Device surface interaction and state sensing system based on distributed piezoelectric sensors

Abstract

The application relates to the technical field of device sensing, and discloses a device surface interaction and state sensing system based on a distributed piezoelectric sensor. The system integrates a plurality of piezoelectric sensing units, a signal processing unit, an evaluation and analysis unit and a state output unit; the piezoelectric sensing units are distributed on the surface of the device, and collect piezoelectric signals generated by surface interaction; the signal processing unit pre-processes the piezoelectric signals, extracts signal features and generates a signal response curve; the evaluation and analysis unit calculates the state sensitivity, the initial state correlation degree, the real state correlation degree and the environmental interference degree of each piezoelectric sensing unit based on the signal response curve and the device state parameters, and screens key sensing positions; and the state output unit constructs a state sensing model based on the signal data of the key sensing positions, and realizes real-time sensing of device surface interaction and state. The system can realize comprehensive sensing of the surface of the device, improve sensing accuracy and efficiency, and adapt to the sensing requirements of devices in multiple scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of device perception, in particular to a device surface interaction and state perception system based on distributed piezoelectric sensors. BACKGROUND

[0002] In the scenarios of industrial device operation and maintenance, consumer electronic interaction control, etc., the capture of interaction behavior and the monitoring of the state of the device surface are important requirements for the intelligent operation of the device. Traditional device surface perception schemes rely on a single type of sensor, such as a capacitive sensor or a resistive sensor. Such sensors have obvious limitations in practical applications. The capacitive sensor is easily affected by environmental humidity and temperature changes. When the device surface is covered with oil or dust, the perception accuracy will be greatly reduced, and it is difficult to stably capture slight surface interaction actions. The resistive sensor needs to be in continuous contact to achieve signal acquisition, and cannot adapt to non-contact interaction scenarios. In addition, it is prone to contact wear after long-term use, resulting in a shortened service life.

[0003] With the increasing complexity of device functions, the comprehensiveness and real-time requirements of surface perception are gradually increasing. In the existing technology, some schemes attempt to use multiple sensor combinations to compensate for the shortcomings of a single sensor. However, the signal formats and transmission protocols of different types of sensors differ, making it difficult to achieve efficient data fusion, resulting in increased overall response delay of the perception system, which cannot meet the demand for real-time monitoring of device state. At the same time, traditional perception systems can only achieve simple recognition of surface interaction actions such as pressing and sliding, and cannot further associate with the running state parameters of the device itself, cannot judge the influence of the interaction action on the device state, and also cannot infer whether the device has potential fault risks through the interaction signal.

[0004] In industrial scenarios, the state monitoring of large mechanical device surfaces is particularly critical. The small deformation and vibration signals on the surface of the device are often related to the running state of the internal components. However, existing sensors are mostly concentrated on the internal key components of the device, and the monitoring coverage of the surface state is limited. In addition, the transmission signal is easily affected by mechanical vibration and environmental noise during device operation, resulting in the effective signal being submerged and making it difficult to accurately extract useful information. Furthermore, traditional perception systems lack an effective screening mechanism for sensing positions, and all sensing units participate in signal acquisition and processing, which not only increases the redundancy of data processing, but also affects the overall perception accuracy due to invalid signals from some non-critical positions, resulting in waste of system resources and further reduction of perception efficiency. These problems make the traditional device surface interaction and state perception scheme unable to meet the current demand for device intelligent development in terms of adaptability, accuracy and efficiency, and a new perception system is needed to break through the existing technical bottlenecks. SUMMARY

[0005] The present application aims to provide a device surface interaction and state sensing system based on distributed piezoelectric sensors to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a device surface interaction and state sensing system based on distributed piezoelectric sensors, which comprises:

[0007] a piezoelectric sensing unit, a signal processing unit, an evaluation and analysis unit, and a state output unit;

[0008] The piezoelectric sensing unit is distributed on the surface of the device for collecting piezoelectric signals generated by surface interaction; the signal processing unit pre-processes the piezoelectric signals, extracts signal features, and generates signal response curves; the evaluation and analysis unit calculates the state sensitivity, initial state correlation, true state correlation, and environmental interference of each piezoelectric sensing unit based on the signal response curves and device state parameters, and screens out key sensing positions; and the state output unit constructs a state sensing model based on the signal data of the key sensing positions and performs real-time sensing of device surface interaction and state.

[0009] Preferably, when calculating the state sensitivity, the evaluation and analysis unit performs the following operations: obtaining the signal response curve of the target piezoelectric sensing unit within a preset time window; determining a group of reference sensing units based on the similarity of device state parameters; extracting the peak response amplitude of all reference sensing units at the target signal position to form an amplitude sequence; calculating the dispersion degree index of the amplitude sequence and performing inverse proportional normalization to obtain a local consistency coefficient; simultaneously analyzing the fluctuation range of the device state parameters corresponding to the reference sensing units to generate a state stability metric; combining the statistical correlation of the local consistency coefficient sequence of all sensing units and the state stability metric sequence to generate a sensitivity adjustment factor; and multiplying the sensitivity adjustment factor and the local consistency coefficient to obtain the state sensitivity of the target sensing unit at the target signal position.

[0010] Preferably, when calculating the initial state correlation, the evaluation and analysis unit performs the following steps: collecting the signal intensity values of all sensing units at the target signal position to construct a signal intensity distribution histogram; performing probability density estimation on the histogram to obtain the probability distribution function of the signal intensity; simultaneously obtaining the device state parameter values of all sensing units to generate a state parameter distribution histogram and convert it into a probability distribution; using a multi-scale distribution matching algorithm to calculate the overall deviation degree between the two probability distributions; converting the deviation degree into a correlation index through a nonlinear mapping function, and taking the index as the initial state correlation of the target signal position.

[0011] Preferably, the evaluation analysis unit calculates the real state correlation degree by performing the following steps: assigning an independent identity identifier to each sensing unit; arranging the identity identifiers in descending order of the device state parameter values to form a state priority sequence; arranging the identity identifiers in ascending order of the signal strength values to generate a signal priority sequence; identifying the identity identifiers with the same position in the two priority sequences by a sequence comparison algorithm to calculate a basic matching degree; extracting the state sensitivity values corresponding to the remaining identity identifiers after removing the matched identity identifiers to construct a sensitivity distribution vector; calculating the spatial similarity measure of the two sensitivity distribution vectors; fusing the basic matching degree and the spatial similarity measure to generate a correlation degree correction coefficient; and combining the correlation degree correction coefficient with the initial state correlation degree to obtain the real state correlation degree.

[0012] Preferably, the evaluation analysis unit calculates the environmental interference degree by performing the following steps: obtaining the device state parameter distribution characteristics of the reference sensing unit group of the target sensing unit; generating the parameter distribution profile of the reference group by a kernel density estimation method; performing multi-resolution comparative analysis on the distribution profile and the global device state parameter distribution profile; calculating the overlapping area ratio of the two distribution profiles at different scales; generating a distribution matching degree based on the weighted average of the overlapping area ratios; and obtaining the environmental interference degree evaluation value of the target signal position by inversely converting and normalizing the distribution matching degree.

[0013] Preferably, the evaluation analysis unit screens the key sensing positions by performing the following steps: establishing a ratio matrix of the real state correlation degree and the environmental interference degree of each signal position; performing adaptive normalization processing on the ratio matrix to generate the importance score of each position; setting a dynamic threshold mechanism that automatically adjusts the threshold boundary based on the statistical characteristics of the importance score; including the positions with importance scores exceeding the dynamic threshold in the key position set; and performing spatial clustering analysis on the key position set to form the final key sensing position group after removing isolated points.

[0014] Preferably, the state output unit constructs the state perception model by performing the following steps: extracting the multi-dimensional signal feature vector of the key sensing position from the historical data; combining the feature vector and the corresponding device state parameter to form a training sample set; selecting the most discriminative feature subset by a recursive feature elimination algorithm; training a regression model on the selected feature subset using a support vector regression algorithm; optimizing the model hyperparameters by cross-validation; and encapsulating the trained regression model and the feature selection rule together as the state perception model.

