Sensor fault prediction method, system and device based on industrial internet of things

Abstract

This application discloses a sensor fault prediction method, system, and device based on the Industrial Internet of Things (IIoT), relating to the technical field of IIoT. The method includes: performing empirical mode decomposition (EMD) on the time-series signal to obtain intrinsic mode function (EMF) components; for homogeneous sensor instances in multiple sensor groups, constructing same-mode correlation vectors at equivalent time points based on the EMF components with the same physical meaning, and combining them to form a same-mode correlation matrix; for heterogeneous sensor instances in the same sensor group, constructing cross-mode correlation vectors at the same time points based on the EMF components representing core physical processes, and combining them to form a cross-mode correlation matrix; and performing mode decomposition on each of the same-mode correlation matrix and the cross-mode correlation matrix to obtain a first dominant mode representation and a second dominant mode representation. This application improves the flexibility of sensor fault prediction.

Description

Technical Field

[0001] This application relates to the technical field of industrial Internet of Things (IIoT), and in particular to sensor fault prediction methods, systems, and devices based on IIoT. Background Technology

[0002] The Industrial Internet of Things (IIoT) technology is increasingly being used in equipment condition monitoring and predictive maintenance. By deploying multiple sensors to collect equipment operating data in real time, it provides a data foundation for early fault warning. Currently, typical fault prediction methods are mainly based on the analysis of single-type sensor data or the simple fusion of multi-sensor data. For example, vibration signal spectrum analysis can be used to identify bearing wear, or temperature threshold-based over-limit alarm mechanisms can be implemented. With the development of deep learning technology, schemes that use recurrent neural networks to process sensor time-series data for anomaly detection have also emerged.

[0003] However, existing methods still face significant limitations in practical industrial applications: First, most methods employ fixed sensor data processing patterns, lacking adaptability to changes in sensor type, quantity, and deployment location within the monitoring environment. When production lines are adjusted or sensor networks are expanded, fault prediction models often need to be rebuilt, resulting in insufficient system flexibility. Second, in complex industrial environments, nonlinear physical correlations exist between different types of sensor data, and existing methods do not adequately explore the deep coupling relationships between these multimodal data, making it difficult to establish a unified correlation analysis framework. Furthermore, traditional methods lack adaptability to changes in sensor network topology, failing to dynamically adjust data analysis strategies according to actual monitoring needs.

[0004] These limitations mean that existing fault prediction systems often exhibit insufficient adaptability and poor deployment flexibility when facing diverse industrial application scenarios, thus hindering the further development and widespread application of industrial IoT systems in predictive maintenance. Summary of the Invention

[0005] To improve the flexibility of sensor fault prediction, this application provides a sensor fault prediction method, system and device based on the Industrial Internet of Things.

[0006] Firstly, this application provides a sensor fault prediction method based on the Industrial Internet of Things, employing the following technical solution:

[0007] A sensor fault prediction method based on the Industrial Internet of Things (IIoT) is applied to an IIoT system, which includes a management platform, a sensor network platform, and an object platform connected in sequence. The method is executed by the management platform and includes:

[0008] Multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment with multiple sets of sensors are acquired. Empirical mode decomposition is performed on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. The test environment of each set of sensors is the same, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor, and sound sensor.

[0009] For isomorphic sensor instances among multiple sensor groups, based on the intrinsic mode function components with the same physical meaning, isomorphic correlation vectors are constructed at equivalent time points and combined to form isomorphic correlation matrices;

[0010] For heterogeneous sensor instances within the same group of sensors, cross-modal correlation vectors are constructed at the same time point based on the intrinsic mode function components representing the core physical processes, and then combined to form a cross-modal correlation matrix.

[0011] Modal decomposition is performed on each of the same-modal correlation matrices to obtain a first dominant mode representation; and modal decomposition is performed on each of the cross-modal correlation matrices to obtain a second dominant mode representation.

[0012] Calculate the same-modal cooperative representation of each of the same-modal association vectors and the first dominant modality representation, and calculate the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modality representation;

[0013] Fault prediction information is generated based on the same-modal cooperative representation and the cross-modal cooperative representation.

[0014] By adopting the above technical solution, multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment using multiple sets of sensors are first acquired. Empirical mode decomposition is then performed on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. The testing environment of each set of sensors is the same, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor, and sound sensor. Then, for homogeneous sensor instances in the multiple sets of sensors, homomodal correlation vectors are constructed at equivalent time points based on the intrinsic mode function components with the same physical meaning, and these vectors are combined to form homomodal correlation matrices. Then, for heterogeneous sensor instances in the same set of sensors, cross-modal correlation vectors are constructed at the same time points based on the intrinsic mode function components representing the core physical processes, and these vectors are combined to form cross-modal correlation matrices. Finally, modal decomposition is performed on each homomodal correlation matrix to obtain the representation of the first dominant mode. Furthermore, modal decomposition is performed on the cross-modal correlation matrix to obtain the second dominant modal representation. Then, the co-modal cooperative representation of each co-modal correlation vector with the first dominant modal representation is calculated, and the cross-modal cooperative representation of each cross-modal correlation vector with the second dominant modal representation is calculated. Finally, fault prediction information is generated based on the co-modal and cross-modal cooperative representations. In this invention, a new data-driven fault monitoring and prediction paradigm is constructed by fusing the cooperative laws of homogeneous sensors with the physical correlation of heterogeneous sensors. First, through feature extraction and correlation analysis of multi-source signals, an analytical framework capable of adaptively characterizing the intrinsic working mode of sensor groups is established. Then, by dynamically learning the cooperative characteristics and coupling relationships of the sensor network, a description of the normal state of the system is formed. Finally, based on the calculation of cooperative degree, early identification and trend prediction of sensor performance degradation are achieved, thereby breaking through the dependence of traditional methods on fixed sensor configurations and preset data patterns, and improving the accuracy and flexibility of the fault prediction system.

[0015] Optionally, the step of performing empirical mode decomposition on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor includes:

[0016] For the timing signal of each sensor, the following screening process is performed:

[0017] Identify the local extrema of the time-series signal;

[0018] Based on the local extreme points, the upper and lower envelopes of the time series signal are constructed respectively, and the mean values ​​of the upper and lower envelopes are calculated to obtain the envelope mean curve.

[0019] The candidate component is obtained by subtracting the envelope mean curve from the time-series signal;

[0020] Check whether the candidate components satisfy the definition conditions of the intrinsic mode function, wherein the definition conditions include that the number of extreme points and the number of zero crossings differ by no more than one, and the mean of the upper envelope formed by local maxima and the lower envelope formed by local minima is close to zero.

[0021] When the candidate component satisfies the defined condition, the candidate component is output as an intrinsic mode function component;

[0022] The intrinsic mode function component is separated from the time-series signal, and the remaining signal is used as a new input signal to repeat the sieving process until the remaining signal becomes a monotonic function or only one extreme point remains. The remaining signal is then output as a residual term.

