Safety tool intelligent management full life cycle tracing system based on internet of things

Abstract

The application discloses an Internet of Things-based safety tool intelligent management full-life-cycle tracing system, and relates to the technical field of warehouse management.The system comprises an Internet of Things data acquisition module, which is used for acquiring environmental parameters and tool use behavior data; a behavior feature extraction module, which is used for feature extraction to obtain a behavior feature sequence; a risk behavior mining module, which is used for statistically obtaining risk behavior features of a fault tool based on the behavior feature sequence, defining a risk behavior graph, and learning embedding vectors and attention weights of the risk behavior by using a graph convolution network; a risk level prediction module, which is used for encoding real-time risk behavior features by using the embedding vectors and the attention weights, and predicting risk levels of the encoded real-time risk behavior features by using an LSTM model; and a risk tracing update module, which is used for generating an event chain and an early warning log when the risk level is greater than or equal to a risk level threshold, and updating the risk behavior graph and the LSTM model, so that potential tool use risks can be found in advance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of warehouse management, in particular to a safety tool intelligent management full life cycle tracing system based on Internet of Things. BACKGROUND

[0002] The safety tool full life cycle management is composed of a vertical interactive platform and a vice cabinet, equipped with a safety tool intelligent management system, covering multiple key fields such as storage layout planning, storage environment monitoring and control, and accurate record of tool in and out of warehouse, effectively solving the problems of incomplete management, lack of standard use, inaccurate and unstable tool information identification, etc. that safety tools face daily, and realizing the change from disordered management to orderly and refined management.

[0003] Chinese patent application with publication number CN116882896A discloses a power supply station tool in and out of warehouse management system based on identity recognition. When the system is running, identity verification and recognition ensure that only authorized personnel use the system, enhancing the security of the system. Differentiation of user roles and permission control effectively manage user permissions, ensuring the integrity and controllability of system operation. Automatic tool information acquisition improves data accuracy and reliability. Extraction and correlation of relevant features help further understand the attributes and status of tools, support subsequent analysis and management decisions, and automated in and out of warehouse operations reduce the risk of human error and inaccurate information, improving data reliability and accuracy. Then generate in and out of warehouse documents and update tool information, which helps to track and record the status and location of tools in real time, providing traceability and management convenience.

[0004] However, in the existing safety tool management system, when a tool is damaged or involved in an accident, the system can only analyze the responsibility chain after the fact and cannot discover potential risks in advance, resulting in a lack of prevention mechanism. SUMMARY

[0005] The present application aims to at least partially solve one of the technical problems in the related art. To this end, one object of the present application is to propose a safety tool intelligent management full life cycle tracing system based on Internet of Things, which can discover potential tool use risks in advance and effectively prevent safety accidents.

[0006] One aspect of the present application provides a safety tool intelligent management full life cycle tracing system based on Internet of Things, comprising:

[0007] An Internet of Things data acquisition module for acquiring environmental parameters and tool use behavior data;

[0008] A behavior feature extraction module for extracting features from historical use behavior data and environmental parameters to obtain a behavior feature sequence.

[0009] a risk behavior mining module configured to count risk behavior features of the faulty tool based on the behavior feature sequence, define a risk behavior graph with risk behaviors as nodes and risk behavior correlations as edges, and learn embedding vectors and attention weights of the risk behaviors in the risk behavior graph using a graph convolution network;

[0010] a risk level prediction module configured to encode real-time risk behavior features using the embedding vectors and the attention weights, and predict risk levels of the encoded real-time risk behavior features using an LSTM model;

[0011] a risk trace updating module configured to set a risk level threshold, generate an event chain and an early warning log when the risk level is greater than or equal to the risk level threshold, and update the risk behavior graph and the LSTM model according to the early warning log.

[0012] The specific method for obtaining the environmental parameters and the use behavior data of the tools is as follows:

[0013] A sensor is installed on each tool to collect the environmental parameters and the use behavior data; the environmental parameters include temperature data, humidity data, and location data; and the use behavior data includes use frequency, use duration, and vibration data.

[0014] The specific method for extracting features from the historical use behavior data and the environmental parameters to obtain the behavior feature sequence is as follows:

[0015] Step S210: A time window length is set, the use frequency in each time window is counted, and a use frequency sequence is obtained.

[0016] Step S220: The N tools with the highest use frequency of personnel in each time window are counted, and a tool vector is formed.

[0017] Step S230: The average duration of use of each tool by personnel in each time window is calculated based on the use duration, and a duration matrix is formed.

[0018] Step S240: The deviation degree of the actual trajectory of the tool from the standard trajectory in each time window is calculated.

[0019] Step S250: The vibration data is synthesized, the vibration amplitude is calculated, a vibration threshold is set, and the proportion of the vibration amplitude greater than or equal to the vibration threshold in each time window is counted.

[0020] Step S260: The safe ranges of temperature and humidity are set, and the proportions of temperature and humidity exceeding the safe ranges in each time window are counted.

