A method and system for analyzing data of equipment fault detection in a digital factory

Abstract

The present invention discloses a method and system for analyzing equipment fault detection data for a digital factory, including: S1: synchronously collecting comprehensive maintenance data of equipment using high-precision clock synchronization technology; S2: constructing a graph neural network with the location data of the equipment as graph nodes and the functional association data and interaction data between the equipment as the edges of the graph, constructing an equipment association model based on the graph neural network, taking the status data of the equipment as the input of the equipment association model, and outputting a status association factor; S3: taking the comprehensive maintenance data and the status association factor as the input of the reinforcement learning agent state, training the agent using the Q-learning algorithm, inputting the real-time status data into the agent, outputting a status score, and setting an abnormal status threshold. When the status score is lower than the abnormal status threshold, the corresponding real-time data is marked as abnormal. In view of the complex and changeable operating status of the equipment, this method can reduce unnecessary downtime and maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and particularly to a method and system for analyzing equipment fault detection data for a digital factory. Background Art

[0002] At present, with the rapid development of digital factories, the stability and reliability of equipment operation have become key factors determining production efficiency and product quality. In digital factories, there are various types of equipment with different functions, and the equipment is interconnected and works collaboratively, forming a complex production system.

[0003] Currently, traditional methods for analyzing equipment fault data in digital factories mostly rely on regular maintenance, simple threshold alarms, or analysis based on a single data type. These methods cannot comprehensively capture the complex operating states of equipment, easily leading to missed detections and false detections of faults. For example, in some factories, for the fault detection of motor equipment, a simple threshold is set only based on the current value, and an alarm is issued when the current exceeds the threshold. However, this analysis method based on a single data type cannot comprehensively capture the complex operating states of equipment. According to statistics, in factories using such traditional methods, the missed detection rate of equipment faults is as high as 25%, and the false detection rate also reaches 20%. This not only leads to frequent sudden shutdowns during the production process but also causes unnecessary waste of maintenance resources. Therefore, a method and system for analyzing equipment fault detection data for a digital factory are proposed herein. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and achieve the above object, the present invention proposes the following technical solutions:

[0005] A method for analyzing equipment fault detection data for a digital factory, the method steps including:

[0006] S1: Synchronously collect the comprehensive maintenance data of the equipment using high-precision clock synchronization technology;

[0007] S2: Use the location data of the equipment as graph nodes, and the functional association data and interaction data between the equipment as edges of the graph to construct a graph neural network, and construct an equipment association model based on the graph neural network. Use the state data of the equipment as the input of the equipment association model, and output a state association factor;

[0008] S3: Use the comprehensive maintenance data and the state association factor as the input of the reinforcement learning agent state, and train the agent using the Q-learning algorithm to obtain a reinforcement learning agent. Input the real-time state data into the reinforcement learning agent, output a state score, and set an abnormal state threshold. When the state score is lower than the abnormal state threshold, mark the corresponding real-time data as abnormal.

[0009] The comprehensive maintenance data includes maintenance time, maintenance type, specified maintenance cost information, and maintenance behavior data.

[0010] The specified maintenance cost information includes parts cost data, labor cost data and special cost data;

[0011] The process of obtaining the special cost data is as follows:

[0012] The pressure sensor measures the pressure at a fixed sampling frequency to obtain pressure data ;

[0013] The wear degree sensor periodically detects the wear condition of the tool and obtains the wear value. ;

[0014] Use blockchain technology to record the full life cycle information of tools ;

[0015] The pressure data , wear value and tool life cycle information Integrate as input data for the cost relationship model and establish a cost relationship model to obtain special cost data .

[0016] The construction process of the cost relationship model is as follows:

[0017] Based on pressure data , wear value and tool life cycle information Construct a new feature combination, expressed as ,in, is the collection time interval, i is the index;

[0018] Through the random forest regression model, the feature combination is used as the input variable and the actual loss cost of the tool is used as the output variable to obtain the cost relationship model.

