An artificial intelligence-based state detection and device fault early warning method

Abstract

The application relates to the technical field of state detection and equipment fault early warning based on artificial intelligence, and discloses a state detection and equipment fault early warning method based on artificial intelligence. Three independent sensors are installed at key positions of equipment to collect temperature, vibration and pressure data respectively, so that the data comprehensiveness and value consistency are ensured. Sampling time discrete serial numbers and original data functions are set, so that the data acquisition time sequence is consistent, and the data completeness in each update cycle is ensured. Local neighbor point mean values are calculated to reduce instantaneous noise. By calculating deviation, the local fluctuation of data is quantified, and abnormal signs are identified. A smoothing function is constructed to nonlinearly compress the deviation, so that the influence of extreme values is reduced. Symmetric continuation is adopted to process boundary data, so that the influence of data loss on smoothing calculation at the start and end of the sampling cycle is avoided. After the minimum and maximum values are calculated, the data is normalized, the data scale is unified, and the accuracy of multi-sensor data fusion and comparison is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of state detection and equipment failure warning based on artificial intelligence, in particular to a state detection and equipment failure warning method based on artificial intelligence. BACKGROUND

[0002] In the current industrial automation and equipment maintenance field, state detection and failure warning technology is an important link to ensure production safety and efficient operation of equipment. The existing technology mainly adopts two methods: one is a real-time monitoring method based on traditional threshold setting, and the other is an intelligent warning method based on statistical analysis or pre-trained machine learning model. The traditional method generally installs temperature, vibration, pressure and other sensors at key parts of the equipment and sets fixed thresholds. When the collected data exceeds the threshold, the warning is triggered. This method is simple in structure and low in cost, but it is sensitive to data noise, inaccurate in threshold setting, difficult to adapt to complex operating conditions of equipment, and has a lag in response when local failure evolves into global failure. The other method uses a large amount of historical data to train the model to predict and judge the state of the equipment. Although there has been some development in recent years, it usually relies on a large amount of data collection and labeling, and the model has a black box and poor generalization ability, making it difficult to cope with the nonlinearity and multi-modal characteristics of the change in the state of the equipment, and the real-time performance and interpretability are low.

[0003] At present, some products using artificial intelligence for fault early warning have appeared in the market. The basic idea of these products is to extract features after data collection, and then use neural networks or other deep learning algorithms for prediction. Although these methods have improved the accuracy of fault detection to some extent, they have limitations such as complex models, long training cycles, sensitivity to abnormal data, and lack of sufficient baseline data in the initial stage of the device. In addition, existing technologies often use conventional smoothing filtering and normalization processing methods in data processing, without fully considering the physical location and real-time dynamic correlation between various sensor data, resulting in information loss or inaccurate data fusion. For example, in centrifugal pumps, numerical control machine tools and other equipment, due to the obvious nonlinear changes and noise interference of vibration, temperature and pressure data during operation, traditional moving average or Kalman filtering methods may mask critical abnormal information when smoothing data, resulting in early warning that is not sensitive enough. At the same time, existing systems generally only judge the data of a single sensor, ignoring the mutual influence between the key parts of the device, making it difficult to reflect local faults to the overall device state in a timely manner. Although some systems attempt to combine multiple sensor data, they usually only use simple logical operations or fixed weight weighted averaging, and fail to adaptively adjust the data fusion strategy according to the actual operation of the device, making it difficult to accurately capture abnormal signals when the device operating conditions change slightly. On the other hand, although the fault early warning system based on machine learning can achieve automatic data recognition and early warning to some extent, the core model often needs large-scale sample data for training, and data preprocessing, feature extraction and model parameter adjustment all rely on traditional algorithm ideas, making it difficult to break through the limitations of data noise and physical interference. In addition, traditional models lack sufficient consideration of the physical location relationship and data coupling between sensors when processing multi-site data of the device, resulting in delayed reaction or false judgments of the model when the device state is abnormal. Therefore, existing technologies still have major defects in real-time performance, adaptability and robustness, and cannot meet the demand for high-reliability early warning systems in industrial sites.

[0004] In summary, existing technologies have problems such as insufficient real-time response, inaccurate data processing, poor model generalization ability, and lack of consideration of the physical coupling between sensors in the field of device state detection and fault early warning. Therefore, the present application aims to develop a new state detection and fault early warning method based on artificial intelligence, which can fully utilize the physical and statistical relationships between data based on multi-sensor data collection, and construct an adaptive topology graph in real time, and use it as the basis for accurate anomaly detection and early warning, providing more reliable and intelligent technical support for industrial equipment maintenance. SUMMARY

[0005] The present application provides a state detection and device fault early warning method based on artificial intelligence, which solves the problems mentioned in the background art.

[0006] The application provides the following technical scheme: a state detection and equipment fault early warning method based on artificial intelligence, comprising:

[0007] Three independent sensors are installed at key positions of the equipment, and are respectively used for collecting temperature, vibration and pressure data; the sensor S1 is a temperature measuring device, the sensor S2 is a vibration measuring device, and the sensor S3 is a pressure measuring device;

[0008] Let the sampling time discrete sequence number be n, n belongs to {1, 2,..., L}, wherein L is the number of sampling points in a single update period;

[0009] Set the original data function X i (n) as the sensor S i The original data collected at the sampling moment n, wherein i belongs to {1, 2, 3}, and is the sensor number;

[0010] Calculate the local neighbor point mean value, specifically as follows:

[0011] Calculate S i (n) and all n satisfying 2≤n≤L-1:

[0012] Wherein, M i (n) is the sensor S i The mean value composed of the left and right neighbor point data at the moment n;

[0013] Set the deviation as: δ i (n)=X i (n)-M i (n);

[0014] Wherein, δ i (n) is the sensor S i The difference between the original data and the left and right neighbor point mean value at the moment n;

[0015] Construct a smoothing function, specifically as follows:

[0016] Wherein, ψ i (n) is the sensor S i The smoothing output data after nonlinear modulation at the moment n; 1+|δ i (n)| is used for nonlinear compression of the deviation;

[0017] For n=1 and n=L, symmetric continuation processing is adopted, specifically as follows:

[0018] X i (0)=X i (2), X i (L+1)=X i(L-1);

[0019] to the sensor S i In the sampling period n∈{1,2,...,L}, calculate:

[0020]

[0021] wherein, m i is the minimum value of ψ i (n); is the maximum value of ψ i (n);

[0022] Set the normalized data as:

[0023] wherein, Z i (n) is the normalized data, and the value range is [0,1].

[0024] Optionally, it further includes real-time data stream integration and node construction, specifically:

[0025] Set the node v i , which represents the data node of the corresponding sensor S i .

[0026] Set the node set as: V={v1,v2,v3};

[0027] At the global update moment t, t≥L, construct the attribute vector of the node v i , specifically:

[0028] A i (t)=[Z i (t-L+1),Z i (t-L+2),...,Z i (t)];

[0029] wherein, A i (t) is the attribute vector of the node v i at the global moment t; t is the current global sampling moment, and the update period is Δt=L;

[0030] In the device design, each sensor is pre-allocated a fixed three-dimensional coordinate, specifically represented as:

[0031] wherein, is the three-dimensional coordinate of the sensor S i ; p i is the x coordinate of the sensor S i ; q i is the y coordinate of the sensor S i ; ri z coordinate of sensor S i ; coordinate unit is unified as millimeter;

[0032] Every time after an update cycle, the vector A i (t) is updated to the corresponding node v i .