[0015] Preferably, when the state output unit performs state sensing, the following steps are performed: real-time acquisition of distributed sensing signals of the to-be-monitored equipment; quality detection and outlier filtering of the signals; extraction of multi-dimensional signal features at key sensing positions; dimension reduction of the feature vectors according to a preset feature selection rule; input of the reduced feature vectors into a state sensing model; acquisition of state estimation values and their confidence levels output by the model; and smoothing processing of the output results by combining a time series filtering algorithm.

[0016] Preferably, when the evaluation analysis unit determines the reference sensing unit group, the following steps are performed: establishment of a dynamic similarity threshold range based on the equipment state parameter value of the target sensing unit; calculation of the absolute difference degree of the equipment state parameters of other sensing units from the reference value; assignment of similarity weights according to the difference degree; selection of sensing units with weights exceeding a preset threshold to form an initial candidate set; density clustering of the initial candidate set, and selection of units in the largest cluster as the final reference sensing unit group.

[0017] Preferably, when the evaluation analysis unit calculates the local consistency coefficient, the following steps are performed: calculation of the coefficient of variation based on the amplitude sequence; acquisition of a peak amplitude stability index of a historical signal response curve at the target signal position; weighted fusion of the coefficient of variation and the stability index to generate a dynamic consistency weight; and adaptive correction of the local consistency coefficient using the dynamic consistency weight.

[0018] Compared with the prior art, the present application has the following advantages:

[0019] The device surface interaction and state sensing system based on distributed piezoelectric sensors integrates multiple piezoelectric sensing units, signal processing units, evaluation analysis units, and state output units to form a complete sensing link, which can effectively address the shortcomings of traditional sensing schemes in adaptability, accuracy, and efficiency. From the perspective of sensing coverage and signal acquisition, the piezoelectric sensing units are distributed on the surface of the equipment, which can achieve comprehensive coverage of different areas of the equipment surface. Compared with the traditional single sensor or local sensor arrangement, it can more completely capture the piezoelectric signals generated by surface interaction, whether it is a slight pressing action or a small vibration signal generated on the surface during equipment operation. It can be effectively acquired, avoiding the sensing blind area caused by insufficient sensing coverage, and is suitable for different structure and size of equipment surface sensing scenarios.

[0020] In the signal processing link, the signal processing unit pre-processes the piezoelectric signal, extracts signal features, and generates signal response curves, which can convert the original piezoelectric signal into more valuable feature information, remove invalid interference components in the signal, and provide high-quality signal basis for subsequent evaluation and analysis. Unlike traditional processing methods that only perform simple filtering, this unit uses feature extraction and response curve generation to make signal information more intuitive and easier to associate with device state parameters, reduce data redundancy in the subsequent analysis process, and improve overall processing efficiency.

[0021] The evaluation and analysis unit calculates the state sensitivity, initial state correlation, real state correlation, and environmental interference based on the signal response curve and the device state parameter, and selects the key sensing position. This process can accurately identify the sensing effectiveness of different sensing units. By calculating various correlation and interference indicators, effective signals and environmental interference signals can be distinguished, and the influence of environmental factors on the sensing result can be avoided. At the same time, the key sensing position is selected, only the position data that has an important effect on the sensing result is retained, the non-key position signal occupies the system resources is reduced, the data processing pressure is reduced, the system can focus on the core sensing task more efficiently, compared with the traditional system that processes all sensing data without distinction, the sensing accuracy and efficiency are greatly improved.

[0022] The state output unit constructs a state sensing model based on the signal data of the key sensing position and performs real-time sensing. On the one hand, by constructing a model based on key position data, the effectiveness and relevance of model input data can be ensured, making the model more accurately reflect the relationship between device surface interaction and state, and avoiding model bias caused by invalid data. On the other hand, the real-time sensing function can capture device surface interaction behavior and state changes in a timely manner, compared with the response delay problem existing in traditional systems, it can quickly feedback the current state of the device, and is suitable for scenarios with high real-time requirements, such as real-time operation and maintenance of industrial equipment, instant interaction control of consumer electronics, etc. Overall, through the cooperative work of each unit, the system realizes the comprehensiveness, accuracy and real-time of device surface interaction and state sensing, can adapt to the sensing needs of different environments and different types of devices, and has wide application scenario adaptation ability. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The working principle diagram of the device surface interaction and state sensing system based on distributed piezoelectric sensors described in the present application;

[0024] Figure 2 The flowchart for the evaluation and analysis unit to calculate the state sensitivity operation;

[0025] Figure 3 The flowchart for the evaluation and analysis unit to calculate the initial state correlation step. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0027] Please refer to Figure 1 The present application provides a device surface interaction and state sensing system based on distributed piezoelectric sensors, which includes a signal processing unit for preprocessing piezoelectric signals, a signal processing unit for extracting signal features and generating signal response curves. An evaluation and analysis unit based on signal response curves and device state parameters, the evaluation and analysis unit calculates the state sensitivity, initial state correlation, real state correlation and environmental interference of each piezoelectric sensing unit, and the evaluation and analysis unit screens out the key sensing position. A state output unit constructs a state sensing model based on the signal data of the key sensing position, and the state output unit realizes real-time sensing of the device surface interaction and state. The system realizes comprehensive coverage through distributed layout, the signal processing unit enhances signal quality by using filtering and amplification technology, the evaluation and analysis unit uses statistical and machine learning methods for deep analysis, and the state output unit integrates regression algorithm and time series processing to realize high-precision sensing.

[0028] Embodiment 1: refer to Figure 2, a set of reference sensing units are determined based on the similarity of the device state parameters, the selection criteria of the reference sensing units focus on the proximity of the device state parameters in the numerical distribution, and the device state parameters include temperature, pressure, vibration amplitude and various physical quantities. Extract the peak value response amplitude of all reference sensing units at the target signal position to form an amplitude sequence, the detection of the peak value response amplitude applies an envelope detection algorithm to eliminate high-frequency noise interference, and the construction of the amplitude sequence arranges the discrete data points in chronological order to form a discrete data point set. Calculate the discrete degree index of the amplitude sequence and perform inverse proportional normalization on the discrete degree index to obtain the local consistency coefficient, specifically, convert the discrete degree index to the local consistency coefficient, and use the reciprocal relationship for mapping when implementing the function. The function input is the discrete degree index value, and the output is the normalized coefficient value, and the function expression is that the local consistency coefficient is equal to one divided by one plus the discrete degree index. The inverse proportional normalization function ensures that the output value range is between zero and one, the larger the discrete degree index value, the smaller the local consistency coefficient, and the function mapping relationship is monotonically decreasing. The calculation of the discrete degree index uses the standard deviation method to reflect the data volatility, the inverse proportional normalization processing maps the standard deviation to the range of zero to one, and the value of the local consistency coefficient is negatively correlated with the discrete degree. Analyze the fluctuation range of the device state parameters corresponding to the reference sensing units to generate a state stability metric, the calculation of the fluctuation range is based on the difference between the maximum and minimum values of the device state parameters within the same period, and the generation of the state stability metric introduces the moving average processing of the historical data to smooth the instantaneous mutation. Combine the statistical correlation of the local consistency coefficient sequence and the state stability metric sequence of all sensing units to generate a sensitivity adjustment factor, the calculation of the statistical correlation uses the Pearson correlation coefficient to capture the linear correlation strength between the two sequences, and the value of the sensitivity adjustment factor is amplified by an exponential function. Multiply the sensitivity adjustment factor and the local consistency coefficient to obtain the state sensitivity of the target sensing unit at the target signal position, the multiplication operation integrates the local consistency and the global adjustment effect, and the higher the output value of the state sensitivity, the stronger the response ability of the piezoelectric sensing unit to the device state change.

