Sensor data reliability evaluation method and system based on industrial internet of things

Abstract

The application discloses a sensor data reliability evaluation method and system based on an industrial Internet of Things, relates to the technical field of the industrial Internet of Things, and comprises the following steps: performing multi-scale decomposition on a time sequence signal to obtain time sequence sub-signals, and giving the time sequence sub-signals corresponding physical meaning labels; for each type of physical meaning label, taking all time sequence sub-signals corresponding to the type of physical meaning label as analysis objects, for each sensor, taking the sensor as a target sensor, performing same-frequency band semantic mining on non-target time sequence sub-signals and target time sequence sub-signals to obtain corresponding same-frequency band semantic features; for each sensor, performing cross-frequency band semantic mining on low-frequency trend components, power frequency components and high-frequency event components to obtain cross-frequency band semantic features, and performing semantic correction on the cross-frequency band semantic features based on the same-frequency band semantic features to obtain semantic correction features. The application has the effect of physical interpretability of sensor data reliability evaluation.

Description

Technical Field

[0001] This application relates to the technical field of industrial Internet of Things (IIoT), and in particular to a method and system for assessing the reliability of sensor data based on IIoT. Background Technology

[0002] As a crucial component of smart manufacturing and industrial automation, the Industrial Internet of Things (IIoT) deploys numerous sensors to monitor and collect data on the operational status of physical systems in real time. The accuracy and reliability of sensor data directly impact the effectiveness of equipment condition monitoring, fault diagnosis, and predictive maintenance. However, in real-world industrial environments, sensors are often affected by factors such as electromagnetic interference, environmental noise, aging, or sudden malfunctions, resulting in abnormal or distorted components in the acquired time-series signals. This seriously threatens the reliability of data-driven decision-making.

[0003] Currently, the methods for evaluating the reliability of sensor data are mainly divided into two categories: one is anomaly detection based on statistical characteristics, such as calculating the mean, variance, and threshold judgment. These methods are simple and easy to implement, but they are difficult to capture dynamic changes under complex working conditions and lack interpretation of the physical meaning of the data; the other is verification methods based on multi-sensor data fusion, which make consistency judgments by comparing the measured values ​​of different sensors on the same physical quantity. However, these methods often assume that the sensors are independent and identically distributed, and fail to fully consider the physical correlation between different frequency band components within the same sensor. For example, low-frequency trends reflect slow system drift, power frequency components characterize periodic driving energy, and high-frequency events correspond to sudden impacts. There is an inherent energy conservation relationship among the three.

[0004] Currently, few methods can simultaneously achieve multi-sensor collaborative verification and single-sensor multi-band physical fusion, making the evaluation results susceptible to local anomalies and difficult to provide interpretable physical evidence for the evaluation conclusions. Therefore, there is an urgent need for a sensor data reliability evaluation method that can deeply integrate multi-source information and physical laws to improve the data quality assurance capabilities of industrial IoT systems. Summary of the Invention

[0005] To improve the physical interpretability of sensor data reliability assessment, this application provides a sensor data reliability assessment method and system based on the Industrial Internet of Things.

[0006] Firstly, this application provides a method for assessing the reliability of sensor data based on the Industrial Internet of Things, employing the following technical solution: A sensor data reliability assessment 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: The time-series signals of multiple sensors deployed on the target physical system are acquired, and the time-series signals are decomposed into multi-scale signals to obtain corresponding time-series sub-signals. Each time-series sub-signal is assigned a corresponding physical meaning label, wherein the physical meaning label includes low-frequency trend components, power frequency components and high-frequency event components. For each type of physical meaning label, all time-series sub-signals corresponding to that type of physical meaning label are taken as the analysis object. For each sensor, that sensor is taken as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, the target time-series sub-signals corresponding to the target sensors are subjected to co-band semantic mining to obtain the corresponding co-band semantic features. For each sensor, cross-band semantic mining is performed on the low-frequency trend component, the power frequency component, and the high-frequency event component corresponding to the sensor to obtain the corresponding cross-band semantic features. Based on the same-band semantic features corresponding to the sensor, the cross-band semantic features are semantically corrected to obtain semantically corrected features. Based on the semantic correction features, the data reliability assessment results of each sensor are obtained.

[0007] By adopting the above technical solution, time-series signals from multiple sensors deployed on the target physical system are acquired. These time-series signals are then decomposed into corresponding time-series sub-signals, each assigned a physical meaning label. These physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components. For each type of physical meaning label, all time-series sub-signals corresponding to that label are used as the analysis object. For each sensor, it is designated as the target sensor. Based on the non-target time-series sub-signals corresponding to non-target sensors, the target time-series sub-signals corresponding to the target sensor undergo co-band semantic mining to obtain corresponding co-band semantic features. Then, for each sensor, cross-band semantic mining is performed on the low-frequency trend components, power frequency components, and high-frequency event components corresponding to that sensor to obtain corresponding cross-band semantic features. Based on the co-band semantic features corresponding to that sensor, the cross-band semantic features are semantically corrected to obtain semantically corrected features. Finally, based on the semantically corrected features, the data reliability assessment results for each sensor are obtained. This invention enables the evaluation results to have clear physical interpretability by performing multi-scale decomposition on time-series signals and assigning them with physical meaning labels. Furthermore, through semantic mining in the same frequency band, semantic mining across frequency bands, and correction, it effectively suppresses the interference of single sensor anomalies on the evaluation results, and realizes information complementarity and fusion between different physical components. This improves the accuracy, robustness, and interpretability of sensor data reliability evaluation, and provides strong support for data quality assurance in the Industrial Internet of Things.

[0008] Optionally, the step of performing multi-scale decomposition on the time-series signal to obtain corresponding time-series sub-signals, and assigning a physical meaning label to each time-series sub-signal, wherein the physical meaning label includes low-frequency trend components, power frequency components, and high-frequency event components, includes: Based on the selected wavelet basis function and the number of decomposition levels, the time-series signal is decomposed into wavelet packet coefficients to obtain multiple wavelet packet coefficients. The wavelet packet coefficients are reconstructed to obtain the corresponding time-series sub-signals; Calculate the energy entropy of each of the time-series sub-signals, and based on the energy entropy, mark the time-series sub-signals that conform to the low-frequency trend characteristics as low-frequency trend components, mark the time-series sub-signals that conform to the power frequency and the harmonic frequency characteristics as power frequency components, and mark the time-series sub-signals that conform to the high-frequency abrupt change energy concentration characteristics as high-frequency event components.

[0009] By adopting the above technical solution, in order to achieve the decomposition of time series signals and the assignment of physical meaning labels, wavelet packet decomposition is performed on the time series signals based on the selected wavelet basis function and the number of decomposition levels to obtain multiple wavelet packet coefficients. Then, the wavelet packet coefficients are reconstructed to obtain the corresponding time series sub-signals. Then, the energy entropy of each time series sub-signal is calculated. Based on the energy entropy, time series sub-signals that conform to low-frequency trend characteristics are marked as low-frequency trend components, time series sub-signals that conform to power frequency and power frequency harmonics are marked as power frequency components, and time series sub-signals that conform to high-frequency abrupt energy concentration characteristics are marked as high-frequency event components.

