Internet of things based instrument whole life cycle management method and system

Abstract

The application provides a kind of instrument full life cycle management method and system based on Internet of Things, it is related to Internet of Things technical field, including the acquisition unique identification of binding RFID label and instrument, collection operating parameter establishes multidimensional data structure;To the time series clustering analysis of data identifies abnormal mode, constructs device group association graph and calculates fault propagation characteristics;Through multi-head self-attention feature extraction constructs performance decay curve and generates predictive maintenance index;Optimize maintenance task allocation scheme and execute;Update prediction index and allocation result.The application realizes real-time monitoring and predictive maintenance to instrument state, improves equipment management efficiency and operation reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things, and in particular to an instrument full life cycle management method and system based on Internet of Things. BACKGROUND

[0002] With the development of science and technology, instruments and equipment in laboratories and industrial environments are increasingly complex and intelligent. These instruments play an important role in scientific research, production and detection. Full life cycle management of instruments and equipment is a key link to ensure efficient operation, prolong service life and reduce maintenance costs. Traditional instrument management mainly relies on manual records and regular inspections. With the development of Internet of Things technology, it is possible to collect instrument operation data in real time through sensor networks, and to apply big data analysis and artificial intelligence technology for intelligent management. Currently, Internet of Things technology has been widely used in equipment management, including automatic identification technology such as RFID tags for asset tracking, condition monitoring technology for real-time parameter collection, and data analysis technology for equipment health assessment.

[0003] However, the existing instrument management method based on Internet of Things has some obvious defects. The existing technology usually only focuses on the state monitoring of a single instrument, lacks analysis of the interrelation between instrument groups, and cannot effectively identify and handle the fault propagation characteristics between devices, resulting in insufficient system risk warning capability. Traditional maintenance strategies are mainly based on fixed time intervals or simple threshold triggers, lack precise modeling of instrument performance degradation rules, and cannot accurately predict the best maintenance time, often leading to resource waste due to premature maintenance or loss due to late maintenance. Existing technologies mostly use manual experience decision-making in maintenance task allocation, lack adaptive scheduling mechanisms considering multi-dimensional resource constraints and task priority, and are difficult to achieve optimal balance between maintenance timeliness and resource utilization efficiency, limiting the improvement of overall maintenance efficiency. SUMMARY

[0004] The present application provides an instrument full life cycle management method and system based on Internet of Things, which can solve the problems in the prior art.

[0005] In a first aspect of the present application, an instrument full life cycle management method based on Internet of Things is provided, comprising:

[0006] Binding the RFID tag with the instrument to obtain unique identification information, collecting the operating parameters of the instrument, and establishing a multi-dimensional data structure;

[0007] Performing time series clustering analysis on the multi-dimensional data structure, identifying abnormal patterns of instrument operation, constructing a device group correlation graph, calculating fault propagation characteristics according to the fault state transition between nodes in the device group correlation graph, and forming a device operation state evaluation result;

[0008] Based on the equipment operation state evaluation result, a performance attenuation curve of the instrument is constructed through multi-head self-attention feature extraction and fitting function optimization, probability calculation is performed in combination with the fault propagation feature, and a predicted maintenance index is generated;

[0009] Based on the predicted maintenance index, a maintenance task allocation scheme is generated by adopting a scheduling optimization algorithm with maintenance timeliness and resource utilization rate as targets;

[0010] Based on the maintenance task allocation result, maintenance work is performed, maintenance process and result data are stored in a maintenance knowledge base, and the predicted maintenance index and the maintenance task allocation result are updated through online analysis.

[0011] In an optional embodiment, time series clustering analysis is performed on the multi-dimensional data structure, abnormal patterns of instrument operation are identified, and a device group association graph is constructed, including:

[0012] Time series feature extraction is performed on the multi-dimensional data structure, statistical features, frequency domain features and time-frequency features in the multi-dimensional data structure are extracted, an optimal feature subset is determined according to feature importance weights, and a dimension-reduced feature matrix is generated;

[0013] Based on the dimension-reduced feature matrix, time series similarity is calculated by dynamic time warping, a dynamic time warping distance matrix is constructed, and time series clustering results are obtained by iteratively calculating the similarity between samples and clustering centers;

[0014] The time series clustering results are compared with a preset normal operation mode benchmark, a clustering deviation value is calculated, and abnormal patterns of instrument operation are generated according to the comparison result of the clustering deviation value and a preset abnormality determination threshold;

[0015] Based on the abnormal patterns of instrument operation, a device group association graph is constructed, a correlation coefficient between devices is calculated based on device operation state parameters, a connection relationship between nodes is determined by setting an association strength threshold, edge weights are calculated based on device physical locations, operation parameters and the abnormal patterns, and a weighted device group association graph is generated.

[0016] In an optional embodiment, the fault propagation feature is calculated based on the fault state transition between nodes in the device group association graph, and the equipment operation state evaluation result is formed, including:

[0017] The number of fault state transitions between adjacent nodes in the device group association graph is counted in a preset time window, the number of fault state transitions is divided by the length of the time window to obtain a fault transition probability between nodes, and a fault propagation probability matrix is formed;

[0018] The shortest propagation path and propagation time delay between each node pair in the fault propagation probability matrix are calculated by gradually updating the propagation distance between nodes, nodes with a fault transfer probability greater than a preset probability threshold are marked as key propagation nodes, and a fault propagation path graph is generated;

[0019] The connection edge number of each node in the fault propagation path graph is counted, the node influence degree is calculated by multiplying the connection edge number and the fault propagation probability of each edge, the fault diffusion speed is obtained by calculating the ratio of the propagation distance between nodes and the sampling interval, and the area where the node influence degree and the fault diffusion speed exceed the corresponding threshold is marked as a propagation danger area, and the fault influence range is determined by calculating the distance from the edge node of the propagation danger area to the key propagation node;

[0020] The single-device state index is obtained by calculating the mean variance of the device operation parameters, the group cooperation index is obtained by calculating the parameter correlation coefficient between devices, the risk index is determined according to the weighted sum of the node influence degree, diffusion speed and fault influence range, and the device operation state evaluation result is obtained by weighted average of the single-device state index, group cooperation index and risk index.

[0021] In an optional embodiment, based on the device operation state evaluation result, the performance decay curve of the instrument is constructed by multi-head self-attention feature extraction and fitting function optimization, which includes:

[0022] The device operation state evaluation result is input into a multi-head self-attention layer, which includes a query-key-value calculation unit, which captures different time sequence scale features through the query-key-value calculation unit, and determines the attention feature;

[0023] The attention feature is input into a time sequence position encoding layer, a sine cosine function is used to represent the relative position of the attention feature, a time sequence dependent relationship is determined, a position encoding feature is generated;

[0024] The position encoding feature is subjected to dimension mapping transformation to generate an intermediate feature, the intermediate feature is subjected to activation function to obtain a transformed feature, the transformed feature and the position encoding feature are subjected to feature superposition, a short circuit connection path is established to maintain the original feature information, and a fusion feature is obtained;

[0025] Based on the fusion feature, a state change rate sequence is constructed, a state change point is determined, periodic change features and long-term evolution features of the state change point are extracted, and multi-scale performance features are generated;

[0026] According to the time interval and amplitude difference of the state change point, grouping is performed, decay feature grouping is constructed, the decay feature grouping is stored in a decay feature database, and an optimal fitting function is selected;

[0027] Calculate the mean square error of the multi-scale performance feature and the optimal fitting function, construct a gradient descent matrix, adjust the fitting coefficients of the optimal fitting function according to the gradient descent matrix, and repeat the iteration until the mean square error is less than a preset termination threshold, to obtain an instrument performance decay curve.