[0021] Step S270: The extracted usage frequency sequence, tool vector, time length matrix, deviation degree, vibration amplitude proportion, temperature proportion, and humidity proportion are taken as features extracted from historical usage behavior data and environmental parameters to obtain a behavior feature segment of each time window;

[0022] Step S280: The behavior feature segments of all time windows are constructed into a behavior feature sequence in chronological order.

[0023] The specific method for the behavior feature sequence to statistically obtain risk behavior features of a faulty tool is:

[0024] A historical accident data set is constructed, which is composed of behavior feature sequences of a faulty tool in M time windows before the occurrence of a fault in an accident case and an accident label of the corresponding accident case; for each accident case, the frequent patterns of its behavior feature sequence are extracted, the frequent patterns of all accident cases are combined to obtain a frequent pattern set; in the frequent pattern set, a frequent pattern with a confidence greater than or equal to a confidence threshold is selected as a risk behavior; for each accident case, its risk behavior features are generated, which are composed of the behavior feature sequence of the corresponding accident case and an indication value sequence indicating whether each risk behavior is contained in the behavior feature sequence.

[0025] The specific method for defining a risk behavior graph with risk behaviors as nodes and risk behavior correlations as edges and using a graph convolution network to learn embedding vectors and attention weights of risk behaviors in the risk behavior graph is:

[0026] Step S310: A risk behavior graph is defined with risk behaviors as nodes and risk behavior correlations as edges, the attribute of a node is the probability of the risk behavior appearing in all accident cases, and the weight of an edge is the probability of two risk behaviors appearing together in the same time window;

[0027] Step S320: A graph convolution network is used to perform feature propagation and learning on the risk behavior graph to obtain an embedding vector of each node of a risk behavior;

[0028] Step S330: An attention mechanism is used to assign an attention weight to each risk behavior.

[0029] The specific method for using embedding vectors and attention weights to encode real-time risk behavior features and using an LSTM model to predict the risk level of the encoded real-time risk behavior features is:

[0030] Step S410: Real-time usage behavior data and environmental parameters are recorded, and a real-time behavior feature sequence thereof divided by time windows is obtained;

[0031] Step S420: obtaining a real-time risk behavior feature based on whether each risk behavior is contained in the real-time behavior feature sequence, and splicing the real-time risk behavior feature sequence and the real-time behavior feature sequence to obtain the real-time risk behavior feature;

[0032] Step S430: encoding the real-time risk behavior feature sequence based on the embedding vector of the risk behavior and the attention weight to obtain the encoded real-time risk behavior feature for each time window;

[0033] Step S440: using the encoded real-time risk behavior feature to constitute a feature sequence, and predicting the risk level of each time window based on the LSTM model.

[0034] The training process of the LSTM model is as follows:

[0035] Step S441: encoding the indicator value in the risk behavior feature of the historical accident data set to obtain an encoded risk behavior feature sequence;

[0036] Step S442: labeling the corresponding risk level label for the encoded risk behavior feature sequence of all accident cases, defining the time window length f, taking the first f-1 encoded risk behavior features in the time window as the input, and taking the fth risk level label in the time window as the output, and combining the encoded risk behavior feature sequence and the risk level label into training data;

[0037] The labeling method of the risk level label is as follows: for each time window, generating a risk level label according to whether an accident occurs in the subsequent time window, if an accident occurs in the subsequent q time windows, the risk level label is 1, otherwise the risk level label is 0;

[0038] Step S443: defining the structure of the LSTM model, including the input layer, the LSTM layer and the output layer, using the binary cross entropy loss function as the loss function for measuring the difference between the predicted risk level and the real risk level label, using the back propagation algorithm and the optimizer to minimize the loss function, and updating the LSTM model parameters;

[0039] Step S444: when the value of the loss function converges to the minimum, a trained LSTM model is obtained.

[0040] The specific method for setting the risk level threshold, generating an event chain and a warning log when the risk level is greater than or equal to the risk level threshold, and updating the risk behavior graph and the LSTM model according to the warning log is as follows:

[0041] Step S510: setting a risk level threshold, triggering a warning when the risk level is greater than or equal to the risk level threshold, and the system automatically setting the tool state to be repaired and prohibiting further borrowing;

[0042] Step S520: Extract the sequence of indication values of the risk behavior of the tool in the last f time windows, associate the risk behavior with the personnel, environment, and task, and generate an event chain;

[0043] Step S530: Record the time of each warning, the tool, the personnel, the sequence of indication values of the risk behavior, and the investigation results, generate a warning log, update the risk behavior graph based on the warning log, and recalculate the embedding vectors and attention weights of the nodes in the risk behavior graph based on the updated risk behavior graph;

[0044] Step S540: Update the encoding of the real-time risk behavior features using the updated embedding vectors and attention weights;

[0045] Step S550: Add the encoded risk behavior feature sequence of the confirmed risk accident case in the warning log to the training data, and update the LSTM model parameters regularly.