[0029] The evaluation analysis unit calculates the coefficient of variation of the amplitude sequence when calculating the local consistency coefficient, and the coefficient of variation is calculated by using the ratio of the standard deviation to the average value to eliminate the dimension effect, and the numerical value of the coefficient of variation reflects the relative dispersion degree of the amplitude sequence. The peak amplitude stability index of the historical signal response curve at the target signal position is obtained, the data of the historical signal response curve comes from the long-term stored database, and the calculation of the peak amplitude stability index is based on the variance statistics in the sliding window, and the smaller the value of the stability index, the more gentle the historical amplitude fluctuation. The dynamic consistency weight is generated by weighting and fusing the coefficient of variation and the stability index, and the weight coefficient of the weighting and fusing is dynamically allocated according to the signal-to-noise ratio characteristics, and the coefficient of variation occupies the dominant weight under the condition of high signal-to-noise ratio, and the generation of the dynamic consistency weight enhances the quantification of the reliability of the current data. The local consistency coefficient is adaptively corrected by using the dynamic consistency weight, the adaptive correction process multiplies the local consistency coefficient and the dynamic consistency weight, and the corrected local consistency coefficient more accurately reflects the consistency level of the actual signal.

[0030] The evaluation analysis unit determines the dynamic similarity threshold range based on the device state parameter value of the target sensing unit when determining the reference sensing unit group, and the lower limit and the upper limit of the dynamic similarity threshold range are dynamically calculated according to the historical distribution quartile of the device state parameter, and the threshold range is adaptively adjusted according to the device operating state. The absolute difference degree of the device state parameters of other sensing units and the reference value is calculated, the absolute difference degree is directly quantified by using the absolute value distance, and the smaller the numerical value of the absolute difference degree, the closer the device state parameters of the sensing unit and the target unit. The similarity weight is allocated according to the difference degree, the similarity weight is calculated by using an inverse proportional function to give a higher weight value to a small difference degree, and the weight allocation curve presents a monotonically decreasing characteristic. The sensing units with a weight exceeding a preset threshold are selected to form an initial candidate set, the value of the preset threshold is determined by cluster analysis to determine the best critical point, and the size of the initial candidate set is affected by the distribution density of the device state parameters. The initial candidate set is subjected to density clustering, the DBSCAN algorithm is applied to identify core points and boundary points based on the spatial density of the device state parameters, and the parameters of the DBSCAN algorithm are optimized according to the data distribution characteristics.

[0031] The collection of signal response curves adopts a sliding window mechanism to ensure data timeliness, the step and width of the sliding window are matched according to signal characteristics, the window width covers enough signal periods, and the step is set to balance the calculation efficiency and real-time requirements. The extraction of peak response amplitude applies an envelope detection algorithm to eliminate noise interference, the envelope detection uses a Hilbert transform method to obtain a signal amplitude envelope line, and the positioning of peak points is realized through local maximum value search. The calculation of the dispersion degree index uses a standard deviation method to reflect data distribution, the calculation of the standard deviation is based on the sample unbiased estimation formula, and the greater the value of the dispersion degree index, the worse the response consistency between the reference sensing units. The inverse proportional normalization maps the index to the zero interval for easy comparison, the inverse proportional function uses an inverse relationship for mapping, and the normalization processing makes the output value have a standard scale. The generation of the state stability measure depends on the historical change trend of the device state parameters, the analysis of the historical change trend uses a time series decomposition method to extract the long-term trend component, and the state stability measure value is quantified by the trend fluctuation amplitude. The calculation of statistical correlation uses Pearson correlation coefficient to capture linear relationship, the calculation of Pearson correlation coefficient involves the ratio of covariance and standard deviation, and a positive correlation coefficient indicates a positive correlation. The generation of the sensitivity adjustment factor fuses multi-dimensional information to enhance robustness, the multi-dimensional information includes the spatial and temporal distribution characteristics of the local consistency coefficient, and the value of the sensitivity adjustment factor is fused by weighted geometric mean of different dimensional contributions. The use of the coefficient of variation standardizes data volatility, the coefficient of variation eliminates dimensional differences for easy cross-parameter comparison, and the standardization process makes the data of different sensing units comparable. The calculation of the peak amplitude stability index is based on long-term observation data, the long-term observation data covers multiple operation cycles of the device, and the calculation of the stability index uses the coefficient of variation method to ensure consistency. The weighted fusion process allocates weights based on signal-to-noise ratio characteristics, data with high signal-to-noise ratio is allocated higher weight, and the weight allocation is realized through heuristic rules. The adaptive correction mechanism dynamically adjusts the coefficient value to match environmental changes, the detection of environmental changes is realized through gradient analysis of device state parameters, and the correction mechanism has feedback adjustment characteristics.

[0032] The establishment of the dynamic similarity threshold range considers the dynamic characteristics of the equipment operating state, which is quantified by autocorrelation analysis of the state parameters, and the threshold range is updated regularly to adapt to the working condition changes. The absolute difference degree is calculated using the absolute value distance to quantify the deviation, and the calculation of the absolute value distance is simple and efficient, and the deviation quantization result directly reflects the difference size. The similarity weight is assigned using an inverse proportional function to give a small difference a high weight, and the coefficient of the inverse proportional function is learned through training data, and the weight distribution curve ensures smooth transition. The preset threshold is set based on empirical values or adaptive learning, and the adaptive learning uses a gradient descent method to optimize the threshold value, and the threshold setting balances the recall rate and accuracy. The density clustering is executed to identify high-density areas using the DBSCAN algorithm, the parameters of the DBSCAN algorithm are optimized by grid search, and the identification of high-density areas is based on a neighbor point number threshold. The selection of the maximum cluster ensures the representativeness of the reference group, and the maximum cluster has high cohesion, and the representativeness is quantified by the average distance within the group. The formation of the reference sensor unit group enhances the context relevance of the state sensitivity calculation, and the context relevance is reflected by the consistency of the equipment state parameters within the group, and the quality of the group is evaluated by variance analysis. The calculation of the absolute difference degree simplifies the comparison process, and the comparison process only needs arithmetic operations, which simplifies the design and reduces the computational complexity. The introduction of the similarity weight smooths the data fluctuation effect, and the weight distribution suppresses the influence of abnormal values, and the smoothing effect is achieved through weighted averaging. The density clustering handles the spatial distribution characteristics, the modeling of the spatial distribution characteristics is based on the Euclidean distance metric, and the density clustering can discover any shape distribution. The maximum cluster screening improves the consistency within the group, and the measurement of the consistency within the group is calculated by the variance within the group, and the screening process eliminates discrete points.

[0033] The evaluation analysis unit determines the reference sensor unit group by establishing a dynamic similarity threshold range based on the equipment state parameter value of the target sensor unit, and calculating the absolute difference degree of the equipment state parameters of other sensor units from the reference value. According to the difference degree, the similarity weight is assigned, and the sensor units with a weight exceeding a preset threshold are selected to form an initial candidate set. The initial candidate set is subjected to density clustering, and the units in the maximum cluster are selected as the final reference sensor unit group. The establishment of the dynamic similarity threshold range considers the dynamic characteristics of the equipment operating state, and the calculation of the absolute difference degree uses the absolute value distance to quantify the deviation. The similarity weight is assigned using an inverse proportional function to give a small difference a high weight, and the preset threshold is set based on empirical values or adaptive learning. The execution of the density clustering identifies high-density areas using the DBSCAN algorithm, and the selection of the maximum cluster ensures the representativeness of the reference group. The formation of the reference sensor unit group enhances the context relevance of the state sensitivity calculation, and the dynamic threshold mechanism adapts to different working conditions. The calculation of the absolute difference degree simplifies the comparison process, and the introduction of the similarity weight smooths the data fluctuation effect. The density clustering handles the spatial distribution characteristics, and the maximum cluster screening improves the consistency within the group. The overall method improves the quality and stability of the reference sensor unit group.