[0033] Optionally, it also includes inter-node association relationship mapping and edge weight calculation, specifically:

[0034] A function specially used for measuring the data similarity of two nodes in the latest update cycle is constructed, and for any i≠j, the similarity value is calculated at the global update time t, specifically:

[0035]

[0036] Wherein, Φ(i,j,t) is the similarity value of the data sequence of node v i and v j at time t;

[0037] The distance between sensors is calculated according to the physical position of the node, specifically:

[0038]

[0039] Wherein, d(i,j) is the Euclidean distance, representing the straight-line distance between sensor S i and S j ;

[0040] A physical interaction factor is constructed, specifically:

[0041]

[0042] Wherein, Θ(i,j) is used to reflect the coupling effect between sensors caused by physical distance;

[0043] The edge weight function is constructed by combining the signal similarity Φ(i,j,t) and the physical interaction factor Θ(i,j), specifically:

[0044] W(i,j,t)=Φ(i,j,t)×Θ(i,j);

[0045] Wherein, W(i,j,t) is the coupling strength between node v i and v j at time t.

[0046] Optionally, it also includes adaptive topology model construction, specifically:

[0047] At each update time t, a time-varying graph is constructed according to the node set V and the calculated edge weight W(i,j,t), specifically: G(t)=(V,E(t));

[0048] Wherein, G(t) is a time-varying topology graph; E(t) is an edge set;

[0049] E(t)={e ij (t)|i≠j and W(i,j,t)>0}, e ij (t) is an edge connecting nodes v i and v j ;

[0050] The connection strength of each node in the graph is calculated, specifically:

[0051] Wherein, D(i,t) is the node degree of node v i at time t.

[0052] Optionally, it also includes global topology index construction and periodic update, specifically:

[0053] The global topology index is calculated, which is set as the arithmetic mean of all node degrees:

[0054] Wherein, T(t) is the global topology index; N is fixed to 3, corresponding to three sensor nodes;

[0055] Every time an update period is passed, that is, when t increases by Δt=L sampling times, W(i,j,t), D(i,t) and T(t) are recalculated using the latest data, thereby forming a continuous time-varying topology graph sequence {G(t)}.

[0056] Optionally, it also includes abnormal feature extraction and state determination, specifically:

[0057] Select the time interval during normal operation of the device, denoted as t∈[t0,t1];

[0058] Calculate the baseline average node degree of node v i , specifically:

[0059]

[0060] Wherein, B(i) is the average node degree of node v i during the baseline period; t0 and t1 are the starting and ending sampling times during normal operation of the device, and the baseline construction interval; t1-t0+1 is the number of sampling periods during the baseline period;

[0061] For any time t>t1, the node deviation value is calculated, specifically:

[0062] Δ(i,t) = D(i,t) - B(i);

[0063] where Δ(i,t) is the difference between the current node degree of node v i the difference between the current node degree at time t and its baseline value;

[0064] During the baseline period, for node v i Calculate the standard deviation of node degree, specifically:

[0065]

[0066] where σ i is the standard deviation of node degree of node v i During the baseline period.

[0067] Optionally, it also includes global anomaly indicator construction and setting of abnormal judgment criteria, specifically:

[0068] Construct a global anomaly indicator, specifically:

[0069]

[0070] where Λ(t) is the global anomaly indicator at time t;

[0071] At the same time, calculate the global baseline mean and standard deviation, specifically:

[0072]

[0073] where is the average value of the global topology indicator during the baseline period; σ G is the standard deviation of the global topology indicator during the baseline period;

[0074] Set the abnormal judgment state criterion as:

[0075] When for any node v i there exists |Δ(i,t)|>σ i , it is judged that the node has abnormal data deviation at time t;

[0076] When the global anomaly indicator satisfies Λ(t)>σ G , it is judged that the entire device system has abnormal state risk at time t.

[0077] Optionally, it also includes fault warning information generation and decision support, specifically:

[0078] At each update time t, for node v i if it satisfies: |Δ(i,t)|>σ iIf Δ(i, t) > σ, a node abnormality report is generated, including: node number i, current node degree D(i, t), baseline node degree B(i), node deviation value Δ(i, t) and baseline standard deviation σ i ;

[0079] The global topology index T(t), the global abnormality index Λ(t) and the global standard deviation σ are arranged for the current time t G , and an overall state report is generated, in which the calculation results of T(t), Λ(t) and σ G are listed in numerical form;

[0080] A fault pre-warning flag is constructed, specifically:

[0081]

[0082] Wherein, FPF(t) is the fault pre-warning flag at time t; FPF(t) = 1 indicates that the equipment is abnormal; FPF(t) = 0 indicates that the equipment is in normal state;

[0083] All numerical data obtained in each update cycle are arranged into a decision support table, which includes: time t; each node data, for i = {1, 2, 3}, D(i, t), B(i), Δ(i, t) and σ i are listed respectively; global index, the numerical values of T(t), Λ(t) and σ G are listed; and the determined value of the fault pre-warning flag FPF(t).

[0084] The present application has the following beneficial effects:

[0085] 1、By installing three independent sensors in the key parts of the device, the problem of incomplete data acquisition in device state monitoring is solved, and multiple physical quantities such as temperature, vibration and pressure can be obtained in real time, thereby providing reliable raw data for subsequent data processing. At the same time, the data units of each sensor are uniformly specified in the design, ensuring the numerical consistency when comparing multi-sensor data, and avoiding misjudgment due to inconsistent units. By setting the sampling time discrete sequence number and the original data function, the problem of inconsistent timing and incomparable data in real-time data acquisition is solved. Each sensor collects raw data at the predetermined sampling time, laying a solid foundation for real-time tracking of device state changes and ensuring data integrity and continuity in each update cycle. In the step of calculating the local neighbor mean, the mean value of the data of the left and right neighbor points of each sensor at the sampling time is calculated, which solves the problem of possible transient noise and local abnormal fluctuations in the data, and makes the obtained mean value more reflect the true data state at that time. This process provides a stable reference value for subsequent calculation of deviation. By setting the deviation, the difference between the original data of the sensor at the time and its local mean value is calculated, which solves the problem of misjudgment caused by local abnormal fluctuations in the data and can timely capture the initial signs of device state abnormalities. The calculation of this deviation provides a basis for subsequent construction of a smoothing function, so that data smoothing processing neither loses abnormal information nor effectively reduces noise. The step of constructing a smoothing function solves the problem of possible excessive output fluctuations when data appears abnormally high deviation. Through this nonlinear modulation measure, the true change trend of the data is retained, and extreme abnormal values are limited, thereby realizing smooth output of data; especially when the data deviation is small, the smoothing function basically restores the original value, ensuring data accuracy. In the boundary data processing, the symmetric continuation processing solves the problem of insufficient data at the beginning and end of the sampling period, thereby ensuring that smoothing calculation can also be performed at the boundary time, making the entire data processing process coherent and seamless. Finally, by calculating the minimum and maximum values of the smoothed data in the sampling period and setting the normalization data step, the problem of multi-sensor data being difficult to compare directly under different numerical scales is solved. Normalization processing unifies different sensor data to the same numerical range, thereby making subsequent multi-data fusion and state comparison more accurate and reliable. The entire process ensures that the data preprocessing stage of the state detection and fault warning system is accurate and stable, laying a solid foundation for subsequent intelligent analysis and warning decision-making.