[0019] The process of acquiring the device's location data, function-related data, and interaction data is as follows:

[0020] The process of obtaining location data is as follows:

[0021] Equip each device in the digital factory with a UWB tag and obtain the 3D coordinates of the device as location data , the location data of each device is represented as a node feature vector ,in Indicates the UWB tag number;

[0022] The process of obtaining functional association data is as follows:

[0023] Extract the functional association relationships between devices from the operation logs of the digital factory, and use a numerical value to represent the weight of the functional association between devices. The value range is , 0 indicates no association, indicates a strong association. Organize the functional association weights between all devices into a functional association adjacency matrix , that is, functional association data;

[0024] The process of obtaining interaction data is as follows:

[0025] Deploy data monitoring software at the data transmission interface of the device to monitor the frequency and data volume of data transmission, and count the number of times and data volume of data transmitted from one device to another within a unit time. After normalization, obtain the data interaction weight , and organize it into an interaction adjacency matrix , that is, interaction data.

[0026] The process of constructing the graph neural network is as follows:

[0027] Use the location data as the nodes of the graph, and the functional association data and interaction data between devices as the edges to obtain the graph;

[0028] Define the graph as , where is the set of nodes, consisting of the nodes corresponding to the device location data, is the set of edges, consisting of the functional association data and interaction data between devices. The weight matrix of the edges is the weighted sum of the adjacency matrices of the functional association data and interaction data, denoted as , where represents the weight of the functional association adjacency matrix , represents the weight of the interaction adjacency matrix ;

[0029] Select the graph convolutional network as the basic architecture of the graph neural network model, and define the message passing and feature update mechanism to update its own features by aggregating the information of neighboring nodes;

[0030] Determine the training objective as the state association factor between devices;

[0031] Train the network by dividing the collected device-related data into training set, validation set and test set;

[0032] After training, obtain the graph neural network.

[0033] The process of constructing the device association model is as follows:

[0034] After preprocessing the status data of the device, it is fused with the graph neural network and used as the input of the graph neural network;

[0035] Using historical device failure data as supervision information, define a loss function Train the device association model and minimize the loss function to complete the training of the model.

[0036] The construction process of the reinforcement learning agent is as follows:

[0037] The reinforcement learning agent includes a state space, an action space, and a reward function;

[0038] Based on the comprehensive maintenance data and the state correlation factor, form the state space;

[0039] Define the action space as the action of judging the device status. The action indicates that the real-time data of the device is judged to be normal, and the action = 1 indicates that the real-time data of the device is judged to be abnormal;

[0040] Define the reward function. When the agent correctly judges the device status, give a positive reward, and when the judgment is wrong, give a negative reward;

[0041] Set the actual status of the device as , indicates normal, indicates abnormal;

[0042] Then train the agent through the Q-learning algorithm to obtain the reinforcement learning agent.

[0043] The process of training the agent through the Q-learning algorithm to obtain the reinforcement learning agent is as follows:

[0044] Determine the state, action space, and reward function of the agent, initialize the Q-network parameters and the experience replay buffer, and set the total number of training time steps and the discount factor;

[0045] At each time step, observe the state of the agent, and through the greedy policy selects an action, and after execution, obtains the reward and the new state feedback from the environment, and stores them in the experience replay buffer;

[0046] Randomly sample from the experience replay buffer, obtain the target Q value, calculate at the same time, and then update the Q-network parameters through the backpropagation algorithm;

[0047] The loss gradually decreases during training, and at the same time, repeat the Q-network training steps until the total number of training time steps is reached, until the Q-learning network converges and the training is completed.

[0048] A device fault detection data analysis system for a digital factory, comprising:

[0049] Synchronous acquisition module: uses high-precision clock synchronization technology to synchronously acquire the comprehensive maintenance data of the device;

[0050] Status correlation acquisition module: uses the location data of the device as graph nodes, and the functional correlation data and interaction data between devices as the edges of the graph to construct a graph neural network, and constructs a device association model based on the graph neural network. Uses the status data of the device as the input of the device association model and outputs a status correlation factor;

[0051] Comprehensive fault detection module: uses the comprehensive maintenance data and the status correlation factor as the input of the reinforcement learning agent state, and trains the agent using the Q-learning algorithm to obtain a reinforcement learning agent. Inputs the real-time status data into the reinforcement learning agent, outputs a status score, and sets an abnormal status threshold , when the status score is lower than the abnormal status threshold , marks the corresponding real-time data as abnormal.

