objective function can be converted into

[0103]

[0104] in Representative node n i The input vector of , that is, its attributes, Representative node n i The output vector of is its embedding vector; V represents the number of all nodes. In the calculation process of i growing to V, the above formula calculates the embedding vectors of all nodes.

[0105] S23: Based on the three-layer perceptron, the node embedding vector and the known label are input to the perceptron model. The perceptron model predicts the node labels of all label positions and evaluates the health of the equipment.

[0106] We use the embedding vectors of all nodes corresponding to devices with known health as the input of the three-layer perceptron, and their health values ​​as the labels to be fitted to train the perceptron model. After the training is completed, the embedding vectors of nodes with unknown labels are input into the perceptron model, and the output obtained is the corresponding node, that is, the predicted health of the corresponding device.

[0107] Each layer in the three-layer perceptron is a fully connected layer, and each neuron follows the following calculation formula.

[0108] output = f(net-θ)

[0109] net=∑z i ·v i

[0110] Among them, z i is the output value of the i-th neuron in the previous layer, v i is the weight of the ith neuron in the previous layer connecting this neuron. θ is the bias value of this neuron, which we set to 0. f(x) is the activation function, which is the sigmoid function.

[0111] This technical solution introduces node embedding based on random walk and Word2Vec into the field of device health assessment, reducing the dimension of feature vectors and reducing the complexity of subsequent calculations. On the other hand, the vector value of a device after embedding is affected by devices that have a strong association with it, and higher-level features can be extracted to achieve more accurate health assessment.

[0112] The device health assessment is based on the graph structure. Through a unified node embedding, the health of all unknown devices can be predicted, which solves the problem that the existing deep learning-based device health assessment method focuses on the prediction of the health of a single device.

[0113] The device health assessment method based on node embedding shown in the above figure is a specific embodiment of the present invention, which has embodied the substantial characteristics and progress of the present invention. According to actual use needs and under the guidance of the present invention, equivalent modifications in shape, structure, etc. can be made to the device health assessment method based on node embedding, which are all within the protection scope of this solution.

Claims

1. A device health assessment method based on node embedding, characterized in that The following steps are involved: 1) Build a device association graph model and define node features in the device association graph through feature extraction; 11) Obtain device information; Equipment information includes basic equipment information, equipment usage information and information for specific equipment types; basic equipment information includes manufacturer, production time and equipment type; equipment usage information includes physical location of equipment, average daily working hours, average daily number of failures and average daily temperature; information for specific equipment types includes working information based on equipment type; 12) Determine the equipment association relationship; Device association refers to the association between device information. If a certain information of two devices is the same, it is considered that there is an association between the two devices. 13) Construct a device association diagram based on the device association relationship; First, the original data is processed, and for each device, its device information vector is generated as the device feature; Then, we build a node for each device, add the device feature vector as an attribute, and if the health is known, add it as a label; Then the weights between nodes are calculated pairwise, and edges with corresponding weights greater than 0 are generated; The device association graph is a type of graph structure, including nodes and edges; node: Each node represents a device, and the node attribute is the feature vector composed of the information of the device, where the continuous value is normalized to Interval, discrete values ​​are processed as one-hot codes; The standardization formula is as follows: in Refers to the field The standardized results, Indicates that this device is in the field The value of and Refers to all devices in the field The maximum and minimum values ​​on ; The node label is the health of the device, and its value is assigned to the corresponding node as a label; side: Each edge in the device association graph connects two nodes. The edge has no direction but has a weight. The construction of the edge follows the following process: a) The weight of the edge between any two nodes is 0; b) For any two nodes, examine the information of their corresponding devices; for every identical field in the device information, the weight of the edge between the two nodes increases by 1; c) Generate all edges between nodes with weight greater than 0; Building a device association graph includes the following steps: 131) Obtain device information and generate device characteristics; 132) Generate nodes based on device characteristics; 133) Determine whether all devices have generated nodes, if not, return to step 131), if yes, proceed to the next step; 134) Calculate the weight between nodes; Calculate the weight between nodes pairwise; 135) Generate edges with corresponding weights greater than 0; 2) Evaluate equipment health based on equipment association diagram; Based on the edge weights of the device association graph, all the associations between devices are captured and the health of all unknown devices is predicted; 21) Based on the device association graph, perform random walks on the graph to obtain the node sequence; 22) Then the node embedding vector is calculated based on the node sequence using the Word2Vec algorithm; 23) Based on the three-layer perceptron, the node embedding vector and the known label are input to the perceptron model. The perceptron model predicts the node labels of all label positions and evaluates the health of the equipment.

2. According to the device health assessment method based on node embedding according to claim 1, it is characterized in that: In step 21), since the device association graph is a static isomorphic graph, the transition probability of the random walk needs to be adjusted: the current node , the next step will be to move to the node The probability of adjustment is: in Represents the node With Node The weight of the edge between Represents the node The sum of the weights of the edges between all neighboring nodes; Refers to each Adjacent nodes, That is, each Adjacent nodes and The sum of the edge weights between them.

3. The device health assessment method based on node embedding according to claim 2 is characterized in that: The number of nodes in the path obtained by the selected walk is 10.

4. The device health assessment method based on node embedding according to claim 3 is characterized in that: When training the perceptron model in step 23), the embedding vectors of the corresponding nodes of all devices with known health are used as the input of the three-layer perceptron, and their health values ​​are used as the labels to be fitted; after the perceptron model training is completed, the embedding vectors of the nodes with unknown labels are input into the perceptron model, and the output obtained is the predicted health of the corresponding node, that is, the corresponding device; Each layer in the three-layer perceptron is a fully connected layer, and each neuron follows the following calculation formula; in, For the previous layer The output value of a neuron, For the previous layer The weight of each neuron connected to this neuron; is the bias value of this neuron, set it to 0; is the activation function, and the activation function is the sigmoid function.

Images