[0023] By adopting the above technical solution, in order to obtain the corresponding intrinsic mode function components, the following screening process is performed for each sensor's time-series signal: First, the local extrema of the time-series signal are identified. Then, based on the local extrema, the upper and lower envelopes of the time-series signal are constructed respectively, and the mean values ​​of the upper and lower envelopes are calculated to obtain the envelope mean curve. Then, the envelope mean curve is subtracted from the time-series signal to obtain candidate components. Then, it is checked whether the candidate components meet the definition conditions of the intrinsic mode function. The definition conditions include that the number of extrema and the number of zero-crossing points do not differ by more than one, and the mean values ​​of the upper envelope formed by local maxima and the lower envelope formed by local minima are close to zero. When the candidate component meets the definition conditions, the candidate component is output as an intrinsic mode function component. Then, an intrinsic mode function component is separated from the time-series signal, and the remaining signal is used as a new input signal to repeat the screening process until the remaining signal becomes a monotonic function or only one extremum remains. The remaining signal is then output as a residual term.

[0024] Optionally, the step of constructing same-mode correlation vectors at equivalent time points based on the eigenmode function components having the same physical meaning, and combining them to form a same-mode correlation matrix, includes:

[0025] Identify the dominant frequency and energy distribution characteristics of the intrinsic mode function components of each of the isomorphic sensor instances;

[0026] Cluster analysis is performed on the dominant frequency and the energy distribution characteristics to obtain multiple modal component mapping relationships, wherein the modal component mapping relationships are used to represent the mapping relationships between the intrinsic mode function components belonging to the same physical source in different sensor groups;

[0027] Based on the modal component mapping relationship, the intrinsic mode function components from the same physical source are grouped into the same group to obtain multiple intrinsic mode function component groups;

[0028] For each intrinsic mode function component group, based on a first preset time window, the same mode features are extracted for each intrinsic mode function component in the intrinsic mode function component group to obtain the corresponding same mode multidimensional features, wherein the same mode multidimensional features include at least one of time domain features, frequency domain features and nonlinear features;

[0029] Time alignment is performed on multiple sets of time-series signals to construct the time-series mapping relationship of each set of sensors;

[0030] Based on the aforementioned time-series mapping relationship, the equivalent time points of each group of sensors are determined;

[0031] Based on the equivalent time point, the same modal multidimensional features within the same intrinsic modal function component group are combined to form the same modal correlation vector corresponding to each of the equivalent time points;

[0032] According to the temporal order of the same-modal correlation vectors, the same-modal correlation vectors at each equivalent time point are arranged and combined to form a same-modal correlation matrix.

[0033] By adopting the above technical solution, in order to construct the same-mode correlation vector and the same-mode correlation matrix, the dominant frequency and energy distribution characteristics of the intrinsic mode function components of each isomorphic sensor instance are first identified. Then, cluster analysis is performed on the dominant frequency and energy distribution characteristics to obtain multiple modal component mapping relationships. These mapping relationships represent the mapping relationships between intrinsic mode function components belonging to the same physical source in different sensor groups. Based on these mapping relationships, the intrinsic mode function components from the same physical source are grouped into the same group, resulting in multiple intrinsic mode function component groups. Then, for each intrinsic mode function component group, based on a first preset time window, the intrinsic mode function... In each intrinsic mode function component of the component group, the same-mode features are extracted to obtain the corresponding same-mode multidimensional features. The same-mode multidimensional features include at least one of time domain features, frequency domain features, and nonlinear features. Then, the time-series signals of multiple groups are time-aligned to construct the time-series mapping relationship of each group of sensors. Then, based on the time-series mapping relationship, the equivalent time point of each group of sensors is determined. Then, based on the equivalent time point, the same-mode multidimensional features in the same intrinsic mode function component group are combined to form the same-mode correlation vector corresponding to each equivalent time point. Then, the same-mode correlation vectors of each equivalent time point are arranged according to the time order of the same-mode correlation vectors and combined to form the same-mode correlation matrix.

[0034] Optionally, the step of constructing cross-modal correlation vectors at the same time point based on the eigenmode function components representing the core physical processes, and combining them to form a cross-modal correlation matrix, includes:

[0035] The core physical processes of each heterogeneous sensor instance in the same group of sensors are determined, wherein the core physical processes include at least one of energy conversion processes, thermodynamic processes and mechanical vibration processes;

[0036] For sensors in the same group, based on the coupling characteristics of the core physical process, the physical correlation between the intrinsic mode function components of each heterogeneous sensor instance is determined;

[0037] Based on the physical correlation, at least one set of physically correlated modal components is determined from the intrinsic modal function components of each of the heterogeneous sensor instances, wherein the physically correlated modal components are composed of the intrinsic modal function components from each of the heterogeneous sensor instances in the same group of sensors that can characterize the same physical process;

[0038] At the same time point, based on the second preset time window, cross-modal correlation features are extracted from the intrinsic mode function components of each heterogeneous sensor instance to obtain cross-modal multidimensional features, wherein the cross-modal multidimensional features include at least one of the following: cross-correlation features for reflecting amplitude correlation, synchronization features for reflecting phase relationship, cross-spectral features for reflecting energy transfer, and mutual information features for reflecting nonlinear dependence.

[0039] For the same group of sensors, the cross-modal multidimensional features are combined according to a predefined sensor order to construct cross-modal correlation vectors at each time point;

[0040] The cross-modal correlation vectors at each time point are arranged according to their temporal order and combined to form a cross-modal correlation matrix.

[0041] By adopting the above technical solution, in order to construct the cross-modal correlation vector and cross-modal correlation matrix, the core physical processes of each heterogeneous sensor instance in the same group of sensors are first determined. These core physical processes include at least one of energy conversion processes, thermodynamic processes, and mechanical vibration processes. Then, for the same group of sensors, based on the coupling characteristics of the core physical processes, the physical correlation between the intrinsic mode function components of each heterogeneous sensor instance is determined. Based on the physical correlation, at least one set of physically correlated modal components is determined from the intrinsic mode function components of each heterogeneous sensor instance. These physically correlated modal components are constructed from the intrinsic mode function components of each heterogeneous sensor instance in the same group that characterize the same physical process. Then, at the same time point, based on the second preset time window, cross-modal correlation features are extracted from the intrinsic mode function components of each heterogeneous sensor instance to obtain cross-modal multidimensional features. The cross-modal multidimensional features include at least one of the following: cross-correlation features reflecting amplitude correlation, synchronization features reflecting phase relationship, cross-spectral features reflecting energy transfer, and mutual information features reflecting nonlinear dependence. Then, for the same group of sensors, the cross-modal multidimensional features are combined according to the predefined sensor order to construct cross-modal correlation vectors at each time point. Then, the cross-modal correlation vectors at each time point are arranged according to the time order of the cross-modal correlation vectors and combined to form a cross-modal correlation matrix.

[0042] Optionally, the step of performing mode decomposition on each of the same-mode correlation matrices to obtain the representation of the first dominant mode includes:

[0043] For each same-modal correlation matrix, a random projection matrix Ω is generated based on a preset target rank parameter;

[0044] Calculate the product of the same-modal correlation matrix and the random projection matrix Ω to obtain the projection matrix Y;

[0045] Perform QR decomposition on the projection matrix Y to obtain an orthogonal matrix Q and an upper triangular matrix R;

[0046] Calculate the product of the transpose of the orthogonal matrix Q and the modal correlation matrix to obtain the intermediate matrix B;

[0047] Singular value decomposition is performed on the intermediate matrix B to obtain the left singular vector matrix U_C1, the singular value matrix Σ_C1, and the right singular vector matrix V_C1;

[0048] Based on a preset variance contribution rate threshold, the top k largest singular values ​​are selected from the singular value matrix Σ_C1, and the top k right singular vectors are used as the representation of the first dominant mode, where k are positive integers determined according to the variance contribution rate threshold.