[0028] In an optional embodiment, a state change rate sequence is constructed based on the fusion feature, a state change point is determined, periodic variation features and long-term evolution features of the state change point are extracted, and the multi-scale performance feature is generated, including:

[0029] The difference value between adjacent time points is calculated according to the fusion feature in time sequence, a state change rate sequence is constructed, the difference value of the state change rate sequence is calculated, and a performance change acceleration sequence is constructed;

[0030] The state change point is determined based on the local maximum and minimum values of the state change rate sequence and the zero points of the performance change acceleration sequence;

[0031] The fluctuation law of the state change point in different time windows is extracted respectively, and a state fluctuation feature sequence is constructed; the periodic variation features are extracted according to the repeated appearance frequency and amplitude change of the state fluctuation feature sequence; the baseline trend of the state change is extracted as the long-term evolution feature through the sliding average processing of the state change point;

[0032] The periodic variation features and the long-term evolution features are combined by weighting, to generate the multi-scale performance feature.

[0033] In an optional embodiment, the scheduling optimization algorithm includes:

[0034] A double-layer coding structure of the maintenance task is constructed based on the predicted maintenance index, the double-layer coding structure including an outer path coding for determining the execution order of the maintenance task and an inner time coding for determining the execution time of the maintenance task;

[0035] The initial pheromone distribution between the maintenance tasks is set in the outer path coding, and a transfer probability matrix between the maintenance tasks is constructed according to the maintenance timeliness index;

[0036] The time allocation space of the maintenance task is established in the inner time coding, and a time allocation probability matrix of the maintenance task is constructed according to the resource utilization rate index;

[0037] The path search of the execution order of the maintenance task is performed based on the transfer probability matrix, and the execution time of the maintenance task is determined based on the time allocation probability matrix;

[0038] The time window constraint and the resource constraint in the prediction maintenance index are taken as temperature parameters to locally adjust the maintenance task allocation scheme searched by the double-layer coding structure, and the initial pheromone distribution is updated based on the target function value after the adjustment.

[0039] Based on the double-layer coding structure, the scheduling optimization algorithm is repeatedly executed until a preset iteration number is reached, and a maintenance task allocation scheme is output.

[0040] In an optional embodiment, path search of the execution order of the maintenance tasks is performed based on the transition probability matrix, and the execution time of the maintenance tasks is determined based on the time allocation probability matrix, and the method further includes:

[0041] In the outer layer path coding, the execution order position of the next maintenance task is selected based on the transition probability matrix.

[0042] In the inner layer time coding, the execution time of the maintenance task corresponding to the execution order position is determined based on the time allocation probability matrix.

[0043] The execution result of the maintenance task at the execution time is fed back to the outer layer path coding, and the transition probability between the maintenance task and other maintenance tasks in the transition probability matrix is updated.

[0044] In a second aspect of the embodiment of the application, an instrument full life cycle management system based on the Internet of Things is provided, including:

[0045] A first unit is configured to bind an RFID tag with an instrument to obtain unique identification information, collect running parameters of the instrument, and establish a multi-dimensional data structure.

[0046] A second unit is configured to perform time series clustering analysis on the multi-dimensional data structure, identify abnormal patterns of instrument operation, construct a device group association graph, calculate fault propagation characteristics between nodes in the device group association graph, and form a device operation state evaluation result.

[0047] A third unit is configured to construct a performance degradation curve of the instrument based on the device operation state evaluation result, through multi-head self-attention feature extraction and fitting function optimization, perform probability calculation in combination with the fault propagation characteristics, and generate a prediction maintenance index.

[0048] A fourth unit is configured to generate a maintenance task allocation scheme by using a scheduling optimization algorithm to take maintenance timeliness and resource utilization rate as targets based on the prediction maintenance index.

[0049] A fifth unit is configured to perform maintenance work based on the maintenance task allocation result, store maintenance process and result data in a maintenance knowledge base, and update the prediction maintenance index and the maintenance task allocation result through online analysis.

[0050] In a third aspect, the present application provides an electronic device, comprising:

[0051] a processor;

[0052] a memory for storing processor-executable instructions;

[0053] wherein the processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0054] In a fourth aspect, the present application provides a computer-readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method described above.

[0055] In the embodiments of the present application, by adopting the Internet of Things technology to bind the RFID tag with the instrument to obtain a unique identifier, and by using multi-dimensional data structure and time series clustering analysis, real-time monitoring of the instrument running state and accurate identification of abnormal patterns are realized, accurate fault warning and diagnosis basis are provided for equipment maintenance, and the device failure rate is significantly reduced; based on the performance decay curve constructed by multi-head self-attention feature extraction and fitting function optimization, combined with the probability calculation of fault propagation characteristics, scientific prediction of instrument maintenance requirements is realized, the accuracy of maintenance decision is improved, unnecessary preventive maintenance is reduced, maintenance cost is reduced, and the service life of the equipment is prolonged; the maintenance task allocation scheme generated by the scheduling optimization algorithm takes into account the timeliness of maintenance and resource utilization, realizes the reasonable allocation of maintenance resources, and at the same time, the establishment of the maintenance knowledge base and the online analysis and updating mechanism form a closed-loop feedback system, continuously optimize the prediction maintenance index and task allocation scheme, and improve the overall equipment management efficiency and reliability. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 FIG. 1 is a flowchart of an instrument full life cycle management method based on the Internet of Things according to an embodiment of the present application;

[0057] Figure 2 FIG. 4 is a flowchart of a predictive maintenance task scheduling optimization algorithm. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0059] The technical solutions of the present application will be described in detail below with specific examples. The following specific examples can be combined with each other, and some examples may not be described in detail for the same or similar concepts or processes.

[0060] Figure 1 The flowchart of the instrument full life cycle management method based on the Internet of Things is shown in the embodiment of the present application, as shown in Figure 1 The method comprises:

[0061] The RFID tag is bound to the instrument to obtain unique identification information, the running parameters of the instrument are collected, and a multi-dimensional data structure is established;

[0062] The multi-dimensional data structure is subjected to time series clustering analysis, abnormal patterns of instrument operation are identified, and a device group correlation graph is constructed. The fault propagation characteristics are calculated according to the fault state transition between the nodes in the device group correlation graph, and the device operation state evaluation result is formed;

[0063] Based on the device operation state evaluation result, the performance degradation curve of the instrument is constructed through multi-head self-attention feature extraction and fitting function optimization, the probability calculation is performed in combination with the fault propagation characteristics, and the predictive maintenance index is generated;

[0064] Based on the predictive maintenance index, a scheduling optimization algorithm is used to generate a maintenance task allocation scheme with timeliness and resource utilization as the target;

[0065] Based on the maintenance task allocation result, the maintenance work is performed, the maintenance process and result data are stored in the maintenance knowledge base, and the predictive maintenance index and the maintenance task allocation result are updated through online analysis.

[0066] In an optional implementation, the time series clustering analysis of the multi-dimensional data structure, the identification of abnormal patterns of instrument operation, and the construction of a device group correlation graph comprise:

[0067] The time series features of the multi-dimensional data structure are extracted, the statistical features, frequency domain features and time-frequency features in the multi-dimensional data structure are extracted, the optimal feature subset is determined according to the feature importance weight, and the feature matrix after dimension reduction is generated;

[0068] Based on the feature matrix after dimension reduction, the time series similarity is calculated by dynamic time warping, the dynamic time warping distance matrix is constructed, and the time series clustering result is obtained by iteratively calculating the similarity between the sample and the clustering center;

[0069] The time series clustering result is compared with the preset normal operation mode benchmark, the clustering deviation value is calculated, and the abnormal patterns of instrument operation are generated according to the comparison result of the clustering deviation value and the preset abnormality determination threshold;

[0070] Based on the abnormal mode of instrument operation, a device group correlation graph is constructed, the correlation coefficient between devices is calculated by extracting the device running state parameters, the connection relationship of nodes is determined by determining the correlation strength threshold, the edge weight is calculated based on the physical location of the device, the running parameter and the abnormal mode, and a weighted device group correlation graph is generated.

[0071] In a specific embodiment, when feature extraction and anomaly identification are performed on multi-dimensional time series data of Internet of Things instrument devices, first, multi-dimensional time series data of instruments are acquired, including multiple parameters such as temperature, pressure, flow, voltage, and current during instrument operation, to form a multi-dimensional data structure. Time series features are extracted from the acquired multi-dimensional data structure, and the extraction process includes the calculation of statistical features, frequency domain features, and time-frequency features.