[0010] Optionally, the step of performing co-frequency band semantic mining on the target time-series sub-signal corresponding to the target sensor based on the non-target time-series sub-signal corresponding to the non-target sensor to obtain the corresponding co-frequency band semantic features includes: The time-series sub-signals are subjected to temporal convolution processing to obtain temporal convolution features; Perform a Fourier transform on the time-series sub-signals to obtain a spectrum. The frequency domain convolution process is performed on the spectrum to obtain frequency domain convolution features; The time-domain convolutional features and the frequency-domain convolutional features are concatenated to obtain the same-frequency band primitive features corresponding to each sensor. Based on the non-target co-frequency band primitive features corresponding to each non-target sensor, attention processing is performed on the target co-frequency band primitive features corresponding to the target sensor to obtain the corresponding co-frequency band semantic features.

[0011] By adopting the above technical solution, in order to achieve semantic mining of the same frequency band, the time-series sub-signals are subjected to time-domain convolution processing to obtain time-domain convolution features. Then, the time-series sub-signals are subjected to Fourier transform to obtain a spectrum. The spectrum is then subjected to frequency-domain convolution processing to obtain frequency-domain convolution features. The time-domain convolution features and frequency-domain convolution features are then concatenated to obtain the same-frequency band primitive features corresponding to each sensor. Then, based on the non-target same-frequency band primitive features corresponding to each non-target sensor, attention processing is performed on the target same-frequency band primitive features corresponding to the target sensor to obtain the corresponding same-frequency band semantic features.

[0012] Optionally, the step of performing frequency domain convolution processing on the spectrogram to obtain frequency domain convolution features includes: The power frequency interference band in the spectrum is masked to obtain a masked spectrum. The spectrogram and the mask spectrogram are convolved respectively to obtain the corresponding global spectral features and mask spectral features; Multi-scale pooling compression is performed on the global spectral features and the mask spectral features respectively to obtain multi-scale global compressed features and multi-scale mask compressed features; For each scale of global compressed features, attention weights are assigned to the global compressed features based on the mask compressed features of that scale to obtain the attention spectrum features of that scale. Attention spectral features at various scales are fused to obtain attention spectral fusion features; Based on a preset spectrum gating matrix, the attention spectrum fusion features are gating mapped to obtain spectrum gating parameters; Based on the aforementioned spectral gating parameters, the global spectral features are multiplied element-wise to obtain frequency domain convolutional features.

[0013] By adopting the above technical solution, in order to achieve frequency domain convolution processing of the spectrum map, the power frequency interference band in the spectrum map is masked to obtain a mask spectrum map. Then, the spectrum map and the mask spectrum map are convolved to obtain the corresponding global spectrum features and mask spectrum features. Then, the global spectrum features and the mask spectrum features are compressed by multi-scale pooling to obtain multi-scale global compressed features and multi-scale mask compressed features. Then, for each scale of global compressed features, attention weights are assigned to the global compressed features based on the mask compressed features of that scale to obtain the attention spectrum features of that scale. Then, the attention spectrum features of each scale are fused to obtain the attention spectrum fusion features. Then, based on a preset spectrum gating matrix, the attention spectrum fusion features are gated and mapped to obtain spectrum gating parameters. Then, based on the spectrum gating parameters, the global spectrum features are multiplied element-wise to obtain frequency domain convolution features.

[0014] Optionally, the step of performing attention processing on the target band primitive features corresponding to the target sensor based on the non-target band primitive features corresponding to each non-target sensor to obtain the corresponding band semantic features includes: For each non-target sensor, the similarity between the target co-band primitive feature and the non-target co-band primitive feature corresponding to the non-target sensor is calculated to obtain the attention score corresponding to the non-target sensor. Based on the attention score corresponding to the non-target sensor, the attention weight corresponding to the non-target sensor is calculated by a normalized exponential function; For each non-target sensor, the non-target co-band primitive features corresponding to the non-target sensor are used as key features and value features, the target co-band primitive features are used as query features, and attention calculation is performed to obtain the attention interaction matrix corresponding to the non-target sensor. Based on the attention weights, the attention interaction matrix corresponding to the non-target sensor is weighted and summed to obtain the same-frequency band semantic features.

[0015] By adopting the above technical solution, in order to obtain the same-frequency band semantic features, for each non-target sensor, the similarity between the target same-frequency band primitive features and the corresponding non-target same-frequency band primitive features of the non-target sensor is calculated to obtain the attention score of the non-target sensor. Then, based on the attention score of the non-target sensor, the attention weight of the non-target sensor is calculated through a normalized exponential function. Then, for each non-target sensor, the corresponding non-target same-frequency band primitive features are used as key features and value features, and the target same-frequency band primitive features are used as query features. Attention is calculated to obtain the attention interaction matrix of the non-target sensor. Then, based on the attention weight, the attention interaction matrix of the non-target sensor is weighted and summed to obtain the same-frequency band semantic features.

[0016] Optionally, the step of performing cross-band semantic mining on the low-frequency trend component, the power frequency component, and the high-frequency event component corresponding to the sensor to obtain the corresponding cross-band semantic features includes: The low-frequency trend component, the power frequency component, and the high-frequency event component are respectively mapped to a high-dimensional feature space to obtain the corresponding low-frequency trend embedding feature, power frequency embedding feature, and high-frequency event embedding feature; Using the power frequency embedding feature as the query criterion, attention interaction is performed on the low-frequency trend embedding feature and the high-frequency event embedding feature respectively to obtain the trend power frequency correlation feature and the event power frequency correlation feature; The trend power frequency correlation features and the event power frequency correlation features are concatenated, and the concatenated features are subjected to nonlinear transformation to obtain cross-band fusion features; The cross-band fusion features are calibrated under energy conservation constraints to form cross-band semantic features.

[0017] By adopting the above technical solution, in order to achieve cross-band semantic mining, the low-frequency trend component, power frequency component, and high-frequency event component are mapped to a high-dimensional feature space to obtain the corresponding low-frequency trend embedding features, power frequency embedding features, and high-frequency event embedding features. Then, using the power frequency embedding features as the query benchmark, attention interaction is performed on the low-frequency trend embedding features and the high-frequency event embedding features to obtain trend power frequency correlation features and event power frequency correlation features. Then, the trend power frequency correlation features and event power frequency correlation features are concatenated, and the concatenated features are subjected to nonlinear transformation to obtain cross-band fusion features. Finally, the cross-band fusion features are calibrated under the constraint of energy conservation to form cross-band semantic features.

