A wireless sensor data transmission and storage method based on distributed nodes

Abstract

This invention relates to the field of wireless communication networks, specifically disclosing a method for wireless sensor data transmission and storage based on distributed nodes. The method includes: nodes extracting feature vectors from local sensor data; reconstructing a local structural response field by coordinating with feature vectors from neighboring nodes, and assessing the anomaly confidence level of their own data; intelligently selecting to transmit either compressed status reports or complete diagnostic data packets based on the field reconstruction quality and anomaly confidence level; and nodes autonomously generating and adjusting future monitoring strategies based on historical data to learn state evolution trends. This invention overcomes the shortcomings of traditional methods, such as high resource consumption, shallow monitoring depth, and rigid strategies, greatly improving resource utilization efficiency, significantly enhancing monitoring depth and reliability, and achieving autonomous optimization and forward-looking operation and maintenance of the system.

Description

Technical Field

[0001] This invention relates to the field of wireless communication network technology, and more specifically to a method for wireless sensor data transmission and storage based on distributed nodes. Background Technology

[0002] Structural health monitoring (SHM) is crucial for ensuring the safety of large infrastructure. With the development of Wireless Smart Sensor Networks (WSSNs), their advantages of flexible deployment and lower cost are gradually leading to their replacement of traditional wired monitoring systems. A typical WSSN node is capable of both data acquisition and wireless transmission.

[0003] However, existing WSSN-based monitoring methods still have significant limitations, restricting their effectiveness in large-scale, long-term monitoring: First, they are inefficient in terms of resources. Nodes typically continuously collect and transmit massive amounts of raw data, leading to rapid depletion of wireless communication bandwidth, node storage space, and battery power, making long-term deployment difficult. Second, they lack monitoring depth. Most existing methods only collect data from discrete points, lacking the ability to perceive the continuous distribution of the structural state space, i.e., the "field," and anomaly detection often relies on simple single-point amplitude thresholds, making them insensitive to weak or localized early damage. Finally, the system lacks autonomy. Monitoring strategies, such as sampling rates and thresholds, are usually statically preset or rely on centralized cloud control, unable to dynamically and proactively adapt to changes in the structure's own state and environmental patterns, resulting in limited intelligence.

[0004] Therefore, there is an urgent need for an innovative data transmission and storage method that can intelligently refine raw data into state information on resource-constrained distributed nodes, and endow the network with collaborative perception and autonomous optimization capabilities, thereby breaking through the existing technological bottlenecks. Summary of the Invention

[0005] The purpose of this invention is to provide a wireless sensor data transmission and storage method based on distributed nodes, so as to solve the defects of high resource consumption, shallow monitoring depth and rigid strategies in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0007] A method for wireless sensor data transmission and storage based on distributed nodes includes the following steps:

[0008] The collected raw time-series data is processed to generate local feature vectors that characterize the local structural response;

[0009] The system receives the local feature vectors from neighboring nodes, combines them with its own vector, reconstructs the local structural response field, generates a local field reconstruction report, and calculates the local anomaly confidence.

[0010] Based on the quality score of the local field reconstruction report and the local anomaly confidence level, a decision is made to either store or transmit the report or package the raw data to generate a high-fidelity diagnostic data package.

[0011] Based on the accumulated local feature vectors, local field reconstruction reports, local anomaly confidence, and metadata of high-fidelity diagnostic data packages, the system analyzes state trends and generates adaptive monitoring strategies, dynamically adjusting execution parameters.

[0012] As a preferred embodiment of the present invention, the processing of the acquired raw time-series data to generate a local feature vector characterizing the local structural response specifically includes:

[0013] Sensor nodes synchronously acquire multidimensional raw time-series data from their own sensing units and perform local fusion processing on the multidimensional raw time-series data to generate a multidimensional fused data stream.

[0014] The sensor node performs real-time feature extraction on the fused data stream to obtain a low-dimensional local feature vector, which is used to characterize the structural response state at the current node's location.

[0015] As a preferred embodiment of the present invention, the step of receiving the local feature vector of the neighboring node, combining it with its own vector, and then reconstructing the local structural response field to generate a local field reconstruction report and calculate the local anomaly confidence level specifically includes:

[0016] Sensor nodes periodically receive data packets broadcast by at least one neighboring node within their communication range via a wireless link. The data packets contain at least the local feature vectors, timestamps, and structural location information of the neighboring nodes.

[0017] The sensor node spatiotemporally aligns and combines its own local feature vector with the received local feature vectors of neighboring nodes to form an extended feature set;

[0018] Based on the extended feature set, the sensor node runs a built-in simplified field reconstruction algorithm to estimate the structural response field distribution of the local area covered by the sensor node and its neighboring nodes, and generates a local field reconstruction report.

[0019] Based on the extended feature set, the sensor node infers the expected feature vector of its own node position; at the same time, the sensor node calculates the local anomaly confidence by comparing its own local feature vector with the inferred expected feature vector.

[0020] As a preferred embodiment of the present invention, forming an extended feature set includes: the sensor node interpolates or aligns its own local feature vectors with those of each neighboring node to the same reference time according to the timestamp, and sorts the aligned feature vectors in space according to the node position information to form a spatial-feature matrix representing the state of discrete points in the local area, which serves as the extended feature set.

[0021] As a preferred embodiment of the present invention, the built-in simplified field reconstruction algorithm includes: taking an extended feature set as input, using a physics-driven or data-driven spatial interpolation method to estimate the structural response between discrete points, generating a local field reconstruction report containing a continuous or quasi-continuous field distribution map of the local region, and performing a self-evaluation on the smoothness or fitting error of the reconstruction result, and outputting a quality score.

[0022] As a preferred embodiment of the present invention, the calculation of local anomaly confidence includes: deriving the expected feature vector of its own node position from the extended feature set through a field reconstruction algorithm or a lightweight regression model; comparing the actual local feature vector extracted by the sensor node with the expected feature vector, calculating the Euclidean distance, cosine similarity or residual norm, and normalizing the calculation result as the local anomaly confidence.

[0023] As a preferred embodiment of the present invention, based on the quality score and local anomaly confidence level of the local field reconstruction report, a decision is made to either store or transmit the report or package the original data to generate a high-fidelity diagnostic data package, specifically including:

[0024] A first threshold and a second threshold are preset; the sensor node compares the quality score with the first threshold in real time, and compares the local anomaly confidence level with the second threshold;

[0025] When the quality score is greater than the first threshold and the local anomaly confidence is less than the second threshold, it is determined that the local structure response state is stable and the local field reconstruction report is reliable and can be effectively reconstructed. Then, the local field reconstruction report and key metadata are compressed and stored in the local memory and transmitted at an opportune time.

[0026] When the quality score is less than or equal to the first threshold or the local anomaly confidence is greater than or equal to the second threshold, it is determined that the local structure response state is abnormal or difficult to reconstruct reliably. Then, the original time-series data segment that triggered the determination, the corresponding fused data stream, and the local field reconstruction report are packaged together to generate a high-fidelity diagnostic data packet, which is preferentially written to the local memory and wireless transmission is started immediately or with high priority.