[0086] 2、Through the real-time data stream integration and node construction step, the problems of data organization confusion and processing time sequence asynchronization after multi-sensor data real-time collection are solved, and unified management and efficient fusion of data are realized. First, by setting independent data nodes for each sensor, the problem of scattered data sources of various sensors in the device and difficult unified expression is solved. Each node corresponds to the data of a single sensor, so that when constructing a dynamic graph structure, each node represents an independent physical collection point, thereby ensuring that the data has a clear source and identification in the processing process. Second, the attribute vector of the node is constructed at the global update moment, and the normalized data of the sensor in a certain number, for example, in an update period, is completely recorded, effectively solving the problems of discontinuous time series, information loss or lag of data. Through this step, the attribute vector of each node contains the current latest state information, ensuring that the system always makes state judgment and fault warning based on the latest data in the dynamic update process. At the same time, the update period and update moment interval are set to ensure that the data stream is updated in real time according to a fixed step, avoiding data accumulation and delay problems. Third, a fixed three-dimensional coordinate is pre-allocated to each sensor in the device design, solving the problem that the sensor data cannot be clearly corresponded in physical position. By associating each node with a fixed physical coordinate, not only does it provide a basis for subsequent edge weight calculation based on physical location, but also makes the data correlation between key parts in the device have a direct physical reference, so that the overall operation state of the device can be more accurately reflected in the state detection process. The fixed coordinate setting is unified in millimeters, avoiding data comparison errors caused by non-uniform units and improving the accuracy and reliability of data processing. Overall, through the real-time data stream integration and node construction step, the system solves the problems of scattered data, time sequence asynchronization, and unclear physical location relationship after multi-sensor data collection, and realizes unified, real-time, and efficient management of device state data. The implementation effect of this step is to ensure that the subsequent graph structure construction and state detection are based on the latest and complete normalized data, and at the same time, combined with the pre-set physical location information, the device state warning system has higher real-time, accuracy and reliability, providing a solid data foundation for subsequent complex data fusion and intelligent fault warning.

[0087] 3、Through the series of steps of inter-node association relationship mapping and edge weight calculation, the problems of information island and data inconsistency in the process of multi-sensor data fusion are solved, and the quantitative description of the mutual relationship between data is realized. Firstly, by constructing a function specially used to measure the data similarity of two nodes in the latest update cycle, the problem of the inability to directly compare and fuse sensor data is solved. Specifically, this step calculates the similarity value of the data sequence of any two nodes at the global update time. The function formula in the form of inverse is used, where the fixed modulation constant 2 in the denominator ensures the stability of the function value, and the square of the data difference is used to reflect the data consistency, ensuring that when the data of two nodes are close, the function value is large, indicating high similarity. At the same time, all data points are averaged to fully reflect the similarity of the entire data sequence. This step solves the problem of how to objectively and accurately quantify the data similarity of each node in the presence of data noise and local deviation, enabling the system to determine the relevance of the running state of each part of the device based on similarity, and achieving scientific processing of data fusion. Secondly, the Euclidean distance between sensors is calculated according to the physical location of the nodes, solving the problem of misjudgment caused by not considering the physical structure factor in data fusion. By calculating the straight-line distance between sensors, the data nodes are not only compared in value, but also combined with physical distance information to reflect the mutual relationship of each key part of the device in the physical space. This process ensures that in state detection, parts that are physically close are more likely to have a linked effect on their data, providing more accurate basis for subsequent fault warning. Thirdly, the physical interaction factor step is constructed by substituting the calculated distance into the formula to construct a quantitative physical factor. This factor ensures that when the distance between sensors is small, the factor value is close to 1, and when the distance is large, it decays rapidly, thereby quantifying the influence of physical location on data coupling. This measure solves the problem of relying only on data similarity to reflect the influence of device structure, enabling edge weight calculation to include not only the similarity of data itself, but also the actual physical distance between device components. Finally, by combining signal similarity and physical interaction factor to construct an edge weight function, the problem of a single indicator being unable to comprehensively reflect the coupling strength between nodes in data fusion is solved. The calculation result of this function is normalized in a specific interval, and the higher the value, the tighter the coupling between nodes, thereby realizing the dual quantization of data similarity and physical location influence. Ultimately, the system can accurately represent the coupling strength between nodes, providing solid data support for subsequent construction of an adaptive topology graph, and providing a reliable basis for device state detection and fault warning, significantly improving the accuracy and robustness of the warning system.

[0088] 4、Through the series of steps of self-adaptive topology model construction, the problem of lack of overall correlation and real-time fusion between multi-sensor data is solved, and the dynamic description of the data coupling relationship of each key part of the device is realized. First, by constructing a time-varying graph at each update time according to the node set and the calculated edge weight, the problem of scattered sensor data in the device and lack of overall collaborative analysis is solved. The step of constructing the time-varying topology graph regards each sensor data node as an independent node in the graph, and uses the previously calculated edge weight to construct the edge set, reflecting the influence of data similarity and physical distance, so as to intuitively display the coupling relationship between nodes in the form of graph structure. The time-varying graph can reflect the data information interaction between each key part of the device at different time points in real time, and provides a unified mathematical model and intuitive geometric expression for overall state monitoring. Secondly, the connection strength of each node in the graph is calculated, that is, the node degree is calculated by formula , which solves the problem that a single data index cannot comprehensively reflect the overall coupling state of the device. The node degree D(i,t) represents the total coupling strength between a certain node and all other nodes, and its value directly reflects the comprehensive influence of the state of the sensor in the device operation. In this way, not only can local data anomalies be captured, but also the overall device state can be judged to be abnormal or not, realizing effective fusion of local and global information. Finally, in the process of constructing the time-varying topology graph, only when the edge weight W(i,j,t)>0, an effective edge is formed, which solves the problem that data noise and invalid connections may introduce false coupling, thereby avoiding the interference of redundant information. The whole construction process ensures the dynamic update of real-time data in the time and space dimensions, so that the data coupling relationship of each part of the device always reflects the current actual running state. The implementation effect of this step is that it realizes seamless conversion from multi-sensor data to graph model, so that device state detection can be based on accurate topology structure for real-time early warning, providing intuitive, quantitative and real-time technical support for subsequent anomaly detection and fault warning.

[0089] 5、Through the series of steps of global topology index construction and periodic updating, the problem of lack of unified quantitative standard in overall state evaluation in multi-sensor data fusion is solved, and the long-term and dynamic monitoring of the coupling relationship of data of each part of the equipment is realized. First, through the step of calculating the global topology index, the arithmetic mean of the degrees of all nodes is used to solve the problem of lack of global perspective when considering the data of each sensor alone. This step integrates the node degrees of each sensor, which is fixed at 3 sensors in this scheme, to form a quantitative index reflecting the overall running state of the equipment, so that local data anomalies can be effectively captured and identified in overall monitoring, realizing the organic integration of local and global information. Second, after each update period, the latest data is used to recalculate the node degrees and the global index to form a continuous time-varying topology graph sequence. This step solves the problem of invalid state judgment in traditional monitoring methods due to data lag or untimely data update. Periodic updating ensures that the system can reflect the changes of the equipment state in real time based on the latest collected data, thereby greatly improving the real-time performance and accuracy of the early warning system. At the same time, this updating process enables the system to automatically adapt to the changes in data coupling caused by changes in load, environmental factors or equipment aging during equipment operation. The global topology index can not only be used for real-time state monitoring, but also provides a solid data foundation and quantitative basis for subsequent anomaly detection, trend analysis and fault warning. When the equipment running state is abnormal, the global index will fluctuate significantly, triggering further fault warning mechanism. Through periodic updating, the continuous change curve of the global index can provide the trend chart of the equipment running state for maintenance personnel, helping to predict possible faults, so as to realize early intervention and maintenance. In summary, through the two steps of calculating the global topology index and periodic updating, the system solves the problems of lack of overall state quantification after multi-sensor data integration and real-time data update lag, realizes the dynamic and real-time monitoring of the overall state of the equipment, and provides intuitive, accurate and continuous data support for fault warning and maintenance decision-making.