[0049] By adopting the above technical solution, in order to obtain the representation of the first dominant mode, for each co-modal correlation matrix, a random projection matrix Ω is generated based on a preset target rank parameter. Then, the product of the co-modal correlation matrix and the random projection matrix Ω is calculated to obtain the projection matrix Y. Then, the projection matrix Y is decomposed into QR to obtain an orthogonal matrix Q and an upper triangular matrix R. Then, the product of the transpose of the orthogonal matrix Q and the co-modal correlation matrix is ​​calculated to obtain an intermediate matrix B. Then, the intermediate matrix B is decomposed into singular values ​​to obtain a left singular vector matrix U_C1, a singular value matrix Σ_C1, and a right singular vector matrix V_C1. Then, according to a preset variance contribution rate threshold, the first k largest singular values ​​are selected from the singular value matrix Σ_C1, and the first k right singular vectors are used as the representation of the first dominant mode, where k is a positive integer determined according to the variance contribution rate threshold.

[0050] Optionally, the step of calculating the co-modal cooperative representation of each of the co-modal association vectors and the representation of the first dominant modality includes:

[0051] For each modality association vector, calculate the first cosine similarity between the modality association vector and the first dominant modality representation, and calculate the first projection length of the modality association vector on the first dominant modality representation;

[0052] The first cosine similarity and the first projection length are weighted and fused to obtain the same-modal collaborative representation.

[0053] By adopting the above technical solution, in order to obtain the same-modal collaborative representation, for each same-modal association vector, the first cosine similarity between the same-modal association vector and the first dominant modality representation is calculated, and the first projection length of the same-modal association vector on the first dominant modality representation is calculated. Then, the first cosine similarity and the first projection length are weighted and fused to obtain the same-modal collaborative representation.

[0054] Optionally, the step of calculating the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modality representation includes:

[0055] For each cross-modal association vector, calculate the second cosine similarity between the cross-modal association vector and the second dominant modality representation, and calculate the second projection length of the cross-modal association vector onto the second dominant modality representation;

[0056] The second cosine similarity and the second projection length are weighted and fused to obtain a cross-modal collaborative representation.

[0057] By adopting the above technical solution, in order to obtain the cross-modal collaborative representation, for each cross-modal association vector, the second cosine similarity between the cross-modal association vector and the second dominant modal representation is calculated, and the second projection length of the cross-modal association vector on the second dominant modal representation is calculated. Then, the second cosine similarity and the second projection length are weighted and fused to obtain the cross-modal collaborative representation.

[0058] Optionally, the step of generating fault prediction information based on the same-modal cooperative representation and the cross-modal cooperative representation includes:

[0059] For the same group of sensors, when the same modal cooperative representation of any homogeneous sensor instance in the group exceeds the first threshold range, and the cross-modal cooperative representation corresponding to the group of sensors exceeds the second threshold range, the group of sensors is determined to be faulty, and fault prediction information containing group identifier and sensor identifier is generated.

[0060] By adopting the above technical solution, in order to generate fault prediction information, for the same group of sensors, when the same modal cooperative representation of any isomorphic sensor instance in the group exceeds the first threshold range, and the cross-modal cooperative representation corresponding to the group of sensors exceeds the second threshold range, it is determined that the group of sensors has failed, and fault prediction information containing group identifier and sensor identifier is generated.

[0061] Secondly, this application also provides a sensor fault prediction system based on the Industrial Internet of Things, which adopts the following technical solution:

[0062] A sensor fault prediction system based on the Industrial Internet of Things (IIoT) includes a management platform, a sensor network platform, and an object platform that are sequentially connected in communication. The management platform is configured with:

[0063] The empirical mode decomposition module is used to acquire multiple sets of time-series signals obtained by sequential testing of the same industrial testing equipment by multiple sets of sensors, and to perform empirical mode decomposition on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. The test environment of each set of sensors is the same, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor and sound sensor.

[0064] The same-mode correlation module is used to construct same-mode correlation vectors at equivalent time points for isomorphic sensor instances in multiple sets of sensors, based on the intrinsic mode function components with the same physical meaning, and combine them to form a same-mode correlation matrix.

[0065] The cross-modal correlation module is used to construct cross-modal correlation vectors at the same time point for heterogeneous sensor instances in the same group of sensors, based on the intrinsic mode function components representing the core physical processes, and combine them to form a cross-modal correlation matrix.

[0066] The mode decomposition module is used to perform mode decomposition on each of the same-modal correlation matrices to obtain a first dominant mode representation; and to perform mode decomposition on each of the cross-modal correlation matrices to obtain a second dominant mode representation;

[0067] The modal cooperative representation module is used to calculate the same-modal cooperative representation of each of the same-modal association vectors and the first dominant modal representation, and to calculate the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modal representation;

[0068] The fault prediction information generation module is used to generate fault prediction information based on the same-modal collaborative representation and the cross-modal collaborative representation.

[0069] Thirdly, this application also provides a computer device, which adopts the following technical solution:

[0070] A computer device includes a memory and a processor, the memory storing a computer program executable on the processor, the processor executing the computer program to implement the method described in the first aspect.

[0071] In summary, this application includes at least the following beneficial technical effects: First, multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment using multiple sets of sensors are acquired. Empirical mode decomposition is then performed on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. The testing environment of each set of sensors is the same, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor, and sound sensor. Then, for homogeneous sensor instances among the multiple sets of sensors, a same-mode correlation vector is constructed at equivalent time points based on the intrinsic mode function components with the same physical meaning, and these vectors are combined to form a same-mode correlation matrix. Then, for heterogeneous sensor instances among the same set of sensors, a cross-mode correlation vector is constructed at the same time point based on the intrinsic mode function components representing the core physical process, and these vectors are combined to form a cross-mode correlation matrix. Finally, mode decomposition is performed on each same-mode correlation matrix to obtain the representation of the first dominant mode. Furthermore, modal decomposition is performed on the cross-modal correlation matrix to obtain the second dominant modal representation. Then, the co-modal cooperative representation of each co-modal correlation vector with the first dominant modal representation is calculated, and the cross-modal cooperative representation of each cross-modal correlation vector with the second dominant modal representation is calculated. Finally, fault prediction information is generated based on the co-modal and cross-modal cooperative representations. In this invention, a new data-driven fault monitoring and prediction paradigm is constructed by fusing the cooperative laws of homogeneous sensors with the physical correlation of heterogeneous sensors. First, through feature extraction and correlation analysis of multi-source signals, an analytical framework capable of adaptively characterizing the intrinsic working mode of sensor groups is established. Then, by dynamically learning the cooperative characteristics and coupling relationships of the sensor network, a description of the normal state of the system is formed. Finally, based on the calculation of cooperative degree, early identification and trend prediction of sensor performance degradation are achieved, thereby breaking through the dependence of traditional methods on fixed sensor configurations and preset data patterns, and improving the accuracy and flexibility of the fault prediction system. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall process of an embodiment of this application.

[0073] Figure 2 This is a structural diagram of one application scenario of the system in this application embodiment.

[0074] Figure 3 This is a structural diagram of another application scenario of the system according to an embodiment of this application.

[0075] Figure 4 This is a structural block diagram of the computer device described in this application. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0077] This application discloses a sensor fault prediction method based on the Industrial Internet of Things.

[0078] Reference Figure 1 A sensor fault prediction method based on the Industrial Internet of Things (IIoT) is applied to an IIoT system, which includes a management platform, a sensor network platform, and an object platform connected in sequence. The method is executed by the management platform and includes:

[0079] Step S11: Obtain multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment with multiple sets of sensors, and perform empirical mode decomposition on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor.