[0027] As a preferred embodiment of the present invention, the step of analyzing state trends and generating adaptive monitoring strategies based on accumulated local feature vectors, local field reconstruction reports, local anomaly confidence levels, and metadata from high-fidelity diagnostic data packets, and dynamically adjusting execution parameters, specifically includes:

[0028] The sensor node maintains a long-term cognitive database locally to accumulate and store metadata such as local feature vectors, local field reconstruction reports, local anomaly confidence scores, and high-fidelity diagnostic data packets generated during each execution.

[0029] Sensor nodes periodically analyze time-series data in a long-term cognitive database to deduce the structural state evolution trend and local anomaly confidence change pattern of the local area under their responsibility.

[0030] Based on evolution trends and change patterns, sensor nodes autonomously generate an adaptive monitoring strategy for a future period.

[0031] As a preferred embodiment of the present invention, the periodic analysis of time-series data in the long-term cognitive database calls the built-in lightweight analysis routine to perform time-series analysis on the local feature vectors, quality scores, and local anomaly confidence in the long-term cognitive database; the evolution trend of the structural state of the local area under the responsibility of the sensor node and the change pattern of the local anomaly confidence are derived by using moving average, trend fitting, or autoregressive models.

[0032] As a preferred embodiment of the present invention, the autonomous generation of an adaptive monitoring strategy for a future period includes: predicting the structural response level in the future period based on the evolution trend, and dynamically adjusting the data acquisition frequency accordingly; updating the neighbor node selection priority list based on the historical communication quality and collaborative reconstruction contribution between nodes; and adaptively adjusting the first threshold and the second threshold used for judgment based on the change pattern and statistical distribution of local anomaly confidence.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] This invention improves resource utilization efficiency: through node-local feature extraction, collaborative field reconstruction, and intelligent data selective transmission, the system normally transmits and stores only highly compressed state perception results, such as feature vectors and field reconstruction reports, and only captures complete original data in abnormal situations. This reduces wireless communication load and local storage usage by more than an order of magnitude, significantly extends the overall network endurance, and alleviates the pressure on cloud data processing.

[0035] This invention significantly enhances the depth and reliability of monitoring: the method enables each node to fuse neighbor information and reconstruct a continuous spatial state field locally, achieving a leap from discrete point monitoring to regional situational awareness. Simultaneously, the anomaly confidence assessment mechanism based on "local consistency" is more sensitive to weak, localized early anomalies, effectively improving the accuracy and timeliness of damage identification.

[0036] This invention enables autonomous optimization and proactive operation and maintenance of the system: by learning from historical data over a long period, nodes can autonomously predict state trends and dynamically adjust key parameters such as sampling rate, collaborative strategies, and discrimination thresholds. This gives the monitoring network self-adaptation and evolution capabilities, allowing it to proactively allocate resources based on the actual operational patterns of the structure and environmental changes, continuously optimizing system performance and energy efficiency while ensuring monitoring effectiveness. Attached Figure Description

[0037] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0038] Figure 1 This is a schematic diagram summarizing the process of the method described in Embodiment 1 of the present invention.

[0039] Figure 2 This is a detailed flowchart illustrating the method described in Embodiment 1 of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0042] Example 1

[0043] like Figure 1 - Figure 2 As shown, this invention provides a method for wireless sensor data transmission and storage based on distributed nodes, comprising the following steps:

[0044] S1. Process the acquired raw time-series data to generate local feature vectors representing the local structural response; specifically including:

[0045] S11. Sensor nodes synchronously acquire multidimensional raw time-series data from their own sensing units and perform local fusion processing on the multidimensional raw time-series data to generate a multidimensional fused data stream, including:

[0046] S111. Sensor nodes synchronize with the network time reference through their internal clocks, including: sensor nodes synchronizing their internal crystal clocks with the UTC time references of anchor nodes in the network through the IEEE 1588 Precision Time Protocol (PTP) or one-way broadcast time synchronization messages, with a synchronization accuracy better than 10 microseconds; at the same time, the nodes use external hardware interrupts as global sampling trigger signals, such as GPS PPS pulses or network synchronization trigger frames, to ensure that all sensing units start analog-to-digital conversion at the same time, eliminating inter-channel phase delay;

[0047] S112. High-precision external interrupts are used to trigger the multi-axis accelerometers and strain gauges in its sensing unit to perform synchronous sampling, thereby obtaining time-aligned multi-dimensional raw time-series data. This includes: the sensing unit includes at least a three-axis MEMS accelerometer, a resistance strain gauge, and a temperature compensation sensor; after receiving the synchronous trigger signal, each sensing unit acquires the structural response signal in parallel at a sampling rate of not less than 1kHz, and writes the data stream to a circular buffer through a direct memory access (DMA) channel; each sampling point is stamped with a 48-bit high-precision timestamp to achieve strict alignment of time-series data across nodes and modes.

[0048] S113. And perform local fusion processing on the multidimensional raw time-series data based on physical models or statistical learning to generate a multidimensional fused data stream, including:

[0049] 1. For multi-axis acceleration data, the sensor node executes lightweight structural dynamics equations in real time in a digital signal processor (DSP) to obtain the composite acceleration in the principal vibration direction. ,in: Acceleration Components along the x, y, and z axes, cross term coefficients , , The structure installation attitude matrix, obtained through offline calibration from local node storage, is used to compensate for non-orthogonal installation deviations and structural coupling effects; the synthesized acceleration... It can directly characterize the energy response in the main vibration direction, reducing three-dimensional data to a one-dimensional fused data stream;

[0050] 2. When the sensor unit dimension exceeds three dimensions or the intermodal correlation is complex, the node enables the incremental principal component analysis (IPCA) engine to reduce the dimensionality of multidimensional time-series data: a covariance matrix is ​​constructed within a sliding time window. For example, with 1024 sampling points, the principal eigenvectors are solved online using the power iteration method, reducing the original... Projecting time series data to Principal component space, This generates a fused data stream; this fusion method significantly reduces the dimensionality of subsequent processing while retaining more than 95% of the variance contribution rate.

[0051] S12. The sensor node performs real-time feature extraction on the fused data stream to obtain a low-dimensional local feature vector, including:

[0052] S121. Extract the root mean square (RMS), peak factor, and kurtosis from the fused data stream in the time domain. The RMS value is used to characterize the energy level of structural vibration and is sensitive to high-frequency energy growth caused by early damage. The peak factor is the ratio of the waveform peak value to the RMS value and is used to capture shock-related abnormal events, such as transient responses caused by bolt loosening. The kurtosis is used to identify the non-Gaussianity of the signal and has the ability to discriminate nonlinear responses caused by local cracks.