[0018] Optionally, the step of performing feature calibration under energy conservation constraints on the cross-band fused features to form cross-band semantic features includes: Extract the energy attenuation rate of the low-frequency trend component, the energy stability of the power frequency component, and the energy impact of the high-frequency event component; Based on the energy decay rate, the energy stability, and the energy impact, calculate the mechanical energy conservation deviation index; Based on the mechanical energy conservation deviation index, the cross-band fusion features are weighted and adjusted so that the adjusted features satisfy the preset energy conservation relationship, thereby obtaining cross-band semantic features.

[0019] By adopting the above technical solution, in order to form cross-band semantic features, the energy attenuation rate of the low-frequency trend component, the energy stability of the power frequency component, and the energy impact of the high-frequency event component are extracted. Then, based on the energy attenuation rate, energy stability, and energy impact, the mechanical energy conservation deviation index is calculated. Then, according to the mechanical energy conservation deviation index, the cross-band fusion features are weighted and adjusted so that the adjusted features satisfy the preset energy conservation relationship, thereby obtaining the cross-band semantic features.

[0020] Optionally, the step of semantically correcting the cross-band semantic features based on the same-band semantic features corresponding to the sensor to obtain semantically corrected features includes: Global average pooling is performed on the semantic features of the same frequency band to obtain a global same frequency band descriptor; Based on the global same-frequency band descriptor, modified weights are generated for each channel of the cross-frequency band semantic features; The corrected weights are multiplied channel by channel by channel with the cross-band semantic features to obtain the recalibrated semantic features; The recalibrated semantic features are residually concatenated with the same frequency band semantic features to form semantic correction features.

[0021] By adopting the above technical solution, in order to obtain semantic correction features, global average pooling is performed on the same-frequency band semantic features to obtain a global same-frequency band descriptor. Then, based on the global same-frequency band descriptor, correction weights for each channel of the cross-frequency band semantic features are generated. The correction weights are then multiplied with the cross-frequency band semantic features channel by channel to obtain recalibrated semantic features. Finally, the recalibrated semantic features are residually connected with the same-frequency band semantic features to form semantic correction features.

[0022] Optionally, the step of generating corrected weights for each channel of the cross-band semantic features based on the global same-band descriptor includes: The global same-frequency band descriptor is input into the first fully connected layer for dimensionality reduction to obtain a compressed descriptor; The compressed descriptor is input into the second fully connected layer for dimensionality increase, and then mapped to a range of 0 to 1 using the Sigmoid activation function to obtain the corrected weights for each channel of the cross-band semantic features.

[0023] By adopting the above technical solution, in order to generate corrected weights, the global same-frequency band descriptor is input into the first fully connected layer for dimensionality reduction to obtain a compressed descriptor. Then, the compressed descriptor is input into the second fully connected layer for dimensionality increase, and mapped to between 0 and 1 through the Sigmoid activation function to obtain corrected weights for each channel of cross-frequency band semantic features.

[0024] Secondly, this application also provides a sensor data reliability assessment system based on the Industrial Internet of Things, which adopts the following technical solution: A sensor data reliability assessment 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: The decomposition module is used to acquire time-series signals from multiple sensors deployed on the target physical system, perform multi-scale decomposition on the time-series signals to obtain corresponding time-series sub-signals, and assign physical meaning labels to each time-series sub-signal. The physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components. The same-frequency band semantic mining module is used to take all the time-series sub-signals corresponding to each type of physical meaning label as the analysis object, and take each sensor as the target sensor as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, the same-frequency band semantic mining is performed on the target time-series sub-signals corresponding to the target sensors to obtain the corresponding same-frequency band semantic features. The cross-band semantic mining and correction module is used to perform cross-band semantic mining on the low-frequency trend component, the power frequency component and the high-frequency event component corresponding to each sensor to obtain the corresponding cross-band semantic features, and to perform semantic correction on the cross-band semantic features based on the same-band semantic features corresponding to the sensor to obtain semantically corrected features. The reliability assessment module is used to obtain the data reliability assessment results of each sensor based on the semantic correction features.

[0025] In summary, this application includes at least the following beneficial technical effects: acquiring time-series signals from multiple sensors deployed on a target physical system; performing multi-scale decomposition on the time-series signals to obtain corresponding time-series sub-signals; assigning physical meaning labels to each time-series sub-signal, wherein the physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components; then, for each type of physical meaning label, taking all time-series sub-signals corresponding to that type of physical meaning label as the analysis object; for each sensor, taking that sensor as the target sensor; based on the non-target time-series sub-signals corresponding to the non-target sensors, performing same-frequency band semantic mining on the target time-series sub-signals corresponding to the target sensors to obtain corresponding same-frequency band semantic features; then, for each sensor, performing cross-frequency band semantic mining on the low-frequency trend components, power frequency components, and high-frequency event components corresponding to that sensor to obtain corresponding cross-frequency band semantic features; and based on the same-frequency band semantic features corresponding to that sensor, performing semantic correction on the cross-frequency band semantic features to obtain semantically corrected features; and finally, based on the semantically corrected features, obtaining the data reliability assessment results for each sensor. This invention enables the evaluation results to have clear physical interpretability by performing multi-scale decomposition on time-series signals and assigning them with physical meaning labels. Furthermore, through semantic mining in the same frequency band, semantic mining across frequency bands, and correction, it effectively suppresses the interference of single sensor anomalies on the evaluation results, and realizes information complementarity and fusion between different physical components. This improves the accuracy, robustness, and interpretability of sensor data reliability evaluation, and provides strong support for data quality assurance in the Industrial Internet of Things. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of an embodiment of this application.

[0027] Figure 2This is a structural diagram of one application scenario of the system in this application embodiment.

[0028] Figure 3 This is a structural diagram of another application scenario of the system according to an embodiment of this application.

[0029] Figure 4 This is a structural block diagram of the computer device described in this application. Detailed Implementation

[0030] 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.

[0031] This application discloses a method for evaluating the reliability of sensor data based on the Industrial Internet of Things.

[0032] Reference Figure 1 A sensor data reliability assessment 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: Step S11: Obtain time-series signals from multiple sensors deployed on the target physical system, perform multi-scale decomposition on the time-series signals to obtain corresponding time-series sub-signals, and assign physical meaning labels to each time-series sub-signal. The physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components.

[0033] It should be noted that step S11 aims to physically separate the raw sensor signals according to frequency scales. Specifically, time-series signals acquired by industrial sensors are often a mixture of multiple physical phenomena, such as long-term system drift, periodic driving energy, and sudden impact events often superimposed. Through multi-scale decomposition (such as wavelet packet decomposition), the signal can be broken down into different frequency band subspaces, allowing subsequent analysis to be performed on their respective independent physical dimensions. Assigning physical meaning labels such as low-frequency trend components, power frequency components, and high-frequency event components ensures that subsequent semantic mining is based on a clear physical context: low-frequency trends reflect slow changes in the system (such as wear and temperature drift), power frequency components characterize stable periodic excitations (such as motor rotation), and high-frequency event components correspond to transient impacts (such as collisions and vibrations). This step provides a physically interpretable data foundation for subsequent multi-perspective analysis.