[0072] When statistical features are extracted, the mean, standard deviation, kurtosis, skewness, and interquartile range of each dimension of data are calculated. Taking temperature data as an example, in a typical 24-hour monitoring period, the mean temperature of a normally operating instrument is 37.5 degrees Celsius, the standard deviation is 1.2, the kurtosis is 2.8, the skewness is 0.3, and the interquartile range is 2.1 degrees Celsius. When frequency domain features are extracted, Fourier transform is performed on the time series data to extract power spectral density, main frequency components and their amplitudes, frequency band energy distribution, and other features. For example, after transforming the voltage fluctuation data, the energy proportion of the low frequency band (0-10 Hz) is 65%, the energy proportion of the medium frequency band (10-50 Hz) is 30%, and the energy proportion of the high frequency band (above 50 Hz) is 5%. When time-frequency features are extracted, wavelet transform method is used to capture the change characteristics of data at different time scales, and wavelet energy and wavelet entropy are calculated.

[0073] After feature extraction, the importance of all extracted features is evaluated, and a tree model-based feature importance calculation method is used to assign a weight value to each feature. According to actual tests, the standard deviation feature weight of flow fluctuation is 0.85, the main frequency component weight of voltage is 0.76, and the temperature wavelet entropy weight is 0.73. These features constitute a high-importance feature subset. According to the set feature importance threshold (such as 0.5), features with importance higher than the threshold are selected to form an optimal feature subset, and a reduced feature matrix is generated.

[0074] Based on the reduced feature matrix, the dynamic time warping algorithm is used to calculate the similarity between time series sequences. The dynamic time warping distance between any two time series sequences is calculated, which measures the optimal matching alignment cost of the two sequences on the time axis. When calculating, a cumulative distance matrix is constructed, and the optimal path is found through dynamic programming to obtain the minimum matching cost as the distance value between the two sequences. The dynamic time warping distance between all samples is calculated, and a dynamic time warping distance matrix is constructed.

[0075] Based on the constructed distance matrix, time series clustering analysis is performed. The partition clustering method is adopted, the cluster centers are randomly selected initially, the similarity between the samples and the cluster centers is iteratively calculated, the samples are assigned to the most similar cluster, and the cluster centers are updated until the clustering result is stable.

[0076] The obtained time series clustering result is compared with the preset normal operation mode benchmark. The normal operation mode benchmark is a standard operation mode obtained by statistical analysis of historical data, which includes the standard position and distribution range of each cluster center. The clustering deviation value of the current clustering result and the benchmark mode is calculated, which includes the cluster center position offset and the cluster distribution form change. Exemplarily, taking the voltage parameter as an example, the cluster center value in the normal operation mode is 220 volts, and the cluster radius is 5 volts; while the current cluster result center value is 228 volts, and the cluster radius is 8 volts, the calculated deviation value is 13.6.

[0077] According to the comparison between the calculated clustering deviation value and the preset abnormal judgment threshold value, the judgment threshold value is set by historical abnormal cases and expert experience, and the typical value is 10.0. If the deviation value exceeds the threshold value, it is determined that the abnormal state exists, and the abnormal type is determined according to the deviation characteristics. In the above case, the voltage parameter deviation value 13.6 is greater than the threshold value 10.0, and it is determined that the voltage is abnormal, and the abnormal mode identification of the instrument operation is generated.

[0078] Based on the detected instrument operation abnormal mode, a device group association graph is constructed. The device operation state parameters are extracted, including the operation parameters and abnormal markers of each device. The correlation coefficient between devices is calculated, and the Pearson correlation coefficient method is adopted to measure the correlation of device operation parameters. For example, the temperature correlation coefficient of two adjacent temperature control devices is 0.86, indicating that their operation states are highly correlated.

[0079] According to the calculated correlation coefficient, the correlation strength threshold value is determined, and the typical value is 0.7. When the correlation coefficient between devices exceeds the threshold value, the node connection relationship is established. Based on the device physical location information, the operation parameter similarity and the abnormal mode consistency, the edge weight is calculated. The edge weight calculation comprehensively considers the physical distance factor (weight 0.3), the parameter correlation (weight 0.5) and the abnormal mode consistency (weight 0.2). For example, two instruments with the same voltage abnormal mode, the physical distance is 10 meters, the parameter correlation is 0.85, and the abnormal mode consistency is 1.0, the calculated edge weight is 0.805.

[0080] Through the above calculation, a weighted device group association graph is generated, nodes represent respective devices, edges represent the association relationship between devices, and the weight of the edge reflects the association strength. For edges with high weight (such as weight greater than 0.8), mark as strong association; for edges with moderate weight (such as weight between 0.7 and 0.8), mark as moderate association; for edges with low weight (such as weight less than 0.7), do not display in the graph. The finally generated association graph intuitively shows the abnormal propagation path and key nodes in the device group, providing a basis for abnormal tracing and fault diagnosis.

[0081] In an optional implementation, the fault propagation feature is calculated according to the fault state transition between nodes in the device group association graph, and the device operation state evaluation result is formed by including:

[0082] In a preset time window, the number of fault state transitions of adjacent nodes in the device group association graph is counted, the number of fault state transitions is divided by the length of the time window to obtain the fault transition probability between nodes, and a fault propagation probability matrix is formed.

[0083] By updating the propagation distance between nodes step by step, the shortest propagation path and propagation time delay between each node pair in the fault propagation probability matrix are calculated, nodes with a fault transition probability greater than a preset probability threshold are marked as key propagation nodes, and a fault propagation path graph is generated.

[0084] The number of connection edges of each node in the fault propagation path graph is counted, the node influence degree is calculated by multiplying the number of connection edges and the fault propagation probability of each edge, the propagation distance of the fault between nodes and the sampling interval are calculated to obtain the fault diffusion speed, the areas where the node influence degree and the fault diffusion speed exceed the corresponding threshold are marked as propagation danger areas, and the distance from the edge node of the propagation danger area to the key propagation node is calculated to determine the fault influence range.

[0085] The single-device state index is calculated by calculating the mean and variance of the device operation parameters, the group coordination index is calculated by calculating the parameter correlation coefficient between devices, the risk index is determined according to the weighted node influence degree, diffusion speed and fault influence range, and the device operation state evaluation result is obtained by weighted average of the single-device state index, group coordination index and risk index.

[0086] In one specific embodiment, device operation data within a preset time window is acquired, for example, the time window can be set to 30 days. During this period, the operation status of each node in the device group association graph is continuously monitored. Assuming that there are five device nodes A, B, C, D, and E in the association graph, the occurrence of failure of each node and its timestamp are recorded. When it is detected that node A fails, if node B also fails within a short period of time (such as within 1 hour), a failure state transition from A to B is recorded. Within the 30-day time window, a total of 15 failure state transitions from A to B are counted, and the failure transition probability from A to B is 15 / 30=0.5. Similarly, the failure transition probabilities between all adjacent node pairs are calculated to form a failure propagation probability matrix. For example, the matrix can be represented as: the probability from node A to node B is 0.5, the probability from node B to node C is 0.4, the probability from node C to node D is 0.6, the probability from node D to node E is 0.3, and the probability from node A to node C is 0.1, and so on.

[0087] Based on the above failure propagation probability matrix, the shortest propagation path and propagation time delay between each node pair are calculated by iteratively updating the propagation distance between nodes. Specifically, an improved shortest path algorithm is used to initialize the propagation distance between all node pairs to infinity, and then the shortest path is found by iteratively updating. For example, the shortest path from node A to node E can be A→B→C→D→E, and the corresponding propagation probability is 0.5×0.4×0.6×0.3=0.036, and the propagation time delay is the sum of the propagation times of each segment, which is assumed to be 4 hours. In actual application, a preset probability threshold of 0.4 can be set, and nodes with a failure transition probability greater than 0.4 are marked as key propagation nodes. In this embodiment, nodes B, C, and D are marked as key propagation nodes. By connecting these key nodes and their propagation paths, a failure propagation path graph is generated.