[0053] S122. The spectral energy, dominant frequency, and spectral entropy within a preset frequency band are extracted in the frequency domain. The spectral energy within the preset frequency band is integrated over the monitoring frequency band defined by the first n natural frequencies of the structure, and is used to track the energy evolution of specific modes. The dominant frequency is located at the maximum value of the spectral line through a peak detection algorithm, and its drift can reflect the structural stiffness degradation. The spectral entropy is used to quantify the spectral complexity and has early warning value for multimodal coupling faults.

[0054] S123. Modal parameter characteristics of the structure are obtained through online parameter identification, including local equivalent natural frequency and damping ratio. The local equivalent natural frequency is extracted from the characteristic equation through an online pole tracking algorithm, and its relative change is a sensitive indicator of structural damage. The damping ratio is estimated online using the logarithmic decay method or the half-power bandwidth method. Interface friction caused by damage will cause the damping ratio to increase abnormally.

[0055] S124. The sensor node combines the extracted time-domain, frequency-domain, and modal parameter features into a low-dimensional local feature vector according to a preset weight. The low-dimensional local feature vector occupies less than 128 bytes of RAM, which meets the real-time requirements of wireless broadcasting. At the same time, the node appends the node ID, timestamp, and sampling configuration code to the metadata of the low-dimensional local feature vector to form a complete data frame for step S2 to perform local field reconstruction and anomaly confidence assessment.

[0056] S2. Receive the local feature vectors of neighboring nodes, combine them with the node's own vector to reconstruct the local structural response field, generate a local field reconstruction report, and calculate the local anomaly confidence level; specifically including:

[0057] S21. Sensor nodes implement periodic neighbor discovery and feature vector sharing through low-power wireless links. Nodes send broadcast data packets on a designated channel at a period of 1 to 5 seconds, using a carrier sense multiple access mechanism to avoid collisions. The transmit power is set to -10 dBmW, and the communication radius is approximately 30 to 50 meters. The receiver activates the promiscuous mode of the radio frequency module to capture all neighbor data packets that conform to the protocol format.

[0058] Each neighbor broadcast data packet includes the following key fields: local feature vector, timestamp, node's structural location information, node's unique identifier, and data freshness indicator sequence number;

[0059] Each node maintains a neighbor feature cache table in static random access memory, storing the latest data of up to 8 neighbors. Each record contains a neighbor identifier, feature vector, timestamp, location information, and data packet lifetime. If a neighbor has not been updated within 3 broadcast cycles, its record is marked as expired and deleted to ensure the timeliness of the expanded feature set.

[0060] S22. The sensor node performs spatiotemporal alignment and combination of its local feature vector with the valid records in the neighbor cache table to form an extended feature set, including:

[0061] S221. Regarding time alignment, based on the difference between the current sampling time and the timestamp of the neighbor's data packet, linear interpolation or zero-order hold method is used to align the neighbor's feature vector to the same reference time. If the time difference exceeds 50 milliseconds, the neighbor's data is determined to be outdated and will not participate in the alignment.

[0062] S222. In terms of spatial sorting, the node arranges its own feature vector and the aligned feature vectors of its neighbors in ascending order of spatial distance to construct a spatial-feature matrix that represents the state of discrete points in the local region, as an extended feature set; the matrix contains feature vectors and corresponding spatial coordinates, stored in columns, with each column representing the spatial-feature joint state of a sensing point.

[0063] S23. Based on the extended feature set, the sensor node runs a built-in simplified field reconstruction algorithm to estimate the structural response field distribution of the local area covered by the sensor node and its neighboring nodes, generates a local field reconstruction report, and self-evaluates the smoothness or fitting error of the reconstruction results, outputting a quality score, including:

[0064] S231. The sensor node runs a lightweight field reconstruction algorithm based on an extended feature set to generate a local continuous response field distribution. The algorithm has two modes: physical-driven and data-driven.

[0065] The physics-driven reconstruction method is suitable for nodes that store local structural stiffness matrices or mode shapes. During the initialization phase, the following parameters are burned into the flash memory protected area through offline finite element analysis: a compact representation of the stiffness matrix, using a sparse storage format that retains only non-zero elements within the bandwidth; a dominant mode shape vector, storing the discrete mode shape values ​​of the first three to five modes in the local region; a geometric topology table, recording the connection relationships between the node and adjacent elements; and a set of material property parameters, including elastic modulus, density, Poisson's ratio, etc. These parameters are loaded into static random access memory at startup, occupying approximately 2KB of space, and all parameters are stored in a 16-bit fixed-point format.

[0066] 1. Construct a polygonal reconstruction domain with the current node as the centroid and the communication radius as the boundary. This region is discretized in memory as a 5×5 grid structure, with the grid density adaptively adjusted: a dense grid with a grid spacing of 0.1 meters is used in the core region less than one-third of the communication radius from the current node; a medium density is used in the transition region between one-third and two-thirds of the radius; and a sparse grid is used in the boundary region greater than two-thirds.

[0067] 2. A Gaussian radial basis function is chosen as the spatial weighting function, the shape of which is determined by distance. The weight is at its maximum value of 1 at the support point, and decreases exponentially with each unit increase in distance. The attenuation coefficient is determined based on the communication radius, and its specific value ensures that the weight decays to half power at the communication boundary; a typical setting is the natural logarithm of 2 divided by the square of the communication radius.

[0068] 3. To improve real-time performance, a three-level optimization approach was implemented: First, a weight threshold of 0.05 was set, and neighbors with values ​​below this threshold were ignored and not included in the calculation; second, a distance-weight lookup table was pre-calculated, transforming complex exponential operations into memory lookup operations; third, the symmetry of the distance matrix was utilized to reduce the computational load by half. Weight calculation took approximately 0.8 milliseconds per sampling period.

[0069] 4. Perform the following operations for each grid point within the reconstruction domain: Read the feature vectors of all support points, i.e., the current node plus its effective neighbors; query the distance from each support point to the grid point and obtain the corresponding weight value; normalize all weights so that the sum of the weights is strictly equal to 1; multiply the normalized weights by the corresponding support point feature vectors and sum them to obtain the response field estimate for the grid point. The calculation adopts a pipelined approach, and the results are directly written to the field distribution buffer.

[0070] 5. The unique verification mechanism of the physics-driven method includes: checking the gradient continuity of the reconstructed field between adjacent grids; if the gradient abrupt change exceeds ±30%, it is marked as an anomaly; statistically analyzing the systematic deviation between the reconstructed field and the original measurement values; if the mean of the persistent deviation is greater than 0.1, it is determined that the physical model is mismatched; monitoring the number of effective neighbors; if it is lower than the minimum number of support points (3) for ten consecutive periods, it is automatically downgraded to the data-driven method. When the quality score is lower than the unreliable threshold multiple times consecutively, a structural parameter recalibration request is triggered.

[0071] Data-driven reconstruction methods are suitable for situations where physical models are lacking. They are based on the assumption of spatial autocorrelation, which states that nodes that are closer to each other have higher similarity in their feature vectors, without relying on structural physical parameters.