[0090] 6、Through the series of steps of abnormal feature extraction and state determination, the problem that the local state deviation is not easy to be found and quantified in time during the equipment operation is solved, and the accurate division of the equipment state abnormality and normal state is realized. First, the time interval during the normal operation of the equipment is selected as the baseline data, and by determining the interval, the problem of how to obtain the stable reference value of the degree of each node of the equipment under normal working conditions is solved. The baseline construction provides a quantitative standard for subsequent anomaly detection, which helps to eliminate the interference caused by environmental and working condition fluctuations. Second, the baseline average node degree of each node is calculated, which solves the problem of how to establish a data feature model for each sensor under the normal running state. This step enables the baseline value of each node to objectively reflect the coupling degree of the equipment under normal state, thereby laying a solid foundation for subsequent deviation detection. Then, for any time, the node deviation value is calculated, which solves the problem of how to quantitatively calculate the difference between the current state of a node and the normal baseline state. The difference value can be positive or negative, directly reflecting whether the sensor has abnormal fluctuations or trend deviation at the current sampling time. Through this calculation, the system can quickly capture the early signs of abnormality of local data in the equipment operation, providing early warning for fault prediction. In addition, the standard deviation of the node degree is calculated for each node during the baseline period, which solves the problem of how to determine the data fluctuation range and tolerance limit. As a quantitative index for judging abnormality, the standard deviation can provide an objective reference. When the deviation value of a node exceeds its standard deviation, it can be judged as an abnormal state. In this way, not only the abnormal situation of a single node can be accurately detected, but also statistical basis is provided for the overall system state monitoring. In summary, through the steps of baseline construction, calculation of baseline average node degree, real-time calculation of node deviation value and standard deviation, the system solves the problem of real-time and quantitative judgment of abnormality of multi-sensor data in equipment operation. This method not only improves the sensitivity and accuracy of fault detection, but also ensures the objectivity of the detection result through quantitative standard, and finally realizes the early warning and accurate determination of the equipment state abnormality.

[0091] 7、Through the series of steps of constructing and setting the abnormality judgment criterion by the global abnormality index, the problem of difficult to unify quantization of each node abnormality deviation in the multi-sensor monitoring system is solved, and the comprehensive judgment of the overall equipment system state abnormality is realized. First, by constructing the global abnormality index, the problem that the single node deviation detection cannot reflect the overall system state is solved. This step integrates the deviation values of each node to form a unified index reflecting the abnormality degree of the entire equipment at time t, thereby realizing the quantitative correlation between local abnormality and global state. The higher the index value, the more obvious the overall abnormality trend of the equipment. Secondly, by calculating the global baseline mean and the global standard deviation, the problem of difficult to determine the global data fluctuation range of the equipment under normal working condition is solved. By calculating the mean and standard deviation of the global topology index in the baseline period, a quantitative reference standard is provided for judging whether the real-time data deviates from the normal state. The baseline mean represents the overall coupling level of the equipment under normal operation, while the standard deviation serves as a measure of fluctuation amplitude, ensuring that the abnormality judgment has an objective basis. Finally, the abnormality judgment state criterion is set, which stipulates that when the deviation value of any node exceeds its corresponding standard deviation, it is judged that the node has an abnormality; at the same time, when the global abnormality index exceeds the global standard deviation, it is judged that the entire equipment system has an abnormal state risk at time t. This step solves the problem of how to establish clear judgment rules between local abnormality and global abnormality, and realizes the hierarchical and comprehensive judgment of the system abnormality warning. Overall, through the construction of global abnormality index, the calculation of global baseline mean and standard deviation, and the setting of clear abnormality judgment criterion, this series of steps effectively solves the problems of non-unification of global state after multi-sensor data fusion and difficulty in quantifying local abnormality, enabling the overall abnormality state of the equipment to be monitored and evaluated in a standardized and quantitative manner, thereby greatly improving the accuracy and response speed of the fault warning system.

[0092] 8、Through the series of steps of failure early warning information generation and decision support, the problem of how to convert a large amount of real-time calculation data into intuitive, quantitative and convenient maintenance personnel judgment decision information in the equipment state monitoring system is solved, and timely early warning and fault positioning of the abnormal state of the equipment are realized. First, at each update time, each data node is judged. When a node meets the preset abnormal condition, for example, the node deviation value exceeds the baseline standard deviation, the system immediately generates a node abnormal report. This step solves the problem of how to quantitatively judge whether a single sensor state deviates from the normal working range by calculating the current node degree, baseline node degree, node deviation value and baseline standard deviation of each node. The generated node abnormal report records the abnormal information of each node in the form of detailed numerical values, providing accurate positioning basis for subsequent maintenance. Secondly, the system sorts the global data at each update time, and the global topology index, global abnormal index and global standard deviation are summarized to generate an overall state report. This step solves the problem that it is difficult to reflect the overall state of the equipment in the traditional monitoring method by relying only on single sensor data. By fusing all sensor data of the equipment to form global indexes, the overall data coupling state is listed in the report in a numerical way, so that the overall operation state of the equipment is quantitatively described, and an objective and unified basis for overall evaluation of the state of the equipment is provided. Thirdly, the system constructs a failure early warning flag, and the calculation result thereof is clearly distinguished according to the preset criterion. When the early warning flag indicates an abnormal state of the equipment, for example, the value is 1, the system timely feeds back this information to the maintenance personnel. This step solves the problem of ambiguous early warning information and untimely response in the traditional monitoring, and through binary output, the early warning result is intuitive and clear, which is convenient for the on-site operator to quickly judge whether the equipment needs to be immediately repaired. Finally, all numerical data obtained in each update period are sorted into a decision support table, and the table content covers the time stamp, detailed data of each node and global indexes and failure early warning flags, such as node degree, baseline node degree, deviation value and standard deviation. The decision support table is directly referred by the equipment maintenance personnel, and the problem of difficult comprehensive judgment of the health state of the equipment by multi-dimensional data is solved. Through the series of steps, the system realizes the full-process closed loop from data acquisition, abnormal judgment to failure early warning information output, and the final effect is that the maintenance personnel can timely judge the state of the equipment based on detailed and quantitative decision information, and quickly take repair measures, thereby greatly improving the accuracy and response speed of the equipment failure early warning. BRIEF DESCRIPTION OF DRAWINGS

[0093] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0094] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0095] Embodiments, with reference to Figure 1 The application discloses an artificial intelligence-based state detection and equipment fault early warning method, which comprises the following steps:

[0096] Three independent sensors are installed at key positions of the equipment for collecting temperature, vibration and pressure data. The sensor S1 is a temperature measuring device, the sensor S2 is a vibration measuring device, and the sensor S3 is a pressure measuring device.

[0097] The key positions of the equipment refer to regions that bear greater load, temperature change, vibration or severe wear during the operation of the equipment, and abnormality often occurs at these positions when a fault occurs. Taking two typical industrial equipment as examples, the specific key positions are described as follows: 1. Centrifugal pump equipment: impeller position, the impeller is an important component for pushing fluid and is prone to abnormality due to liquid impact cavitation and wear; bearing seat, the position for supporting the rotating shaft, often bears vibration and temperature change, and vibration abnormality may be caused by bearing wear or insufficient lubrication; sealing cavity, the installation region of the sealing element, leakage may be caused by the failure of the sealing element, thereby causing system failure; 2. Numerical control machine tool: spindle position, the spindle bears high-speed rotation and machining load, and temperature rise, vibration or bearing wear may be embodied here; guide rail system, the guide rail bears cutting force and reciprocating motion, and is prone to wear or deformation; hydraulic system interface, the hydraulic driving part, such as the hydraulic cylinder and the oil path interface, and if leakage or pressure abnormality occurs, the machining precision will be directly affected. By installing independent sensors at these specific key positions, the present application can collect temperature, vibration and pressure data in real time, which is helpful to capture abnormal changes of the equipment in the regions most prone to failure, thereby realizing accurate state detection and fault early warning.