[0088] The number of connection edges of each node in the failure propagation path graph is counted. For example, node C is connected to an edge from B and an edge to D, a total of 2 connection edges. For node C, the influence degree is calculated as the number of connection edges multiplied by the average failure propagation probability of each edge, i.e. 2×(0.4+0.6) / 2=1.0. At the same time, the ratio of the propagation distance between nodes to the sampling interval is calculated to obtain the fault diffusion speed. Assuming that the physical distance from B to C is 100 meters, the failure propagation time is 30 minutes, and the system sampling interval is 5 minutes, then the fault diffusion speed is (100 / 30)×5=16.67 meters / minute. The area where the node influence degree exceeds 1.0 and the fault diffusion speed exceeds 15 meters / minute is marked as a propagation danger area. In this embodiment, the area where node C is located is marked as a propagation danger area. By calculating the distance from the edge nodes (such as nodes B and D) of the propagation danger area to the key propagation nodes (node C), the fault influence range is determined to be 100 meters.

[0089] To comprehensively evaluate the equipment running state, single equipment state indicators, group coordination indicators and risk indicators are calculated. The single equipment state indicators are obtained by calculating the mean and variance of the equipment running parameters. For example, for node C, the mean of its temperature sensor is 60°C and the variance is 2.5 when it is running normally, while the current measured mean is 65°C and the variance is 4.0, indicating that the equipment may be in an abnormal state. The group coordination indicators are obtained by calculating the correlation coefficient of the parameters between the devices. For example, the correlation coefficient of the temperature parameters of nodes B and C is 0.85, indicating that the two devices have a strong cooperative relationship. The risk indicators are determined according to the weights of node influence degree, diffusion speed and fault influence range. Assuming that the weights of the three are 0.4, 0.3 and 0.3 respectively, the risk indicator of node C is calculated as 0.4x1.0+0.3x(16.67 / 15)+0.3x(100 / 100)=0.4+0.333+0.3=1.033.

[0090] The single equipment state indicators, group coordination indicators and risk indicators are weighted and averaged to obtain the equipment running state evaluation result. Assuming that the weights of the three are 0.3, 0.3 and 0.4 respectively, for node C, assuming that its single equipment state indicator is 0.8 (considering temperature anomaly), group coordination indicator is 0.85, and risk indicator is 1.033, then the final running state evaluation result of device C is 0.3x0.8+0.3x0.85+0.4x1.033=0.24+0.255+0.4132=0.9082. The value is close to 1, indicating that the equipment C has a high running risk and needs to be paid attention to.

[0091] In this embodiment, by analyzing the fault propagation characteristics in the device group association graph, considering the single device state, group coordination and risk factors, a comprehensive and objective basis is provided for the equipment running state evaluation, which helps to prevent fault propagation and systemic risk, and improves the overall reliability and safety of the device group.

[0092] In an alternative embodiment, based on the equipment running state evaluation result, a performance degradation curve of the instrument is constructed by multi-head self-attention feature extraction and fitting function optimization, comprising:

[0093] The equipment running state evaluation result is input into a multi-head self-attention layer, which includes a query-key-value calculation unit, which captures different time sequence scale features through the query-key-value calculation unit, and determines the attention features;

[0094] The attention features are input into a time sequence position encoding layer, which uses sine and cosine functions to represent the relative position of the attention features, determines the time sequence dependence relationship, and generates position encoding features;

[0095] The position coding feature is subjected to a dimension mapping transformation to generate an intermediate feature, the intermediate feature is subjected to a transformation through an activation function to obtain a transformation feature, the transformation feature is subjected to feature superposition with the position coding feature, a short circuit connection path is established to maintain original feature information, and a fusion feature is obtained;

[0096] Based on the fusion feature, a state change rate sequence is constructed, a state change point is determined, a periodic change feature and a long-term evolution feature of the state change point are extracted, and a multi-scale performance feature is generated;

[0097] According to the time interval and amplitude difference of the state change point, grouping is performed, a decay feature group is constructed, the decay feature group is stored in a decay feature database, and an optimal fitting function is selected;

[0098] The mean square error of the multi-scale performance feature and the optimal fitting function is calculated, a gradient descent matrix is constructed, the fitting coefficient of the optimal fitting function is adjusted according to the gradient descent matrix, and iteration is repeated until the mean square error is less than a preset termination threshold, and an instrument performance decay curve is obtained.

[0099] In a specific embodiment, in the instrument full life cycle management process, after obtaining the device running state evaluation result, it is input into a multi-head self-attention layer for processing. The multi-head self-attention layer includes 8 parallel query-key-value calculation units, each unit has a 64-dimensional feature space, and generates a query vector, a key vector and a value vector through linear transformation. Specifically, the device running state evaluation result is represented as a sequence of 512-dimensional vectors, each vector contains the evaluation values of temperature, humidity, voltage, current and other parameters. For the temperature parameter, the normal operating state value is 36.5, the evaluation result is 35.8, and the difference value is 0.7; for the voltage parameter, the normal value is 220, the evaluation result is 218, and the difference value is 2. These difference values form part of the evaluation result vector.

[0100] In the multi-head self-attention processing process, each query-key-value calculation unit is responsible for capturing features of different time scales. The first calculation unit focuses on short-term time series features, with an attention window of 1 hour; the second calculation unit focuses on medium-term time series features, with an attention window of 6 hours; and the third calculation unit focuses on long-term time series features, with an attention window of 24 hours. By calculating the dot product of the query vector and the key vector and performing scaling processing, the attention score is obtained. For the temperature parameter, the attention score in the short-term window is 0.82, in the medium-term window is 0.65, and in the long-term window is 0.43, indicating that the temperature change is more strongly correlated in the short term. The attention score is applied to the value vector to obtain a weighted feature representation. The output features of the 8 heads are combined, and the final attention feature is obtained through linear transformation, with a dimension of 512.

[0101] The attention feature is input into a time position encoding layer, which uses sine and cosine functions to represent the relative position of the attention feature. Position encoding uses sine and cosine functions of different frequencies to generate a unique encoding value for each position in the sequence. In a 512-dimensional feature space, even dimensions use the sine function and odd dimensions use the cosine function. The period of the position encoding varies from 2π to 10000×2π, covering various time scales from short-term to long-term. For example, for the 10th position in the sequence, the position encoding value is 0.841 in dimension 0, 0.540 in dimension 1, 0.005 in dimension 510, and 0.999 in dimension 511. By adding the position encoding to the attention feature, the temporal dependence is determined, and the position encoding feature is generated.

[0102] The position encoding feature is subjected to a dimension mapping transformation, which maps the 512-dimensional feature to a 2048-dimensional space through a fully connected layer to generate an intermediate feature. A parameterized rectified linear unit is used as the activation function to perform a nonlinear transformation on the intermediate feature. This activation function maintains a linear relationship when input is positive and scales the value to a small value close to 0 but not 0 when input is negative, with a slope parameter of 0.1. After processing by the activation function, the transformed feature is obtained, still with a dimension of 2048. The transformed feature is mapped back to a 512-dimensional space through another fully connected layer and added element-wise to the position encoding feature to establish a short-cut connection path that preserves the original feature information, resulting in a fused feature. The short-cut connection allows gradients to flow directly back to early layers, alleviating the vanishing gradient problem and improving model training stability.

[0103] Based on the fused feature, a state change rate sequence is constructed by calculating the difference between adjacent time points. For the temperature parameter, the average value of the change rate sequence within 24 hours under normal conditions is 0.1 / hour, and the standard deviation is 0.05; while under abnormal conditions, the average value of the change rate rises to 0.3 / hour, and the standard deviation increases to 0.15. By setting a change rate threshold (such as 0.25 / hour), the state change point is determined. The periodic variation features of the state change point are extracted, including the change frequency, amplitude, and phase, such as the temperature parameter appearing a peak value every 24 hours under normal operating conditions, with a peak deviation of no more than 0.5; long-term evolution features are extracted, including the trend line slope and long-term modulation effect of periodic variation, such as the voltage parameter showing a -0.02 / day downward trend during device aging. Combining the periodic variation features and long-term evolution features, multi-scale performance features are generated.