[0046] One aspect of the present application provides a safety tool intelligent management full life cycle tracing method based on Internet of Things, comprising:

[0047] Step S100: Obtain the usage behavior data of the tool and the environmental parameters;

[0048] Step S200: Feature extraction is performed on the historical usage behavior data and environmental parameters to obtain a behavior feature sequence;

[0049] Step S300: Based on the behavior feature sequence, the risk behavior features of the faulty tool are counted, a risk behavior graph is defined with the risk behavior as the node and the risk behavior association as the edge, and the embedding vectors and attention weights of the risk behavior in the risk behavior graph are learned using a graph convolution network;

[0050] Step S400: The real-time risk behavior features are encoded using the embedding vectors and attention weights, and the risk level of the encoded real-time risk behavior features is predicted using an LSTM model;

[0051] Step S500: Set a risk level threshold, generate an event chain and a warning log when the risk level is greater than or equal to the risk level threshold, and update the risk behavior graph and the LSTM model according to the warning log.

[0052] The safety tool intelligent management full life cycle tracing system based on Internet of Things proposed in the present application has the following advantages compared with the prior art:

[0053] The application realizes automatic collection, remote transmission and cloud storage of tool data by deploying various sensors on tools and using Internet of Things technology. By setting a time window, the continuous tool data is discretized, and the behavior features are extracted from multiple dimensions to obtain a structured behavior feature sequence, which provides a suitable data representation for risk behavior analysis.

[0054] The application uses frequent pattern mining technology to mine risk behaviors leading to accidents from massive historical accident data, reveals the behavioral root of accidents, and provides key objects for risk prevention and control.

[0055] The application constructs a global risk behavior graph by constructing a risk behavior graph, taking the mined risk behaviors as nodes and the association between behaviors as edges, and systematically analyzes the association network of risk behaviors.

[0056] The application uses graph convolution network technology to reveal the internal mechanism of risk evolution and automatically learn the deep features of risk behaviors; at the same time, through the attention mechanism, the importance difference of different risk behaviors is learned, providing a new method for risk quantitative evaluation.

[0057] The application applies the embedding vector and attention weight of risk behavior to the coding of real-time data, highlights the influence of key risk behaviors, obtains a risk feature representation considering the correlation and importance of risk behaviors, provides information-rich feature input for risk prediction, and uses an LSTM model for risk level prediction, fully mines the time dependence of risk behaviors, and makes a prediction on the evolution trend of risk, realizes dynamic estimation and trend warning of risk.

[0058] The application generates an event chain and a warning log to provide data support for accident cause analysis and responsibility tracing, moves the safety management forward, realizes early detection, early prevention and early disposal of risk, and improves the timeliness and refinement level of safety management.

[0059] The application introduces an online updating mechanism of model and knowledge, gives the system the ability of self-learning and continuous evolution, realizes a dynamic closed loop of “knowledge-data-knowledge”, adapts to the change of tool using environment, and continuously maintains the effectiveness and accuracy of risk warning. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A timing structure diagram of the safety tool intelligent management full life cycle tracing system based on Internet of Things provided by the application;

[0061] Figure 2 A risk behavior feature mining flowchart provided by the application;

[0062] Figure 3A structural schematic of a risk behavior graph provided for the present application;

[0063] Figure 4 A method flowchart of a safety tool intelligent management full life cycle tracing method based on Internet of Things provided for the present application. DETAILED DESCRIPTION

[0064] For a better understanding of the present application, various aspects of the present application will be described in more detail with reference to the accompanying drawings. It is to be understood that the detailed description is merely descriptive of illustrative embodiments of the present application and is not intended to limit the scope of the present application in any way. Throughout the description, like reference numerals refer to like elements. The expression “and / or” includes any and all combinations of one or more of the associated listed items.

[0065] In the drawings, the size, dimensions, and shapes of the elements have been slightly adjusted for ease of illustration. The drawings are merely illustrative and not strictly drawn to scale. As used in this document, the terms “substantially,” “approximately,” and similar terms are used as terms of approximation and not as terms of degree, and are intended to account for the inherent deviations in measuring or calculating values that would be recognized by those of ordinary skill in the art. Additionally, in the present application, the order of the steps of the process descriptions does not necessarily indicate the order in which the processes occur in actual operation, unless otherwise explicitly defined or derivable from context.

[0066] It should also be understood that expressions such as “include”, “including”, “have”, “has”, “contain”, and / or “containing” are open-ended terms that are intended to mean the presence of the stated feature, element, and / or component, but do not preclude the presence or addition of one or more other features, elements, components, and / or combinations thereof. Furthermore, when such expressions as “at least one of” appear in a list of two or more items, it means that any combination of one or more of the listed items is acceptable. In addition, when describing embodiments of the present application, the use of “can” means “one or more embodiments of the present application”. Also, the use of the term “exemplary” is intended to present an example or an illustration.

[0067] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and not be interpreted in an overly formal or overly literal sense unless expressly so defined herein.

[0068] It should be noted that the embodiments and the features in the embodiments in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0069] Embodiment 1:

[0070] As shown in the present application, the safety tool intelligent management full life cycle tracing system based on Internet of Things comprises: Figure 1

[0071] The Internet of Things data acquisition module is used to acquire environmental parameters and use behavior data of the tool;

[0072] The specific method for acquiring the environmental parameters and the use behavior data of the tool is:

[0073] A sensor is installed on each tool to collect environmental parameters and use behavior data; the environmental parameters include temperature data, humidity data and position data; the use behavior data includes use frequency, use duration and vibration data;

[0074] The use frequency means the number of uses within a period of time recorded by a counter.