[0034] Example 2: refer to Figure 3 When the evaluation analysis unit calculates the initial state correlation degree, it collects the signal intensity values of all sensing units at the target signal position to construct a signal intensity distribution histogram. The collection of signal intensity values is based on the voltage amplitude or energy integral output by the piezoelectric sensing unit. The construction of the distribution histogram uses an equal-width binning method to divide the signal intensity range into multiple intervals and count the frequency of each interval. The histogram is subjected to probability density estimation to obtain the probability distribution function of the signal intensity. The probability density estimation uses the kernel density estimation method to smooth the histogram data using a Gaussian kernel function to generate a continuous probability distribution curve. At the same time, the device state parameter values of all sensing units are obtained to generate a state parameter distribution histogram and convert it into a probability distribution. The device state parameter values include temperature, pressure, or vibration values. The construction of the state parameter distribution histogram follows the same binning strategy to ensure consistency. The conversion of the probability distribution is achieved by normalization processing to make the sum of the histogram frequencies equal to one. A multi-scale distribution matching algorithm is used to calculate the overall deviation between the two probability distributions. The multi-scale distribution matching algorithm decomposes the probability distribution into sub-band components at different resolutions through wavelet transform. At each scale, the difference between the sub-band distributions is calculated and weighted to fuse. The overall deviation is quantified using the following formula:

[0035]

[0036] where D represents the overall deviation, N represents the total number of scales, i represents the scale index, wi represents the weight coefficient of the scale, Sj represents the sub-band distribution of the signal intensity probability distribution at scale j, Sj represents the sub-band distribution of the device state parameter probability distribution at scale j, and the integral operation calculates the area of the absolute difference between the two sub-band distributions. The weight coefficient is assigned according to the importance of the scale, with higher scales corresponding to macro features being assigned a larger weight. The deviation is converted to a correlation index through a nonlinear mapping function. The nonlinear mapping function uses a sigmoid function to map the deviation to the range of zero to one. The value of the correlation index is negatively related to the deviation.

[0037] The evaluation analysis unit assigns each sensing unit an independent identity identifier when calculating the real state correlation degree, the identity identifier adopts a unique digital code or a string identifier, and the assignment of the identity identifier ensures that each sensing unit has a traceable identity. The identity identifiers are arranged in descending order of device state parameter values to form a state priority sequence, and the arrangement of the device state parameter values is based on the numerical value, and the descending arrangement places the sensing units with high state parameter values in the front of the sequence. At the same time, the identity identifiers are arranged in ascending order of signal strength values to generate a signal priority sequence, and the ascending arrangement of the signal strength values emphasizes the priority of the low signal strength area, and the sequence generation maintains the correspondence of the identity identifiers. The basic matching degree is calculated by identifying the identity identifiers with the same position in the two priority sequences through a sequence comparison algorithm, and the sequence comparison algorithm applies the longest common subsequence method to find the occurrence position of the same identity identifier in the two sequences. The calculation of the basic matching degree obtains a ratio value by dividing the number of matching identity identifiers by the total number. After removing the matching identity identifiers, the state sensitivity values corresponding to the remaining identity identifiers are extracted to construct a sensitivity distribution vector, and the extraction of the state sensitivity values is based on the pre-calculated state sensitivity results. The construction of the sensitivity distribution vector arranges the state sensitivity values of the remaining identity identifiers in sequence order to form a vector. The spatial similarity measure of the two sensitivity distribution vectors is calculated, and the calculation of the spatial similarity measure uses the cosine similarity method to measure the consistency of the vector direction, and the value range of the cosine similarity is between negative one and positive one. The correlation degree correction coefficient is generated by fusing the basic matching degree and the spatial similarity measure, and the fusion process uses a weighted average method to assign weights based on the reliability of the matching degree and the similarity. The value of the correlation degree correction coefficient is used to adjust the initial state correlation degree. The real state correlation degree is obtained by combining the correlation degree correction coefficient with the initial state correlation degree, and the combination operation adopts a multiplication or linear interpolation method. The output of the real state correlation degree more accurately reflects the real correlation strength between the signal and the state.

[0038] The construction of the signal strength distribution histogram uses an equal-width binning method to ensure consistent interval widths, and the determination of the number of bins is based on the Sturges rule or the square root selection method. After the histogram is constructed, the distribution shape is verified to meet the normality or uniformity assumption. The kernel density estimation method for probability density estimation selects the bandwidth parameter to affect the smoothing degree, and the bandwidth parameter is optimized through cross-validation or plug-in method. The generation of the probability distribution function provides a continuous probability model for subsequent analysis. The generation of the device state parameter distribution histogram is carried out synchronously, the collection of the device state parameter value is time-aligned with the signal strength value, and the transformation of the probability distribution makes the parameters of different dimensions comparable. The wavelet transform of the multi-scale distribution matching algorithm uses the Daubechies wavelet basis function, specifically, based on its standard application in signal processing, it can decompose the probability distribution into sub-band components at different scales, and the decomposition process realizes the multi-scale representation of the probability distribution through wavelet transform. The total deviation degree formula uses numerical integration methods such as the trapezoidal rule for integral operation, and the integral quantization of absolute difference quantifies the overlap degree of distribution, and the allocation of weight coefficient is based on the contribution of scale pair discrimination ability.

[0039] The allocation of the identity identifier is completed in the system initialization stage, and the identity identifier is stored in the sensor unit metadata. The uniqueness of the identity identifier is maintained through the central database. The generation of the state priority sequence uses the quicksort algorithm, and the sequence arrangement considers the dynamic change of the device state parameter. The sequence update period is synchronized with data acquisition. The construction of the signal priority sequence is similar, and the common subsequence search of the sequence comparison algorithm uses the dynamic programming method. The calculation of the basic matching degree introduces normalization processing. The extraction of the state sensitivity value accesses the intermediate results of the evaluation analysis unit. The dimension of the sensitivity distribution vector is consistent with the number of remaining identity identifiers. After the vector is constructed, the standardization processing eliminates the dimension. The cosine similarity calculation of the spatial similarity measure involves vector dot product and module length. The higher the similarity value, the more consistent the sensitivity distribution. The generation of the correlation degree correction coefficient uses heuristic rules to set the weighted average weight, and the correction coefficient range is controlled within zero to one. When combined with the initial state correlation degree, the multiplication operation emphasizes the scaling effect of the correction coefficient, and the output of the real state correlation degree is used for subsequent key position screening. The implementation of the multi-scale distribution matching algorithm involves multi-resolution analysis. The sub-band distribution at each scale is obtained through filtering and downsampling, and the difference calculation is performed independently at each scale. The smaller the value of the overall deviation degree, the more similar the two probability distributions are. The normalization processing of the deviation degree avoids the influence of scale difference. The sigmoid function expression of the nonlinear mapping function is where is a tunable parameter, and the mapping function converts the deviation degree into a correlation index. The permutation sequence of identity identifiers is stored in a memory array, and the time complexity of the sequence alignment algorithm is optimized for real-time processing. The spatial similarity measure of the sensitivity distribution vector supplements the information of sequence alignment, and the fusion process enhances the robustness of the correlation evaluation.

[0040] The collection period of signal strength values matches the device operation period, and the dynamic update of the histogram construction adapts to signal changes. The kernel function selection of the probability density estimation affects the distribution smoothness, and the variance parameter of the Gaussian kernel function is adjusted according to the data variance. The shape analysis of the device state parameter distribution detects skewness or kurtosis, and the multi-scale matching captures the distribution detail features. The scale weight of the overall deviation degree formula is set according to experience or learning, and the integral calculation uses discrete approximation to realize in digital systems. The sigmoid function of the nonlinear mapping provides a smooth transition, and the correlation index is mapped to a standard range. The sequence generation of identity identifiers uses a stable sorting algorithm, and the sequence alignment process handles the same value case. The basic matching degree calculation ignores sequence position shifts, and the construction of the sensitivity distribution vector considers vector dimension consistency. The cosine calculation of the spatial similarity measure optimizes numerical stability, and the fusion weight of the correlation correction coefficient balances the contributions of matching degree and similarity. The estimation accuracy of the probability distribution function is verified by the goodness-of-fit test, and the resolution level of multi-scale analysis covers from macro to micro. The use of absolute difference in overall deviation degree calculation is sensitive to outliers, and alternatives such as squared difference can be considered. The parameter control of the nonlinear mapping function controls the steepness of the mapping curve, and the parameter setting offset. The sequence permutation of identity identifiers maintains the order integrity, and the sequence alignment algorithm uses a divide-and-conquer strategy when processing large-scale data. The similarity measure of the sensitivity distribution vector has an alternative solution of Euclidean distance, but the cosine similarity pays more attention to direction. The generation of the correlation correction coefficient introduces a nonlinear fusion function, and the adjustable coefficient of the combination operation adapts to different needs. The modular design of the real state correlation calculation process facilitates expansion and maintenance, and the calculation results are stored in the system database.