[0080] Each group of sensors is tested in the same environment, and each group of sensors contains at least two of the following: a temperature sensor, a vibration sensor, and a sound sensor.

[0081] It should be noted that in step S11, multiple sensor sets refer to the repeated deployment of multiple identical sensor combinations at exactly the same critical locations on the same industrial testing equipment. Their connection lies in the fact that all sensor sets monitor the same physical location of the equipment, and each set contains the same type and number of sensors. For example, each set includes a temperature sensor, a vibration sensor, and a sound sensor, with identical models and installation locations. The difference is that these sensor sets operate sequentially and in turn over time; they do not run simultaneously but rather collect data at different time intervals through sequential testing. Sequential testing specifically means that these identical sensor sets are tested sequentially and in turn under the same stable operating conditions. For example, the first set of sensors collects signals for a period of time, then the second set of sensors is installed or switched to collect signals under the same conditions, and so on. The core purpose of this is to obtain multiple sets of parallel data from the same location, under the same environment, and generated by different sensor instances. The identical test environment means that the operating status and environmental conditions of the industrial equipment remain consistent throughout all these sequential tests. This ensures that any systematic differences between the acquired time-series signals can be primarily attributed to variations in sensor performance or potential faults, rather than to fluctuations in equipment status or the external environment. Subsequently, empirical mode decomposition (IMF) is performed on the time-series signals acquired by each sensor. This adaptive signal processing method decomposes complex, non-stationary raw signals into a series of intrinsic mode function components (IMFs) ranging from high to low frequencies. Each IMF component represents an oscillation mode at a specific time scale inherent in the signal, providing corresponding data for subsequent analysis.

[0082] Step S12: For isomorphic sensor instances among multiple sensor groups, construct isomorphic correlation vectors at equivalent time points based on eigenmode function components with the same physical meaning, and combine them to form isomorphic correlation matrices.

[0083] It should be noted that in step S12, a homogeneous sensor instance refers to a sensor that is completely identical in type, model, and installation location within the group defined in this application. Specifically, in the context of this solution, it refers to sensors of the exact same model that are deployed at exactly the same location on the equipment in different sensor groups. For example, all temperature sensors installed at "point A of the machine tool spindle bearing housing" constitute a set of homogeneous sensor instances. These sensors are physically multiple independent individuals, but they play exactly the same role in the monitoring system. After performing empirical mode decomposition on the time-series signals collected by all these homogeneous sensors, a series of IMF components are obtained. Having the same physical meaning means that the IMF components decomposed from these signals have the same ordinal number or similar frequency scale. For example, the first IMF component decomposed from each sensor signal usually represents the highest frequency oscillation, and they have similar physical meanings. The second IMF component represents the second highest frequency oscillation, and so on. These components describe the dynamic behavior of the equipment at the same physical location and the same physical dimension, at the same scale. Equivalent time points refer to the same phase or operating condition stage in the industrial equipment's operating cycle when different sensor groups are tested sequentially. For example, they all correspond to the 10th second of a machine tool completing a full machining cycle. At each such equivalent time point, we extract features from the IMF components with the same physical meaning of all isomorphic sensor instances and arrange the extracted features into a vector. This vector is the modal correlation vector, which identifies the cooperative state of all similar sensors at a certain moment, clearly showing the consistency or difference of all sensors with identical functions and positions when characterizing the same physical process. Then, the modal correlation vectors generated at each equivalent time point are combined in chronological order to form a matrix, namely the modal correlation matrix. Therefore, this matrix is ​​a temporal feature matrix characterizing the evolution of the cooperative state of a group of isomorphic sensors in the time dimension. By analyzing the modal correlation matrix, we can capture the dynamic change law of the cooperative state of a group of isomorphic sensors.

[0084] Step S13: For heterogeneous sensor instances in the same group of sensors, construct cross-modal correlation vectors at the same time point based on the intrinsic mode function components representing the core physical processes, and combine them to form a cross-modal correlation matrix.

[0085] It should be noted that in step S13, heterogeneous sensor instances refer to different types of sensors within the same group, such as a temperature sensor, a vibration sensor, and a sound sensor; "Intrinsic Mode Function Components Representing the Core Physical Process" refers to selecting the IMF components most relevant to the core operating mechanism of the equipment from these heterogeneous sensor signals, such as the components in vibration and sound signals that are at the same frequency as the spindle speed, and the low-frequency components in temperature signals that reflect the overall thermal balance; at "the same point in time", combining these IMF component values ​​from different physical domains but reflecting the same core process forms a cross-modal correlation vector, which describes how different physical quantities jointly reflect the operating state of the equipment at this moment. The cross-modal correlation vectors at all time points form the cross-modal correlation matrix.

[0086] Step S14: Perform mode decomposition on each same-modal correlation matrix to obtain the first dominant mode representation; and perform mode decomposition on each cross-modal correlation matrix to obtain the second dominant mode representation.

[0087] It should be noted that in step S14, by performing modal decomposition on the same-modal correlation matrix, a first dominant modal representation that can represent the cooperative change modes of the vast majority of isomorphic sensors can be extracted; similarly, by decomposing the cross-modal correlation matrix, a second dominant modal representation that can summarize the core correlation modes between different physical quantities can be obtained. These two dominant modal representations serve as templates for the cooperative behavior of isomorphic and heterogeneous sensor groups, respectively.

[0088] Step S15: Calculate the same-modal collaborative representation of each same-modal association vector and the first dominant modal representation, and calculate the cross-modal collaborative representation of each cross-modal association vector and the second dominant modal representation.

[0089] Step S16: Generate fault prediction information based on the same-modal cooperative representation and cross-modal cooperative representation.

[0090] In the above implementation, multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment using multiple sets of sensors are first acquired. Empirical mode decomposition is then performed on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. Each set of sensors has the same testing environment, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor, and sound sensor. Then, for homogeneous sensor instances in the multiple sets of sensors, homomodal correlation vectors are constructed at equivalent time points based on the intrinsic mode function components with the same physical meaning, and these vectors are combined to form homomodal correlation matrices. Then, for heterogeneous sensor instances in the same set of sensors, cross-modal correlation vectors are constructed at the same time points based on the intrinsic mode function components representing the core physical processes, and these vectors are combined to form cross-modal correlation matrices. Finally, modal decomposition is performed on each homomodal correlation matrix to obtain the representation of the first dominant mode. Furthermore, modal decomposition is performed on the cross-modal correlation matrix to obtain the second dominant modal representation. Then, the co-modal cooperative representation of each co-modal correlation vector with the first dominant modal representation is calculated, and the cross-modal cooperative representation of each cross-modal correlation vector with the second dominant modal representation is calculated. Finally, fault prediction information is generated based on the co-modal and cross-modal cooperative representations. In this invention, a new data-driven fault monitoring and prediction paradigm is constructed by fusing the cooperative laws of homogeneous sensors with the physical correlation of heterogeneous sensors. First, through feature extraction and correlation analysis of multi-source signals, an analytical framework capable of adaptively characterizing the intrinsic working mode of sensor groups is established. Then, by dynamically learning the cooperative characteristics and coupling relationships of the sensor network, a description of the normal state of the system is formed. Finally, based on the calculation of cooperative degree, early identification and trend prediction of sensor performance degradation are achieved, thereby breaking through the dependence of traditional methods on fixed sensor configurations and preset data patterns, and improving the accuracy and flexibility of the fault prediction system.