[0075] The use duration means the tool use time obtained by recording the start and end time of each use by a timer.

[0076] The vibration data means three-axis vibration data collected by an acceleration sensor.

[0077] The temperature data means the environmental temperature collected by a temperature sensor.

[0078] The humidity data means the environmental humidity collected by a humidity sensor.

[0079] The position data means the position coordinates collected by a GPS or indoor positioning module.

[0080] Further, the environmental parameters and the use behavior data of the tool are sampled at a fixed frequency by setting a sampling interval.

[0081] Further, the collected environmental parameters and use behavior data are uploaded to a cloud server in real time through RFID to realize remote transmission and centralized storage of data.

[0082] The present application can collect the use behavior data and the environmental parameters of the tool in real time by installing a sensor on each tool, and the sensor is sampled at a fixed frequency to ensure the continuity of the collected data. The data acquisition method based on Internet of Things can greatly improve the timeliness and reliability of the data, and provides a data basis for subsequent feature extraction, data analysis and risk assessment. ​

[0083] a behavior feature extraction module configured to extract features from historical use behavior data and environmental parameters to obtain a behavior feature sequence;

[0084] The specific method for extracting features from the historical use behavior data and the environmental parameters to obtain the behavior feature sequence is as follows:

[0085] Step S210: setting a time window length, counting the use frequency in each time window to obtain a use frequency sequence;

[0086] The time window length is , and the number of sampling points contained in each time window is The use frequency in M time windows is counted to obtain a use frequency sequence represented as , wherein is the use frequency in the mth time window, m = 1, 2,..., M, is a single sampling duration, represents rounding down;

[0087] Step S220: counting N tools with the highest use frequency in each time window to form a tool vector;

[0088] The tool vector is represented as , wherein represents the nth tool with the highest use frequency, N represents the number of tools with the highest use frequency, and n = 1, 2,..., N;

[0089] Step S230: calculating the average duration of each tool used by personnel in each time window based on the use duration to form a duration matrix;

[0090] The duration matrix is , wherein represents the duration matrix of each tool used by personnel in the mth time window, represents the average duration of the tool used by personnel;

[0091] Step S240: calculating the deviation degree of the actual trajectory of the tool from the standard trajectory in each time window;

[0092] The deviation degree is calculated using a dynamic time warping algorithm;

[0093] Step S250: synthesizing the vibration data, calculating the vibration amplitude, setting a vibration threshold, and counting the proportion of the vibration amplitude greater than or equal to the vibration threshold in each time window; ​

[0094] The vibration data In , , respectively represent the vibration data on the x, y, z axes; the vibration threshold Set by the person skilled in the art according to experience.

[0095] The vibration amplitude The calculation formula is: ;

[0096] The calculation formula of the vibration amplitude ratio is: , wherein represents the vibration amplitude ratio in the mth time window, is the number of sampling points greater than or equal to the vibration threshold;

[0097] Step S260: Set the safety range of temperature and humidity, and count the temperature ratio and humidity ratio in each time window that exceed the safety range;

[0098] The calculation formula of the temperature ratio is: , wherein is the temperature ratio in the mth time window that exceeds the safety range of temperature, is the number of sampling points that exceed the safety range of temperature;

[0099] The calculation formula of the humidity ratio is: , wherein is the humidity ratio in the mth time window that exceeds the safety range of humidity, is the number of sampling points that exceed the safety range of humidity;

[0100] Step S270: Take the extracted usage frequency sequence, tool vector, time length matrix, deviation degree, vibration amplitude ratio, temperature ratio, and humidity ratio as features extracted from historical usage behavior data and environmental parameters, and obtain the behavior feature segment of each time window;

[0101] The behavior feature segment can be represented as ;

[0102] Step S280: Construct a behavior feature sequence according to the time sequence of the behavior feature segments of all time windows.

[0103] The above steps extract features from the collected historical data. By setting time windows, continuous time series data is divided into discrete time segments, each corresponding to a behavioral feature segment. This time window-based feature extraction method in this application considers both the time dependence of the data and facilitates subsequent feature analysis. The extracted behavioral features comprehensively depict the usage status and environment of the tools and equipment, providing data support for subsequent risk behavior mining.

[0104] The risk behavior mining module is used to statistically analyze the risk behavior features of faulty tools based on the behavior feature sequence. It defines a risk behavior graph with risk behavior as nodes and risk behavior associations as edges, and uses a graph convolutional network to learn the embedding vectors and attention weights of risk behaviors in the risk behavior graph.