[0041] Example 3: The analysis unit evaluates the environmental disturbance degree of the target sensing unit when calculating the device state parameter distribution characteristics of its reference sensing unit group. The determination of the reference sensing unit group is based on the clustering analysis results of the device state parameters, and the extraction of the device state parameter distribution characteristics uses the statistical moment method to calculate the mean, variance, skewness, and kurtosis indicators. The parameter distribution profile of the reference group is generated by the kernel density estimation method, which uses a Gaussian kernel function to smooth the device state parameter data. The bandwidth is automatically determined based on Scott's rule, and the parameter distribution profile is represented in the form of a continuous probability density function. The distribution profile is compared with the global device state parameter distribution profile, which contains the device state parameter information of all sensing units, through multi-resolution analysis. The multi-resolution analysis decomposes the two distribution profiles into approximation coefficients and detail coefficients at different scales through wavelet transform. The overlap area ratio of the two distribution profiles at different scales is calculated, and the formula is:

[0042]

[0043] wherein: represents the overlap area ratio at scale represents the reconstruction function of the reference group parameter distribution profile at scale represents the reconstruction function of the global device state parameter distribution profile at scale The integral operation calculates the area of the function in the domain. The numerator calculates the minimum overlap area of the two distribution profiles at scale , and the denominator calculates the maximum coverage area of the two distribution profiles at scale . The distribution matching degree is generated based on the weighted average of the overlap area ratio, and the calculation uses the arithmetic weighted average method. The weight coefficient is assigned according to the importance of the scale, and a higher scale corresponds to a smaller weight for macro features. The environmental disturbance degree evaluation value of the target signal position is obtained by inverse conversion and normalization of the distribution matching degree. The inverse conversion uses one minus the distribution matching degree, and the normalization linearly maps the result to the range of zero to one.

[0044] ​​The evaluation analysis unit establishes a ratio matrix of the real state correlation degree and the environmental interference degree of each signal position when screening the key sensing position, the real state correlation degree is derived from the previous calculation result, and the environmental interference degree is from the current environmental interference evaluation value. The rows of the ratio matrix correspond to different signal positions, and the columns correspond to different time segments, and the matrix element value is the ratio of the real state correlation degree and the environmental interference degree. The importance score of each position is generated by adaptive normalization processing on the ratio matrix, the adaptive normalization processing uses the minimum-maximum normalization method, and the normalization range is dynamically adjusted according to the statistical characteristics of the ratio matrix. A dynamic threshold mechanism is set to automatically adjust the threshold boundary based on the statistical characteristics of the importance score, the calculation of the dynamic threshold mechanism uses a statistical method based on Gaussian distribution, and the threshold boundary is set to a multiple of the mean value of the importance score plus a standard deviation offset. The positions with importance scores exceeding the dynamic threshold are included in the key position set, the inclusion operation is realized by element-by-element comparison, and the key position set is initially empty. Spatial clustering analysis is performed on the key position set to remove isolated points to form the final key sensing position group, the spatial clustering analysis uses the DBSCAN algorithm to cluster based on the geographical coordinates of the sensing unit, and the determination of the isolated point is based on the minimum point number requirement in the neighborhood.

[0045] The distribution feature extraction of the reference sensing unit group uses high-order statistics to capture the distribution form, and the smoothing degree of the kernel density estimation affects the detail performance of the distribution profile. The wavelet decomposition layer number of the multi-resolution contrast analysis is determined according to the data length, and the reconstruction function is obtained by inverse wavelet transform. The minimum and maximum functions in the formula for calculating the overlap area ratio ensure that the ratio range is between zero and one, and the integral calculation is realized by using the numerical integral method. Specifically, the numerical integral method realizes the integral of continuous functions into the summation calculation of discrete points, by dividing the domain into multiple small intervals, approximating the function value in each interval and accumulating the area. The weighted average of the distribution matching degree considers the contribution difference of different scales, and the reverse conversion converts the distribution matching degree into the interference degree index.

[0046] The construction of the ratio matrix maintains the time dimension information, and the adaptive normalization processing enhances the comparability of different time segment data. The generation of the importance score reflects the comprehensive importance of the location, and the adaptive adjustment of the dynamic threshold mechanism avoids the limitations of the fixed threshold. The spatial clustering analysis of the key location set considers the actual layout of the sensing unit, and the removal of isolated points improves the spatial coherence of the key location group. The formation of the final key sensing location group optimizes the sensor network layout. The bandwidth parameter optimization of the kernel density estimation is realized through cross-validation, and the comparative analysis of the distribution profile is carried out at multiple resolutions. The calculation of the overlapping area ratio uses the discrete approximation method, and the weighted average weight is set according to the scale resolution. The normalization processing of the environmental interference degree evaluation value uses the linear scaling method. The update of the ratio matrix is synchronized with data acquisition, and the adaptive normalization parameter is dynamically updated over time. The Gaussian distribution parameter of the dynamic threshold mechanism is calculated through the sliding window, and the parameter of the spatial clustering algorithm is set according to the site size.

[0047] The real-time update of the device state parameter distribution feature adapts to system changes, and the multi-resolution analysis provides a multi-granularity perspective. The overlapping area ratio reflects the distribution similarity, and the distribution matching degree integrates multi-scale information. The environmental interference degree evaluation value is used to quantify the noise influence. The ratio matrix integrates multi-dimensional evaluation indicators, and the importance score dynamically reflects the location value. The dynamic threshold mechanism ensures the flexibility of screening, and the spatial clustering ensures the spatial rationality of the key location. The output of the key sensing location group provides optimized input for state perception. The device state parameter distribution feature extraction period of the reference sensing unit group is synchronized with the system sampling period, and the calculation efficiency of the kernel density estimation is improved through algorithm optimization. The calculation complexity of the multi-resolution comparative analysis is linearly related to the number of scales, and the calculation accuracy of the overlapping area ratio affects the accuracy of the environmental interference degree evaluation. The weighted average weight of the distribution matching degree can be optimized through machine learning methods, and the time series analysis of the environmental interference degree evaluation value detects the interference pattern. The storage of the ratio matrix uses a ring buffer structure, and the range adjustment of the adaptive normalization processing avoids the influence of extreme values. The statistical feature calculation of the importance score uses robust statistics, and the parameter of the dynamic threshold mechanism can be adjusted through feedback control. The update trigger condition of the key location set is based on the importance score change rate, and the distance metric of the spatial clustering algorithm uses the Euclidean distance.

[0048] The generation of the distribution profile requires a sufficient number of samples, and the selection of the wavelet basis function affects the feature extraction effect. The numerator and denominator of the overlapping area ratio calculation formula are scaled simultaneously to keep the ratio stable, and the distribution matching degree calculation considers the correlation between scales. The calibration of the environmental interference degree evaluation value is verified by field measurement, and the dimension expansion of the ratio matrix can include more evaluation indexes. The dynamic change of the importance score reflects the system state evolution, and the sensitivity of the dynamic threshold mechanism can be adjusted. The spatial distribution characteristics of the key position group affect the monitoring effect, and the parameters of the isolated point removal algorithm need to be adjusted for specific applications. The convolution operation of the kernel density estimation uses a fast algorithm, and the scale selection of the multi-resolution analysis covers the main feature frequency band. The calculation of the overlapping area ratio optimizes the numerical stability, and the generation of the distribution matching degree introduces nonlinear fusion. The normalization method of the importance score affects the result distribution, and the adaptive ability of the dynamic threshold mechanism should be able to adapt to the working condition changes. The clustering radius of the spatial clustering affects the intra-group spacing, and the maintenance period of the key sensor position group is set according to the system requirements.