[0034] Step S12: For each type of physical meaning label, take all the time-series sub-signals corresponding to that type of physical meaning label as the analysis object. For each sensor, take that sensor as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, perform co-frequency band semantic mining on the target time-series sub-signals corresponding to the target sensors to obtain the corresponding co-frequency band semantic features.

[0035] It's important to note that the core idea of ​​this step is semantic comparison among multiple sensors. Specifically, under the same physical meaning label (e.g., both being power frequency components), different sensors should be measuring the same physical phenomenon, thus exhibiting consistency. One sensor is designated as the target sensor, and signals from other sensors (non-target sensors) in the same frequency band are used as references. Attention mechanisms and other methods are employed to uncover common semantic features between the target sensor signal and the reference signal. The resulting semantic features within the same frequency band represent the "consensus" information consistent with other sensors in the current frequency band, while isolated or anomalous individual biases are suppressed. This mechanism effectively identifies anomalies caused by individual sensor malfunctions while retaining valid information reflecting the overall system state.

[0036] Step S13: For each sensor, perform cross-band semantic mining on the low-frequency trend component, power frequency component and high-frequency event component corresponding to the sensor to obtain the corresponding cross-band semantic features. Based on the same-band semantic features corresponding to the sensor, perform semantic correction on the cross-band semantic features to obtain semantically corrected features.

[0037] It should be noted that this step involves two progressive operations. First, cross-band semantic mining is performed on different physical components (low-frequency trends, power frequency, and high-frequency events) within the same sensor. The aim is to reveal the physical relationships between these components, such as how the energy driven by the power frequency is transformed into trend changes or sudden impacts in the system. This reflects physical laws such as the conservation of mechanical energy. Cross-band semantic features are formed through attention interaction and energy constraint calibration. Second, the cross-band semantic features are corrected using the same-band semantic features obtained in step S12. The physical meaning of this is: using "group consensus" to calibrate "individual physical fusion results." That is, if the internal energy relationship of a sensor in multiple frequency bands conforms to physical laws, but its overall consistency with other sensors is poor, its confidence level is reduced through correction. This step achieves a dual fusion of "horizontal consensus" and "vertical physical laws."

[0038] Step S14: Based on semantic correction features, obtain the data reliability assessment results of each sensor.

[0039] It should be noted that this step maps the semantically corrected features, after multi-sensor lateral verification and physical law longitudinal calibration, to specific reliability conclusions. The semantically corrected features have integrated consensus information from the same frequency band and physical energy relationship information across frequency bands, thus comprehensively characterizing the reliability of sensor data. Through classifiers or regression models, a reliability level (e.g., normal, requiring attention, faulty) or reliability score for each sensor can be output, providing a quantitative basis for data quality monitoring, sensor fault early warning, and subsequent decision analysis in industrial IoT systems. This assessment result is physically interpretable, facilitating on-site maintenance personnel to understand the source of anomalies and take appropriate measures.

[0040] In the above implementation, time-series signals from multiple sensors deployed on the target physical system are acquired. These time-series signals are then decomposed into corresponding time-series sub-signals, and each sub-signal is assigned a physical meaning label. These physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components. For each type of physical meaning label, all time-series sub-signals corresponding to that label are used as the analysis object. For each sensor, that sensor is designated as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, the target time-series sub-signals corresponding to the target sensor undergo co-band semantic mining to obtain corresponding co-band semantic features. Then, for each sensor, the low-frequency trend components, power frequency components, and high-frequency event components corresponding to that sensor undergo cross-band semantic mining to obtain corresponding cross-band semantic features. Based on the co-band semantic features corresponding to that sensor, the cross-band semantic features are semantically corrected to obtain semantically corrected features. Finally, based on the semantically corrected features, the data reliability assessment results for each sensor are obtained. This invention enables the evaluation results to have clear physical interpretability by performing multi-scale decomposition on time-series signals and assigning them with physical meaning labels. Furthermore, through semantic mining in the same frequency band, semantic mining across frequency bands, and correction, it effectively suppresses the interference of single sensor anomalies on the evaluation results, and realizes information complementarity and fusion between different physical components. This improves the accuracy, robustness, and interpretability of sensor data reliability evaluation, and provides strong support for data quality assurance in the Industrial Internet of Things.

[0041] As a further implementation of the method, the time-series signal is decomposed into multiple scales to obtain corresponding time-series sub-signals, and each time-series sub-signal is assigned a corresponding physical meaning label, wherein the physical meaning label includes low-frequency trend components, power frequency components, and high-frequency event components. This includes the following steps: Step S21: Based on the selected wavelet basis function and the number of decomposition levels, perform wavelet packet decomposition on the time series signal to obtain multiple wavelet packet coefficients.

[0042] Step S22: Reconstruct the wavelet packet coefficients to obtain the corresponding time-series sub-signals.

[0043] Step S23: Calculate the energy entropy of each time series sub-signal, and based on the energy entropy, mark the time series sub-signals that conform to the low-frequency trend characteristics as low-frequency trend components, mark the time series sub-signals that conform to the power frequency and power frequency harmonics as power frequency components, and mark the time series sub-signals that conform to the high-frequency abrupt change energy concentration characteristics as high-frequency event components.

[0044] In the above implementation, in order to decompose the time series signal and assign physical meaning labels, the time series signal is decomposed into wavelet packets based on the selected wavelet basis function and the number of decomposition levels to obtain multiple wavelet packet coefficients. Then, the wavelet packet coefficients are reconstructed to obtain the corresponding time series sub-signals. Then, the energy entropy of each time series sub-signal is calculated. Based on the energy entropy, the time series sub-signals that conform to the low-frequency trend characteristics are marked as low-frequency trend components, the time series sub-signals that conform to the power frequency and the harmonic frequency characteristics are marked as power frequency components, and the time series sub-signals that conform to the high-frequency abrupt change energy concentration characteristics are marked as high-frequency event components.

[0045] As a further implementation of the method, the step of performing co-band semantic mining on the target time-series sub-signal corresponding to the target sensor based on the non-target time-series sub-signal corresponding to the non-target sensor to obtain the corresponding co-band semantic features includes: Step S31: Perform temporal convolution processing on the temporal sub-signals to obtain temporal convolution features.

[0046] Step S32: Perform Fourier transform on the time-series sub-signals to obtain the spectrum.

[0047] Step S33: Perform frequency domain convolution processing on the spectrogram to obtain frequency domain convolution features.

[0048] Step S34: Concatenate the time-domain convolutional features and the frequency-domain convolutional features to obtain the same-frequency band primitive features corresponding to each sensor.