[0105] like Figure 2 The diagram shown is a flowchart of the risk behavior feature mining process provided in this application. The specific method for statistically analyzing the risk behavior features of faulty tools based on behavior feature sequences is as follows: A historical accident dataset is constructed, consisting of the behavior feature sequences of faulty tools within M time windows prior to the fault in accident cases and the accident labels of the corresponding accident cases; for each accident case, frequent patterns of its behavior feature sequences are extracted, and the frequent patterns of all accident cases are merged to obtain a frequent pattern set; frequent patterns with confidence levels greater than or equal to a confidence threshold are selected from the frequent pattern set as risk behaviors; for each accident case, its risk behavior features are generated, consisting of the behavior feature sequence of the corresponding accident case and an indicator value sequence indicating whether the behavior feature sequence contains each risk behavior;

[0106] The historical accident dataset can be represented as follows: Where G represents the total number of accident cases. Let M be the sequence of behavioral characteristics of the faulty tool in the g-th accident case within M time windows prior to the accident, where... This represents the behavioral characteristic segment within the m-th time window prior to the occurrence of the g-th accident case, where m = 1, 2, ..., M. For accident labels, =0 indicates that no accident occurred in the g-th accident case. =1 indicates that an accident has occurred in the g-th accident case.

[0107] The frequent patterns refer to combinations of behavioral features with a support greater than or equal to a support threshold in the behavioral feature sequence of accident cases; the frequent patterns are extracted using data mining algorithms; preferably, the data mining algorithm is the Apriori algorithm;

[0108] The formula for calculating the support is: ,in, a behavior feature combination indicating that support degree needs to be calculated, a u-th behavior feature segment in the behavior feature combination indicating that support degree needs to be calculated, a number of times of occurrence of the behavior feature combination in the behavior feature sequence; the support degree threshold value is 0.1-0.5;

[0109] The frequent pattern can be represented as , wherein, a number of behavior feature combinations in the behavior feature sequence of the g-th accident case whose support degrees are greater than or equal to the support degree threshold value, a u-th behavior feature combination in the frequent pattern;

[0110] The frequent pattern set can be represented as:

[0111] The confidence degree is calculated according to the following formula: ; wherein is a number of accident cases containing the frequent pattern

[0112] The confidence degree threshold value is 0.5-0.9, the frequent pattern whose confidence degree is greater than or equal to the confidence degree threshold value is selected as a risk behavior, and is recorded as K represents a total number of risk behaviors mined from the historical accident data set, a K-th risk behavior;

[0113] The risk behavior feature is represented as , wherein, is a sequence of indication values indicating whether each risk behavior is contained in the behavior feature sequence, , an indication value indicating whether the k-th risk behavior is contained in the behavior feature sequence, k = 1, 2,..., K, , = 0 indicates that the k-th risk behavior is not contained in the behavior feature sequence, = 1 indicates that the k-th risk behavior is contained in the behavior feature sequence.

[0114] The above steps provide data support and interpretability for subsequent risk assessment and embedding vector learning of the graph convolution network by mining the association pattern of the risk behavior from the historical accident data set.

[0115] The specific method for defining a risk behavior graph with the risk behavior as a node and the risk behavior association as an edge and learning embedding vectors and attention weights of the risk behavior in the risk behavior graph using the graph convolution network is as follows:

[0116] ​​​Step S310: defining a risk behavior graph, taking risk behaviors as nodes and risk behavior correlations as edges, the attribute of a node being the probability of the risk behavior appearing in all accident cases, and the weight of an edge being the probability of two risk behaviors appearing in the same time window;

[0117] As shown in FIG. 1, a structure diagram of a risk behavior graph provided by the present application is shown, the risk behavior graph being a graph structure composed of nodes and edges between the nodes in the graph; Figure 3

[0118] The calculation formula of the attribute of the node is: wherein, is the attribute of the node corresponding to the ki-th risk behavior, indicates the indication value of whether the ki-th risk behavior is contained in the m-th time window of the g-th accident case;

[0119] The calculation formula of the weight of the edge is: wherein, is the weight of the edge between the nodes corresponding to the ki-th and kj-th risk behaviors, indicates the indication value of whether the kj-th risk behavior is contained in the m-th time window of the g-th accident case;

[0120] Step S320: using a graph convolution network to perform feature propagation and learning on the risk behavior graph to obtain an embedding vector of each node of the risk behavior graph;

[0121] The graph convolution network includes an input layer, a convolution layer and an output layer. The adjacency matrix and the node feature matrix of the risk behavior graph are input to the input layer. The convolution layer is stacked by multiple graph convolution layers. Each graph convolution layer performs forward propagation on the risk behavior features of the node to obtain an embedding representation of the node. The output layer is the last graph convolution layer, which outputs the final embedding vector of the node.

[0122] Step S330: using an attention mechanism to assign an attention weight to each risk behavior;

[0123] The calculation formula of the attention weight is: wherein, , respectively indicate the attention scores of the risk behavior and the risk behavior , , indicate the ki-th and kj-th risk behaviors, is an exponential function;

[0124] The calculation formula of the attention score is: v, , ​Let be the learnable parameters of the graph convolutional network, and let represent the attention vector, weight matrix, and bias term, respectively. This is the transpose of the attention vector. For risky behavior Embedded vector, This is the activation function.