[0104] The state change points are grouped according to the time interval and amplitude difference. The time interval is classified into short-term (1-24 hours), medium-term (1-7 days) and long-term (more than 7 days); the amplitude difference is classified into slight (change rate <0.1 / hour), medium (change rate 0.1-0.3 / hour) and significant (change rate >0.3 / hour). Nine kinds of feature grouping types are obtained by combination, and the decay feature grouping is constructed. For example, for the temperature parameter, the short-term-significant group contains 5 change points, the medium-term-medium group contains 12 change points, and the long-term-slight group contains 8 change points. The decay feature grouping is stored in the decay feature database, which contains historical decay feature grouping data of different instrument types and different use environments.

[0105] The historical data matching the current instrument type and use environment are retrieved from the decay feature database. By comparing the similarity of the current decay feature grouping and the historical data, the optimal fitting function is selected. The candidate fitting functions include linear function, exponential function, power function and logarithmic function. For motor instruments, the exponential function usually provides the best fit; for sensor instruments, the power function usually provides the best fit; for electronic components, the linear function provides good fit in the early stage, and the exponential function is better in the later stage. According to the historical data statistics, the current instrument selects the exponential function as the optimal fitting function, and the initial parameters are set to the starting value 1.0 and the decay rate 0.005 / day.

[0106] The mean square error of the multi-scale performance feature and the optimal fitting function is calculated as the fitting quality evaluation index. The mean square error after initial fitting is 0.285, which exceeds the preset termination threshold 0.05. The gradient descent matrix is constructed, which contains the partial derivative values of the mean square error to each fitting parameter. The fitting coefficients of the optimal fitting function are adjusted according to the gradient descent matrix, and the learning rate is set to 0.01. After the first iteration, the starting value is adjusted to 0.98, the decay rate is adjusted to 0.006 / day, and the mean square error is reduced to 0.215; after the fifth iteration, the starting value is 0.95, the decay rate is 0.0075 / day, and the mean square error is reduced to 0.108; after the tenth iteration, the starting value is 0.93, the decay rate is 0.0082 / day, and the mean square error is reduced to 0.063; after the fifteenth iteration, the starting value is 0.92, the decay rate is 0.0085 / day, and the mean square error is reduced to 0.048, which is less than the preset termination threshold 0.05, and the iteration is terminated.

[0107] The obtained instrument performance decay curve is expressed as an exponential function with an initial value of 0.92 and a decay rate of 0.0085 / day. The curve shows that the instrument performance will decrease to 50% of the initial performance on the 82nd day and to 25% of the initial performance on the 165th day. According to the prediction of the decay curve, the instrument needs to be maintained or replaced when the performance decreases to 30% (about the 140th day), and a warning message should be sent 30 days in advance (i.e., the 110th day). Through accurate modeling of the performance decay curve, scientific management of the entire life cycle of the instrument is achieved, the maintenance strategy is optimized, the service life of the instrument is prolonged, and the operation efficiency is improved.

[0108] In an alternative embodiment, a state change rate sequence is constructed based on the fusion features, state change points are determined, periodic variation features and long-term evolution features of the state change points are extracted, and multi-scale performance features are generated, including:

[0109] The difference values between adjacent time points are calculated for the fusion features according to the time sequence to construct a state change rate sequence, and the difference values are calculated for the state change rate sequence to construct a performance change acceleration sequence;

[0110] State change points are determined based on the local maxima and minima of the state change rate sequence and the zero points of the performance change acceleration sequence;

[0111] State fluctuation features are constructed by extracting the fluctuation patterns of the state change points in different time windows, periodic variation features are extracted according to the repetition frequency and amplitude variation of the state fluctuation feature sequence, and the baseline trend of the state change is extracted by performing a moving average process on the state change points as a long-term evolution feature;

[0112] The periodic variation features and long-term evolution features are combined by weighting to generate multi-scale performance features.

[0113] In a specific embodiment, the difference values between adjacent time points are calculated for the fusion features according to the time sequence to construct a state change rate sequence. Specifically, assume that the fusion feature sequence is F(t), where t represents the time point, t = 1, 2,..., T, and T is the total time length. The difference values between adjacent time points are calculated as ΔF(t) = F(t+1) - F(t), where t = 1, 2,..., T-1. The ΔF(t) obtained in this way is the state change rate sequence, which reflects the change rate of the system state at each time point. For example, for the device temperature fusion feature [35.2, 35.4, 35.8, 36.1, 36.5, 36.7, 36.8, 36.7, 36.5, 36.2]℃, the calculated state change rate sequence is [0.2, 0.4, 0.3, 0.4, 0.2, 0.1, -0.1, -0.2, -0.3]℃ / time unit.

[0114] Calculate the difference values ​​for the rate of change of state to construct a performance change acceleration sequence. Specifically, calculate the difference value of ΔF(t): Δ 2 F(t) = ΔF(t+1) - ΔF(t), where t = 1, 2, ..., T-2. The resulting ΔF(t)... 2 F(t) is the performance change acceleration sequence, reflecting the rate of change of the system state. Continuing the example above, the calculated performance change acceleration sequence is [0.2, -0.1, 0.1, -0.2, -0.1, -0.2, -0.1, -0.1, -0.1]℃ / time unit. 2 .

[0115] Based on the local maxima and local minima of the rate of change sequence and the zeros of the acceleration sequence of performance change, the state change points are determined. A local maximum represents the moment when the rate of increase in the state reaches its peak, a local minimum represents the moment when the rate of decrease in the state reaches its peak, and the zeros of the acceleration sequence represent the inflection points where the rate of change in the state changes from increasing to decreasing or from decreasing to increasing. Specifically, traversing the rate of change sequence ΔF(t), for any time t, if ΔF(t) > ΔF(t-1) and ΔF(t) > ΔF(t+1), then time t is a local maximum point; if ΔF(t) < ΔF(t-1) and ΔF(t) < ΔF(t+1), then time t is a local minimum point. For the acceleration sequence of performance change Δ²F(t), if ΔF(t) > ΔF(t-1) and ΔF(t+1), then ΔF(t) > ΔF(t+1), then ΔF(t) > ΔF(t+1), then ΔF(t) > ΔF(t+1). 2 F(t) and Δ 2 If F(t+1) has opposite signs, then there is a zero point between t and t+1. By combining these three types of characteristic points, the key moments of change in the system state can be identified. In the example above, the local maxima are located at time points 3 (corresponding to a rate of change of 0.4) and 5 (corresponding to a rate of change of 0.4), the local minima are located at time point 9 (corresponding to a rate of change of -0.3), and the zero points of the acceleration sequence are located between time points 2 and 4, and between time points 6 and 7.

[0116] The fluctuation patterns of state change points within different time windows are extracted to construct a state fluctuation characteristic sequence. Multiple time windows of different sizes are set, such as 24 hours, 7 days, and 30 days, and the distribution and characteristics of state change points within each window are analyzed. For each window w, the number of state change points n_w within the window is counted, and the average time interval μ_w and standard deviation σ_w between change points, as well as the average amplitude a_w and standard deviation s_w, are calculated. These statistics constitute the state fluctuation characteristic sequence W(w) = [n_w, μ_w, σ_w, a_w, s_w]. For example, within a 24-hour window, the number of temperature change points may be 12, with an average interval of 2 hours, a standard deviation of 0.5 hours, an average amplitude of 0.3℃, and an amplitude standard deviation of 0.1℃.