[0098] Let the sampling time discrete sequence number be n, n∈{1,2,...,L}, wherein L is the number of sampling points in a single update period;

[0099] Set the original data function X i (n) as the sensor S i The original data collected at the sampling moment n, wherein i∈{1,2,3}, is the sensor number;

[0100] The numerical units of each X i (n) are determined according to the sensor category, and all units are uniformly specified in the present application at the design time;

[0101] The local neighbor mean is calculated, specifically:

[0102] For S i and all n satisfying 2≤n≤L-1, calculate:

[0103] Where M i (n) is the mean value of the left and right neighbors of sensor S i at time n;

[0104] The deviation is set as: i (n) = X i (n) - M i (n);

[0105] Where δ i (n) is the difference between the original data of sensor S i at time n and the mean value of the left and right neighbors; it reflects local abnormal fluctuations;

[0106] The smoothing function is constructed, specifically:

[0107] Where ψ i (n) is the smoothed output data of sensor S i at time n after nonlinear modulation; 1+|δ i (n)| nonlinearly compresses the deviation, so that larger deviations are limited within a finite range, thereby avoiding excessive output fluctuations caused by abnormally high deviations; when |δ i (n)| is small, ψ i (n) ≈ X i (n);

[0108] For n=1 and n=L, symmetric extension processing is used, specifically:

[0109] X i (0) = X i (2), X i (L+1) = X i (L-1); this processing ensures that smoothing calculation can be performed at the boundary time;

[0110] For sensor S i , calculate in the sampling period n∈{1,2,...,L}:

[0111]

[0112] Where m i is the minimum value of ψ i (n); is the maximum value of ψ i (n);

[0113] The normalized data is set as:

[0114] wherein Z i (n) is the normalized data, whose value range is [0, 1]; the normalization ensures that different sensor data are compared in the same numerical scale.

[0115] By installing three independent sensors at the key parts of the device, the problem of incomplete data collection in device state monitoring is solved, and various physical quantities such as temperature, vibration and pressure can be obtained in real time, thereby providing reliable raw data for subsequent data processing. At the same time, the data units of each sensor are uniformly specified in the design, ensuring the numerical consistency when comparing multi-sensor data, and avoiding misjudgment caused by inconsistent units. By setting the sampling time discrete sequence number and the original data function, the problem of inconsistent timing and incomparable data in real-time data collection is solved. Each sensor collects original data at the predetermined sampling time, laying a solid foundation for real-time tracking of device state changes and ensuring data integrity and continuity in each update period. In the step of calculating the local neighbor mean, the mean value of the data of the left and right neighbors of each sensor at the sampling time is calculated, which solves the problem of possible transient noise and local abnormal fluctuations in the data, and makes the obtained mean value more reflect the true data state at that time; This process provides a stable reference value for subsequent calculation of deviation. By setting the deviation, the difference between the original data of the sensor at the time and its local mean value is calculated, which solves the problem of misjudgment caused by local abnormal fluctuations in the data and can timely capture the initial signs of device state abnormalities. The calculation of this deviation provides a basis for the subsequent construction of a smoothing function, so that data smoothing processing neither loses abnormal information nor effectively reduces noise. In the step of constructing a smoothing function, the deviation is nonlinearly compressed, which solves the problem that when the data appears an abnormal high deviation, it may cause the output to fluctuate too much. Through this nonlinear modulation measure, the true change trend of the data is retained, and the extreme abnormal value is limited, thereby realizing smooth output of the data; Especially when the data deviation is small, the smoothing function basically restores the original value, ensuring the data accuracy. In the boundary data processing, the symmetric continuation processing is used to solve the problem of insufficient data at the beginning and end of the sampling period, thereby ensuring that the smoothing calculation can also be performed at the boundary time, making the entire data processing process coherent and seamless. Finally, by calculating the minimum and maximum values of the smoothed data in the sampling period and setting the normalization data step, the problem of multi-sensor data being difficult to compare directly under different numerical scales is solved. Normalization processing unifies different sensor data to the same numerical interval, thereby making the subsequent multi-data fusion and state comparison more accurate and reliable. The whole process ensures that the data preprocessing stage of the state detection and fault warning system is accurate and stable, laying a solid foundation for subsequent intelligent analysis and warning decision-making.

[0116] It also includes real-time data stream integration and node construction, specifically:

[0117] Set node: v i for i∈{1,2,3}, which represents the corresponding sensor S idata nodes of the sensor S

[0118] The node set is set as: V={v1, v2, v3}; each v i represents a single sensor data node, and subsequent figure construction is taken as a basic unit of v i ;

[0119] At the global update moment t, t≥L, the attribute vector of the node v i is constructed, specifically as follows:

[0120] A i (t)=[Z i (t-L+1), Z i (t-L+2),..., Z i (t)];

[0121] Wherein, A i (t) is the attribute vector of the node v i at the global moment t, which completely records the normalized data Z i (n) of the sensor S i in the last L; t is the current global sampling moment, the update period is Δt=L, and the update moment interval is equal to a sampling period;

[0122] In device design, each sensor is pre-allocated a fixed three-dimensional coordinate, which is specifically expressed as:

[0123] Wherein, is the three-dimensional coordinate of the sensor S i ; p i is the x coordinate of the sensor S i ; q i is the y coordinate of the sensor S i ; r i is the z coordinate of the sensor S i ; and the coordinate unit is unified as millimeter;

[0124] Every time an update period is passed, the vector A i (t) formed by the latest sampling data is updated into the corresponding node v i , which provides a real-time data basis for subsequent construction of the graph structure.

[0125] By integrating real-time data streams and building nodes, the problem of disorganized data and asynchronous processing time sequence after real-time collection of multi-sensor data is solved, and unified management and efficient fusion of data are achieved. First, by setting independent data nodes for each sensor, the problem of scattered data sources of sensors in the device and difficulty in unified expression is solved. Each node corresponds to the data of a single sensor, so that when constructing a dynamic graph structure, each node represents an independent physical collection point, ensuring that the data has a clear source and identification during processing. Second, the attribute vector of the node is constructed at the global update time, which records the normalized data of the sensor in a certain number, for example, in an update period, effectively solving the problem of discontinuous time series, information loss or lag. Through this step, the attribute vector of each node contains the latest state information, ensuring that the system always makes state judgments and fault warnings based on the latest data during dynamic updating. At the same time, the update period and update time interval are set to ensure that the data stream is updated in real time at a fixed step, avoiding data accumulation and delay. Third, a fixed three-dimensional coordinate is assigned to each sensor in the device design, solving the problem that sensor data cannot be clearly corresponded in physical location. By associating each node with a fixed physical coordinate, it not only provides a basis for subsequent edge weight calculation based on physical location, but also makes the data correlation between key parts in the device have a direct physical reference, so that the overall running state of the device can be more accurately reflected during state detection. The fixed coordinate setting is unified in millimeters, avoiding data comparison errors caused by non-uniform units and improving the accuracy and reliability of data processing. Overall, through this step of real-time data stream integration and node construction, the system solves the problems of scattered data, asynchronous time sequence, and unclear physical location relationship after multi-sensor data collection, achieving unified, real-time, and efficient management of device state data. The implementation effect of this step is to ensure that the subsequent graph structure construction and state detection are based on the latest and complete normalized data, and at the same time, combined with the pre-set physical location information, the device state warning system has higher real-time, accuracy and reliability, providing a solid data foundation for subsequent complex data fusion and intelligent fault warning.

[0126] It also includes mapping of inter-node association relationships and edge weight calculation, specifically:

[0127] A function is constructed to measure the data similarity of two nodes in the latest update period. For any i≠j, the similarity value is calculated at the global update time t, specifically:

[0128]

[0129] where Φ(i,j,t) is the similarity value of nodes v i and v jThe similarity value of the data sequence at time t, Φ(i,j,t), the larger the value, the stronger the similarity of S. i With S j The higher the similarity of the data sequences; the 2 in the denominator is a fixed modulation constant used to stabilize the function value; [Z i (n)-Z j (n)] 2 The square of the data difference is used to reflect data consistency; the reciprocal form ensures that the function value is larger when the data of two nodes are close; averaging over all n ensures that the overall similarity of the data sequence is reflected; the numerical range of this function is automatically adjusted according to the data difference, and the larger the value, the more similar the two data sequences are.