[0091] As a further implementation of the method, the step of performing empirical mode decomposition on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor includes:

[0092] Step S21: For the timing signal of each sensor, perform the following screening process:

[0093] Step S22: Identify the local extrema of the time series signal.

[0094] Step S23: Based on the local extreme points, construct the upper envelope and lower envelope of the time sequence signal respectively, and calculate the mean of the upper envelope and lower envelope to obtain the envelope mean curve.

[0095] Step S24: Subtract the envelope mean curve from the time series signal to obtain the candidate component.

[0096] Step S25: Check whether the candidate components satisfy the definition conditions of the intrinsic mode function. The definition conditions include that the number of extreme points and the number of zero crossings differ by no more than one, and that the mean of the upper envelope formed by local maxima and the lower envelope formed by local minima is close to zero.

[0097] Step S26: When a candidate component satisfies the defined conditions, the candidate component is output as an eigenmode function component.

[0098] Step S27: Separate an intrinsic mode function component from the time-series signal, and repeat the sieving process with the remaining signal as a new input signal until the remaining signal becomes a monotonic function or only one extreme point remains, and output the remaining signal as a residual term.

[0099] In the above implementation, in order to obtain the corresponding intrinsic mode function components, the following screening process is performed for each sensor's time-series signal: First, the local extrema of the time-series signal are identified. Then, based on the local extrema, the upper and lower envelopes of the time-series signal are constructed respectively, and the mean values ​​of the upper and lower envelopes are calculated to obtain the envelope mean curve. Then, the envelope mean curve is subtracted from the time-series signal to obtain candidate components. Then, it is checked whether the candidate components meet the definition conditions of the intrinsic mode function. The definition conditions include that the number of extrema and the number of zero crossings do not differ by more than one, and the mean values ​​of the upper envelope formed by local maxima and the lower envelope formed by local minima are close to zero. When the candidate component meets the definition conditions, the candidate component is output as an intrinsic mode function component. Then, an intrinsic mode function component is separated from the time-series signal, and the remaining signal is used as a new input signal to repeat the screening process until the remaining signal becomes a monotonic function or only one extrema remains. The remaining signal is then output as a residual term.

[0100] As a further implementation of the method, the step of constructing same-mode correlation vectors at equivalent time points based on eigenmode function components with the same physical meaning, and combining them to form same-mode correlation matrices, includes:

[0101] Step S31: Identify the dominant frequency and energy distribution characteristics of the intrinsic mode function components of each isomorphic sensor instance.

[0102] Step S32: Cluster analysis is performed on the dominant frequency and energy distribution characteristics to obtain multiple modal component mapping relationships. The modal component mapping relationship is used to represent the mapping relationship between intrinsic mode function components belonging to the same physical source in different sensor groups.

[0103] Step S33: Based on the modal component mapping relationship, the intrinsic mode function components from the same physical source are divided into the same group to obtain multiple intrinsic mode function component groups.

[0104] Step S34: For each intrinsic mode function component group, based on the first preset time window, perform same-mode feature extraction on each intrinsic mode function component in the intrinsic mode function component group to obtain the corresponding same-mode multidimensional features, wherein the same-mode multidimensional features include at least one of time domain features, frequency domain features and nonlinear features.

[0105] Step S35: Time alignment is performed on multiple sets of timing signals to construct the timing mapping relationship of each set of sensors.

[0106] Step S36: Based on the time-series mapping relationship, determine the equivalent time point of each group of sensors.

[0107] Step S37: Based on the equivalent time point, combine the same modal multidimensional features within the same intrinsic modal function component group to form the same modal correlation vector corresponding to each equivalent time point.

[0108] Step S38: Arrange the same-modal correlation vectors at each equivalent time point according to the time order of the same-modal correlation vectors, and combine them to form the same-modal correlation matrix.

[0109] In the above implementation, to construct the same-mode correlation vector and the same-mode correlation matrix, the dominant frequency and energy distribution characteristics of the intrinsic mode function components of each isomorphic sensor instance are first identified. Then, cluster analysis is performed on the dominant frequency and energy distribution characteristics to obtain multiple modal component mapping relationships. These mapping relationships represent the mapping relationships between intrinsic mode function components belonging to the same physical source in different sensor groups. Based on these mapping relationships, the intrinsic mode function components from the same physical source are grouped into the same group, resulting in multiple intrinsic mode function component groups. Then, for each intrinsic mode function component group, based on a first preset time window, the intrinsic mode function components are... In each intrinsic mode function component of the signal set, the same-mode features are extracted to obtain the corresponding same-mode multidimensional features. The same-mode multidimensional features include at least one of time domain features, frequency domain features, and nonlinear features. Then, the time series signals of multiple sets are time-aligned to construct the time series mapping relationship of each set of sensors. Then, based on the time series mapping relationship, the equivalent time point of each set of sensors is determined. Then, based on the equivalent time point, the same-mode multidimensional features in the same intrinsic mode function component set are combined to form the same-mode correlation vector corresponding to each equivalent time point. Then, the same-mode correlation vectors of each equivalent time point are arranged according to the time order of the same-mode correlation vectors and combined to form the same-mode correlation matrix.

[0110] As a further implementation of the method, the step of constructing cross-modal correlation vectors at the same time point based on the eigenmode function components representing the core physical processes, and combining them to form a cross-modal correlation matrix, includes:

[0111] Step S41: Determine the core physical processes of each heterogeneous sensor instance in the same group of sensors, wherein the core physical processes include at least one of energy conversion processes, thermodynamic processes, and mechanical vibration processes.

[0112] It should be noted that, for step S41, the operation of industrial equipment is essentially the result of the interplay of multiple physical processes. For example, the operation of a rotating machine simultaneously involves mechanical vibration processes (such as vibration generated by bearing rotation), thermodynamic processes (such as heat generated by friction and loss), and energy conversion processes (such as the conversion of electrical energy into mechanical energy and acoustic energy). The purpose of this step is to identify these dominant underlying physical processes in the monitored equipment that can be perceived by existing sensors (temperature, vibration, sound).

[0113] Step S42: For sensors in the same group, based on the coupling characteristics of the core physical processes, determine the physical correlation between the intrinsic mode function components of each heterogeneous sensor instance.

[0114] It should be noted that the core of step S42 lies in deeply analyzing the interaction mechanisms between different physical processes, thereby establishing the intrinsic connections between heterogeneous sensor data. Based on identifying the core physical processes, this step needs to explore how these processes are coupled through physical laws. For example, bearing wear during mechanical vibration can affect thermodynamic processes through tribological principles, leading to temperature increases; simultaneously, the transfer of vibration energy can affect acoustic signal characteristics through acoustic radiation principles. This coupling analysis ensures that sensor data of different physical quantities are no longer isolated entities, but rather form a network reflecting the state of the same device. By establishing the correlation between specific frequency band components of the vibration signal and trend components of the temperature signal, or the correspondence between vibration characteristics and the acoustic spectrum, a basis is provided for subsequent screening of physically meaningful IMF components.

[0115] Step S43: Based on physical correlation, at least one set of physically correlated mode components are determined from the intrinsic mode function components of each heterogeneous sensor instance. The physically correlated mode components are composed of intrinsic mode function components from each heterogeneous sensor instance in the same group of sensors that can characterize the same physical process.