[0072] 1. An exponential variogram is used to describe spatial correlation. The correlation length is taken as one-fifth of the structural feature size; for example, if the structural feature size is 10 meters, the correlation length is set to 2 meters. The field variance is standardized to 1 to eliminate the influence of dimensions. Euclidean distance is used for distance calculation and three-dimensional space is supported.

[0073] 2. For each grid point to be solved, a linear system consisting of a variogram matrix needs to be solved. The size of the variogram matrix is ​​equal to the number of support points plus 1, and the matrix elements reflect the spatial correlation between each support point. The right-hand side vector reflects the correlation between the point to be solved and each support point. The weight vector is obtained by solving a constrained linear system, with the constraint that the sum of all weights equals 1 to ensure unbiasedness. An optimized solution strategy is adopted: when the set of support points remains unchanged, the matrix inverse is directly reused; the inverse matrix is ​​recalculated only when the neighbors change; for ill-conditioned matrices, the conjugate gradient method is used for iterative solution, with a maximum number of iterations limited to 20.

[0074] 3. The initial solution takes approximately 2.5 milliseconds, which decreases to 0.5 milliseconds when the matrix is ​​reused. Temporary matrix storage requires approximately 1.2KB of static random access memory. When the condition number of the mutation function matrix exceeds one million, a regularization term is automatically added to improve numerical stability.

[0075] S232. Both reconstruction methods share a unified mesh framework. The mesh resolution is 0.1 meters in the core region and 0.5 meters in the boundary region, generating a total of 25 mesh points. The coordinate system adopts the structural local coordinate system, and for plate and shell structures, the vertical direction can be compressed to two dimensions. Each mesh point stores coordinates and estimated feature vectors, with a total storage space of feature dimension plus 3 multiplied by 4 bytes.

[0076] The computation order is based on spatial proximity, prioritizing the computation of grids closest to the current node. An outlier filtering mechanism is implemented; if the estimated value of a grid point exceeds a physically reasonable range, such as a natural frequency greater than 500 Hz, it is marked as an invalid point. The results are organized into a structured data packet, containing grid metadata, feature field data, and a validity mask identifying trustworthy grids.

[0077] A hierarchical computing strategy is adopted: high priority is used to calculate the position of the current node and the four nearest grids, with a calculation cycle of 1 millisecond; medium priority is used to calculate the remaining grids, which are completed in the idle time slots, with a total time of no more than 5 milliseconds; low priority is used to calculate the boundary area grids, which can be postponed to the next cycle.

[0078] S233. The local field reconstruction report includes grid point coordinates and corresponding eigenvector estimates, as well as a quality self-assessment score.

[0079] Quality self-assessment provides a basis for decision-making in step S3, avoiding erroneous data processing decisions based on low-quality reconstruction results.

[0080] For each node in the support point set, perform verification: temporarily remove the data of that node, reconstruct the field value at its location based on the remaining support points, and calculate the deviation between the reconstructed value and the true value. Summarize the deviation results after traversing all support points.

[0081] The quality score is calculated based on the normalization of cross-validation errors. The numerator is the sum of squares of all leave-one-out prediction errors, and the denominator is the sum of squares of the total deviations of each support point from the mean. The score ranges from negative infinity to 1, with negative values ​​indicating that the reconstruction performance is worse than simple mean prediction.

[0082] A three-tier quality assessment system is established: the high-quality level requires a quality score greater than or equal to 0.7, under which the reconstruction result is reliable and data compression can be performed; the medium-quality level requires a score between 0.5 and 0.7, at which point it is necessary to increase the number of neighboring nodes or shorten the broadcast period; the low-quality level is a score below 0.5, which is marked as unreliable reconstruction and immediately triggers the high-fidelity data transmission process.

[0083] The quality score and the outlier confidence level together form a two-dimensional decision plane, driving the dynamic switching of the four data strategies in step S3. When the quality score is at the unreliable level, regardless of the outlier confidence level, the original time-series data is forcibly preserved to ensure information integrity.

[0084] S24. Based on the extended feature set, the sensor node infers the expected feature vector of its own node position; simultaneously, the sensor node calculates the local anomaly confidence by comparing its own local feature vector with the inferred expected feature vector, including:

[0085] S241. Based on the local field reconstruction model constructed in step S23, the sensor node estimates the response field at its location. This estimation process uses a known set of support points, including the node itself and its effective neighboring nodes, to calculate the predicted value of the feature vector at the node's location through a reconstruction algorithm.

[0086] This prediction is essentially a spatially smoothed estimate of the node's state by its neighboring nodes, reflecting a consensus expectation of the local structural response. The calculations are preferentially performed using the results of the physics-driven model; if the physics model is unavailable, the output of the data-driven model is used.

[0087] The node stores the calculated expected feature vector in a temporary buffer, the size of which is the feature dimension multiplied by a single byte width, to ensure fast access for subsequent comparison operations.

[0088] The S242 node provides three optional anomaly indicators for quantifying the degree of deviation, which can be flexibly configured according to the specific monitoring scenario and structure type:

[0089] The first method is Euclidean distance residual calculation: directly measuring the linear distance between the actual feature vector and the expected feature vector. This method treats all feature dimensions equally and is suitable for scenarios where the importance of each dimension is relatively balanced, such as vibration monitoring of homogeneous isotropic structures.

[0090] The second method is cosine similarity deviation calculation: it focuses on the directional consistency of feature vectors rather than their absolute magnitude. When a structure is damaged, changes in the direction of feature vectors are often more sensitive than changes in their amplitude, making it particularly suitable for identifying modal shape change-type faults.

[0091] The third method is weighted residual norm calculation, which allows for differentiated weights to be assigned to different feature dimensions. For example, the weight of the structural natural frequency dimension can be set to 30%, the vibration energy dimension to 20%, and the damping ratio dimension to 25%. This weighting method can highlight key damage-sensitive indicators and improve the targeting of anomaly detection.

[0092] During the initialization phase, a node selects one of the methods by configuring parameters, and the selection strategy can be dynamically adjusted based on long-term cognitive results during subsequent operation.

[0093] S243. The original residual values ​​are converted into standardized local anomaly confidence scores through normalization, with the values ​​strictly controlled between zero and one.

[0094] The normalized baseline is derived from the statistical results of the baseline period of the health status. The first thousand samples after node deployment serve as a learning period, and the average residual and three times the standard deviation are calculated. The sum of these two values ​​serves as the upper limit of the normal fluctuation range. The ratio of subsequent real-time residuals to this upper limit is the anomaly confidence level. A very small constant is added as a protection against division-by-zero errors.

[0095] The established threshold determination mechanism is as follows: when the anomaly confidence level is below 0.3, it is determined to be a normal state; when it is between 0.3 and 0.7, it is determined to be a suspicious state, and the sampling frequency needs to be increased; when it exceeds 0.7, it is marked as a high-confidence anomaly.