[0049] In Example 4, the state output unit extracts a multi-dimensional signal feature vector of the key sensor position from the historical data when constructing the state perception model. The historical data is sourced from the archived database collected by the distributed piezoelectric sensor over a long period of time, and the key sensor position is selected by the evaluation and analysis unit through real state correlation degree and environmental interference degree calculation. The construction of the multi-dimensional signal feature vector covers three dimensions of time domain features, frequency domain features, and time-frequency domain features. The time domain features include signal mean, variance, peak factor, and pulse index. The frequency domain features include power spectral density, center frequency, and frequency band energy ratio calculated by fast Fourier transform. The time-frequency domain features include wavelet coefficient energy distribution and entropy value extracted by continuous wavelet transform. The dimension of the feature vector is dynamically adjusted according to the device type and signal sampling rate, and each feature dimension is standardized by Z-score to eliminate dimensional differences. The standardization parameters are calculated from the mean and standard deviation of the historical data. The feature extraction process uses a sliding window mechanism to cover different working condition periods, and the window length is synchronized with the device running period. The overlap rate is set to avoid information loss. Referring to Table 1, a typical example of the signal feature vector of the key sensor position is shown:

[0050] Table 1: Signal feature vector table of key sensor position

[0051] Feature type Feature name Position A value Position B value Position C value Unit Time domain feature Signal mean 0.45 0.38 0.52 Time domain feature Signal variance 0.12 0.09 0.15 Frequency domain feature Center frequency 125.6 130.2 120.8 Frequency domain feature Band energy ratio 0.67 0.72 0.61 Dimensionless Time-frequency domain feature Wavelet energy entropy 1.34 1.28 1.41 Dimensionless

[0052] The eigenvectors are combined with corresponding device state parameters to form a training sample set, including device temperature, working pressure and vibration amplitude, and the parameter values are derived from the synchronous records of the device monitoring system. The construction of the training sample set ensures that each eigenvector is time-aligned with the device state parameters, and the number of samples is determined according to the length of the historical data. The sample division uses stratified sampling method to maintain balanced working condition distribution. Recursive feature elimination algorithm is used to select the most discriminative feature subset. Recursive feature elimination algorithm removes low importance features based on the weight coefficients of support vector regression model. The feature importance ranking is determined by the absolute value of the model coefficient, and the number of features removed in each iteration is set to a fixed proportion of the total number of features. The feature subset selection criterion uses cross-validation precision maximization. The termination condition of recursive feature elimination algorithm is based on feature number threshold or precision plateau. The final feature subset contains the most discriminative feature dimensions, and the feature selection process records the feature weight history. Support vector regression algorithm is used to train the regression model on the selected feature subset. Support vector regression algorithm uses radial basis function kernel to handle nonlinear relationships, and kernel function parameters include bandwidth coefficient and penalty factor. The model training input is the standard value of the selected feature subset, and the output is the regression value of the device state parameter. The training process uses sequential minimal optimization algorithm to accelerate convergence. The support vector of the support vector regression model is selected based on the size of the Lagrange multiplier, and the model complexity is controlled by the boundary tolerance parameter. The training error is calculated using the mean square error index.

[0053] The model hyperparameters are optimized by cross-validation, which divides the training sample set into training subsets and validation subsets using the k-fold method. The hyperparameter optimization grid includes kernel bandwidth, penalty factor and boundary tolerance. The cross-validation evaluation index uses mean absolute error and determination coefficient. The optimal hyperparameter combination is selected based on the performance of the validation set, and the optimization process uses grid search or random search strategy. The cross-validation results record the mapping relationship between hyperparameters and performance, and the optimal hyperparameter combination is applied to the final model training. The trained regression model and feature selection rules are packaged together as a state-aware model, including model parameter files, feature scaling rules and feature subset indexes. The deployment of the state-aware model uses lightweight container technology, and the model interface defines the standard input and output format. Version management is identified by hash value. The packaging process integrates model verification steps, and the verification data comes from an independent test set. Model performance benchmarking covers a variety of working conditions.

[0054] The state output unit collects distributed sensing signals of the equipment to be monitored in real time when performing state sensing. The distributed sensing signals are collected by using a high-speed data acquisition card to synchronize multiple piezoelectric sensing units. The sampling frequency is set according to the highest frequency component of the signal. The transmission of the distributed sensing signals adopts an industrial Ethernet protocol. The data packet contains a time stamp and a sensor identifier. The size of the acquisition buffer is designed to avoid data loss. The real-time acquisition process is synchronized with the equipment clock. The acquisition trigger condition is based on the device state change event or the timing cycle. The signal is subjected to quality detection and outlier filtering. The quality detection includes signal-to-noise ratio calculation, signal integrity check and baseline drift correction. The outlier filtering adopts statistical methods such as isolation forest algorithm or Z-score threshold method. The signal quality index is calculated and recorded in real time. The quality threshold is dynamically adjusted according to the historical signal characteristics. The abnormal signal is marked and isolated for processing. The quality detection result is fed back to the signal acquisition module to adaptively adjust the sampling parameters. The signal filtered by the outlier filtering is stored in a ring buffer. Multi-dimensional signal features at key sensing positions are extracted. The feature extraction algorithm is consistent with that in the training stage, including time domain feature calculation, frequency domain transformation and time-frequency analysis. The feature calculation adopts a sliding window for real-time processing. The window length is matched with that in the training stage. The feature values are standardized in real time using historical parameters.

[0055] The feature vector is reduced in dimension according to a preset feature selection rule. The feature selection rule is derived from the recursive feature elimination result in the training stage, including a feature subset index and a scaling parameter. The dimension reduction operation directly selects the value of the selected feature dimension. The dimension of the feature vector is reduced to the size of the optimal subset. The reduced feature vector remains in a standardized format. The reduced feature vector is input into the state sensing model. The model input interface verifies the format and dimension of the feature vector. The model inference adopts batch processing or streaming processing mode. The state sensing model outputs an estimated value of the device state parameter. The estimated value is converted to a physical unit based on a calibration curve. The model inference delay is controlled within milliseconds. The state estimate value and its confidence are obtained. The confidence is calculated based on the prediction interval or variance estimation of the support vector regression model. The confidence interval is calibrated through historical error distribution. The state estimate value and the confidence are displayed in real time on the monitoring interface. The data is stored in a time series database. The estimated value exceeding the threshold triggers an alarm. The confidence calculation considers the model uncertainty and input noise. Low confidence results are marked as to be verified. The output results are smoothed by combining the time series filtering algorithm. The time series filtering algorithm adopts exponential weighted moving average or Kalman filtering. The filtering parameters are adjusted according to the signal noise characteristics. The smoothing process reduces the short-term fluctuations of the estimated value. The filtering window length is matched with the dynamic response time of the equipment. The filtered results are output as the final state sensing value. The smoothing algorithm is suitable for non-stationary signal conditions. The filtered state is updated in real time based on new observations. The construction of the state sensing model and the state sensing process form a closed-loop system. The model is periodically retrained based on new acquisition data. The feature selection rule and hyperparameters are updated online.