[0116] It should be noted that, for step S43, based on clarifying the physical correlation, the screening and grouping of modal components are specifically implemented. This step selects those components from the intrinsic modal function components of each sensor that are physically related and coordinated in terms of time-frequency characteristics, and classifies them into physically related modal component groups. For example, the characteristic modes reflecting bearing impact in vibration signals, the trend modes corresponding to frictional temperature rise in temperature signals, and the radiated noise modes originating from the same vibration in sound signals are grouped together. This grouping ensures that the features of subsequent cross-modal analysis all point to the same physical phenomenon, avoids false correlations without physical meaning, and improves the accuracy and interpretability of fault feature extraction.

[0117] Step S44: At the same time point, based on the second preset time window, cross-modal correlation features are extracted from the intrinsic mode function components of each heterogeneous sensor instance to obtain cross-modal multidimensional features. The cross-modal multidimensional features include at least one of the following: cross-correlation features reflecting amplitude correlation, synchronization features reflecting phase relationship, cross-spectral features reflecting energy transfer, and mutual information features reflecting nonlinear dependence.

[0118] Step S45: For the same group of sensors, combine the cross-modal multidimensional features according to the predefined sensor order to construct the cross-modal correlation vector at each time point.

[0119] Step S46: Arrange the cross-modal correlation vectors at each time point according to the time order of the cross-modal correlation vectors, and combine them to form a cross-modal correlation matrix.

[0120] In the above implementation, to construct the cross-modal correlation vector and cross-modal correlation matrix, the core physical processes of each heterogeneous sensor instance in the same group of sensors are first determined. These core physical processes include at least one of energy conversion processes, thermodynamic processes, and mechanical vibration processes. Then, for the same group of sensors, based on the coupling characteristics of the core physical processes, the physical correlation between the intrinsic mode function components of each heterogeneous sensor instance is determined. Based on these physical correlations, at least one set of physically correlated modal components is determined from the intrinsic mode function components of each heterogeneous sensor instance. These physically correlated modal components consist of intrinsic mode function components from each heterogeneous sensor instance in the same group that characterize the same physical process. Then, at the same time point, based on the second preset time window, cross-modal correlation features are extracted from the intrinsic mode function components of each heterogeneous sensor instance to obtain cross-modal multidimensional features. The cross-modal multidimensional features include at least one of the following: cross-correlation features reflecting amplitude correlation, synchronization features reflecting phase relationship, cross-spectral features reflecting energy transfer, and mutual information features reflecting nonlinear dependence. Then, for the same group of sensors, the cross-modal multidimensional features are combined according to the predefined sensor order to construct cross-modal correlation vectors at each time point. Then, the cross-modal correlation vectors at each time point are arranged according to the time order of the cross-modal correlation vectors and combined to form a cross-modal correlation matrix.

[0121] As a further implementation of the method, the step of performing mode decomposition on each homomodal correlation matrix to obtain the representation of the first dominant mode includes:

[0122] Step S51: For each same modal correlation matrix, generate a random projection matrix Ω based on a preset target rank parameter.

[0123] Step S52: Calculate the product of the same modal correlation matrix and the random projection matrix Ω to obtain the projection matrix Y.

[0124] Step S53: Perform QR decomposition on the projection matrix Y to obtain the orthogonal matrix Q and the upper triangular matrix R.

[0125] Step S54: Calculate the product of the transpose of the orthogonal matrix Q and the modal incidence matrix to obtain the intermediate matrix B.

[0126] Step S55: Perform singular value decomposition on the intermediate matrix B to obtain the left singular vector matrix U_C1, the singular value matrix Σ_C1, and the right singular vector matrix V_C1.

[0127] Step S56: Based on the preset variance contribution rate threshold, select the top k largest singular values ​​from the singular value matrix Σ_C1, and use the top k right singular vectors as the representation of the first dominant mode, where k are positive integers determined according to the variance contribution rate threshold.

[0128] It should be noted that the calculation method of the second dominant mode representation is basically the same as that of the first dominant mode representation. The calculation method of the second dominant mode representation can be referred to steps S51 to S56.

[0129] In the above implementation, in order to obtain the representation of the first dominant mode, for each modal correlation matrix, a random projection matrix Ω is generated based on a preset target rank parameter. Then, the product of the modal correlation matrix and the random projection matrix Ω is calculated to obtain the projection matrix Y. Then, the projection matrix Y is decomposed into QR to obtain an orthogonal matrix Q and an upper triangular matrix R. Then, the product of the transpose of the orthogonal matrix Q and the modal correlation matrix is ​​calculated to obtain an intermediate matrix B. Then, the intermediate matrix B is decomposed into singular values ​​to obtain a left singular vector matrix U_C1, a singular value matrix Σ_C1, and a right singular vector matrix V_C1. Then, according to a preset variance contribution rate threshold, the top k largest singular values ​​are selected from the singular value matrix Σ_C1, and the top k right singular vectors are used as the representation of the first dominant mode, where k is a positive integer determined according to the variance contribution rate threshold.

[0130] As a further implementation of the method, the step of calculating the co-modal cooperative representation of each co-modal association vector and the representation of the first dominant modality includes:

[0131] Step S61: For each modal association vector, calculate the first cosine similarity between the modal association vector and the first dominant modality representation, and calculate the first projection length of the modal association vector on the first dominant modality representation.

[0132] Step S62: Weighted fusion of the first cosine similarity and the first projection length is performed to obtain the same-modal collaborative representation.

[0133] In the above implementation, in order to obtain the same-modal cooperative representation, for each same-modal association vector, the first cosine similarity between the same-modal association vector and the first dominant modality representation is calculated, and the first projection length of the same-modal association vector on the first dominant modality representation is calculated. Then, the first cosine similarity and the first projection length are weighted and fused to obtain the same-modal cooperative representation.

[0134] As a further implementation of the method, the step of calculating the cross-modal cooperative representation of each cross-modal correlation vector and the second dominant modality representation includes:

[0135] Step S71: For each cross-modal association vector, calculate the second cosine similarity between the cross-modal association vector and the second dominant modality representation, and calculate the second projection length of the cross-modal association vector on the second dominant modality representation.

[0136] Step S72: Weighted fusion of the second cosine similarity and the second projection length is performed to obtain the cross-modal collaborative representation.

[0137] In the above implementation, in order to obtain the cross-modal collaborative representation, for each cross-modal association vector, the second cosine similarity between the cross-modal association vector and the second dominant modal representation is calculated, and the second projection length of the cross-modal association vector on the second dominant modal representation is calculated. Then, the second cosine similarity and the second projection length are weighted and fused to obtain the cross-modal collaborative representation.

[0138] As a further implementation of the method, the step of generating fault prediction information based on intramodal cooperative representation and cross-modal cooperative representation includes:

[0139] For the same group of sensors, when the same modal cooperative representation of any isomorphic sensor instance in the group exceeds the first threshold range, and the cross-modal cooperative representation of the corresponding sensor group exceeds the second threshold range, the sensor group is determined to be faulty, and fault prediction information containing group identifier and sensor identifier is generated.

[0140] In the above embodiments, in order to generate fault prediction information, for the same group of sensors, when the same modal cooperative representation of any isomorphic sensor instance in the group exceeds the first threshold range, and the cross-modal cooperative representation corresponding to the group of sensors exceeds the second threshold range, it is determined that the group of sensors has failed, and fault prediction information containing group identifier and sensor identifier is generated.