[0096] Once a high-confidence anomaly is triggered, the node immediately performs two operations: first, it marks the current time-series data segment as high priority; second, it sends an anomaly coordination request to neighboring nodes, initiating a high-density monitoring mode for the local area. This marking is maintained for five sampling cycles to ensure that the anomaly is fully captured.

[0097] S244. Anomaly confidence calculation is implemented in a digital signal processor, with a single calculation taking no more than 0.5 milliseconds. Each node allocates a circular queue in flash memory to store historical confidence values, with a capacity of one thousand records, for offline analysis and trend prediction. When the storage queue reaches its capacity limit, a first-in, first-out (FIFO) strategy is used to overwrite the oldest data.

[0098] The reconstruction quality score and local anomaly confidence generated in S21 to S24 above will be directly input into the adaptive decision engine in step S3 to drive the dynamic switching of data storage and transmission strategies.

[0099] S3. Based on the quality score of the local field reconstruction report and the local anomaly confidence level, decide whether to store or transmit only the report or package the raw data to generate a high-fidelity diagnostic data package; specifically including:

[0100] S31. Preset a first threshold and a second threshold, including:

[0101] 1. Upon initial deployment, sensor nodes receive initial configurations of a first threshold and a second threshold from the gateway node. The first threshold measures the quality and reliability of local field reconstruction reports, with a default initial value of 0.5; the second threshold determines the sensitivity of local anomaly confidence, with a default initial value of 0.7. Both thresholds support three levels of dynamic adjustment: basic, alert, and sensitive, corresponding to different stages of structural health status.

[0102] 2. During normal operation, after every 100 status determination cycles, the node automatically optimizes the threshold based on statistical results from the long-term cognitive database. If historical data shows a false alarm rate higher than 5%, the first threshold is appropriately relaxed or the second threshold is increased to reduce false alarms; if the false negative rate exceeds 3%, the opposite strategy is adopted to improve monitoring sensitivity.

[0103] The sensor nodes compare the quality score with a first threshold in real time, and compare the local anomaly confidence level with a second threshold, including:

[0104] 1. At the end of each sampling period, the node compares the quality score generated in step S23 with the first threshold in real time, and simultaneously compares the local anomaly confidence level calculated in step S24 with the second threshold. The comparison operation is implemented in the digital signal processor using integer arithmetic to avoid floating-point operation delays and ensure that the determination is completed within 0.1 milliseconds.

[0105] 2. When the quality score is greater than the first threshold and the local anomaly confidence level is less than the second threshold, the system enters a stable state branch; when the quality score is less than or equal to the first threshold, or the local anomaly confidence level is greater than or equal to the second threshold, or both conditions are met simultaneously, the system enters an anomaly state branch. A state-keeping counter is set up for each node to avoid data processing jitter caused by frequent switching. The corresponding data processing flow is only triggered when the judgment results are consistent three times consecutively.

[0106] S32. When the quality score is greater than the first threshold and the local anomaly confidence level is less than the second threshold, after determining that the state is stable, the sensor node performs the following operations:

[0107] S321. The node performs lossy compression on the field distribution data in the local field reconstruction report. Specifically, a feature dimension truncation strategy is adopted: the complete accuracy of the first three dominant feature components is retained, and the remaining secondary components are quantized at fixed points, with the quantization step size adaptively adjusted according to the dynamic range of the features.

[0108] S322. Simultaneously perform key parameter extraction, saving only the statistical characteristics of the field distribution rather than the complete grid data, including the coordinates of the maximum, minimum, mean, and gradient extreme points of the local response. The extracted data size is compressed from approximately 500 bytes to less than 50 bytes.

[0109] S323. The compressed field data and core metadata are combined into a complete data packet, including: a six-byte timestamp, twelve bytes of node location coordinates, a one-byte quality score, a one-byte local anomaly confidence score, up to fifty bytes of compressed field data, and a two-byte checksum. The total packet length is controlled within seventy-two bytes to adapt to the single-frame transmission capability of the wireless link.

[0110] S324. Critical metadata packets are written to a low-priority storage area of ​​local memory. This area uses ordinary flash memory blocks and supports 100,000 erase / write cycles. Storage operations use a batch write method, performing a physical write once every sixteen data packets to reduce flash memory wear. Storage addresses are used cyclically, automatically overwriting the oldest data when the area's limit is reached. This area is designed to retain at least 72 hours of historical data.

[0111] S325. Wireless transmission is performed in a non-real-time, best-effort mode. Nodes monitor channel status and only initiate transmission when carrier sensing indicates the channel is idle for more than twenty milliseconds. Data packets are transmitted at a low power of -10^30 dBmW to conserve energy, and acknowledgment retransmission is not enabled. If three transmissions fail due to channel congestion, the transmission is abandoned, and the node waits for the next cycle to retry. The transmission target is a gateway node or data aggregation node, using a unicast address but without end-to-end acknowledgment.

[0112] S33. When the quality score is less than or equal to the first threshold or the local anomaly confidence level is greater than or equal to the second threshold, after determining an abnormal state, the sensor node immediately performs the following operations:

[0113] S331. The node immediately extracts the original time-series data segment within the complete time window before and after the trigger moment from the circular buffer. The time window length is set according to the structural dynamics characteristics, and the default is two seconds before and one second after the trigger point, covering the complete transient process of the structural response. The buffer segment is locked during extraction to prevent new data from overwriting it. The corresponding fused data stream is extracted synchronously. This data stream was generated and cached in step S11 and is strictly aligned with the original data in time. The fused data stream has a lower dimension than the original data and is used for fast playback analysis.

[0114] S332. The data packet encapsulation adopts a layered structure: the first layer is the original time-series data segment, stored as a 16-bit signed integer, preserving the original precision of the analog-to-digital converter; the second layer is the fused data stream, using a single-precision floating-point format; the third layer is a complete local field reconstruction report, containing all grid point estimates and support point information; the fourth layer is detailed diagnostic metadata, including intermediate values ​​of anomaly confidence calculations, snapshots of neighbor node states, threshold configuration versions, and hardware status register values. During the encapsulation process, a 32-bit cyclic redundancy check (CRC) code is calculated for each data block, and synchronization marker bytes are padded between blocks to ensure that the data can still be parsed and recovered after transmission or storage corruption.

[0115] S333. High-fidelity diagnostic data packets are preferentially written to a protected storage area in local memory. This area uses high-endurance flash memory blocks, supports millions of erase / write cycles, and enables hardware-level error checking and correction. Storage operations are executed atomically: a backup copy is written first, and its integrity is verified. Only after confirmation of accuracy is the data committed to the main storage area to prevent data corruption due to power failure. This area is designed to store at least fifty sets of high-fidelity data packets and employs a hybrid FIFO and priority management strategy: high-priority data packets can overwrite lower-priority older data in advance, ensuring that records of extreme events are not overwritten by ordinary anomalies. Storage addresses are managed independently and are physically isolated from low-priority areas.