[0130] The distance between sensors is calculated based on the physical location of the nodes, specifically as follows:

[0131]

[0132] Where d(i,j) is the Euclidean distance, representing the distance between sensor S and d(i,j). i With S j The straight-line distance between them;

[0133] Construct the physical interaction factors, specifically:

[0134]

[0135] Θ(i,j) is used to reflect the coupling effect between sensors due to physical distance. The smaller the distance, the closer the factor value is to 1. It decays rapidly when the distance is large. The 1 in the denominator ensures that the factor value is equal to 1 when d(i,j)=0.

[0136] Combining signal similarity Φ(i,j,t) and physical interaction factor Θ(i,j), the edge weight function is constructed as follows:

[0137] W(i,j,t)=Φ(i,j,t)×Θ(i,j);

[0138] Where W(i,j,t) represents node v i and v j The coupling strength between timestamps; this function ensures that it reflects both data similarity and the influence of physical location; the calculation result is normalized to the interval (0,1], and the higher the value, the tighter the coupling.

[0139] By the series of steps of inter-node association relationship mapping and edge weight calculation, the problems of information island and data inconsistency in the process of multi-sensor data fusion are solved, and the quantitative description of the mutual relationship between data is realized. Firstly, by constructing a function specially used to measure the data similarity of two nodes in the latest update cycle, the problem of the inability to directly compare and fuse sensor data is solved. Specifically, this step calculates the similarity value of the data sequence of any two nodes at the global update time. The function formula in the form of reciprocal is used, where the fixed modulation constant 2 in the denominator ensures the stability of the function value, and the square of the data difference is used to reflect the data consistency, ensuring that when the data of two nodes are close, the function value is large, indicating high similarity. At the same time, all data points are averaged to fully reflect the similarity of the entire data sequence. This step solves the problem of how to objectively and accurately quantify the data similarity of each node in the presence of data noise and local bias, enabling the system to determine the relevance of the running state of each part of the device based on similarity, and achieving scientific processing of data fusion. Secondly, the Euclidean distance between sensors is calculated according to the physical location of the nodes, solving the problem of misjudgment caused by not considering the physical structure factor in data fusion. By calculating the straight-line distance between sensors, the data nodes are not only compared in value, but also combined with physical distance information to reflect the mutual relationship of each key part of the device in the physical space. This process ensures that in state detection, parts that are physically close are more likely to have a linked effect on their data, providing more accurate basis for subsequent fault warning. Thirdly, the physical interaction factor step is constructed, and a quantitative physical factor is constructed by substituting the calculated distance into the formula. This factor ensures that when the distance between sensors is small, the factor value is close to 1, and when the distance is large, it decays rapidly, thereby quantifying the influence of physical location on data coupling. This measure solves the problem of relying only on data similarity to reflect the influence of device structure, enabling edge weight calculation to include not only the similarity of data itself, but also the actual physical distance between device components. Finally, by combining signal similarity and physical interaction factor to construct edge weight function, the problem of single indicator in data fusion that cannot fully reflect the coupling strength between nodes is solved. The calculation result of this function is normalized in a specific interval, and the higher the value, the tighter the coupling between nodes, thereby realizing the dual quantization of data similarity and physical location influence. Finally, the system can accurately represent the coupling strength between nodes in numerical form, providing a solid data support for subsequent construction of adaptive topology graph, and providing a reliable basis for device state detection and fault warning, significantly improving the accuracy and robustness of the warning system.

[0140] It also includes adaptive topology model construction, specifically:

[0141] At each update time t, a time-varying graph is constructed according to the node set V and the calculated edge weight W(i, j, t), specifically: G(t) = (V, E(t));

[0142] Wherein, G(t) is a time-varying topology graph, describing the coupling relationship between each sensor data of the device at time t; E(t) is an edge set;

[0143] E(t) = {e ij (t)|i≠j and W(i, j, t) > 0}, e ij (t) is an edge connecting nodes v i and v j ; only when W(i, j, t) > 0, an effective edge is formed, avoiding the introduction of invalid connection;

[0144] The connection strength of each node in the graph is calculated, specifically:

[0145] Wherein, D(i, t) is the node degree of node v i at time t, indicating the total coupling between the node and other nodes.

[0146] Through the series of steps of constructing the adaptive topology model, the problems of lack of overall correlation and real-time fusion between multi-sensor data are solved, and the dynamic description of the coupling relationship of the data of each key part of the device is realized. First, by constructing a time-varying graph according to the node set and the calculated edge weight at each update time, the problem of scattered sensor data in the device and lack of overall collaborative analysis is solved. The step of constructing the time-varying topology graph regards each sensor data node as an independent node in the graph, and uses the previously calculated edge weight to construct the edge set, reflecting the influence of data similarity and physical distance, so as to intuitively display the coupling relationship between each node in the form of graph structure. The time-varying graph can reflect the data information interaction between each key part of the device at different time points in real time, providing a unified mathematical model and intuitive geometric expression for overall state monitoring. Secondly, the connection strength of each node in the graph is calculated, that is, through the formula The node degree solves the problem that a single data index cannot comprehensively reflect the overall coupling state of the equipment. The node degree D(i, t) represents the total coupling strength between a certain node and all other nodes, and its value directly reflects the comprehensive influence of the sensor on the state of the equipment during operation. In this way, not only can local data anomalies be captured, but the overall equipment state can also be judged to be abnormal or not, effectively integrating local and global information. Finally, in the process of constructing the time-varying topology graph, only when the edge weight W(i, j, t) > 0 is an effective edge, solving the problem that data noise and invalid connections can introduce false coupling, thereby avoiding the interference of redundant information. The entire construction process ensures the dynamic updating of real-time data in the time and space dimensions, so that the data coupling relationship of each part of the equipment always reflects the current actual operating state. The implementation effect of this step is to achieve seamless conversion from multi-sensor data to a graph model, enabling equipment state detection to provide real-time early warning based on accurate topology, providing intuitive, quantitative, and real-time technical support for subsequent anomaly detection and fault warning.

[0147] It also includes global topology index construction and periodic updating, specifically:

[0148] The global topology index is calculated as the arithmetic mean of all node degrees:

[0149] Where T(t) is the global topology index, reflecting the overall level of data coupling in the entire equipment system at time t; N is fixed at 3, corresponding to the three sensor nodes;

[0150] Every time an update cycle is completed, that is, when t increases by At = L sampling times, the latest data is used to recalculate W(i, j, t), D(i, t), and T(t), thereby forming a continuous sequence of time-varying topology graphs {G(t)}.

[0151] The problem of lack of unified quantitative standard for overall state evaluation in multi-sensor data fusion is solved by the series of steps of constructing and periodically updating the global topology index, and the long-term and dynamic monitoring of the coupling relationship of data of each part of the equipment is realized. Firstly, by calculating the global topology index step, the arithmetic mean of the degrees of all nodes is used to solve the problem of lack of global perspective when considering the data of each sensor alone. This step integrates the node degrees of each sensor, which is fixed at 3 sensors in this scheme, to form a quantitative index reflecting the overall running state of the equipment, so that local data anomalies can be effectively captured and identified in overall monitoring, realizing the organic integration of local and global information. Secondly, after each update period, the latest data is used to recalculate the node degrees and the global index to form a continuous time-varying topology graph sequence. This step solves the problem of state judgment failure caused by data lag or untimely data update in traditional monitoring methods. Periodic updating ensures that the system can reflect the changes of the equipment state in real time based on the latest collected data, thereby greatly improving the real-time performance and accuracy of the early warning system. At the same time, this updating process enables the system to automatically adapt to the changes in data coupling caused by load changes, environmental factors or equipment aging during equipment operation. The global topology index can not only be used for real-time state monitoring, but also provides a solid data foundation and quantitative basis for subsequent anomaly detection, trend analysis and fault warning. When the equipment running state is abnormal, the global index will fluctuate significantly, triggering further fault warning mechanism. Through periodic updating, the continuous change curve of the global index can provide the trend graph of the equipment running state for maintenance personnel, helping to predict possible faults and thus realize early intervention and maintenance. In summary, through the two steps of calculating the global topology index and periodic updating, the system solves the problems of lack of overall state quantification after multi-sensor data integration and real-time data update lag, realizes dynamic and real-time monitoring of the overall state of the equipment, and provides intuitive, accurate and continuous data support for fault warning and maintenance decision-making.