[0056] In Example 5, the analysis unit determines a dynamic similarity threshold range based on a device state parameter value of a target sensor unit, which can be a wind turbine gearbox bearing temperature value, and establishes the dynamic similarity threshold range with a reference value of 85.3 degrees Celsius read by the temperature sensor in the current sampling period. The dynamic similarity threshold range is established based on historical statistical characteristics of the device state parameter, including the mean and standard deviation of the temperature data in the last 100 sampling periods. The lower threshold is calculated as the reference value minus twice the standard deviation, and the upper threshold is calculated as the reference value plus twice the standard deviation, forming a dynamically changing temperature interval range. The update frequency of the dynamic similarity threshold range is synchronized with the device state parameter sampling period, and the new sampling data is input into the sliding window to recalculate the statistical characteristics, and the threshold range is adjusted adaptively with the device operating state. The absolute difference degree of the device state parameter of other sensor units, including temperature sensors distributed at different positions of the gearbox, is calculated, and the absolute difference degree is directly quantified by absolute value operation. The temperature read by the sensor closest to the bearing is 87.1 degrees Celsius, and the absolute difference degree is 1.8 degrees Celsius; the temperature read by the sensor at the inlet of the oil cooling pipeline is 82.4 degrees Celsius, and the absolute difference degree is 2.9 degrees Celsius; the temperature read by the gearbox shell sensor is 79.6 degrees Celsius, and the absolute difference degree is 5.7 degrees Celsius. The calculation of the absolute difference degree covers all available sensor units, and the difference degree value reflects the temperature deviation of each sensor unit from the target sensor unit, which reflects the internal thermal field distribution characteristics of the device. Similarity weights are assigned according to the difference degree, and the similarity weight is calculated using an inverse proportional function to achieve high weight allocation for small difference degrees. The inverse proportional function coefficient is set to twice the temperature measurement error, the weight value is one when the difference degree is zero, and the weight value decreases by 15% for every one degree Celsius increase in difference degree. The sensor closest to the bearing obtains a weight of 0.73, the sensor at the inlet of the oil cooling pipeline obtains a weight of 0.57, and the gearbox shell sensor obtains a weight of 0.36. The similarity weight distribution curve presents a continuous and smooth characteristic, and the weight distribution reflects the working condition closeness of the sensor unit to the target sensor unit.

[0057] The sensor units with weight exceeding a preset threshold form an initial candidate set, the preset threshold value is determined by clustering analysis to find the best split point, and the threshold value is set to 0.5. The sensor units with weight value greater than 0.5 are included in the initial candidate set, the sensors closest to the bearing and the oil cooling pipeline inlet meet the conditions, and the gear box shell sensor is excluded due to too low weight. The size of the initial candidate set is affected by the distribution density of the equipment state parameters, the number of candidate units in the dense distribution area is more, and the number of candidate units in the sparse distribution area is less. The formation of the initial candidate set is based on single-dimensional weight screening, which retains potential reference units with high similarity to the target sensor unit. The initial candidate set is subjected to density clustering, and the execution of density clustering applies the DBSCAN algorithm to identify core points based on the spatial density of the equipment state parameter values. The neighborhood radius parameter of the DBSCAN algorithm is set to three times the temperature measurement accuracy, and the minimum point parameter is dynamically adjusted according to the size of the initial candidate set. The sensors closest to the bearing and the oil cooling pipeline inlet form a high-density area in the parameter space and are identified as the same cluster; the peripheral discrete points are marked as noise points. The density clustering process can find clusters of any shape and overcome the limitations of traditional clustering based on distance threshold. The units in the largest cluster are selected as the final reference sensor unit group, and the largest cluster is determined according to the number of sensor units contained in the cluster. When the number is the same, the cluster with higher density is selected. The cluster composed of the sensors closest to the bearing and the oil cooling pipeline inlet is selected as the largest cluster, and the two sensor units form the final reference sensor unit group. The equipment state parameter values of the final reference sensor unit group have high similarity, and the temperature difference in the group is controlled within three degrees Celsius.

[0058] In the scenario of industrial robot joint torque monitoring, the device state parameter value of the target sensor unit is the servo motor output shaft torque value, and the reference value is taken from the real-time reading of the torque sensor, which is 135 Nm. The dynamic similarity threshold range is established based on the historical fluctuation characteristics of the torque value, and the threshold boundary is automatically adjusted with the change of load. The torque sensor readings of other joints are compared with the reference value, the input end torque of the harmonic reducer is 128 Nm, and the output end torque is 142 Nm, and the absolute difference degrees are 7 Nm and 7 Nm respectively. The similarity weight distribution considers the torque transmission characteristics, and the input end and output end sensors obtain higher weights. The preset threshold filters two sensors into the initial candidate set, and the density clustering identifies the sensor unit group with continuous torque characteristics, and finally the reference sensor unit group contains the sensors at both ends of the harmonic reducer. For hydraulic system pressure monitoring applications, the device state parameter value of the target sensor unit is the main pump outlet pressure value, and the reference value is set to 21.5 MPa. The dynamic similarity threshold range is adjusted according to the system pressure fluctuation period, covering the normal working pressure range. The pressure sensor readings distributed at the entrances of each actuator are compared with the reference value, the pressure in the rodless cavity of the hydraulic cylinder is 20.8 MPa, the pressure in the rod cavity is 18.3 MPa, and the pressure before the proportional valve is 22.1 MPa. The similarity weight distribution reflects the pressure correlation degree, and the sensors near the main oil way obtain higher weights. The preset threshold filters three sensors to form the initial candidate set, and the density clustering identifies the main oil way sensor as the core cluster, and the sensor before the proportional valve is excluded due to abnormal pressure value, and finally the reference sensor unit group contains the pressure sensors at both ends of the hydraulic cylinder.

[0059] The establishment of the dynamic similarity threshold range considers the dynamic characteristics of the equipment operating state. The daily fluctuation range of the wind turbine gearbox bearing temperature value during normal operation can reach fifteen degrees Celsius. The threshold range needs to include this normal fluctuation. The calculation of the absolute difference degree uses absolute value operation to avoid directional influence. The absolute value of the temperature difference reflects the relative deviation degree and is not affected by the high and low directions. The allocation curve of the similarity weight needs to maintain the monotonic decreasing characteristic. The weight value decreases with the increase of the difference degree but never becomes zero. The setting of the preset threshold is based on the elbow rule in cluster analysis to determine the best segmentation point. The threshold value is recalculated regularly to adapt to system changes. The execution of density clustering applies the OPTICS algorithm to improve the variable density processing capability of DBSCAN. The OPTICS algorithm generates a reachable distance map to identify different density areas. The selection of the maximum cluster considers the balance of cluster quality and scale. The cluster quality is evaluated by the silhouette coefficient. The scale weight is set to zero point seven. The distribution of the equipment state parameter values of the reference sensor unit group needs to meet the normality assumption. The normality test is verified by the Shapiro-Wilk method. Specifically, all the equipment state parameter values of the reference sensor unit group are collected. These values come from real-time data collected by distributed piezoelectric sensors or historical databases. The equipment state parameter values include temperature, pressure or vibration values. The data collection process ensures that the sample size is sufficient for statistical testing. The test process calculates the Shapiro-Wilk statistic. This statistic is based on the comparison of the order statistics of the equipment state parameter value sequence and the expected value of the theoretical normal distribution. The calculation involves the ordering and weighted sum calculation of the parameter values. The closer the statistic value is to one, the closer the distribution is to normality. The number of sensor units in the group affects the stability of the state sensitivity calculation. When the number is insufficient, the similarity weight threshold value is expanded.

[0060] In the spindle vibration monitoring of a numerical control machine tool, the device state parameter value of the target sensor unit is the vibration acceleration value, and the reference value is taken from the vibration sensor of the spindle front bearing seat with a value of 56 m / s2. The dynamic similarity threshold range is established based on the vibration intensity historical data, and the upper and lower limits of the threshold are adjusted in proportion to the speed change. The vibration value of the spindle rear bearing seat is 52 m / s2, and the vibration value of the tool holder is 61 m / s2. The absolute difference is 4 m / s2 and 5 m / s2, respectively. The similarity weight distribution considers the mechanical transmission path, and the spindle bearing seat sensor obtains a higher weight. The preset threshold filters out two bearing seat sensors into the initial candidate set, and the density clustering identifies the spindle bearing vibration feature cluster. Finally, the reference sensor unit group contains the front and rear bearing seat vibration sensors. For the photovoltaic inverter heat dissipation system monitoring, the device state parameter value of the target sensor unit is the IGBT module junction temperature value, and the reference value is taken from the temperature sensor reading of 97 degrees Celsius. The dynamic similarity threshold range is established in combination with the temperature distribution of the heat sink, and the threshold boundary changes adaptively with the ambient temperature. The inlet temperature of the heat sink is 89 degrees Celsius, the outlet temperature is 93 degrees Celsius, and the air flow rate sensor conversion temperature is 86 degrees Celsius. The similarity weight distribution considers the heat conduction relationship, and the heat sink inlet and outlet sensors obtain a higher weight. The preset threshold filters out the heat sink inlet and outlet sensors to form the initial candidate set, and the density clustering identifies the sensors on the heat flow path as the same cluster. The air flow rate sensor is excluded due to temperature conversion error, and the final reference sensor unit group contains the heat sink inlet and outlet temperature sensors.