[0141] This application also discloses a sensor fault prediction system based on the Industrial Internet of Things.

[0142] refer to Figure 2 The sensor fault prediction system based on the Industrial Internet of Things includes a management platform, a sensor network platform, and an object platform that are connected in sequence via communication. The management platform is configured with:

[0143] The Empirical Mode Decomposition (EMD) module is used to acquire multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment using multiple sets of sensors. The EMD module is then used to perform EMD on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. Each set of sensors is tested in the same environment, and each set of sensors includes at least two of the following: a temperature sensor, a vibration sensor, and a sound sensor.

[0144] The same-mode correlation module is used to construct same-mode correlation vectors at equivalent time points for isomorphic sensor instances in multiple sets of sensors, based on eigenmode function components with the same physical meaning, and combine them to form a same-mode correlation matrix;

[0145] The cross-modal correlation module is used to construct cross-modal correlation vectors at the same time point for heterogeneous sensor instances in the same group of sensors, based on the intrinsic mode function components representing the core physical processes, and combine them to form a cross-modal correlation matrix.

[0146] The mode decomposition module is used to perform mode decomposition on each same-mode correlation matrix to obtain the first dominant mode representation; and to perform mode decomposition on each cross-mode correlation matrix to obtain the second dominant mode representation.

[0147] The modal cooperative representation module is used to calculate the same-modal cooperative representation of each same-modal association vector and the first dominant modal representation, and to calculate the cross-modal cooperative representation of each cross-modal association vector and the second dominant modal representation;

[0148] The fault prediction information generation module is used to generate fault prediction information based on the same-modal cooperative representation and cross-modal cooperative representation.

[0149] The overall framework of another application scenario of the sensor fault prediction system based on the Industrial Internet of Things in this application is as follows: Figure 3 As shown, it can include a user platform, service platform, management platform, sensor network platform, and object platform that interact sequentially, forming a five-platform architecture based on the Industrial Internet of Things. The management platform includes an empirical mode decomposition module, a same-modal association module, a cross-modal association module, a mode decomposition module, a modal collaborative representation module, and a fault prediction information generation module. The sensor network platform includes several sensor network sub-platforms, each with its own sensor sub-database.

[0150] Specifically, in another application scenario mentioned above, the sensor fault prediction system based on the Industrial Internet of Things includes a management platform. The management platform is configured to: acquire multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment using multiple sets of sensors; perform empirical mode decomposition on each time-series signal to obtain the intrinsic mode function components corresponding to each sensor; wherein each set of sensors has the same testing environment, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor, and sound sensor; for isomorphic sensor instances among the multiple sets of sensors, construct isomodal correlation vectors at equivalent time points based on intrinsic mode function components with the same physical meaning, and group them... A homomodal correlation matrix is ​​formed. For heterogeneous sensor instances within the same group of sensors, cross-modal correlation vectors are constructed at the same time point based on the intrinsic mode function components representing the core physical processes, and these vectors are combined to form a cross-modal correlation matrix. Modal decomposition is performed on each homomodal correlation matrix to obtain a first dominant mode representation. Modal decomposition is also performed on each cross-modal correlation matrix to obtain a second dominant mode representation. Homomodal cooperative representations of each homomodal correlation vector and the first dominant mode representation are calculated, as are cross-modal cooperative representations of each cross-modal correlation vector and the second dominant mode representation. Fault prediction information is generated based on the homomodal cooperative representations and the cross-modal cooperative representations.

[0151] By leveraging the interaction between the various functional platforms of the industrial IoT-based sensor fault prediction system, which is based on the aforementioned three or five platforms, a complete closed-loop information operation logic is established, ensuring the orderly operation of sensing and control information and realizing intelligent equipment management.

[0152] The sensor fault prediction system based on the Industrial Internet of Things of the present invention can implement any of the sensor fault prediction methods based on the Industrial Internet of Things, and the specific working process of the sensor fault prediction system based on the Industrial Internet of Things of the present invention can refer to the corresponding process in the above-mentioned sensor fault prediction methods based on the Industrial Internet of Things.

[0153] This application also discloses a computer device.

[0154] refer to Figure 4 A computer device includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement any of the above-described methods for sensor fault prediction based on the Industrial Internet of Things.