[0152] Further comprising abnormal feature extraction and state determination, specifically:

[0153] Selecting the time interval during normal operation of the equipment, denoted as t∈[t0,t1];

[0154] For node v i Calculating the baseline average node degree, specifically:

[0155]

[0156] Where B(i) is the average node degree of node v i During the baseline period; t0, t1 are the start and end sampling time during normal operation of the equipment, and the baseline construction interval; t1-t0+1 is the number of sampling periods during the baseline period;

[0157] For any time t > t1, the node deviation value is calculated, specifically:

[0158] Δ(i, t) = D(i, t) - B(i);

[0159] Where Δ(i, t) is the node v i The difference between the current node degree at time t and its baseline value can be positive or negative, directly reflecting the degree of deviation;

[0160] During the baseline period, for node v i The node degree standard deviation is calculated, specifically:

[0161]

[0162] Where σ i is the node v i The standard deviation of the node degree within the baseline period is a quantitative index for judging abnormalities.

[0163] By the steps of abnormal feature extraction and state determination, the problem that the local state deviation is not easy to be found and quantified in time during the equipment operation is solved, and the accurate division of the equipment state abnormality and normal state is realized. First, the time interval during the normal operation of the equipment is selected as the baseline data, and by determining the interval, the problem of how to obtain the stable reference value of the degree of each node of the equipment under normal working conditions is solved. The baseline construction provides a quantitative standard for subsequent anomaly detection, which helps to eliminate the interference caused by environmental and working condition fluctuations. Second, the baseline average node degree of each node is calculated, which solves the problem of how to establish a data feature model for each sensor under the normal running state. This step enables the baseline value of each node to objectively reflect the coupling degree of the equipment under normal state, thereby laying a solid foundation for subsequent deviation detection. Then, for any time, the node deviation value is calculated, which solves the problem of how to quantitatively reflect the difference between the current state and the normal baseline state of a node. The difference value can be positive or negative, directly reflecting whether the sensor has abnormal fluctuations or trend deviation at the current sampling time. Through this calculation, the system can quickly capture the initial signs of abnormal local data in the equipment operation, providing early warning signals for fault prediction. In addition, the standard deviation of the node degree is calculated for each node during the baseline period, which solves the problem of how to determine the data fluctuation range and tolerance limit. As a quantitative indicator for judging abnormality, the standard deviation can provide an objective reference. When the deviation value of a node exceeds its standard deviation, it can be judged as an abnormal state. In this way, not only the abnormal situation of a single node can be accurately detected, but also statistical basis is provided for the overall system state monitoring. In summary, through the steps of baseline construction, calculation of baseline average node degree, real-time calculation of node deviation value and standard deviation, the system solves the problem of real-time and quantitative judgment of abnormality of multi-sensor data in the equipment operation. This method not only improves the sensitivity and accuracy of fault detection, but also ensures the objectivity of the detection results through quantitative standards, and finally realizes the early warning and accurate determination of the equipment state abnormality.

[0164] It also includes global abnormality index construction and setting of abnormality judgment criteria, specifically:

[0165] The global abnormality index is constructed to quantify the overall deviation of all nodes, specifically:

[0166]

[0167] Wherein, Λ(t) is the global abnormality index at time t, reflecting the overall abnormality degree;

[0168] The global baseline mean and standard deviation are also calculated, specifically:

[0169]

[0170] Wherein, is the average value of the global topology index in the baseline period; σ G is the standard deviation of the global topology index in the baseline period, as a measure of the overall fluctuation amplitude;

[0171] The abnormality judgment state criterion is set as:

[0172] When for any node v i there exists |Δ(i, t)|>σ i , it is judged that the node has abnormal data deviation at time t;

[0173] When the global abnormality index satisfies Λ(t)>σ G , it is judged that the entire device system has abnormal state risk at time t.

[0174] Through the series of steps of constructing the global abnormality index and setting the abnormality judgment criterion, the problem of difficult unified quantification of each node abnormal deviation in the multi-sensor monitoring system is solved, and the comprehensive judgment of the abnormal state of the entire device system is realized. First, by constructing the global abnormality index, the problem that the single node deviation detection cannot reflect the overall system state is solved. This step integrates the deviation values of each node to form a unified index reflecting the abnormality degree of the entire device at time t, thereby realizing the quantitative correlation between local abnormality and global state. The higher the index value, the more obvious the overall abnormal trend of the device. Second, by calculating the global baseline mean and the global standard deviation, the problem of difficult determination of the global data fluctuation range under normal working conditions of the device is solved. By calculating the mean and standard deviation of the global topology index in the baseline period, a quantitative reference standard is provided for judging whether the real-time data deviates from the normal state. The baseline mean represents the overall coupling level of the device under normal operation, and the standard deviation serves as a measure of the fluctuation amplitude, ensuring that the abnormality judgment has an objective basis. Finally, the abnormality judgment state criterion is set, which stipulates that when the deviation value of any node exceeds its corresponding standard deviation, it is judged that the node has an abnormality; at the same time, when the global abnormality index exceeds the global standard deviation, it is judged that the entire device system has an abnormal state risk at time t. This step solves the problem of how to establish a clear judgment rule between local abnormality and global abnormality, and realizes the hierarchical and comprehensive judgment of the system abnormality warning. Overall, through the construction of the global abnormality index, the calculation of the global baseline mean and the standard deviation, and the setting of the clear abnormality judgment criterion, this series of steps effectively solves the problems of non-uniform global state after multi-sensor data fusion and difficult quantification of local abnormality, enabling the overall abnormal state of the device to be monitored and evaluated in a standardized and quantitative manner, thereby greatly improving the accuracy and response speed of the fault warning system.

[0175] It also includes fault warning information generation and decision support, specifically:

[0176] At each update time t, for node v i If |Δ(i, t)|>σ i is generated, and the report contains: node number i, current node degree D(i, t), baseline node degree B(i), node deviation value Δ(i, t) and baseline standard deviation σ i ; Each item of data is a specific calculation result, which is used to quantitatively judge whether the node deviates from the normal state;

[0177] The global topology index T(t), the global anomaly index Λ(t) and the global standard deviation σ G at the current time t are arranged, and an overall state report is generated, which lists the calculation results of T(t), Λ(t) and σ G in numerical form, ensuring that the overall data coupling state of the equipment is quantitatively described;

[0178] A fault warning flag is constructed, specifically:

[0179]

[0180] Wherein, FPF(t) is the fault warning flag at time t; FPF(t)=1 indicates that the equipment is abnormal; FPF(t)=0 indicates that the equipment is normal;

[0181] All numerical data obtained in each update period are arranged into a decision support table, which includes: time t; node data, for i={1, 2, 3}, D(i, t), B(i), Δ(i, t) and σ i are listed respectively; global index, the numerical values of T(t), Λ(t) and σ G ; The determined value of the fault warning flag FPF(t); This decision support table is directly used for reference by equipment maintenance personnel to judge whether to start the maintenance program.