[0125] In the above steps, the graph convolutional network performs feature propagation and aggregation on the risk behavior graph, learns the correlations and implicit patterns between risk behaviors, identifies risk behaviors that may lead to accidents, and provides a more comprehensive and in-depth feature representation for risk assessment. This effectively mines the risk behavior patterns in historical accident datasets, improving the accuracy and sensitivity of risk perception. Then, a risk behavior graph is constructed using risk behaviors as nodes, and the graph convolutional network is used to learn features from the risk behavior graph, obtaining an embedding vector for each risk behavior node. This embedding vector contains the semantic information and contextual relationships of the risk behavior. Furthermore, this application uses an attention mechanism to assign attention weights to each risk behavior, highlighting the impact of key risk behaviors. This fully utilizes the representational capabilities of the graph convolutional network to automatically learn the risk behavior patterns during tool usage, greatly improving the intelligence level of risk assessment.

[0126] The risk level prediction module is used to encode real-time risk behavior features using embedding vectors and attention weights, and then use an LSTM model to predict the risk level of the encoded real-time risk behavior features.

[0127] The specific method for encoding real-time risk behavior features using embedding vectors and attention weights, and then using an LSTM model to predict the risk level of the encoded real-time risk behavior features is as follows:

[0128] Step S410: Record real-time usage behavior data and environmental parameters, and obtain a real-time behavior feature sequence divided by time windows;

[0129] The real-time behavioral feature sequence can be represented as: ,in, This represents the behavioral feature segment within the m-th time window of the real-time behavioral feature sequence, where m = 1, 2, ..., M;

[0130] Step S420: Based on whether the real-time behavior feature sequence contains each risk behavior, a real-time indication value sequence is obtained. The real-time risk behavior feature is obtained by concatenating the real-time indication value sequence and the real-time behavior feature sequence.

[0131] The real-time risk behavior characteristics can be represented as follows: Where b is the real-time indicator value sequence of whether the real-time behavioral feature sequence contains each risky behavior. , an indication value indicating whether the kth risk behavior is contained in the real-time behavior feature sequence, k = 1, 2, …, K, , = 0 indicates that the kth risk behavior is not contained in the real-time behavior feature sequence, = 1 indicates that the kth risk behavior is contained in the real-time behavior feature sequence;

[0132] Step S430: encoding the real-time indication value sequence based on the embedding vector of the risk behavior and the attention weight for each time window to obtain the encoded real-time risk behavior feature;

[0133] The encoded real-time risk behavior feature is represented as ; wherein, is the indication value of the kth risk behavior in the mth time window;

[0134] Step S440: using the encoded real-time risk behavior feature to constitute a feature sequence, and predicting the risk level of each time window based on an LSTM model;

[0135] The risk level is represented as , indicating the risk level of the mth time window, when = 1, indicating high risk, when = 0, indicating no risk;

[0136] The training process of the LSTM model is as follows:

[0137] Step S441: encoding the indication values in the risk behavior features of the historical accident data set to obtain an encoded risk behavior feature sequence;

[0138] The encoded risk behavior feature sequence can be represented as ;

[0139] Step S442: labeling the corresponding risk level label for the encoded risk behavior feature sequence of all accident cases, defining the time window length f, taking the first f-1 encoded risk behavior features in the time window as input, and taking the fth risk level label in the time window as output, and combining the encoded risk behavior feature sequence and the risk level label into training data;

[0140] Preferably, the time window length f is 3;

[0141] The labeling method of the risk level label is as follows: for each time window, generating a risk level label according to whether an accident occurs in the subsequent time windows, if an accident occurs in the subsequent q time windows, the risk level label is 1, otherwise the risk level label is 0;

[0142] Preferably, q = 3.

[0143] Step S443: defining the structure of the LSTM model, including the input layer, the LSTM layer and the output layer, using the binary cross-entropy loss function as the loss function for measuring the difference between the predicted risk level and the real risk level label, using the back propagation algorithm and the optimizer to minimize the loss function, and updating the LSTM model parameters;

[0144] Step S444: obtaining the trained LSTM model when the value of the loss function converges to the minimum.

[0145] The above steps learn the long-term dependence relationship and risk evolution law in the time feature sequence by training the LSTM model on the encoded risk behavior feature sequence, thereby predicting the risk level of each time window. The trained LSTM model is used to perform real-time risk assessment on the real-time collected usage behavior data and environmental parameters, thereby providing support for early warning and decision-making.

[0146] For example, when new usage behavior data and environmental parameters are uploaded to the cloud, the behavior feature sequence thereof is first extracted, and then the risk behavior embedding vector and the attention weight are used to encode the risk behavior indicator value to obtain a comprehensive real-time risk behavior feature. Finally, the encoded real-time risk behavior feature sequence is input into the pre-trained LSTM model, and the LSTM model is used to predict the risk level of each time window by automatically learning the long-term dependence relationship of the feature sequence, thereby capturing the time evolution law of the risk behavior and realizing the automation and real-time of risk assessment.