[0052] The present invention has the following beneficial effects:

[0053] In the present invention, first, by using high-precision clock synchronization technology to collect comprehensive maintenance data, including maintenance time, type, cost, and behavior data. Special cost data is obtained through the Internet of Things and blockchain technology, and maintenance behavior data is integrated through a handheld terminal and sensors. These multi-dimensional data provide a more comprehensive basis for fault judgment, avoiding missed and misdetected faults caused by single data;

[0054] Secondly, taking the device location as nodes and the functional correlation data and interactions as edges to construct a graph neural network and a device association model, and outputting a status correlation factor to deeply explore the complex relationships between devices. In a complex digital factory, a device failure may trigger a chain reaction. Through this model, the source of the failure and the affected devices can be quickly located, avoiding the expansion of the failure range;

[0055] Finally, using the comprehensive maintenance data and the status correlation factor as the input of the reinforcement learning agent, training with the Q-learning algorithm, outputting a status score and comparing it with the abnormal threshold to judge the device status. This method can dynamically adapt to the changes in device operation. Compared with traditional fixed-threshold alarms, it can detect faults more flexibly and accurately, reducing the missed and misdetected fault rates. For the complex and changeable operating status of the device, this agent can accurately judge faults based on real-time data, reducing unnecessary downtime and maintenance. Description of the Drawings

[0056] Figure 1 is a method step diagram of a device fault detection data analysis method and system for a digital factory proposed by the present invention;

[0057] Figure 2System block diagram of a device fault detection data analysis method and system for a digital factory proposed by the present invention. Detailed implementation manners

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] Embodiment 1: As Figure 1 shown, a device fault detection data analysis method for a digital factory proposed by the present invention includes: S1: Using high-precision clock synchronization technology to synchronously collect comprehensive maintenance data of devices;

[0060] Using high-precision clock synchronization technology to extract comprehensive maintenance data from the device maintenance process in a digital factory;

[0061] The comprehensive maintenance data includes maintenance time, maintenance type (preventive maintenance, fault repair maintenance), specified maintenance cost information, and maintenance behavior data;

[0062] The process of obtaining the specified maintenance cost information is:

[0063] The maintenance cost includes component cost data, labor cost data, and special cost data;

[0064] The process of obtaining the special cost data is:

[0065] For the special cost (tool loss) used in the maintenance process, it is obtained by deploying and applying blockchain technology through Internet of Things devices;

[0066] By deploying Internet of Things devices:

[0067] Install micro Internet of Things sensors on the tools used in the digital factory equipment, including pressure sensors and wear degree sensors;

[0068] The pressure sensor measures the pressure at a fixed sampling frequency, and each sampling obtains a pressure data , and at the same time records the sampling moment, expressed as , where is the frequency number of the collected pressure data;

[0069] The wear degree sensor periodically detects the wear condition of the tool, and each detection obtains a value representing the wear amount , and records the corresponding time, expressed as , where, is the cycle number of the collected wear data;

[0070] Specifically, the pressure sensor can monitor the magnitude and duration of the pressure borne by the tool during use in real time, and the wear degree sensor uses the principle of optical or electromagnetic induction to monitor the wear condition of the tool in real time;

[0071] Obtained by applying blockchain technology:

[0072] Use blockchain technology to record the full life cycle information of the tool, including the relevant operation information of the tool use (user, use time, use location), and record it on the distributed ledger in the form of a blockchain smart contract. At the same time, the real-time status data of the tool transmitted by the Internet of Things platform will also be packaged into the transaction records of the blockchain, expressed as , where, represents the number of data points in the full life cycle, , represents the user, represents the use time, represents the use location;

[0073] Based on the pressure data , wear amount value and the full life cycle information of the tool are integrated as the input data of the cost relationship model, and a cost relationship model of the tool loss cost is established to obtain special cost data ;

[0074] The construction process of the cost relationship model is as follows:

[0075] Based on the pressure data , wear amount value and the full life cycle information of the tool build a new feature combination, expressed as , where, is the acquisition time interval and i is the index;

[0076] Through the random forest regression model, use the feature combination as the input variable and the actual loss cost of the tool as the output variable (which can be determined by historical tool loss records and cost accounting data). Divide the training set and the test set according to a ratio of 70% - 30%. Use the training set data to train the random forest regression model and adjust the hyperparameters of the model (the number of decision trees, the maximum depth);

[0077] Use the test set data to evaluate the trained model, and use the mean squared error (MSE) as the evaluation index. The formula is expressed as:

[0078]

[0079] Among them, is the number of samples in the test set, is the actual tool wear cost, is the tool wear cost predicted by the model;