[0117] Periodic variation features are extracted according to the frequency of recurrence and amplitude variation of state fluctuation sequences. A sliding window method is used to calculate the similarity S(i, j) of state fluctuation sequences of adjacent windows for each window size w, where i and j are window indices. When the similarity is higher than a preset threshold θ (such as 0.85), it is considered that the two windows have similar fluctuation patterns. By detecting the time interval between similar windows, the periodic patterns of the system are identified. For each detected period p, its period length T_p, frequency f_p and stability index s_p (number of consecutive occurrences) are recorded. These features constitute the periodic variation features P=[{T_p, f_p, s_p}]. For example, a 24-hour temperature fluctuation period with a frequency of 1 time / day and a stability index of 30 (consecutive occurrence for 30 days) can be detected.

[0118] The state change points are subjected to a moving average process to extract the baseline trend of state change as long-term evolution features. A larger window size W (such as 30 days) is selected to perform a moving average on the amplitude values of state change points to obtain a smoothed trend sequence T(t). The slope variation of the trend sequence is calculated to identify key trend transition points. For each trend segment k, its start time t_start_k, end time t_end_k, duration d_k, average slope s_k and trend direction dir_k (up or down) are recorded. These features constitute the long-term evolution features L=[{t_start_k, t_end_k, d_k, s_k, dir_k}]. For example, a 45-day rising trend with an average slope of 0.02°C / day, starting on January 1st and ending on February 15th, can be identified.

[0119] The periodic variation features and long-term evolution features are combined by weighting to generate multi-scale performance features. According to the requirements of the application scenario, the periodic feature weight w_p and the long-term feature weight w_l are set to satisfy w_p + w_l = 1. For each time point t, the periodic feature value F_p(t) and the long-term evolution feature value F_l(t) are calculated, and the final multi-scale performance feature M(t) = w_p·F_p(t) + w_l·F_l(t) is obtained by weighted summation. For example, for the application scenario of predicting equipment failure, w_p = 0.4 and w_l = 0.6 can be set, because the long-term evolution features may be more important for failure prediction. The generated multi-scale performance features can reflect both the periodic variation law and the long-term evolution trend of the system, providing effective feature input for subsequent state assessment, fault diagnosis and predictive maintenance.

[0120] In an alternative embodiment, the scheduling optimization algorithm comprises:

[0121] A double-layer coding structure of maintenance tasks is constructed based on the predictive maintenance indicators, the double-layer coding structure including an outer path coding for determining the execution order of the maintenance tasks and an inner time coding for determining the execution time of the maintenance tasks;

[0122] An initial pheromone distribution between the maintenance tasks is set in the outer path coding, and a transition probability matrix between the maintenance tasks is constructed according to the maintenance timeliness indicators;

[0123] A time allocation space of the maintenance tasks is established in the inner time coding, and a time allocation probability matrix of the maintenance tasks is constructed according to the resource utilization indicators;

[0124] Path search of the execution order of the maintenance tasks is performed based on the transition probability matrix, and the execution time of the maintenance tasks is determined based on the time allocation probability matrix;

[0125] The time window constraint and the resource constraint in the predictive maintenance indicators are taken as temperature parameters, a local adjustment is performed on the maintenance task allocation scheme searched by the double-layer coding structure, and the initial pheromone distribution is updated based on the target function value after the adjustment;

[0126] Based on the double-layer coding structure, the scheduling optimization algorithm is repeatedly executed until a preset iteration number is reached, and a maintenance task allocation scheme is output.

[0127] In a specific embodiment, the outer path coding sets an initial pheromone distribution between the maintenance tasks to represent the execution priority relationship between different maintenance tasks. Taking laboratory analysis instruments connected by Internet of Things as an example, it is assumed that there are 5 instrument devices that need to be maintained, which are marked as task 1 to task 5. The initial pheromone distribution can be set as a uniform distribution, that is, the initial pheromone concentration between all tasks is 0.2. A transition probability matrix between the maintenance tasks is constructed according to the maintenance timeliness indicators obtained by instrument state monitoring, which reflects the fault risk degree of each instrument. For example, according to historical data analysis, the transition probability from task 1 to task 2 is 0.35, the transition probability from task 2 to task 3 is 0.28, and so on to form a complete 5x5 transition probability matrix.

[0128] The inner layer time coding establishes a time allocation space for the maintenance tasks, which is composed of available time points in the current scheduling period. Taking an 8-hour workday as an example, the time allocation space can be divided into 48 time slices of 15 minutes each. A time allocation probability matrix for the maintenance tasks is constructed according to the resource utilization index, which reflects the efficiency of performing maintenance tasks at different time points. For each maintenance task, its allocation probability in each time slice is calculated according to its expected maintenance duration and the availability of technicians. For example, the allocation probability of task 1 in the time slice from 10:00 to 10:15 is 0.08, while the allocation probability in the time slice from 12:00 to 12:15 is only 0.02, which reflects the differences in resource utilization efficiency at different times.

[0129] The execution process of the scheduling optimization algorithm is implemented using an improved ant colony algorithm. Path search for the execution order of maintenance tasks is performed based on the transition probability matrix in the outer layer coding. Taking the above 5 maintenance tasks as an example, suppose the optimal path calculated through 100 iterations is 4-1-3-5-2, indicating that maintenance work should be performed in the order of task 4, task 1, task 3, task 5, and task 2. At the same time, the execution time of each maintenance task is determined based on the time allocation probability matrix in the inner layer coding. For example, for task 4, the algorithm may determine its best execution time to be from 8:30 to 9:45 in the morning; for task 1, the best execution time is from 10:00 to 11:15 in the morning, and so on to complete the time allocation of all tasks.

[0130] In actual scheduling, various constraints need to be considered. The time window constraint and resource constraint in the predicted maintenance index are taken as temperature parameters to locally adjust the maintenance task allocation scheme obtained by searching the double-layer coding structure. The time window constraint means that a specific maintenance task must be completed within a specified time range, for example, task 3 must be completed before 2:00 pm on the same day; the resource constraint refers to limitations such as the number of available maintenance personnel and spare parts. The introduction of temperature parameters enables the algorithm to accept suboptimal solutions appropriately while meeting the constraint conditions, avoiding being trapped in local optimization. When the allocation scheme of a certain maintenance task violates the constraint condition, it is adjusted to the nearest feasible time point, and the objective function value is recalculated. The objective function considers maintenance timeliness, resource utilization, and constraint satisfaction, which can be represented as the weighted sum of the three, and the weights can be set according to actual needs, for example, the weight of maintenance timeliness is 0.5, the weight of resource utilization is 0.3, and the weight of constraint satisfaction is 0.2. The initial pheromone distribution is updated based on the adjusted objective function value, which increases the pheromone concentration on high-quality paths and reduces the pheromone concentration on low-quality paths.

[0131] To obtain a globally optimal maintenance task allocation scheme, a scheduling optimization algorithm is repeatedly executed based on the double-layer encoding structure until a preset number of iterations, such as 200 times, is reached. During the iteration process, the algorithm continuously optimizes the execution order and time arrangement of maintenance tasks, gradually improving the quality of the scheduling scheme. The final output maintenance task allocation scheme includes the execution order and specific time arrangement of each instrument maintenance task, such as "Task 4 is executed at 8:30-9:45, Task 1 is executed at 10:00-11:15, Task 3 is executed at 13:00-14:30, Task 5 is executed at 15:00-16:15, and Task 2 is executed at 16:30-17:45."

[0132] This embodiment calculates the predictive maintenance index through real-time data collected by sensors and intelligently schedules based on the double-layer encoding structure, which can significantly improve the efficiency and accuracy of instrument maintenance in the Internet of Things environment. Not only can it realize real-time monitoring and prediction of instrument status, but also can automatically generate an optimal maintenance scheduling scheme based on monitoring results, greatly improving the availability and service life of laboratory instruments and reducing maintenance costs, providing effective technical support for the whole life cycle management of laboratory instruments.

[0133] As shown in Figure 2 , a predictive maintenance task scheduling optimization algorithm flowchart is shown.