[0147] The risk trace updating module is configured to set a risk level threshold, generate an event chain and an early warning log when the risk level is greater than or equal to the risk level threshold, and update the risk behavior graph and the LSTM model according to the early warning log;

[0148] The specific method of setting the risk level threshold, generating the event chain and the early warning log when the risk level is greater than or equal to the risk level threshold, and updating the risk behavior graph and the LSTM model according to the early warning log is as follows:

[0149] Step S510: setting the risk level threshold When the risk level is greater than or equal to the risk level threshold, an early warning is triggered, and the system automatically sets the tool state to be under repair and prohibits further borrowing;

[0150] The value of the risk level threshold is 1;

[0151] Step S520: extracting the indicator value sequence of the risk behavior of the tool in the last f time windows, associating the risk behavior with personnel, environment and task, and generating an event chain;

[0152] Step S530: record the time of each warning, the warning tool, the personnel, the indication value sequence of the risk behavior, the subsequent investigation result, generate a warning log, update the risk behavior graph based on the warning log, and recalculate the embedding vector and the attention weight of the node in the risk behavior graph based on the updated risk behavior graph;

[0153] The investigation result is divided into confirmed risks and false positives;

[0154] Step S540: update the real-time risk behavior features using the updated embedding vector and attention weight;

[0155] Step S550: add the encoding risk behavior feature sequence of the accident case of the confirmed risk in the warning log to the training data, and update the LSTM model parameters regularly.

[0156] The above steps realize the continuous updating and optimization of the risk behavior graph and the LSTM model through human-computer interaction and model updating.

[0157] The safety tool intelligent management full life cycle tracing system based on the Internet of Things provided in the application fully utilizes the Internet of Things technology for data acquisition, transmission and storage, forming a complete data closed loop. In addition, data mining is performed using the data collected based on the Internet of Things technology, and the intelligent analysis and risk assessment of massive historical data and real-time data are realized by combining a graph neural network and an LSTM model, thereby realizing an automatic and optimizable safety tool intelligent management tracing system.

[0158] Embodiment 2

[0159] As shown in Figure 4 The safety tool intelligent management full life cycle tracing method based on the Internet of Things provided in the application comprises:

[0160] Step S100: acquire environmental parameters and usage behavior data of the tool;

[0161] Step S200: extract features from historical usage behavior data and environmental parameters to obtain a behavior feature sequence;

[0162] Step S300: statistically obtain risk behavior features of a faulty tool based on the behavior feature sequence, define a risk behavior graph with risk behaviors as nodes and risk behavior correlations as edges, and learn embedding vectors and attention weights of the risk behaviors in the risk behavior graph using a graph convolution network;

[0163] Step S400: encode real-time risk behavior features using the embedding vectors and the attention weights, and predict risk levels of the encoded real-time risk behavior features using an LSTM model;

[0164] Step S500: set a risk level threshold, generate an event chain and a warning log when the risk level is greater than or equal to the risk level threshold, update the risk behavior graph and the LSTM model according to the warning log.

[0165] In addition, the part of the above technical solutions provided in the embodiments of the present application which is consistent with the implementation principle of the corresponding technical solutions in the prior art is not described in detail to avoid excessive repetition.

[0166] The specific embodiments described above further illustrate the objects, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An Internet of Things-based safety tool intelligent management full-life cycle tracing system, characterized in that, The method comprises the following steps: An Internet of Things data acquisition module is used to acquire environmental parameters and usage behavior data of tools and instruments; A behavior feature extraction module is used to extract features from historical usage behavior data and environmental parameters to obtain a behavior feature sequence; A risk behavior mining module is used to count risk behavior features of faulty tools and instruments based on the behavior feature sequence, define a risk behavior graph with risk behaviors as nodes and risk behavior correlations as edges, and use a graph convolution network to learn embedding vectors and attention weights of risk behaviors in the risk behavior graph; A risk level prediction module is used to encode real-time risk behavior features using embedding vectors and attention weights, and use an LSTM model to predict risk levels of the encoded real-time risk behavior features; A risk trace updating module is used to set a risk level threshold, generate an event chain and an early warning log when the risk level is greater than or equal to the risk level threshold, and update the risk behavior graph and the LSTM model according to the early warning log; The specific method for extracting features from historical usage behavior data and environmental parameters to obtain a behavior feature sequence is as follows: A time window length is set, the number of uses in each time window is counted to obtain a usage frequency sequence, the N tools and instruments with the highest usage frequency of personnel in each time window are counted to form a tool and instrument vector, the average duration of use of each tool and instrument by personnel in each time window is calculated based on the duration of use to form a duration matrix, the deviation degree of the actual trajectory of a tool and instrument from a standard trajectory in each time window is calculated, vibration data is synthesized, vibration amplitudes are calculated, a vibration threshold is set, and the proportion of vibration amplitudes greater than or equal to the vibration threshold in each time window is counted; the safe ranges of temperature and humidity are set, the proportions of temperatures and humidities that exceed the safe ranges in each time window are counted; the extracted usage frequency sequence, tool and instrument vector, duration matrix, deviation degree, vibration amplitude proportion, temperature proportion, and humidity proportion are used as features extracted from historical usage behavior data and environmental parameters to obtain a behavior feature segment for each time window; and the behavior feature segments of all time windows are arranged in chronological order to construct a behavior feature sequence; The specific method for counting risk behavior features of faulty tools and instruments based on the behavior feature sequence is as follows: a historical accident data set is constructed, the historical accident data set is composed of behavior feature sequences of faulty tools and instruments in M time windows before the occurrence of an accident in an accident case and an accident label of the corresponding accident case; for each accident case, the frequent patterns of the behavior feature sequence thereof are extracted, the frequent patterns of all accident cases are combined to obtain a frequent pattern set; in the frequent pattern set, frequent patterns with a confidence degree greater than or equal to a confidence threshold are selected as risk behaviors; and for each accident case, its risk behavior features are generated, the risk behavior features are composed of the behavior feature sequence of the corresponding accident case and an indication value sequence indicating whether each risk behavior is contained in the behavior feature sequence.