[0080] After the evaluation index is completed, a trained cost relationship model is obtained

[0081] Input the tool usage data obtained in real time into the trained cost relationship model, and the model outputs the predicted special cost data (tool wear) ;

[0082] The process of obtaining maintenance behavior data is as follows:

[0083] Use a handheld terminal to scan the QR code or RFID tag on the device to obtain the unique identifier of the device , then record the tools used and the information of replaced parts, store this information in the form of structured data, and at the same time collect sensor data. Combine the data collected by the sensor and the unique identifier recorded by the handheld terminal According to the device identifier and maintenance information, integrate data from different sources. For the maintenance behavior data at a certain moment is represented as a multi-tuple , among which, the multi-tuple consists of the measured values of the sensors at the moment , the maintenance operation content, tool usage information, replaced part information and unique identifier recorded at the moment

[0084] S2: Use the location data of the device as the graph nodes, and the functional association data and interaction data between devices as the edges of the graph to construct a graph neural network, and construct a device association model based on the graph neural network. Use the status data of the device as the input of the device association model and output the status association factor;

[0085] The process of constructing the graph neural network is as follows:

[0086] Equip each device in the digital factory with a UWB tag to obtain the three-dimensional coordinates of the device in real time as the location data , and represent the location data of each device as a node feature vector , where represents the UWB tag number;

[0087] Extract the functional association relationships between devices from the operation logs of the digital factory. For example, if device 1 provides processing raw materials for device 2, then there is a functional association between device 1 and device 2, which is represented by a numerical value​​ Represents the weight of the functional association between devices, and the value range is , 0 indicates no association, Indicates a strong association. Organize the functional association weights between all devices into a functional association adjacency matrix , this functional association adjacency matrix That is, the functional association data;

[0088] Deploy data monitoring software at the data transmission interface of the device to monitor the data interaction between devices in real time, including the frequency and volume of data transmission. Count the number of times and the volume of data transmitted from device to device within a unit time, and obtain the data interaction weight after normalization processing , and organize it into an interaction adjacency matrix , this interaction adjacency matrix That is, the interaction data;

[0089] Define the graph with the location data as the nodes of the graph and the functional association data and interaction data between devices as the edges of the graph. Among them, among them, Is the node set, composed of the nodes corresponding to the device location data, Is the edge set, composed of the functional association data and interaction data between devices. The weight matrix of the edge is the weighted sum of the weight matrices of the functional association data and interaction data, expressed as , among which, Represents the weight of the functional association adjacency matrix Of, Represents the interaction adjacency matrix Of the weight;

[0090] Select the graph convolutional network as the basic architecture of the graph neural network model, and define the message passing and feature update mechanism, that is, the node updates its own features by aggregating the information of neighbor nodes. For node , its feature At the The update formula for layer is:

[0091]

[0092] Among them, Is the layer index, Is the normalization constant, Is the feature of the Layer, Is the Layer weight, Is the Layer bias term, Is the node The number of neighbor nodes;

[0093] Determine the training objective. During the training process, set the objective as the state correlation factor between devices with accurate outputs.

[0094] Train using the cross - entropy loss function, and divide the collected device - related data into a training set, a validation set, and a test set. The division ratio is 70% as the training set, 15% as the validation set, and 10% - 15% as the test set.

[0095] Graph neural network training process: Use the training set data to train the graph neural network. Calculate the gradients of the loss function with respect to the model parameters (weight matrix and bias vector) through the backpropagation algorithm, and then use the optimizer to update the parameters to minimize the loss function. Regularly evaluate the model performance on the validation set, adjust the hyperparameters according to the evaluation results. When the performance of the model on the validation set no longer improves (the loss function no longer decreases), it is considered that the model converges, stop training, and obtain a graph neural network constructed with the device - based location data as the nodes of the graph and the functional association data and interaction data between devices as the edges of the graph.

[0096] The process of constructing the device association model is as follows:

[0097] After pre - processing the state data of the devices, fuse it with the graph neural network as the input of the graph neural network. Let the device state data be , where is the number of data points.

[0098] Specifically, the operating parameters collected through sensor data are the state data. The specific operating parameters include electrical parameters and mechanical parameters. The electrical parameters include voltage, current, power, and resistance, and the mechanical parameters include temperature, vibration, rotational speed, and pressure.

[0099] Using the historical device failure data as supervision information, define the loss function Train the device association model and minimize the loss function to complete the training of the model.