[0116] S334. Wireless transmission starts immediately or executes in the next high-priority communication slot. Nodes transmit at a maximum transmit power of ±5 dBmW to ensure reliable arrival under complex channel conditions. Carrier sensing is performed before transmission; if the channel is busy, a priority backoff algorithm is activated to preempt channel resources. Data packet transmission uses an acknowledgment and retransmission mechanism, retrying a maximum of five times. If all five attempts fail, the transmission is marked as suspended, and retrying every ten seconds until successful. The transmission target can be configured as multiple redundant gateway nodes to improve the transmission success rate. During transmission, nodes can temporarily increase the processor clock frequency to shorten the transmission preparation time, ensuring that abnormal data reaches the monitoring system within 500 milliseconds.

[0117] S34. To prevent frequent fluctuations near the critical threshold, nodes implement a delayed switching strategy. A transition from a stable state to an abnormal state requires three consecutive successful judgments; however, a return from an abnormal state to a stable state requires five consecutive successful judgments, ensuring the completeness of abnormal state capture.

[0118] When high-fidelity data packet transmission and low-priority data storage simultaneously request resources, nodes prioritize handling anomalous data. Digital signal processors suspend non-critical tasks, such as long-term cognitive analysis, allocating over 90% of their computing resources to anomalous data encapsulation and transmission. Static random access memory (SRAM) is prioritized for high-fidelity data packet buffers, while low-priority data packets are temporarily stored in the flash memory swap area.

[0119] The high-power mode in abnormal states has a limited duration, not exceeding thirty seconds. If the abnormal state persists, the node automatically switches to a degraded mode: reducing the sampling rate to one-third of the normal state, reducing neighbor broadcast frequency, and retaining only key feature extraction and transmission to ensure that battery-powered nodes can still operate for at least seventy-two hours under continuous abnormal conditions.

[0120] S4. Based on the accumulated local feature vectors, local field reconstruction reports, local anomaly confidence levels, and metadata from high-fidelity diagnostic data packages, analyze state trends and generate adaptive monitoring strategies, dynamically adjusting execution parameters; specifically including:

[0121] S41. The sensor node allocates at least 32 kilobytes of dedicated storage area in its local non-volatile memory, such as ferroelectric memory or high-endurance flash memory. This area is physically isolated from the conventional data storage area to prevent interference from erase and write operations. The storage area employs a wear leveling algorithm to evenly distribute write operations across different storage blocks, ensuring an overall erase and write lifespan of over one million cycles. The area is divided into four logical sub-areas: a feature vector archive area, a quality score log area, a confidence time series area, and a diagnostic event metadata area. Each sub-area manages read and write pointers independently.

[0122] The database stores data cyclically and cumulatively in chronological order, with timestamps serving as unique index keys. Local feature vectors are stored in 16-bit fixed-point format, with each vector occupying approximately 16 to 32 bytes of storage. The quality score of the local field reconstruction report is stored as an 8-bit unsigned integer, and the local anomaly confidence level also uses an 8-bit format. The trigger time and type metadata of the high-fidelity diagnostic data packet adopt a compact bit-field structure, with each record occupying only six bytes, including the absolute trigger time, anomaly type encoding, and data packet size level.

[0123] When the storage area reaches 90% capacity, an aging and eviction mechanism is activated: feature vectors of stable states exceeding 72 hours are compressed and archived, retaining only the hourly statistical average; data related to abnormal events are permanently retained until manually deleted. Nodes periodically calculate and store the cyclic redundancy check code for each data block, verifying integrity during reading, and automatically attempting to restore from the backup area if corruption is detected.

[0124] Write operations employ an atomic transaction mechanism: first, a temporary buffer is written, and a checksum is calculated. Once verified, the entire operation is committed to the main storage area. In the event of a sudden power outage, a supercapacitor maintains power for the last ten milliseconds to save the currently written page. Upon restart, the storage area is automatically scanned, and incomplete records are repaired based on timestamps and checksum information to restore database consistency.

[0125] S42. The sensor node has a built-in real-time operating system that triggers a lightweight analysis routine every ten minutes, or immediately after accumulating one thousand new records. Analysis tasks are allocated to idle time slots of the digital signal processor, with a priority set lower than real-time sampling but higher than regular data transmission, ensuring no impact on core monitoring functions. The time taken for a single analysis is controlled within twenty milliseconds, and memory usage does not exceed four bytes.

[0126] Each component of the local feature vector is analyzed independently over time. A moving average is calculated using a sliding window of fifty points to suppress high-frequency measurement noise. The trend line of the most recent two hundred sampling points is fitted using the linear least squares method to identify the slow rise or fall of the structural response level. A first-order autoregressive model is constructed to predict the next time step using the values ​​of the previous ten time steps. The model coefficients are updated online using the recursive least squares method.

[0127] The quality score time series is segmented and statistically analyzed. The average quality of the most recent fifty reconstructions is calculated. If the quality score remains below a threshold, the local field reconstruction model is considered to have failed. The standard deviation of the quality score is calculated to assess the reconstruction stability. The local anomaly confidence time series is classified into different modes: sporadic single-point anomalous pulses, periodic anomalous fluctuations, and continuously rising anomalous trends are identified, corresponding to impact events, periodic disturbances, and progressive damage, respectively.

[0128] The analysis routine outputs a confidence score for each trend item, based on factors including the continuity of data points, goodness of fit, and conformity to physical laws. If the confidence score of a trend is below 60%, that trend will not be used as a basis for strategy adjustment to prevent noise interference from leading to erroneous decisions. All analysis results are organized into three levels: feature dimension, time window, and confidence level, for use in step S43.

[0129] S43. Based on the long-term evolution trend of structural state parameters, nodes autonomously generate adaptive sampling plans for future time periods. If the trend indicates that the structural response level is predicted to increase by more than 20% within the next 30 minutes, the sampling frequency in step S1 is proactively increased from the default 1 kHz to 2 kHz; if the predicted response level remains below 50% of the normal range, the sampling frequency is reduced to 500 Hz to save energy. The adjustment is gradual, increasing or decreasing by 20% every five sampling periods to avoid signal distortion introduced by frequency abrupt changes.

[0130] Each node maintains a neighbor contribution score table, recording each neighbor's participation, data freshness, and quality improvement effect in the last one hundred collaborative reconstructions. Based on historical communication quality, nodes with a received signal strength consistently above -70 dBmW and a packet loss rate below 5% are prioritized. Based on collaborative reconstruction contribution, nodes with a feature vector correlation coefficient higher than 0.8 with the current node are prioritized. The priority list is refreshed every fifty periods; the top three neighbor nodes will be frequently requested to participate in reconstruction, while nodes ranked below fifth are temporarily removed when resources are scarce.