2. The IoT-based safety tool intelligent management full-life-cycle tracing system according to claim 1, wherein, The specific method of acquiring the environmental parameters and the use behavior data of the tools comprises: installing sensors on each tool to collect the environmental parameters and the use behavior data; the environmental parameters comprise temperature data, humidity data and position data; and the use behavior data comprises use frequency, use duration and vibration data. 3.The IoT-based safety tool intelligent management full-life-cycle tracing system of claim 2, wherein, The specific method of defining a risk behavior graph with risk behaviors as nodes and risk behavior correlations as edges, and learning embedding vectors and attention weights of the risk behaviors in the risk behavior graph using a graph convolution network comprises: defining a risk behavior graph with risk behaviors as nodes and risk behavior correlations as edges, the attribute of the node being the probability of the risk behavior occurring in all accident cases, and the weight of the edge being the probability of the two risk behaviors occurring together in the same time window; using a graph convolution network to perform feature propagation and learning on the risk behavior graph to obtain the embedding vector of each node of the risk behavior; using an attention mechanism to assign an attention weight to each risk behavior. 4.The IoT-based safety tool intelligent management full-life-cycle tracing system of claim 3, wherein, The specific method of encoding real-time risk behavior features using the embedding vectors and the attention weights, and predicting the risk level of the encoded real-time risk behavior features using an LSTM model comprises: recording real-time use behavior data and environmental parameters to obtain a real-time behavior feature sequence divided by a time window; obtaining a real-time indication value sequence based on whether each risk behavior is included in the real-time behavior feature sequence, and splicing the real-time indication value sequence and the real-time behavior feature sequence to obtain a real-time risk behavior feature; encoding the real-time indication value sequence based on the embedding vectors and the attention weights of the risk behaviors for each time window to obtain an encoded real-time risk behavior feature; using the encoded real-time risk behavior features to form a feature sequence, and predicting the risk level of each time window based on an LSTM model. 5.The IoT-based safety tool intelligent management full-life-cycle tracing system according to claim 4, wherein, The training process of the LSTM model comprises: encoding the indication values in the risk behavior features of the historical accident data set to obtain an encoded risk behavior feature sequence; annotating the encoded risk behavior feature sequence of all accident cases with a corresponding risk level label, defining a time window length f, taking the first f-1 encoded risk behavior features in the time window as input, and taking the fth risk level label in the time window as output, combining the encoded risk behavior feature sequence and the risk level label into training data; defining the structure of the LSTM model, including an input layer, an LSTM layer and an output layer, using a binary cross-entropy loss function as a loss function for measuring the difference between the predicted risk level and the real risk level label, using a back propagation algorithm and an optimizer to minimize the loss function, and updating the parameters of the LSTM model; when the value of the loss function converges to the minimum, a trained LSTM model is obtained. 6.The IoT-based safety tool intelligent management full-life-cycle tracing system according to claim 5, wherein, The annotation method of the risk level label comprises: for each time window, generating a risk level label according to whether an accident occurs in the subsequent time windows, if an accident occurs in the subsequent q time windows, the risk level label is 1, otherwise the risk level label is 0.

7. The IoT-based safety appliance intelligent management full-life-cycle tracing system of claim 6, wherein, The risk level threshold is set, and the event chain and the early warning log are generated when the risk level is greater than or equal to the risk level threshold. The specific method for updating the risk behavior graph and the LSTM model according to the early warning log is: The risk level threshold is set, and the early warning is triggered when the risk level is greater than or equal to the risk level threshold. The system automatically sets the tool state to be repaired and prohibits further borrowing; Extract the risk behavior indication value sequence of the tool in the last f time windows, associate the risk behavior with the personnel, environment and task, and generate an event chain; Record the time, tool, personnel, risk behavior indication value sequence and troubleshooting result of each early warning, generate an early warning log, update the risk behavior graph based on the early warning log, and recalculate the embedding vector and attention weight of the nodes in the risk behavior graph based on the updated risk behavior graph; Update the real-time risk behavior feature using the updated embedding vector and attention weight; Add the encoding risk behavior feature sequence of the confirmed risk accident case in the early warning log to the training data, and update the LSTM model parameters regularly.

Images