[0100] The device association model outputs the state correlation factor between devices ;

[0101] Specifically, inside the device association model, the graph neural network performs calculations based on the pre - trained parameters and the defined message - passing and feature - updating mechanisms. The graph neural network will calculate according to the functional association weight matrix between devices (reflecting the relationships of dependence, cooperation, etc. in function between devices. For example, if device A provides raw materials for device B, there is a functional association between them), the data interaction weight matrix (reflecting the interaction situations such as the frequency and data volume of data transmission between devices) and the device location information (represented by the three - dimensional coordinates obtained from UWB tags and other devices equipped on the equipment), process the input device status data. The node continuously updates its own features by aggregating the information of neighbor nodes (according to specific aggregation rules, such as considering neighbor node features, edge weights, and normalization constants in graph convolutional networks, etc.). After multiple layers of calculations and feature transformations, finally obtain a result that can reflect the degree of state association between devices;

[0102] S3: Use the comprehensive maintenance data and the state association factor as the input of the reinforcement learning agent state, and adopt the Q - learning algorithm to train the agent to obtain the reinforcement learning agent. Input the real - time status data into the reinforcement learning agent, output the status score, and set the abnormal status threshold. When the status score is lower than the abnormal status threshold, mark the corresponding real - time data as abnormal;

[0103] The construction process of the reinforcement learning agent is as follows:

[0104] The reinforcement learning agent includes a state space, an action space, and a reward function;

[0105] Construct the state space: The state space consists of comprehensive maintenance data and state association factors. Represent the comprehensive maintenance data as an encoded vector, denoted as , where is the maintenance time vector, is the encoded vector of the maintenance type, is the maintenance cost vector, is the maintenance behavior data;

[0106] Then the state space of the agent is ;

[0107] Specifically, the state space covers comprehensive maintenance data (such as maintenance time, maintenance type, maintenance cost, etc.) and device state association factors and other information that can reflect the operating conditions of devices in the digital factory;

[0108] Construct the action space: The action space is defined as the action of judging the device state, that is, marking whether the real - time data of the device is abnormal. It can be represented by discrete actions. Define a discrete action , the action means judging that the real - time data of the device is normal, and the action = 1 means judging that the real - time data of the device is abnormal;

[0109] Specifically, the action space is the action of the agent to judge the device state (judging normal or abnormal);

[0110] Define the reward function: When the agent correctly judges the device state (that is, the judgment result is consistent with the actual situation), give a positive reward, and when the judgment is wrong, give a negative reward;

[0111] Set the actual state of the device to , indicating normal,[[]] indicating abnormal, and the judgment action of the agent is , and the reward function can be defined as:

[0112]

[0113] The process of training the agent through the Q-learning algorithm is as follows:

[0114] Determine the state, action space and reward function of the agent, initialize the Q-network parameters and the experience replay buffer, and set the total number of training time steps and the discount factor;

[0115] At each time step, observe the state of the agent, and through the greedy policy to select an action, and after execution, obtain the reward and the new state feedback from the environment, and store them in the experience replay buffer;

[0116] Randomly sample from the experience replay buffer, and obtain the Q value of the target. At the same time, calculate the loss (number of samples), and then update the Q-network parameters through the backpropagation algorithm;

[0117] The loss (number of samples) gradually decreases during training. At the same time, repeat the Q-network training steps until the total number of training time steps is reached until the Q-learning network converges. Finally, obtain an agent trained by the Q-learning network;

[0118] Specifically, the Q-learning network is a neural network used to estimate the value of state-action pairs. The initial value of its parameters will affect the learning starting point of the agent. At the same time, create an experience replay buffer, which is used to store the experience data generated by the interaction between the agent and the environment (including information such as state, action, reward and new state, etc.). This helps to break the correlation between data and improve the training effect. In addition, set the total number of training time steps and the discount factor. The total number of training time steps determines the training duration of the agent, and the discount factor (the value range is [0, 1]) is used to measure the importance of future rewards. The closer it is to 1, the more the agent pays attention to long-term rewards;

[0119] Determine an abnormal state threshold , and input the real-time device data into the trained reinforcement learning agent. The agent outputs a state score , when , the corresponding real-time device data is abnormal;

[0120] Specifically, the method for determining the abnormal state threshold is as follows:

[0121] In a digital factory, continuously collect the full - life - cycle data of equipment from the time of commissioning to the current stage, including the operating duration, cumulative workload, maintenance records, and operating parameters of the equipment, and construct a performance decay curve based on the full - life - cycle data;

[0122] For example: taking the operating duration as the abscissa and the performance index as the ordinate, in the initial operation of 0 - 500 hours, the average machining accuracy deviation is 0.05mm, and the tool wear is 0.1mm. When running to 1000 hours, the machining accuracy deviation reaches 0.08mm, and the tool wear is 0.2mm; at 2000 hours, the machining accuracy deviation reaches 0.12mm, and the tool wear is 0.4mm. Connect the data points to form a decay curve;

[0123] According to the reaction efficiency decay curve, obtain the decay efficiency value when the decline degree of the equipment performance index is abnormal within a specific time period, that is, within a specific time period of equipment use, when the decline degree of the equipment performance index (reaction efficiency) exceeds the normal decline degree of the performance index that should be within this time period as preset or obtained through statistical analysis, the corresponding decay efficiency value is the abnormal state threshold 。

[0124] Embodiment 2: As Figure 2 shown, a data analysis system for equipment fault detection in a digital factory includes:

[0125] Synchronous acquisition module: Use high - precision clock synchronization technology to synchronously acquire the comprehensive maintenance data of the equipment;

[0126] State - association acquisition module: Take the location data of the equipment as graph nodes, and the functional association data and interaction data between equipment as the edges of the graph to construct a graph neural network, and construct an equipment association model based on the graph neural network. Take the state data of the equipment as the input of the equipment association model and output the state - association factor;

[0127] Comprehensive fault detection module: Take the comprehensive maintenance data and the state - association factor as the input of the reinforcement learning agent state, and use the Q - learning algorithm to train the agent to obtain the reinforcement learning agent. Input the real - time state data into the reinforcement learning agent, output the state score, and set the abnormal state threshold. When the state score is lower than the abnormal state threshold, mark the corresponding real - time data as abnormal.

[0128] In the application, several formulas involved are calculated by taking their numerical values after dimensionless, and the establishment of the formulas is obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so no more details will be elaborated here.

[0129] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those of ordinary skill in the art will appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution.

[0130] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art will understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing device fault detection data in a digital factory, characterized in that, The method steps include: S1: Synchronously collect the comprehensive maintenance data of the device using high-precision clock synchronization technology; S2: Use the location data of the device as graph nodes, and the functional association data and interaction data between devices as the edges of the graph to construct a graph neural network. Based on the graph neural network, construct a device association model. Use the status data of the device as the input of the device association model and output the status association factor; The acquisition processes of the location data, functional association data, and interaction data of the device are as follows: The acquisition process of the location data is: Equip each device in the digital factory with a UWB tag and obtain the three-dimensional coordinates of the device as location data Represent the location data of each device as a node feature vector where represents the UWB tag number; The acquisition process of the functional association data is: Extract the functional association relationships between devices from the operation logs of the digital factory, and use numerical values to represent the weights of the functional associations between devices. The value range is , where 0 indicates no association, represents a strong association. Organize the weights of the functional associations between all devices into a functional association adjacency matrix , that is, functional association data; The acquisition process of the interaction data is: Deploy data monitoring software at the data transmission interface of the device to monitor the frequency and volume of data transmission, and count the number of times and the volume of data transmitted from one device to another within a unit time. After normalization, obtain the data interaction weight , and organize it into an interaction adjacency matrix , that is, interaction data; The process of constructing the graph neural network is: Use the location data as the nodes of the graph, and the functional association data and interaction data between devices as the edges of the graph to obtain the graph; Define the graph as , where is a set of nodes, which consists of nodes corresponding to device location data, is a set of edges, which consists of functional association data and interaction data between devices. The weight matrix of the edges is the weighted sum of the adjacency matrices of the functional association data and the interaction data, expressed as , where represents the weight of the functional association adjacency matrix , represents the weight of the interaction adjacency matrix ; Select the graph convolutional network as the basic architecture of the graph neural network model, and define the message passing and feature update mechanism, that is, the node updates its own features by aggregating the information of neighboring nodes. For node , its feature at the -th layer is updated according to the following formula: ； wherein, is the layer index, is the normalization constant, is the feature of the th layer, is the weight of the th layer, is the bias term of the th layer, is the number of neighbor nodes of node ; Determine the training objective as the status association factor between devices; Perform network training by dividing the collected device-related data into a training set, a validation set, and a test set; After training is completed, obtain the graph neural network; S3: Use the comprehensive maintenance data and the status association factor as the input of the reinforcement learning agent state, and use the Q-learning algorithm to train the agent to obtain the reinforcement learning agent. Input the real-time status data into the reinforcement learning agent, output the status score, and set the abnormal status threshold. When the status score is lower than the abnormal status threshold, mark the corresponding real-time data as abnormal.