[0134] In an optional implementation, path search of the execution order of maintenance tasks is performed based on the transition probability matrix, and the execution time of the maintenance tasks is determined based on the time allocation probability matrix, further comprising:

[0135] In the outer layer path encoding, the execution order position of the next maintenance task is selected based on the transition probability matrix;

[0136] In the inner layer time encoding, the execution time of the maintenance task corresponding to the execution order position is determined based on the time allocation probability matrix;

[0137] The execution result of the maintenance task at the execution time is fed back to the outer layer path encoding, and the transition probability between the maintenance task and other maintenance tasks in the transition probability matrix is updated.

[0138] In one specific implementation, the outer layer path encoding selects the execution order position of the next maintenance task based on a transition probability matrix. The transition probability matrix is a two-dimensional array representing the transition relationship between different maintenance tasks, where the matrix elements represent the probability of transitioning from one maintenance task to another. Taking five main analytical instruments in a laboratory as an example, these instruments realize real-time acquisition and transmission of state data through Internet of Things modules. The dimension of the transition probability matrix is 5x5, and each row represents the probability distribution of selecting the next maintenance instrument from the current instrument. Assuming that the maintenance of the first analytical instrument has been completed, the transition probability matrix constructed based on the state monitoring data shows that the probability of transitioning to the second analytical instrument is 0.42, the probability of transitioning to the third detection instrument is 0.28, the probability of transitioning to the fourth measurement instrument is 0.18, and the probability of transitioning to the fifth monitoring device is 0.12. The calculation of transition probability takes into account factors such as instrument health status, usage frequency, maintenance urgency, and importance weight. The next maintenance object is selected from these probabilities through a roulette selection algorithm, and the selection process is implemented as follows: a random number between 0 and 1, 0.37, is generated, and since this value is less than the transition probability from the first analytical instrument to the second analytical instrument, 0.42, the second analytical instrument is selected as the next instrument to perform maintenance. This probability-based selection mechanism ensures that instruments with high maintenance urgency have a higher selection probability, while introducing randomness to avoid falling into a local optimal solution.

[0139] After determining the maintenance order, the inner layer time encoding determines the execution time of the maintenance task corresponding to the execution order position based on a time allocation probability matrix. The time allocation probability matrix represents the probability distribution of performing maintenance tasks at different time points. Taking an 8-hour work schedule as an example, the working hours are divided into 16 30-minute time segments. For the second analytical instrument selected, a time allocation probability matrix with a dimension of 1x16 is established. This matrix is calculated based on technician availability, equipment usage plan, and maintenance resource status, for example, the allocation probability from 9:00 to 9:30 is 0.12, the allocation probability from 9:30 to 10:00 is 0.15, the allocation probability from 10:00 to 10:30 is 0.18, and the allocation probability from 12:00 to 12:30 is only 0.03. Similarly, the roulette selection method is used to generate a random number, 0.45, to determine that the maintenance of the second analytical instrument starts at 10:00. Considering that the maintenance of this instrument is expected to take 90 minutes, combined with the division of time segments, its maintenance time period is from 10:00 to 11:30. The time allocation takes into account the estimated duration of the maintenance task, the workload balance of the technicians, and the usage plan of the instruments, ensuring that the maintenance task is completed efficiently without affecting normal experiments.

[0140] After completing the maintenance task of the second analyzer, the execution result is fed back to the outer path encoding, and the transition probabilities between the instrument and other instruments in the transition probability matrix are updated. The feedback information includes multiple dimensions: abnormal conditions found during maintenance, performance improvement after maintenance, deviation between actual maintenance time and planned time, maintenance resource consumption, etc. The Internet of Things sensors continuously collect instrument state data before and after maintenance for comparative analysis, such as the second analyzer's stability improvement of 18%, background noise reduction of 25%, and precision improvement of 12% after maintenance. At the same time, the Internet of Things platform monitors the abnormal fluctuations of the key parameters of the third detection instrument, indicating that its state has deteriorated. Based on this feedback information, the transition probabilities of the second analyzer to other instruments in the transition probability matrix are updated: the transition probability from the second analyzer to the third detection instrument increases from 0.28 to 0.45, the transition probability to the fourth measurement instrument remains 0.18, the transition probability to the fifth monitoring device decreases from 0.12 to 0.07, and the transition probability to the first analyzer decreases from 0.42 to 0.30.

[0141] The calculation process of transition probability update includes four main steps: state evaluation, weight calculation, transition tendency calculation, and probability normalization. State evaluation calculates the health index of each instrument based on Internet of Things data, integrating the deviation of temperature, pressure, current, flow, etc. For example, the health index of the third detection instrument decreases from 85 to 67, indicating that its state has deteriorated significantly. Weight calculation considers the importance and frequency of use of instruments, assigning basic weights to different instruments, such as the importance weight of the third detection instrument being 0.25, higher than other instruments. Transition tendency calculation combines health index and importance weight to obtain transition tendency values from the current instrument to each target instrument. Finally, through probability normalization processing, the sum of all transition probabilities is ensured to be 1, forming an effective probability distribution.

[0142] The updated transition probability matrix is used for the selection of the next round of maintenance tasks, continuing the task selection process in the outer path encoding. According to the new transition probability, the third detection instrument is selected as the next maintenance object. Then the time allocation step of the inner time encoding is executed again to determine the maintenance time of the third detection instrument as 13:30-15:30. This cycle continues until all instrument tasks that need maintenance are completed or the preset scheduling period ends.

[0143] The combination of the double-layer coding structure and the feedback mechanism enables the maintenance scheduling to dynamically adapt to the instrument state changes, and realizes intelligent and adaptive maintenance task scheduling. The outer path coding solves the problems of "what to maintain" and "maintenance sequence", the inner time coding solves the problem of "when to maintain", and the feedback mechanism enables the whole system to continuously learn and optimize, forming a closed loop control. The application of Internet of Things technology enables the collection and transmission of instrument state data to be automated and real-time, providing accurate data support for the double-layer coding structure, making the maintenance decision more accurate and timely.

[0144] The instrument full life cycle management system based on the Internet of Things includes:

[0145] The first unit is used for binding the RFID tag with the instrument to obtain unique identification information, collecting the running parameters of the instrument, and establishing a multi-dimensional data structure.

[0146] The second unit is used for time series clustering analysis on the multi-dimensional data structure, identifying abnormal patterns of instrument operation, and constructing a device group association graph. The fault propagation characteristics are calculated according to the fault state transition between the nodes in the device group association graph, and the device operation state evaluation result is formed.

[0147] The third unit is used for constructing the performance decay curve of the instrument based on the device operation state evaluation result, through multi-head self-attention feature extraction and fitting function optimization, combining the fault propagation characteristics for probability calculation, and generating a predictive maintenance index.

[0148] The fourth unit is used for generating a maintenance task allocation scheme based on the predictive maintenance index, using a scheduling optimization algorithm to target maintenance timeliness and resource utilization.

[0149] The fifth unit is used for executing maintenance work based on the maintenance task allocation result, storing the maintenance process and result data into a maintenance knowledge base, and updating the predictive maintenance index and the maintenance task allocation result through online analysis.

[0150] In a third aspect of the embodiments of the present application, an electronic device is provided, including:

[0151] A processor;

[0152] A memory for storing processor-executable instructions;

[0153] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0154] In a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0155] The present application can be a method, an apparatus, a system, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for performing various aspects of the present application.