[0049] It should be noted that the same-frequency band primitive features in step S34 refer to the fused feature vector formed by concatenating the original features extracted from two complementary perspectives in the time and frequency domains under the same physical meaning label (such as both being power frequency components or both being high-frequency event components). Specifically, the time-domain convolutional features retain the local morphological information of the signal on the time axis (such as impact shape and oscillation mode), while the frequency-domain convolutional features characterize the energy distribution information of the signal on the frequency axis (such as peak frequency and harmonic structure). The primitive feature formed by concatenating the two contains both a time-domain description of "what the waveform looks like" and a frequency-domain description of "how the frequency components are distributed," together constituting a complete digital representation of the sensor's signal in that frequency band. This feature is called a "primitive" because it has not yet undergone correlation screening between multiple sensors; it is only a preliminary feature extraction result of the original signal from a single sensor, serving as the basic input unit for subsequent multi-sensor attention processing.

[0050] Step S35: Based on the non-target co-frequency band primitive features corresponding to each non-target sensor, attention processing is performed on the target co-frequency band primitive features corresponding to the target sensor to obtain the corresponding co-frequency band semantic features.

[0051] In the above implementation, in order to achieve co-band semantic mining, the time-series sub-signals are subjected to time-domain convolution processing to obtain time-domain convolution features. Then, the time-series sub-signals are subjected to Fourier transform to obtain a spectrum. The spectrum is then subjected to frequency-domain convolution processing to obtain frequency-domain convolution features. The time-domain convolution features and frequency-domain convolution features are then concatenated to obtain the co-band primitive features corresponding to each sensor. Then, based on the non-target co-band primitive features corresponding to each non-target sensor, attention processing is performed on the target co-band primitive features corresponding to the target sensor to obtain the corresponding co-band semantic features.

[0052] As a further implementation of the method, the step of performing frequency domain convolution processing on the spectrogram to obtain frequency domain convolution features includes: Step S41: Mask the power frequency interference band in the spectrum diagram to obtain the masked spectrum diagram.

[0053] Step S42: Perform convolution processing on the spectrogram and the mask spectrogram respectively to obtain the corresponding global spectral features and mask spectral features.

[0054] Step S43: Perform multi-scale pooling compression on the global spectral features and the mask spectral features respectively to obtain multi-scale global compressed features and multi-scale mask compressed features.

[0055] Step S44: For each scale of global compressed features, based on the mask compressed features of that scale, attention weights are assigned to the global compressed features to obtain the attention spectrum features of that scale.

[0056] Step S45: The attention spectrum features at each scale are fused to obtain the attention spectrum fusion features.

[0057] Step S46: Based on the preset spectrum gating matrix, perform gating mapping on the attention spectrum fusion features to obtain spectrum gating parameters.

[0058] Step S47: Based on the spectral gating parameters, the global spectral features are multiplied element-wise to obtain the frequency domain convolutional features.

[0059] It should be noted that the frequency domain convolution processing from steps S41 to S47 is fundamentally different from general frequency domain convolution. Specifically, general frequency domain convolution usually directly performs two-dimensional convolution operations on the spectrogram to extract local pattern features. Essentially, it treats the spectrogram as an ordinary image for feature learning. Steps S41 to S4, however, introduce masking and multi-scale attention mechanisms to achieve refined filtering and enhancement of the spectrogram. This is specifically reflected in the following aspects: First, by masking the power frequency interference bands in the spectrogram, a masked spectrogram is generated. Convolution is then performed on both the original and masked spectrograms to obtain global spectral features and masked spectral features. The purpose of this design is to separate the deterministic components such as the power frequency and its harmonics from other frequency band information, enabling the model to learn features from two perspectives: "the complete spectrum containing the power frequency" and "the pure spectrum after removing the power frequency," laying the foundation for subsequent comparative learning. Second, by performing multi-scale pooling compression on the global spectral features and masked spectral features, the following is obtained: The mechanism involves obtaining compressed features under different receptive fields and assigning attention weights based on the mask compressed features at each scale for the global compressed features. Essentially, this mechanism uses the mask features (information after removing the power frequency) as a "query" to guide the global features (containing all information) to focus on key frequency bands related to the mask features. In other words, through contrastive learning, the model automatically focuses on frequency band components outside the power frequency but with important physical significance, rather than passively processing all frequency bands. Third, the global spectral features are multiplied element-wise using spectral gating parameters to achieve soft filtering of frequency bands. Unlike general frequency domain convolution that directly outputs convolutional feature maps, this step first generates gating parameters that reflect the importance of each frequency band, and then uses these parameters to weight the original global spectral features. This enables the model to dynamically suppress redundant or interfering frequency bands and enhance key frequency bands, thereby obtaining more discriminative and robust frequency domain convolutional features. From step S41 to step S47, through mask comparison, multi-scale attention interaction and gating screening, the frequency domain convolution processing is upgraded from passive feature extraction to active frequency band selection and enhancement, which can more accurately capture the spectral features related to the reliability of sensor data.

[0060] In the above implementation, in order to achieve frequency domain convolution processing of the spectrogram, the power frequency interference band in the spectrogram is masked to obtain a masked spectrogram. Then, the spectrogram and the masked spectrogram are convolved to obtain the corresponding global spectral features and masked spectral features. Then, the global spectral features and the masked spectral features are compressed by multi-scale pooling to obtain multi-scale global compressed features and multi-scale mask compressed features. Then, for each scale of global compressed features, attention weights are assigned to the global compressed features based on the mask compressed features of that scale to obtain attention spectral features of that scale. Then, the attention spectral features of each scale are fused to obtain attention spectral fusion features. Then, based on a preset spectral gating matrix, the attention spectral fusion features are gated and mapped to obtain spectral gating parameters. Then, based on the spectral gating parameters, the global spectral features are multiplied element-wise to obtain frequency domain convolution features.

[0061] As a further implementation of the method, the step of performing attention processing on the target band semantic features corresponding to the target sensor based on the non-target band primitive features corresponding to each non-target sensor to obtain the corresponding band semantic features includes: Step S51: For each non-target sensor, calculate the similarity between the target co-band primitive features and the corresponding non-target co-band primitive features of the non-target sensor to obtain the attention score corresponding to the non-target sensor.

[0062] Step S52: Based on the attention scores corresponding to the non-target sensors, the attention weights corresponding to the non-target sensors are calculated using a normalized exponential function.

[0063] Step S53: For each non-target sensor, the non-target co-band primitive features corresponding to the non-target sensor are used as key features and value features, the target co-band primitive features are used as query features, and attention is calculated to obtain the attention interaction matrix corresponding to the non-target sensor.

[0064] Step S54: Based on the attention weights, the attention interaction matrices corresponding to the non-target sensors are weighted and summed to obtain the semantic features of the same frequency band.