[0182] The series of steps of fault early warning information generation and decision support solves the problem of how to convert a large amount of real-time calculation data into intuitive, quantitative and convenient maintenance personnel judgment decision information in the equipment state monitoring system, and realizes the timely early warning and fault positioning of the abnormal state of the equipment. First, at each update time, each data node is judged. When a node meets the preset abnormal condition, for example, the node deviation value exceeds the baseline standard deviation, the system immediately generates a node abnormal report. This step solves the problem of how to quantitatively judge whether a single sensor state deviates from the normal working range by calculating the current node degree, baseline node degree, node deviation value and baseline standard deviation of each node. The generated node abnormal report records the abnormal information of each node in the form of detailed numerical values, providing accurate positioning basis for subsequent maintenance. Secondly, the system sorts the global data at each update time, and summarizes the global topology index, global abnormal index and global standard deviation to generate an overall state report. This step solves the problem that it is difficult to reflect the overall state of the equipment in the traditional monitoring method by relying only on single sensor data. By fusing all sensor data of the equipment to form global indexes, the report lists the overall data coupling state in numerical form, so that the overall operation state of the equipment is quantitatively described, providing an objective and unified basis for overall evaluation of the state of the equipment. Thirdly, the system constructs a fault early warning flag, and its calculation result is clearly distinguished between the abnormal and normal states of the equipment according to the preset criteria. When the early warning flag indicates that the equipment is abnormal, for example, the value is 1, the system timely feeds back this information to the maintenance personnel. This step solves the problem of ambiguous early warning information and delayed response in traditional monitoring, and through binary output, the early warning result is intuitive and clear, which is convenient for the on-site operator to quickly judge whether the equipment needs immediate repair. Finally, all numerical data obtained in each update period are sorted into a decision support table, which covers the time stamp, detailed data of each node, and global indexes and fault early warning flags, such as node degree, baseline node degree, deviation value, and standard deviation. The decision support table is directly used for reference by equipment maintenance personnel, solving the problem of difficult comprehensive judgment of equipment health status by multi-dimensional data. Through this series of steps, the system realizes the whole process closed loop from data acquisition, abnormal judgment to fault early warning information output, and the final effect is that the maintenance personnel can judge the state of the equipment based on detailed and quantitative decision information in time, and quickly take repair measures, thereby greatly improving the accuracy and response speed of equipment fault early warning.

[0183] It is to be noted that, in the present text, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0184] The above description is merely preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, some improvements and refinements can be made without departing from the technical principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for condition detection and equipment fault early warning based on artificial intelligence, characterized in that, include: Three independent sensors are installed at key parts of the equipment to collect temperature, vibration, and pressure data, respectively. For temperature measurement devices, sensors For vibration measurement devices, sensors For pressure measurement; Let the discrete index of the sampling time be . ,in, The number of sampling points within a single update cycle; Set the function for raw data For sensors At sampling time The collected raw data, among which, , is the sensor number; The local neighbor mean is calculated as follows: right and all satisfy calculate: ; in, For sensors At any moment The mean of the data from the left and right neighboring points; Set the deviation as follows: ; in, For sensors At any moment The difference between the original data and the mean of its left and right neighboring points; Construct a smoothing function, specifically as follows: ; in, For sensors At any moment Smoothed output data after nonlinear modulation; Nonlinear compression is applied to the deviation; right and At that time, a symmetrical extension process is adopted, specifically as follows: ， ； For sensors During the sampling period Internal calculation: ， ； in, for The minimum value; for The maximum value; Set the normalized data as follows: ; in, The normalized data has a range of values. ; Real-time data stream integration and node construction; Mapping relationships between nodes and calculating edge weights; Perform adaptive topology model construction; The adaptive topology model is constructed as follows: At each update moment Based on the set of nodes and the calculated edge weights Constructing a time-varying graph is as follows: ; in, A time-varying topological graph; Let it be the set of edges; For nodes and At any moment The coupling strength between them; , For connecting nodes and The edge; Calculate the connection strength of each node in the graph, specifically as follows: ;in, For nodes At any moment The degree of the node; Perform global topology indicator construction and periodic updates; Anomaly feature extraction and status determination; The abnormal feature extraction and state determination are specifically as follows: Select the time interval during normal equipment operation, denoted as ; For nodes The baseline average nodal degree is calculated as follows: ； in, For nodes Average nodal degree during the baseline period; The baseline construction interval is defined as the start and end sampling times during normal equipment operation. The number of sampling periods during the baseline period; For any time Calculate the node offset value, specifically: ； in, For nodes At any moment The difference between the current node degree and its baseline value; During the baseline period, for nodes The standard deviation of the nodal degree is calculated as follows: ； in, For nodes Standard deviation of nodality during the baseline period; Construct global anomaly indicators and set anomaly judgment criteria; The construction and setting of the global anomaly index and anomaly judgment criteria are as follows: Construct a global anomaly indicator, specifically as follows: ； in, At any moment Global abnormal indicators; Simultaneously calculate the global baseline mean and standard deviation, specifically: ， ； in, This represents the average value of global topology metrics during the baseline period; This represents the standard deviation of the global topology index over the baseline period. For global topology metrics; The criteria for determining anomalies are set as follows: For any node exist At that time, determine the node at time [time]. There are abnormal data deviations; When the global anomaly index meets At that time, determine the entire equipment system at time 10:

00. There is a risk of abnormal conditions.

2. The method for state detection and equipment fault early warning based on artificial intelligence according to claim 1, characterized in that, It also includes real-time data stream integration and node construction, specifically: right Configure nodes: , indicating the corresponding sensor Data nodes; Set the node set as follows: ; At the global update time Construct nodes The attribute vector is as follows: ； in, For nodes At global time The attribute vector within; The current global sampling time has an update period of 1. ; In the device design, each sensor is pre-assigned fixed three-dimensional coordinates, specifically represented as follows: ;in, For sensors 3D coordinates; For sensors of coordinate; For sensors of coordinate; For sensors of Coordinates; the unit of coordinates is uniformly millimeters; After each update cycle, the vector formed by the latest sampled data is... Update to the corresponding node Inside.

3. The method for state detection and equipment fault early warning based on artificial intelligence according to claim 2, characterized in that, It also includes mapping of relationships between nodes and calculating edge weights, specifically: Construct a function specifically for measuring the data similarity between two nodes within the latest update period, for any... At the global update time The similarity score is calculated as follows: ； in, For nodes and At any moment The similarity value of the data sequences; The distance between sensors is calculated based on the physical location of the nodes, specifically as follows: ； in, The Euclidean distance represents the sensor distance. and The straight-line distance between them; Construct the physical interaction factors, specifically: ； in, Used to reflect the coupling effect between sensors due to physical distance; Combining signal similarity Interaction factors with physics Construct the edge weight function, specifically: ； in, For nodes and At any moment The coupling strength between them.

4. The method for state detection and equipment fault early warning based on artificial intelligence according to claim 3, characterized in that, It also includes the construction and periodic updating of global topology indicators, specifically: Calculate the global topology metric, set as the arithmetic mean of the degrees of all nodes: ;in, For global topology metrics; The value is fixed at 3, corresponding to three sensor nodes; Each update cycle, that is, when Increase At each sampling time, recalculate using the latest data. , and This forms a continuous time-varying topological graph sequence. .

5. The method for state detection and equipment fault early warning based on artificial intelligence according to claim 4, characterized in that, It also includes fault warning information generation and decision support, specifically: At each update moment For nodes If the following conditions are met: Then a node anomaly report will be generated, and the report content includes: node number. Current node degree Baseline node degree Node deviation value and baseline standard deviation ; For the current moment Organize global topology indicators Global anomaly indicators and global standard deviation Generate an overall status report, which lists the data in numerical form. , and The calculation results; Construct fault warning signs, specifically as follows: ； in, For at any time Fault warning signs; This indicates a device malfunction; This indicates that the equipment is in normal condition; All numerical data obtained in each update cycle are compiled into a decision support table, which includes: time period. Data for each node List them separately , , and Global metrics, listed , and The numerical value; fault warning sign A definite value.

Images