[0131] Based on the statistical change pattern of local anomaly confidence, the node autonomously adjusts the first and second thresholds in step S3. If the anomaly confidence distribution in the past two hundred judgments shows a long-tailed pattern, it indicates that the current threshold is too lenient, and the node automatically lowers the second threshold by 0.05 to improve detection sensitivity. If, after a high-confidence anomaly is triggered, the proportion of false alarms verified by subsequent high-fidelity data exceeds 10%, the second threshold is increased by 0.05 to reduce the false alarm rate. The adjustment of the first threshold is linked to the average quality score. If the average quality score is consistently higher than 0.8, the first threshold is moderately increased to 0.6 to reduce unnecessary raw data storage.

[0132] The newly generated adaptive monitoring strategy is stored in the strategy cache of random access memory, with a version number marked by a timestamp. The strategy does not take effect globally immediately, but is first A / B tested on a single feature dimension or a few neighbor nodes to compare the effects of the old and new strategies. If the test shows that the new strategy improves the data compression rate by more than 10% and does not increase the anomaly miss rate, it will be fully implemented; if the anomaly miss rate increases during the testing period, it will automatically roll back to the previous version of the strategy, and the reason for the failure will be recorded for long-term cognitive module learning. The strategy version history retains the most recent ten versions, and manual rollback to any historical version is supported.

[0133] Example 2

[0134] 1. Implementation Scenario and System Configuration: A long-span suspension bridge was selected as the test scenario. Thirty intelligent sensor nodes, as described in this invention, were deployed on the main girder of the bridge, forming a wireless sensor network. Each node integrated a triaxial MEMS accelerometer and a strain gauge. A gateway node was deployed at the bridge tower. The verification period was 30 consecutive days.

[0135] 2. Implementation process of the method of the present invention

[0136] S1. Each node synchronously acquires raw acceleration and strain data at a sampling rate of 100Hz. Principal component analysis (PCA) is used to fuse the triaxial accelerations to obtain a fused data stream in the principal vibration direction. From this fused data stream, a local feature vector containing the following three elements is extracted once per second: energy in the 0.5–2Hz frequency band, energy in the 2–5Hz frequency band, and the current principal strain value.

[0137] S2. Each node broadcasts its local feature vector and location code once per second. Simultaneously, each node receives the vectors of its three nearest neighbors. The node uses an inverse distance weighted method to spatially interpolate the frequency band energy of the four points, generating a local field reconstruction report of vibration energy in a local area of ​​the bridge deck, and calculates the fitting residual as a quality score. The node uses its neighbor data to predict its own characteristics through linear regression, and after normalizing the deviation between the predicted and actual values, obtains the local anomaly confidence level.

[0138] S3. Set the quality score threshold to 0.8 and the local anomaly confidence threshold to 0.7. Under most stable conditions, nodes only store and intermittently upload compressed local field reconstruction reports, containing only key parameters, with a data volume of approximately 0.5% of the original data.

[0139] S4. The node analyzes the local feature vectors and local anomaly confidence levels over the past 24 hours daily. For example, the node discovers that the local anomaly confidence level rises slightly and regularly during the daily evening traffic peak. Therefore, it autonomously generates a strategy: during the predicted traffic peak period, it temporarily increases the sampling rate of S1 from 100Hz to 200Hz and fine-tunes the local anomaly confidence threshold of S3 from 0.7 to 0.65 to improve monitoring sensitivity.

[0140] 3. Verification Tests and Results Comparison

[0141] To verify the effectiveness of this invention, the following three typical scenarios were simulated and compared with the traditional periodic full-sampling transmission method:

[0142] Scenario 1: Steady-state monitoring, continuous normal operation for 7 days

[0143] The advantages of this invention are: it operates in a "stationary state" most of the time, with nodes only exchanging local feature vectors and sparsely uploading local field reconstruction reports. The average daily wireless communication data volume is 15MB, and the average daily write volume to local storage on each node is 20MB. The theoretical battery life is extended to 180 days.

[0144] Traditional method results: All nodes continuously sample at 100Hz full rate and send all raw data back to the gateway via wireless network. The daily average wireless communication data volume reaches 3.1GB, with frequent writes to local storage as a cache. The theoretical battery life is only about 20 days.

[0145] Comparative conclusion: Under long-term steady-state monitoring, this invention achieves a reduction of over 99% in wireless data transmission volume and a nearly 9-fold increase in battery life, demonstrating extremely significant advantages in resource efficiency.

[0146] Scenario 2: Localized Abnormal Event - On the 8th day, a sudden impact occurred on a section of the bridge surface due to cargo falling from a vehicle.

[0147] The effect of this invention: The local anomaly confidence levels of the three nodes A, B, and C closest to the impact point instantly exceed the threshold. The quality score of node A drops sharply due to the drastic response space gradient. All three immediately trigger the S3 anomaly process, packaging the original time-series data segment (5 seconds before and after the impact), the fused data stream, and a detailed local field reconstruction report into a high-fidelity diagnostic data packet, which is transmitted to the gateway with high priority within 2 seconds. The gateway receives complete diagnostic information with precise spatiotemporal labels, facilitating accurate location and analysis.

[0148] Traditional methods involve all nodes continuously uploading massive amounts of data. Data from impact events is overwhelmed by the data stream, requiring full data backtracking and analysis in the cloud to discover and locate the event. This process can take hours and consumes significant computing and storage resources.

[0149] Comparative conclusion: This invention realizes the transformation from "finding a needle in a haystack" to "delivering a needle with intelligent tags", significantly improving the real-time capture of abnormal events, data correlation, and diagnostic completeness.

[0150] Scenario 3: Long-term trend learning and adaptation, validating the entire cycle.

[0151] The advantages of this invention are: Nodes gradually learn the differences in vibration patterns between weekdays and holidays, and between daytime and nighttime through S4, and automatically adjust their monitoring strategies. For example, during the low-activity period in the early morning, some nodes automatically reduce their sampling rate to 50Hz, further saving energy. When the local anomaly confidence baseline of a node shows a slow but continuous upward trend over several days, even if it does not exceed the threshold, the system marks that area as a "key area of ​​concern" in advance.

[0152] Traditional methods are ineffective because their monitoring strategies remain consistent and fail to detect or adapt to long-term structural changes and environmental cycles. They are also insensitive to slowly developing damage trends until they trigger significant threshold exceedance events.

[0153] Comparative conclusion: This invention endows the system with preventative sensing capabilities and intelligent resource allocation on demand, while traditional methods are merely rigid, one-size-fits-all monitoring.

[0154] 4. Overall Effect Comparison Table

[0155] The following table 1 shows a comparison of the effects of conventional wireless sensor network methods in the prior art and the method described in this invention:

[0156] Table 1: Comparison of Results

[0157]

[0158] 5. Conclusion: This embodiment demonstrates that the method proposed in this invention enables the system to continuously collect data, analyze its status, make autonomous decisions, and optimize its operation, forming a complete intelligent cycle. This fundamentally changes the working paradigm of wireless sensor networks in structural health monitoring. It not only achieves orders-of-magnitude improvements in fundamental performance indicators such as communication, storage, and energy consumption, but more importantly, it endows the network with unprecedented deep state perception capabilities, intelligent autonomous decision-making capabilities, and forward-looking adaptive capabilities. This provides a practical and innovative solution to the challenge of long-term, efficient, and intelligent health monitoring of large-scale infrastructure.