[0065] In the above implementation, in order to obtain the same-frequency band semantic features, for each non-target sensor, the similarity between the target same-frequency band primitive features and the corresponding non-target same-frequency band primitive features of the non-target sensor is calculated to obtain the attention score of the non-target sensor. Then, based on the attention score of the non-target sensor, the attention weight of the non-target sensor is calculated through a normalized exponential function. Then, for each non-target sensor, the corresponding non-target same-frequency band primitive features are used as key features and value features, and the target same-frequency band primitive features are used as query features. Attention is calculated to obtain the attention interaction matrix of the non-target sensor. Then, based on the attention weight, the attention interaction matrix of the non-target sensor is weighted and summed to obtain the same-frequency band semantic features.

[0066] As a further implementation of the method, the step of performing cross-band semantic mining on the low-frequency trend component, power frequency component, and high-frequency event component corresponding to the sensor to obtain the corresponding cross-band semantic features includes: Step S61: Map the low-frequency trend component, power frequency component, and high-frequency event component to the high-dimensional feature space to obtain the corresponding low-frequency trend embedding feature, power frequency embedding feature, and high-frequency event embedding feature.

[0067] Step S62: Using the power frequency embedding feature as the query benchmark, attention interaction is performed on the low-frequency trend embedding feature and the high-frequency event embedding feature respectively to obtain the trend power frequency correlation feature and the event power frequency correlation feature.

[0068] Step S63: Concatenate the trend power frequency correlation features and the event power frequency correlation features, and perform a nonlinear transformation on the concatenated features to obtain cross-band fusion features.

[0069] Step S64: Perform feature calibration under energy conservation constraints on the cross-band fusion features to form cross-band semantic features.

[0070] It should be noted that the core of steps S61 to S64 lies in the introduction of a feature calibration mechanism under the constraint of energy conservation. This upgrades cross-band semantic mining from simple data-driven fusion to interpretable fusion guided by physical laws. Compared with conventional cross-band fusion methods, steps S61 to S64 have the following significant differences: First, conventional cross-band fusion usually involves splicing, adding, or performing simple attention interactions on features from different frequency bands. The resulting fused features only reflect the statistical correlation between data and lack a deep understanding of the physical world. In contrast, this step, after completing the initial feature concatenation and nonlinear transformation, further introduces the constraint of energy conservation, that is, based on the principle of conservation of mechanical energy, requiring... The system maintains a physical balance between its power frequency drive energy, low-frequency trend dissipation energy, and high-frequency event impact energy. This constraint imbues the fused features with a physical meaning of energy conservation, rather than merely a numerical combination. Secondly, this step introduces a mechanical energy conservation deviation index as the basis for feature calibration, enabling quantitative control over the degree of deviation between the data and physical laws. This mechanism ensures that the final cross-band semantic features retain the effective information from the original data while actively aligning with the physical prior of energy conservation, significantly improving the robustness and physical interpretability of the features. When sensor data appears statistically reasonable but contradicts physical laws, the calibration mechanism effectively suppresses the impact of such anomalies. Cross-band semantic features calibrated with energy conservation can more realistically reflect the inherent operating laws of the physical system, providing a feature foundation that conforms to both data distribution and physical constraints for subsequent reliability assessments.

[0071] In the above implementation, in order to achieve cross-band semantic mining, the low-frequency trend component, power frequency component, and high-frequency event component are mapped to a high-dimensional feature space to obtain the corresponding low-frequency trend embedding features, power frequency embedding features, and high-frequency event embedding features. Then, using the power frequency embedding features as the query benchmark, attention interaction is performed on the low-frequency trend embedding features and the high-frequency event embedding features to obtain trend power frequency correlation features and event power frequency correlation features. Then, the trend power frequency correlation features and event power frequency correlation features are concatenated, and the concatenated features are subjected to nonlinear transformation to obtain cross-band fusion features. Finally, the cross-band fusion features are calibrated under the constraint of energy conservation to form cross-band semantic features.

[0072] As a further implementation of the method, the step of performing feature calibration under energy conservation constraints on the cross-band fusion features to form cross-band semantic features includes: Step S71: Extract the energy attenuation rate of the low-frequency trend component, the energy stability of the power frequency component, and the energy impact of the high-frequency event component.

[0073] Step S72: Calculate the mechanical energy conservation deviation index based on the energy decay rate, energy stability, and energy impact.

[0074] Step S73: Based on the mechanical energy conservation deviation index, the cross-band fusion features are weighted and adjusted so that the adjusted features satisfy the preset energy conservation relationship, thereby obtaining the cross-band semantic features.

[0075] It should be noted that from steps S71 to S73, the calibration and optimization of the cross-band fusion characteristics are achieved by introducing physical laws. First, step S71 extracts three physical quantities from the original components: energy decay rate quantifies the dissipation speed of low-frequency trends, energy stability characterizes the fluctuation degree of power frequency drive, and energy impact measures the instantaneous intensity of high-frequency events. These three physical quantities together constitute a digital description of the system's energy state. Second, step S72 calculates the mechanical energy conservation deviation index based on these three physical quantities. This index quantitatively reflects the degree of deviation between the actual observed value and the ideal energy balance state, that is, the severity of the data violating physical laws. Finally, step S73 performs weighted adjustment on the cross-band fusion characteristics according to the deviation index, so that the adjusted characteristics actively approach the preset energy conservation relationship in terms of numerical relationship. This calibration is not a destructive modification of the features, but rather a soft correction guided by physical laws while preserving the original feature information. The calibration has almost no effect when the data conforms to physical laws, and only intervenes moderately when the data deviates significantly from the law of energy conservation. This results in cross-band semantic features that contain both the original information and conform to physical laws, providing a more realistic and robust feature foundation for subsequent reliability assessment.

[0076] In the above implementation, in order to form cross-band semantic features, the energy attenuation rate of the low-frequency trend component, the energy stability of the power frequency component, and the energy impact of the high-frequency event component are extracted. Then, based on the energy attenuation rate, energy stability, and energy impact, the mechanical energy conservation deviation index is calculated. Then, according to the mechanical energy conservation deviation index, the cross-band fusion features are weighted and adjusted so that the adjusted features satisfy the preset energy conservation relationship, thereby obtaining the cross-band semantic features.

[0077] As a further implementation of the method, the step of semantically correcting the cross-band semantic features based on the same-band semantic features corresponding to the sensor to obtain semantically corrected features includes: Step S81: Perform global average pooling on the semantic features of the same frequency band to obtain the global same frequency band descriptor.

[0078] Step S82: Based on the global same-band descriptor, generate corrected weights for each channel of cross-band semantic features.

[0079] Step S83: Multiply the corrected weights and cross-band semantic features channel by channel to obtain the recalibrated semantic features.

[0080] Step S84: Perform residual connection between the recalibrated semantic features and the semantic features of the same frequency band to form semantic correction features.