[0061] The determination process of the reference sensor unit group needs to meet the real-time requirement, and the algorithm complexity is controlled within the linear range. The calculation of the dynamic similarity threshold range uses the incremental update method, which updates the statistics without recalculating all historical data when new data points are input. The calculation of the absolute difference degree is parallel processed for multiple sensors, and the multi-core processor is used to accelerate the operation. The similarity weight distribution uses the lookup table method instead of real-time calculation, and the weight value table is pre-calculated and stored in the memory. The density clustering is executed using the incremental clustering algorithm, which locally updates the clustering structure when a new sensor unit is added. The output format of the final reference sensor unit group contains the sensor identifier list and the similarity index in the group, which is used for the reliability evaluation of the subsequent state sensitivity calculation. The update trigger conditions of the reference sensor unit group include significant changes in the device state parameter and the expiration of the time period, and the update frequency balances the calculation overhead and adaptability requirements. The spatial distribution information of the sensor units in the group is used for anomaly detection, and the outlier analysis identifies the faulty sensors. The determination method of the reference sensor unit group is universal for different device types, and the parameter settings are adjusted according to the specific application scenario.

[0062] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A device surface interaction and state sensing system based on distributed piezoelectric sensors, characterized in that, The system integrates multiple piezoelectric sensing units, signal processing units, evaluation and analysis units, and status output units; The piezoelectric sensing units are distributed on the device surface to collect piezoelectric signals generated by surface interactions. The signal processing unit preprocesses the piezoelectric signals, extracts signal features, and generates signal response curves. The evaluation and analysis unit calculates the state sensitivity, initial state correlation, actual state correlation, and environmental interference of each piezoelectric sensing unit based on the signal response curves and device state parameters, and identifies key sensing locations. The state output unit constructs a state perception model based on the signal data from the key sensing locations and performs real-time perception of device surface interactions and states. When calculating state sensitivity, the evaluation and analysis unit performs the following operations: acquires the signal response curve of the target piezoelectric sensing unit within a preset time window; determines a set of reference sensing units based on the similarity of device state parameters; extracts the peak response amplitudes of all reference sensing units at the target signal location to form an amplitude sequence; calculates the dispersion index of the amplitude sequence and performs inverse proportional normalization to obtain a local consistency coefficient; simultaneously analyzes the fluctuation range of the device state parameters corresponding to the reference sensing units to generate a state stability metric; and combines the statistical correlation between the local consistency coefficient sequence of all sensing units and the state stability metric sequence to generate a sensitivity adjustment factor. Multiplying the sensitivity adjustment factor by the local consistency coefficient yields the state sensitivity of the target sensing unit at the target signal location; When the evaluation and analysis unit calculates the correlation degree of the real state, it performs the following steps: assigning an independent identifier to each sensing unit; arranging the identifiers in descending order of the device state parameter values ​​to form a state priority sequence; simultaneously arranging the identifiers in ascending order of the signal strength values ​​to generate a signal priority sequence; and identifying the identifiers with the same position in the two priority sequences through a sequence comparison algorithm to calculate the basic matching degree. After removing the matched identity identifiers, extract the state sensitivity values ​​corresponding to the remaining identity identifiers and construct a sensitivity distribution vector; calculate the spatial similarity measure of the two sensitivity distribution vectors; fuse the basic matching degree and the spatial similarity measure to generate the correlation correction coefficient; combine the correlation correction coefficient with the initial state correlation to obtain the true state correlation. When the state output unit constructs the state awareness model, it performs the following steps: extracting multi-dimensional signal feature vectors of key sensing locations from historical data; The feature vectors and corresponding device state parameters are combined to form a training sample set; a recursive feature elimination algorithm is used to select the most discriminative feature subset; a support vector regression algorithm is used to train a regression model on the selected feature subset; the model hyperparameters are optimized through cross-validation; and the trained regression model and feature selection rules are encapsulated together into a state-aware model.

2. The device surface interaction and state sensing system based on distributed piezoelectric sensors according to claim 1, characterized in that, When the evaluation and analysis unit calculates the initial state correlation degree, it performs the following steps: collecting the signal strength values ​​of all sensing units at the target signal location and constructing a signal strength distribution histogram; estimating the probability density of the histogram to obtain the probability distribution function of the signal strength; simultaneously acquiring the device state parameter values ​​of all sensing units, generating a state parameter distribution histogram and converting it into a probability distribution; using a multi-scale distribution matching algorithm to calculate the overall deviation between the two probability distributions; converting the deviation into a correlation degree index through a nonlinear mapping function, and using this index as the initial state correlation degree of the target signal location.

3. The device surface interaction and state sensing system based on distributed piezoelectric sensors according to claim 2, characterized in that, When calculating the environmental interference degree, the evaluation and analysis unit performs the following steps: for the target sensing unit, obtain the device state parameter distribution characteristics of its reference sensing unit group; generate the parameter distribution profile of the reference group using the kernel density estimation method; perform multi-resolution comparative analysis of this distribution profile and the global device state parameter distribution profile; calculate the overlap area ratio of the two distribution profiles at different scales; generate the distribution matching degree based on the weighted average of the overlap area ratio; and obtain the environmental interference degree evaluation value of the target signal location by performing inverse phase transformation and normalization on the distribution matching degree.

4. The device surface interaction and state sensing system based on distributed piezoelectric sensors according to claim 1, characterized in that, When the evaluation and analysis unit screens key sensing locations, it performs the following steps: establishing a ratio matrix of the true state correlation degree to the environmental interference degree for each signal location; The comparison value matrix is ​​adaptively normalized to generate importance scores for each position; A dynamic threshold mechanism is set up, which automatically adjusts the threshold boundary based on the statistical characteristics of importance scores; locations with importance scores exceeding the dynamic threshold are included in the key location set; spatial clustering analysis is performed on the key location set to remove outliers and form the final key sensing location group.

5. The device surface interaction and state sensing system based on distributed piezoelectric sensors according to claim 1, characterized in that, When the state output unit performs state perception, it executes the following steps: real-time acquisition of distributed sensing signals from the device under monitoring; quality detection and outlier filtering of the signals; extraction of multi-dimensional signal features at key sensing locations; dimensionality reduction of the feature vector according to preset feature selection rules; input of the dimensionality-reduced feature vector into the state perception model; acquisition of the state estimate and its confidence level output by the model; and smoothing of the output result using a time series filtering algorithm.

6. The device surface interaction and state sensing system based on distributed piezoelectric sensors according to claim 1, characterized in that, When the evaluation and analysis unit determines the reference sensing unit group, it performs the following steps: establish a dynamic similarity threshold range based on the device state parameter value of the target sensing unit; calculate the absolute difference between the device state parameters of other sensing units and the benchmark value; assign similarity weights according to the difference; select sensing units with weights exceeding a preset threshold to form an initial candidate set; perform density clustering on the initial candidate set, and select the units in the largest cluster as the final reference sensing unit group.

7. The device surface interaction and state sensing system based on distributed piezoelectric sensors according to claim 1, characterized in that, When calculating the local consistency coefficient, the evaluation and analysis unit performs the following steps: calculating the coefficient of variation based on the amplitude sequence; Obtain the peak amplitude stability index of the historical signal response curve at the target signal location; weight and fuse the coefficient of variation with the stability index to generate a dynamic consistency weight; use the dynamic consistency weight to adaptively correct the local consistency coefficient.

Images