[0156] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the present application; although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the technical solutions recorded in the above-mentioned embodiments can be modified, or some or all of the technical features thereof can be replaced equivalently; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for managing the entire lifecycle of instruments based on the Internet of Things, characterized in that, include: Bind RFID tags to instruments to obtain unique identification information, collect instrument operating parameters, and establish a multi-dimensional data structure; Temporal clustering analysis is performed on the multidimensional data structure to identify abnormal operating patterns of the instrument, and an equipment group association graph is constructed. Based on the fault state transition between nodes in the equipment group association graph, fault propagation characteristics are calculated to form equipment operating status assessment results. Based on the equipment operation status assessment results, the instrument's performance degradation curve is constructed through multi-head self-attention feature extraction and fitting function optimization, including: The device operation status evaluation results are input into the multi-head self-attention layer, which includes a query-key-value calculation unit. The query-key-value calculation unit captures features at different time scales and determines the attention features. The attention features are input into the temporal position coding layer, and the relative position of the attention features is represented by sine and cosine functions to determine the temporal dependency and generate position coding features. The location encoding features are transformed by dimensional mapping to generate intermediate features. The intermediate features are then passed through an activation function to obtain transformed features. The transformed features are then superimposed with the location encoding features to establish short-circuit connection paths that preserve the original feature information, resulting in fused features. Based on the fusion features, a state change rate sequence is constructed, state change points are determined, and periodic change features and long-term evolution features of the state change points are extracted to generate multi-scale performance features. Based on the time interval and amplitude difference of the state change points, the attenuation feature groups are grouped to construct attenuation feature groups, which are then stored in the attenuation feature database, and the optimal fitting function is selected. Calculate the mean square error between the multi-scale performance characteristics and the optimal fitting function, construct a gradient descent matrix, adjust the fitting coefficients of the optimal fitting function according to the gradient descent matrix, repeat the iteration until the mean square error is less than a preset termination threshold, and obtain the instrument performance degradation curve. Based on the aforementioned fault propagation characteristics, probability calculations are performed to generate predictive maintenance indicators; Based on the predicted maintenance indicators, a scheduling optimization algorithm is used to generate a maintenance task allocation scheme with the goals of maintenance timeliness and resource utilization. Maintenance work is performed based on the maintenance task allocation results, and the maintenance process and result data are stored in the maintenance knowledge base. The predicted maintenance indicators and the maintenance task allocation results are updated through online analysis.

2. The method according to claim 1, characterized in that, Performing time-series clustering analysis on the multidimensional data structure to identify abnormal instrument operation patterns and constructing an equipment group association graph includes: Temporal feature extraction is performed on multidimensional data structures, extracting statistical features, frequency domain features, and time-frequency features from the multidimensional data structures. The optimal feature subset is determined based on the feature importance weights, and a dimension-reduced feature matrix is ​​generated. Based on the dimensionality-reduced feature matrix, dynamic time warping is used to calculate the similarity of time series sequences, a dynamic time warping distance matrix is ​​constructed, and the time series clustering results are obtained by iteratively calculating the similarity between samples and cluster centers; The time-series clustering results are compared with a preset normal operation mode benchmark to calculate the clustering deviation value. Based on the comparison result of the clustering deviation value and the preset anomaly judgment threshold, an abnormal mode of instrument operation is generated. Based on the abnormal operation patterns of the instruments, a device group association graph is constructed. The device operation status parameters are extracted to calculate the correlation coefficient between devices, the association strength threshold is determined to establish node connection relationships, and the edge weights are calculated based on the device physical location, operation parameters and the abnormal patterns to generate a weighted device group association graph.

3. The method according to claim 1, characterized in that, Based on the fault state transition between nodes within the equipment group association diagram, the fault propagation characteristics are calculated to form the equipment operating status assessment results, including: Within a preset time window, the number of fault state transitions between adjacent nodes in the device group association diagram is counted. The number of fault state transitions is divided by the length of the time window to obtain the fault transition probability between nodes, forming a fault propagation probability matrix. By gradually updating the propagation distance between nodes, the shortest propagation path and propagation time delay between each pair of nodes in the fault propagation probability matrix are calculated. Nodes with a fault transfer probability greater than a preset probability threshold are marked as key propagation nodes, and a fault propagation path map is generated. The number of connecting edges of each node in the fault propagation path graph is counted. The number of connecting edges is multiplied by the fault propagation probability of each edge to calculate the node influence degree. The ratio of the fault propagation distance between nodes to the sampling interval is calculated to obtain the fault spread rate. Areas where the node influence degree and the fault spread rate exceed the corresponding threshold are marked as propagation danger zones. The distance from the edge nodes of the propagation danger zone to the key propagation nodes is calculated to determine the fault impact range. The mean and variance of equipment operating parameters are calculated to obtain the single equipment status index, the correlation coefficient between parameters of equipment is calculated to obtain the group coordination index, the risk index is determined by weighting the node influence degree, diffusion speed and fault impact range, and the weighted average of the single equipment status index, group coordination index and risk index is used to obtain the equipment operating status assessment result.

4. The method according to claim 1, characterized in that, Based on the fused features, a state change rate sequence is constructed, state change points are determined, and periodic change features and long-term evolution features of the state change points are extracted to generate multi-scale performance features, including: The difference values ​​between adjacent time points are calculated for the fused features according to the time series to construct a state change rate sequence. The difference values ​​are then calculated for the state change rate sequence to construct a performance change acceleration sequence. The state change point is determined based on the local maxima and local minima of the state change rate sequence and the zero point of the performance change acceleration sequence. The fluctuation patterns of the state change points within different time windows are extracted to construct a state fluctuation feature sequence; periodic change features are extracted based on the recurrence frequency and amplitude changes of the state fluctuation feature sequence; and a moving average is applied to the state change points to extract the baseline trend of state changes as a long-term evolution feature. By weighting and combining periodic change characteristics and long-term evolution characteristics, multi-scale performance characteristics are generated.

5. The method according to claim 1, characterized in that, The scheduling optimization algorithm includes: A two-layer coding structure for maintenance tasks is constructed based on predictive maintenance indicators. The two-layer coding structure includes an outer path coding for determining the execution order of maintenance tasks and an inner time coding for determining the execution time of maintenance tasks. In the outer path encoding, the initial pheromone distribution between maintenance tasks is set, and the transition probability matrix between maintenance tasks is constructed based on the maintenance timeliness index; A time allocation space for maintenance tasks is established in the inner time encoding, and a time allocation probability matrix for maintenance tasks is constructed based on the resource utilization index. The path search for the execution order of maintenance tasks is performed based on the transition probability matrix, and the execution time of maintenance tasks is determined based on the time allocation probability matrix. Using the time window constraint and resource constraint in the predicted maintenance index as temperature parameters, the maintenance task allocation scheme obtained by the search of the dual-layer coding structure is locally adjusted, and the initial pheromone distribution is updated based on the adjusted objective function value. Based on the two-layer coding structure, the scheduling optimization algorithm is repeatedly executed until the preset number of iterations is reached, and a maintenance task allocation scheme is output.

6. The method according to claim 5, characterized in that, The path search for the execution order of maintenance tasks is based on the transition probability matrix, and the execution time of maintenance tasks is determined based on the time allocation probability matrix. This also includes: In the outer path encoding, the execution order position of the next maintenance task is selected based on the transition probability matrix; In the inner time encoding, the execution time of the maintenance task corresponding to the execution order position is determined based on the time allocation probability matrix; The execution result of the maintenance task at the specified execution time is fed back to the outer path encoding to update the transition probability between the maintenance task and other maintenance tasks in the transition probability matrix.

7. An Internet of Things-based instrument lifecycle management system, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to bind RFID tags to instruments to obtain unique identification information, collect instrument operating parameters, and establish a multi-dimensional data structure. The second unit is used to perform time-series clustering analysis on the multidimensional data structure, identify abnormal operating patterns of the instrument, construct an equipment group association diagram, calculate fault propagation characteristics based on the fault state transfer between nodes in the equipment group association diagram, and form an equipment operating status assessment result. The third unit is used to construct the instrument's performance degradation curve based on the equipment operation status assessment results by extracting multi-head self-attention features and optimizing fitting functions, and to perform probability calculations by combining the fault propagation features to generate predictive maintenance indicators. The fourth unit is used to generate a maintenance task allocation scheme based on the predicted maintenance indicators and using a scheduling optimization algorithm with the goals of maintenance timeliness and resource utilization. The fifth unit is used to perform maintenance work based on the maintenance task allocation results, store the maintenance process and result data in the maintenance knowledge base, and update the predicted maintenance indicators and the maintenance task allocation results through online analysis.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images