[0081] It should be noted that steps S81 to S84 introduce a channel attention mechanism to achieve soft correction of cross-band semantic features by semantic features in the same frequency band. Its core design lies in using multi-sensor consensus to guide the optimization of individual physical features. Specifically, step S81 performs global average pooling on the same-frequency band semantic features, compressing them into a global same-frequency band descriptor. This descriptor aggregates the "collective consensus" information that the sensor shares with other sensors in the same frequency band, eliminating local detail interference and forming a compact representation of the overall reliability of the sensor. Step S82 generates corrected weights for each channel of the cross-frequency band semantic features based on this descriptor. These weights reflect the degree of trust that the collective consensus places in different physical feature channels. A high weight is assigned when the physical phenomenon represented by a certain channel is consistent with the collective consensus, and vice versa. Step S83 multiplies the corrected weights with the cross-frequency band semantic features channel by channel to recalibrate the channel dimensions, enhancing the components in the cross-frequency band features that conform to the collective consensus and weakening the components that deviate from the consensus. Finally, step S84 performs a residual connection between the recalibrated features and the original same-frequency band semantic features, preserving the corrected cross-frequency band information while ensuring that the same-frequency band consensus information is not lost during the correction process, forming the final semantically corrected features. This mechanism, through "group consensus guiding individual corrections," enables feature representations to retain physical laws while possessing robustness through multi-sensor collaborative verification.

[0082] In the above implementation, in order to obtain semantic correction features, global average pooling is performed on the same-frequency band semantic features to obtain a global same-frequency band descriptor. Then, based on the global same-frequency band descriptor, correction weights for each channel of the cross-frequency band semantic features are generated. The correction weights are then multiplied with the cross-frequency band semantic features channel by channel to obtain recalibrated semantic features. Finally, the recalibrated semantic features are residually concatenated with the same-frequency band semantic features to form semantic correction features.

[0083] As a further implementation of the method, the step of generating corrected weights for each channel of cross-band semantic features based on the global co-band descriptor includes: Step S91: Input the global same-frequency band descriptor into the first fully connected layer for dimensionality reduction to obtain the compressed descriptor.

[0084] In step S92, the compressed descriptor is input into the second fully connected layer for dimensionality increase, and mapped to a range of 0 to 1 through the Sigmoid activation function to obtain the corrected weights for each channel of the cross-band semantic features.

[0085] In the above implementation, in order to generate corrected weights, the global same-band descriptor is input into the first fully connected layer for dimensionality reduction to obtain a compressed descriptor. Then, the compressed descriptor is input into the second fully connected layer for dimensionality increase and mapped to between 0 and 1 through the Sigmoid activation function to obtain corrected weights for each channel of cross-band semantic features.

[0086] This application also discloses a sensor data reliability assessment system based on the Industrial Internet of Things.

[0087] refer to Figure 2 A sensor data reliability assessment 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: The decomposition module is used to acquire time-series signals from multiple sensors deployed on the target physical system, perform multi-scale decomposition on the time-series signals to obtain corresponding time-series sub-signals, and assign physical meaning labels to each time-series sub-signal. The physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components. The same-frequency band semantic mining module is used to analyze all the time-series sub-signals corresponding to each type of physical meaning label, and to analyze each sensor as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, the same-frequency band semantic mining is performed on the target time-series sub-signals corresponding to the target sensors to obtain the corresponding same-frequency band semantic features. The cross-band semantic mining and correction module is used to perform cross-band semantic mining on the low-frequency trend component, power frequency component and high-frequency event component corresponding to each sensor to obtain the corresponding cross-band semantic features. Based on the same-band semantic features corresponding to the sensor, the cross-band semantic features are semantically corrected to obtain semantically corrected features. The reliability assessment module is used to obtain the data reliability assessment results of each sensor based on semantic correction features.

[0088] The overall framework of another application scenario of the sensor data reliability assessment 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 sensor network platform includes n sensor network sub-platforms, each with its own sensor sub-database. The service platform includes a main service database, n service sub-databases, and n service sub-platforms.

[0089] Specifically, in the aforementioned application scenario, the sensor data reliability assessment method based on the Industrial Internet of Things (IIoT) includes a management platform configured to: acquire time-series signals from multiple sensors deployed on a target physical system; decompose the time-series signals into multi-scale components to obtain corresponding time-series sub-signals; and assign physical meaning labels to each time-series sub-signal, whereby the physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components; for each type of physical meaning label, use all time-series sub-signals corresponding to that type of physical meaning label as the analysis object; for each sensor, use that sensor as the target sensor; based on the non-target time-series sub-signals corresponding to non-target sensors, perform same-frequency band semantic mining on the target time-series sub-signals corresponding to the target sensor to obtain corresponding same-frequency band semantic features; for each sensor, perform cross-frequency band semantic mining on the low-frequency trend components, power frequency components, and high-frequency event components corresponding to that sensor to obtain corresponding cross-frequency band semantic features; and based on the same-frequency band semantic features corresponding to that sensor, perform semantic correction on the cross-frequency band semantic features to obtain semantically corrected features; and based on the semantically corrected features, obtain the data reliability assessment results for each sensor.

[0090] By leveraging the interaction between the various functional platforms of the industrial IoT-based sensor data reliability assessment 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.

[0091] The sensor data reliability assessment system based on the Industrial Internet of Things (IIoT) of the present invention can implement any of the sensor data reliability assessment methods based on the Industrial Internet of Things, and the specific working process of the sensor data reliability assessment system based on the Industrial Internet of Things of the present invention can refer to the corresponding process in the above-mentioned sensor data reliability assessment methods based on the Industrial Internet of Things.

[0092] 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 assessing the reliability of sensor data based on the Industrial Internet of Things, characterized in that, 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 and communicatively connected. The method is executed by the management platform and includes: The time-series signals of multiple sensors deployed on the target physical system are acquired, and the time-series signals are decomposed into multi-scale signals to obtain corresponding time-series sub-signals. Each time-series sub-signal is assigned a corresponding physical meaning label, wherein the physical meaning label includes low-frequency trend components, power frequency components and high-frequency event components. For each type of physical meaning label, all time-series sub-signals corresponding to that type of physical meaning label are taken as the analysis object. For each sensor, that sensor is taken as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, the target time-series sub-signals corresponding to the target sensors are subjected to co-band semantic mining to obtain the corresponding co-band semantic features. For each sensor, cross-band semantic mining is performed on the low-frequency trend component, the power frequency component, and the high-frequency event component corresponding to the sensor to obtain the corresponding cross-band semantic features. Based on the same-band semantic features corresponding to the sensor, the cross-band semantic features are semantically corrected to obtain semantically corrected features. Based on the semantic correction features, the data reliability assessment results of each sensor are obtained.

2. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 1, characterized in that, The step of performing multi-scale decomposition on the time-series signal to obtain corresponding time-series sub-signals, and assigning a physical meaning label to each time-series sub-signal, wherein the physical meaning label includes low-frequency trend components, power frequency components, and high-frequency event components, includes: Based on the selected wavelet basis function and the number of decomposition levels, the time-series signal is decomposed into wavelet packet coefficients to obtain multiple wavelet packet coefficients. The wavelet packet coefficients are reconstructed to obtain the corresponding time-series sub-signals; Calculate the energy entropy of each of the time-series sub-signals, and based on the energy entropy, mark the time-series sub-signals that conform to the low-frequency trend characteristics as low-frequency trend components, mark the time-series sub-signals that conform to the power frequency and the harmonic frequency characteristics as power frequency components, and mark the time-series sub-signals that conform to the high-frequency abrupt change energy concentration characteristics as high-frequency event components.

3. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 1, characterized in that, The step of performing co-frequency band semantic mining on the target time-series sub-signal corresponding to the target sensor based on the non-target time-series sub-signal corresponding to the non-target sensor to obtain the corresponding co-frequency band semantic features includes: The time-series sub-signals are subjected to temporal convolution processing to obtain temporal convolution features; Perform a Fourier transform on the time-series sub-signals to obtain a spectrum. The frequency domain convolution process is performed on the spectrum to obtain frequency domain convolution features; The time-domain convolutional features and the frequency-domain convolutional features are concatenated to obtain the same-frequency band primitive features corresponding to each sensor. Based on the non-target co-frequency band primitive features corresponding to each non-target sensor, attention processing is performed on the target co-frequency band primitive features corresponding to the target sensor to obtain the corresponding co-frequency band semantic features.

4. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 3, characterized in that, The step of performing frequency domain convolution processing on the spectrogram to obtain frequency domain convolution features includes: The power frequency interference band in the spectrum is masked to obtain a masked spectrum. The spectrogram and the mask spectrogram are convolved respectively to obtain the corresponding global spectral features and mask spectral features; Multi-scale pooling compression is performed on the global spectral features and the mask spectral features respectively to obtain multi-scale global compressed features and multi-scale mask compressed features; For each scale of global compressed features, attention weights are assigned to the global compressed features based on the mask compressed features of that scale to obtain the attention spectrum features of that scale. Attention spectral features at various scales are fused to obtain attention spectral fusion features; Based on a preset spectrum gating matrix, the attention spectrum fusion features are gating mapped to obtain spectrum gating parameters; Based on the aforementioned spectral gating parameters, the global spectral features are multiplied element-wise to obtain frequency domain convolutional features.

5. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 3, characterized in that, The step of performing attention processing on the target band primitive features corresponding to the target sensor based on the non-target band primitive features corresponding to each non-target sensor to obtain the corresponding band semantic features includes: For each non-target sensor, the similarity between the target co-band primitive feature and the non-target co-band primitive feature corresponding to the non-target sensor is calculated to obtain the attention score corresponding to the non-target sensor. Based on the attention score corresponding to the non-target sensor, the attention weight corresponding to the non-target sensor is calculated by a normalized exponential function; For each non-target sensor, the non-target co-band primitive features corresponding to the non-target sensor are used as key features and value features, the target co-band primitive features are used as query features, and attention calculation is performed to obtain the attention interaction matrix corresponding to the non-target sensor. Based on the attention weights, the attention interaction matrix corresponding to the non-target sensor is weighted and summed to obtain the same-frequency band semantic features.

6. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 1, characterized in that, The step of performing cross-band semantic mining on the low-frequency trend component, the power frequency component, and the high-frequency event component corresponding to the sensor to obtain the corresponding cross-band semantic features includes: The low-frequency trend component, the power frequency component, and the high-frequency event component are respectively mapped to a high-dimensional feature space to obtain the corresponding low-frequency trend embedding feature, power frequency embedding feature, and high-frequency event embedding feature; Using the power frequency embedding feature as the query criterion, attention interaction is performed on the low-frequency trend embedding feature and the high-frequency event embedding feature respectively to obtain the trend power frequency correlation feature and the event power frequency correlation feature; The trend power frequency correlation features and the event power frequency correlation features are concatenated, and the concatenated features are subjected to nonlinear transformation to obtain cross-band fusion features; The cross-band fusion features are calibrated under energy conservation constraints to form cross-band semantic features.

7. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 6, characterized in that, The step of performing feature calibration under energy conservation constraints on the cross-band fused features to form cross-band semantic features includes: Extract the energy attenuation rate of the low-frequency trend component, the energy stability of the power frequency component, and the energy impact of the high-frequency event component; Based on the energy decay rate, the energy stability, and the energy impact, calculate the mechanical energy conservation deviation index; Based on the mechanical energy conservation deviation index, the cross-band fusion features are weighted and adjusted so that the adjusted features satisfy the preset energy conservation relationship, thereby obtaining cross-band semantic features.

8. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 1, characterized in that, The step of semantically correcting the cross-band semantic features based on the same-band semantic features corresponding to the sensor to obtain semantically corrected features includes: Global average pooling is performed on the semantic features of the same frequency band to obtain a global same frequency band descriptor; Based on the global same-frequency band descriptor, modified weights are generated for each channel of the cross-frequency band semantic features; The corrected weights are multiplied channel by channel by channel with the cross-band semantic features to obtain the recalibrated semantic features; The recalibrated semantic features are residually concatenated with the same frequency band semantic features to form semantic correction features.

9. The sensor data reliability assessment method based on the Industrial Internet of Things according to claim 8, characterized in that, The step of generating corrected weights for each channel of the cross-band semantic features based on the global same-band descriptor includes: The global same-frequency band descriptor is input into the first fully connected layer for dimensionality reduction to obtain a compressed descriptor; The compressed descriptor is input into the second fully connected layer for dimensionality increase, and then mapped to a range of 0 to 1 using the Sigmoid activation function to obtain the corrected weights for each channel of the cross-band semantic features.

10. A sensor data reliability assessment system based on the Industrial Internet of Things, characterized in that, 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 decomposition module is used to acquire time-series signals from multiple sensors deployed on the target physical system, perform multi-scale decomposition on the time-series signals to obtain corresponding time-series sub-signals, and assign physical meaning labels to each time-series sub-signal. The physical meaning labels include low-frequency trend components, power frequency components, and high-frequency event components. The same-frequency band semantic mining module is used to take all the time-series sub-signals corresponding to each type of physical meaning label as the analysis object, and take each sensor as the target sensor as the target sensor. Based on the non-target time-series sub-signals corresponding to the non-target sensors, the same-frequency band semantic mining is performed on the target time-series sub-signals corresponding to the target sensors to obtain the corresponding same-frequency band semantic features. The cross-band semantic mining and correction module is used to perform cross-band semantic mining on the low-frequency trend component, the power frequency component and the high-frequency event component corresponding to each sensor to obtain the corresponding cross-band semantic features, and to perform semantic correction on the cross-band semantic features based on the same-band semantic features corresponding to the sensor to obtain semantically corrected features. The reliability assessment module is used to obtain the data reliability assessment results of each sensor based on the semantic correction features.

Images