2. The method for analyzing equipment fault detection data for a digital factory according to claim 1, wherein The comprehensive maintenance data includes maintenance time, maintenance type, specified maintenance cost information, and maintenance behavior data.

3. The method for analyzing equipment fault detection data for a digital factory according to claim 2, wherein, The specified maintenance cost information includes parts cost data, labor cost data, and special cost data; The acquisition process of the special cost data is: The pressure is measured based on a pressure sensor at a fixed sampling frequency to obtain pressure data ; Periodically detect the wear condition of the tool based on the wear degree sensor to obtain the wear amount value ; Record the full life cycle information of the tool using blockchain technology ; Integrate the pressure data , the wear amount value and the tool full life cycle information as the input data of the cost relationship model, and establish a cost relationship model to obtain special cost data .

4. A method for analyzing device fault detection data for a digital factory according to claim 3, characterized in that The construction process of the cost relationship model is: Based on pressure data , wear amount value and tool full life cycle information Construct a new feature combination, denoted as , where is the acquisition time interval and i is the index; Through the random forest regression model, use the feature combination as the input variable and the actual loss cost of the tool as the output variable to obtain the cost relationship model.

5. A method for analyzing device fault detection data for a digital factory according to claim 1, characterized in that, The construction process of the device association model is: After preprocessing the status data of the device, fuse it with the graph neural network and use it as the input of the graph neural network; Define a loss function using historical equipment failure data as supervision information Train the equipment association model and minimize the loss function to complete the training of the model.

6. A method for analyzing equipment fault detection data for a digital factory according to claim 1, characterized in that, The construction process of the reinforcement learning agent is: The reinforcement learning agent includes a state space, an action space, and a reward function; Based on the comprehensive maintenance data and the status association factor, form the state space; Define the action space as the action of judging the device state, and the action indicates that the real-time data of the device is judged to be normal, and the action = 1 indicates that the real-time data of the device is judged to be abnormal; Define the reward function. When the agent correctly judges the device status, give a positive reward; when the judgment is wrong, give a negative reward; Set the actual status of the device to , indicating normal,[ indicating abnormal; Then train the agent through the Q-learning algorithm to obtain the reinforcement learning agent.

7. A method for analyzing device fault detection data for a digital factory according to claim 1, characterized in that, The process of training the agent through the Q-learning algorithm to obtain the reinforcement learning agent is: Determine the state, action space, and reward function of the agent, initialize the Q-network parameters and the experience replay buffer, and set the total number of training time steps and the discount factor; At each time step, observe the state of the agent, and through the greedy policy to select an action. After execution, obtain the reward and new state feedback from the environment, and store them in the experience replay buffer; Randomly sample from the experience replay buffer, obtain the target Q value, and calculate at the same time. Then update the Q-network parameters through the backpropagation algorithm; The loss gradually decreases during training. At the same time, repeat the Q-network training steps until the total number of training time steps is reached until the Q-learning network converges and the training is completed.

8. An equipment fault detection data analysis system for a digital factory, using the method according to any one of claims 1 to 7, characterized in that, It includes: Synchronous acquisition module: Synchronously collect the comprehensive maintenance data of the device using high-precision clock synchronization technology; Status association acquisition module: Construct a graph neural network with the location data of the device as graph nodes and the functional association data and interaction data between devices as the edges of the graph, construct a device association model based on the graph neural network, use the status data of the device as the input of the device association model, and output the status association factor; Comprehensive fault detection module: Using the comprehensive maintenance data and status correlation factors as the input of the reinforcement learning agent state, and training the agent with the Q-learning algorithm to obtain the reinforcement learning agent. Input the real-time state data into the reinforcement learning agent, output the state score, and set the abnormal state threshold. , when the state score is lower than the abnormal state threshold , mark the corresponding real-time data as abnormal.

Images