[0155] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for predicting sensor failure based on industrial internet of things, characterized by, Applied to an industrial Internet of Things (IIoT) system, the IIoT system includes a management platform, a sensor network platform, and an object platform that are sequentially connected in communication. The method is executed by the management platform and includes: Multiple sets of time-series signals obtained by sequentially testing the same industrial testing equipment with multiple sets of sensors are acquired. Empirical mode decomposition is performed on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. The test environment of each set of sensors is the same, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor, and sound sensor. For isomorphic sensor instances among multiple sensor groups, based on the intrinsic mode function components with the same physical meaning, isomorphic correlation vectors are constructed at equivalent time points and combined to form isomorphic correlation matrices; For heterogeneous sensor instances within the same group of sensors, cross-modal correlation vectors are constructed at the same time point based on the intrinsic mode function components representing the core physical processes, and then combined to form a cross-modal correlation matrix. Modal decomposition is performed on each of the same-modal correlation matrices to obtain a first dominant mode representation; and modal decomposition is performed on each of the cross-modal correlation matrices to obtain a second dominant mode representation. Calculate the same-modal cooperative representation of each of the same-modal association vectors and the first dominant modality representation, and calculate the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modality representation; Based on the same-modal cooperative representation and the cross-modal cooperative representation, fault prediction information is generated; The step of performing mode decomposition on each of the same-mode correlation matrices to obtain the representation of the first dominant mode includes: For each same-modal correlation matrix, a random projection matrix Ω is generated based on a preset target rank parameter; Calculate the product of the same-modal correlation matrix and the random projection matrix Ω to obtain the projection matrix Y; Perform QR decomposition on the projection matrix Y to obtain an orthogonal matrix Q and an upper triangular matrix R; Calculate the product of the transpose of the orthogonal matrix Q and the modal correlation matrix to obtain the intermediate matrix B; Singular value decomposition is performed on the intermediate matrix B to obtain the left singular vector matrix U_C1, the singular value matrix Σ_C1, and the right singular vector matrix V_C1; According to the preset variance contribution rate threshold, the top k largest singular values ​​are selected from the singular value matrix Σ_C1, and the top k right singular vectors are used as the first dominant mode representation, where k are positive integers determined according to the variance contribution rate threshold. The step of calculating the co-modal cooperative representation of each of the co-modal association vectors and the representation of the first dominant modality includes: For each modality association vector, calculate the first cosine similarity between the modality association vector and the first dominant modality representation, and calculate the first projection length of the modality association vector on the first dominant modality representation; The first cosine similarity and the first projection length are weighted and fused to obtain the same-modal collaborative representation; The step of calculating the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modality representation includes: For each cross-modal association vector, calculate the second cosine similarity between the cross-modal association vector and the second dominant modality representation, and calculate the second projection length of the cross-modal association vector onto the second dominant modality representation; The second cosine similarity and the second projection length are weighted and fused to obtain a cross-modal collaborative representation. 2.The industrial Internet of things based sensor failure prediction method according to claim 1, characterized in that, The step of performing empirical mode decomposition on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor includes: For the timing signal of each sensor, the following screening process is performed: Identify the local extrema of the time-series signal; Based on the local extreme points, the upper and lower envelopes of the time series signal are constructed respectively, and the mean values ​​of the upper and lower envelopes are calculated to obtain the envelope mean curve. The candidate component is obtained by subtracting the envelope mean curve from the time-series signal; Check whether the candidate components satisfy the definition conditions of the intrinsic mode function, wherein the definition conditions include that the number of extreme points and the number of zero crossings differ by no more than one, and the mean of the upper envelope formed by local maxima and the lower envelope formed by local minima is close to zero. When the candidate component satisfies the defined condition, the candidate component is output as an intrinsic mode function component; The intrinsic mode function component is separated from the time-series signal, and the remaining signal is used as a new input signal to repeat the sieving process until the remaining signal becomes a monotonic function or only one extreme point remains. The remaining signal is then output as a residual term. 3.The industrial Internet of things based sensor failure prediction method according to claim 1, characterized in that, The step of constructing same-mode correlation vectors at equivalent time points based on the intrinsic mode function components with the same physical meaning, and combining them to form same-mode correlation matrices, includes: Identify the dominant frequency and energy distribution characteristics of the intrinsic mode function components of each of the isomorphic sensor instances; Cluster analysis is performed on the dominant frequency and the energy distribution characteristics to obtain multiple modal component mapping relationships, wherein the modal component mapping relationships are used to represent the mapping relationships between the intrinsic mode function components belonging to the same physical source in different sensor groups; Based on the modal component mapping relationship, the intrinsic mode function components from the same physical source are grouped into the same group to obtain multiple intrinsic mode function component groups; For each intrinsic mode function component group, based on a first preset time window, the same mode features are extracted for each intrinsic mode function component in the intrinsic mode function component group to obtain the corresponding same mode multidimensional features, wherein the same mode multidimensional features include at least one of time domain features, frequency domain features and nonlinear features; Time alignment is performed on multiple sets of time-series signals to construct the time-series mapping relationship of each set of sensors; Based on the aforementioned time-series mapping relationship, the equivalent time points of each group of sensors are determined; Based on the equivalent time point, the same modal multidimensional features within the same intrinsic modal function component group are combined to form the same modal correlation vector corresponding to each of the equivalent time points; According to the temporal order of the same-modal correlation vectors, the same-modal correlation vectors at each equivalent time point are arranged and combined to form a same-modal correlation matrix. 4.The industrial Internet of things based sensor failure prediction method according to claim 1, characterized in that, The step of constructing cross-modal correlation vectors at the same time point based on the intrinsic mode function components representing the core physical processes, and combining them to form a cross-modal correlation matrix, includes: The core physical processes of each heterogeneous sensor instance in the same group of sensors are determined, wherein the core physical processes include at least one of energy conversion processes, thermodynamic processes and mechanical vibration processes; For sensors in the same group, based on the coupling characteristics of the core physical process, the physical correlation between the intrinsic mode function components of each heterogeneous sensor instance is determined; Based on the physical correlation, at least one set of physically correlated modal components is determined from the intrinsic modal function components of each of the heterogeneous sensor instances, wherein the physically correlated modal components are composed of the intrinsic modal function components from each of the heterogeneous sensor instances in the same group of sensors that can characterize the same physical process; At the same time point, based on the second preset time window, cross-modal correlation features are extracted from the intrinsic mode function components of each heterogeneous sensor instance to obtain cross-modal multidimensional features, wherein the cross-modal multidimensional features include at least one of the following: cross-correlation features for reflecting amplitude correlation, synchronization features for reflecting phase relationship, cross-spectral features for reflecting energy transfer, and mutual information features for reflecting nonlinear dependence. For the same group of sensors, the cross-modal multidimensional features are combined according to a predefined sensor order to construct cross-modal correlation vectors at each time point; The cross-modal correlation vectors at each time point are arranged according to their temporal order and combined to form a cross-modal correlation matrix.

5. The sensor fault prediction method based on the Industrial Internet of Things according to claim 1, characterized in that, The step of generating fault prediction information based on the same-modal cooperative representation and the cross-modal cooperative representation includes: For the same group of sensors, when the same modal cooperative representation of any homogeneous sensor instance in the group exceeds the first threshold range, and the cross-modal cooperative representation corresponding to the group of sensors exceeds the second threshold range, the group of sensors is determined to be faulty, and fault prediction information containing group identifier and sensor identifier is generated.

6. A sensor failure prediction system based on industrial internet of things, characterized by, It includes a management platform, a sensor network platform, and an object platform that are connected in sequence. The management platform is configured with: The empirical mode decomposition module is used to acquire multiple sets of time-series signals obtained by sequential testing of the same industrial testing equipment by multiple sets of sensors, and to perform empirical mode decomposition on the time-series signals to obtain the intrinsic mode function components corresponding to each sensor. The test environment of each set of sensors is the same, and each set of sensors includes at least two of the following: temperature sensor, vibration sensor and sound sensor. The same-mode correlation module is used to construct same-mode correlation vectors at equivalent time points for isomorphic sensor instances in multiple sets of sensors, based on the intrinsic mode function components with the same physical meaning, and combine them to form a same-mode correlation matrix. The cross-modal correlation module is used to construct cross-modal correlation vectors at the same time point for heterogeneous sensor instances in the same group of sensors, based on the intrinsic mode function components representing the core physical processes, and combine them to form a cross-modal correlation matrix. The mode decomposition module is used to perform mode decomposition on each of the same-modal correlation matrices to obtain a first dominant mode representation; and to perform mode decomposition on each of the cross-modal correlation matrices to obtain a second dominant mode representation; The modal cooperative representation module is used to calculate the same-modal cooperative representation of each of the same-modal association vectors and the first dominant modal representation, and to calculate the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modal representation; The fault prediction information generation module is used to generate fault prediction information based on the same-modal cooperative representation and the cross-modal cooperative representation; The step of performing mode decomposition on each of the same-mode correlation matrices to obtain the representation of the first dominant mode includes: For each same-modal correlation matrix, a random projection matrix Ω is generated based on a preset target rank parameter; Calculate the product of the same-modal correlation matrix and the random projection matrix Ω to obtain the projection matrix Y; Perform QR decomposition on the projection matrix Y to obtain an orthogonal matrix Q and an upper triangular matrix R; Calculate the product of the transpose of the orthogonal matrix Q and the modal correlation matrix to obtain the intermediate matrix B; Singular value decomposition is performed on the intermediate matrix B to obtain the left singular vector matrix U_C1, the singular value matrix Σ_C1, and the right singular vector matrix V_C1; According to the preset variance contribution rate threshold, the top k largest singular values ​​are selected from the singular value matrix Σ_C1, and the top k right singular vectors are used as the first dominant mode representation, where k are positive integers determined according to the variance contribution rate threshold. The step of calculating the co-modal cooperative representation of each of the co-modal association vectors and the representation of the first dominant modality includes: For each modality association vector, calculate the first cosine similarity between the modality association vector and the first dominant modality representation, and calculate the first projection length of the modality association vector on the first dominant modality representation; The first cosine similarity and the first projection length are weighted and fused to obtain the same-modal collaborative representation; The step of calculating the cross-modal cooperative representation of each of the cross-modal association vectors and the second dominant modality representation includes: For each cross-modal association vector, calculate the second cosine similarity between the cross-modal association vector and the second dominant modality representation, and calculate the second projection length of the cross-modal association vector onto the second dominant modality representation; The second cosine similarity and the second projection length are weighted and fused to obtain a cross-modal collaborative representation.

7. A computer device, comprising: The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method of any one of claims 1 to 5.

Images