[0159] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects:

[0160] This invention achieves a fundamental shift from raw data transmission to high-value information exchange, improving system resource efficiency and data availability. Before data transmission, the nodes in this invention perform local feature extraction (S1) and collaborative field reconstruction (S2), refining massive, low-information-density raw sensor readings into local feature vectors representing structural states and local field reconstruction reports reflecting local continuous states. Based on this, through an adaptive data selective transmission mechanism (S3), the system transmits highly compressed cognitive results only when the state is stable, triggering full data capture only in abnormal situations. This series of operations fundamentally changes the objects of wireless network transmission and local storage. Effectively, this directly results in a reduction of over 95% in conventional wireless data traffic and an increase in effective storage space utilization by more than an order of magnitude, thereby extending the sensor network's endurance several times over. More importantly, the data transmitted to the backend is itself structured information that has undergone preliminary analysis and cognition, such as features, field distribution, and anomaly confidence, rather than raw byte streams requiring secondary processing. This significantly reduces the cloud processing load and allows monitoring personnel to understand and apply the data more quickly and intuitively.

[0161] This invention breaks through the limitations of traditional point-based monitoring, endowing it with the capabilities of collaborative sensing and intelligent anomaly diagnosis in cyberspace, significantly improving monitoring quality and reliability. By introducing the collaborative field reconstruction and anomaly assessment S2 step, this invention solves the long-standing problem of difficulty in sensing the structural "state field" under sparse monitoring. Each node utilizes the local feature vectors of its neighbors to estimate a continuous local field reconstruction report locally, achieving a leap from discrete point monitoring to local continuous field sensing. Simultaneously, by calculating the deviation between its own data and the collaborative expected data, the local anomaly confidence level is obtained, upgrading the judgment criterion from the traditional single amplitude threshold to a physical logic based on local spatial consistency. This makes the system highly sensitive to weak anomalies inconsistent with surrounding patterns, while effectively reducing false alarms caused by environmental noise. In terms of results, the system can provide a spatial distribution map of structural response and detect potential problems earlier and more accurately, providing an unprecedented high-quality data foundation for damage localization and state assessment.

[0162] An autonomous system with memory, learning, and predictive capabilities was constructed, enabling dynamic optimal matching and proactive adjustment of monitoring strategies. This invention elevates the system's intelligence from immediate response to continuous evolution through the addition of a predictive adaptive monitoring strategy generation S4 step. Nodes, by maintaining a long-term cognitive database locally and analyzing historical trends, can autonomously deduce the evolution trend and anomaly patterns of the structural state, thereby generating adaptive monitoring strategies and dynamically adjusting future data collection frequency, neighbor collaboration priority, and discrimination thresholds. This solves the deep-seated problems of strategy rigidity and static resource allocation in long-term monitoring. Like an experienced engineer, the system can proactively enter a low-power mode during stable periods and actively increase the monitoring level before predicted risk windows, based on the structure's own operating rules and environmental changes. This makes the entire network an organic whole capable of self-learning, self-optimization, and on-demand resource allocation, achieving global minimization of lifecycle costs and energy consumption while ensuring or even improving monitoring efficiency.

[0163] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the present invention, but these should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0164] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for wireless sensor data transmission and storage based on distributed nodes, characterized in that, Includes the following steps: The collected raw time-series data is processed to generate local feature vectors that characterize the local structural response; The system receives the local feature vectors from neighboring nodes, combines them with its own vector to reconstruct the local structural response field, generates a local field reconstruction report, and calculates the local anomaly confidence level; specifically including: Sensor nodes periodically receive data packets broadcast by at least one neighboring node within their communication range via a wireless link. The data packets contain at least the local feature vectors, timestamps, and structural location information of the neighboring nodes. The sensor node spatiotemporally aligns and combines its own local feature vector with the received local feature vectors of neighboring nodes to form an extended feature set; Based on the extended feature set, the sensor node runs a built-in simplified field reconstruction algorithm to estimate the structural response field distribution of the local area covered by the sensor node and its neighboring nodes, and generates a local field reconstruction report. The built-in simplified field reconstruction algorithm includes: taking the extended feature set as input, using a physics-driven or data-driven spatial interpolation method to estimate the structural response between discrete points, generating a local field reconstruction report containing a continuous or quasi-continuous field distribution map of the local area, and performing a self-evaluation of the smoothness or fitting error of the reconstruction result, and outputting a quality score. Based on the extended feature set, the sensor node infers the expected feature vector of its own node position; at the same time, the sensor node calculates the local anomaly confidence by comparing its own local feature vector with the inferred expected feature vector. Based on the quality score of the local field reconstruction report and the local anomaly confidence level, a decision is made to either store or transmit the report or package the raw data to generate a high-fidelity diagnostic data package. Based on accumulated local feature vectors, local field reconstruction reports, local anomaly confidence levels, and metadata from high-fidelity diagnostic data packages, the system analyzes state trends and generates adaptive monitoring strategies, dynamically adjusting execution parameters; specifically including: The sensor node maintains a long-term cognitive database locally to accumulate and store metadata such as local feature vectors, local field reconstruction reports, local anomaly confidence scores, and high-fidelity diagnostic data packets generated during each execution. Sensor nodes periodically analyze time-series data in a long-term cognitive database to deduce the evolution trend of the structural state of the local area under their responsibility and the changing pattern of local anomaly confidence. Based on evolution trends and change patterns, sensor nodes autonomously generate an adaptive monitoring strategy for a future period.

2. The wireless sensor data transmission and storage method based on distributed nodes according to claim 1, characterized in that, The process of processing the acquired raw time-series data to generate a local feature vector representing the local structural response specifically includes: Sensor nodes synchronously acquire multidimensional raw time-series data from their own sensing units and perform local fusion processing on the multidimensional raw time-series data to generate a multidimensional fused data stream. The sensor node performs real-time feature extraction on the fused data stream to obtain a low-dimensional local feature vector, which is used to characterize the structural response state at the current node's location.

3. The wireless sensor data transmission and storage method based on distributed nodes according to claim 1, characterized in that, The process of forming an extended feature set includes: the sensor node interpolates or aligns its own local feature vectors with those of each neighboring node to the same reference time based on the timestamp, and sorts the aligned feature vectors in space according to the node position information to form a spatial-feature matrix representing the state of discrete points in the local area, which serves as the extended feature set.

4. The wireless sensor data transmission and storage method based on distributed nodes according to claim 3, characterized in that, The calculation of local anomaly confidence includes: deriving the expected feature vector of its own node position from the extended feature set through a field reconstruction algorithm or a lightweight regression model; comparing the actual local feature vector extracted by the sensor node with the expected feature vector, calculating the Euclidean distance, cosine similarity or residual norm, and normalizing the calculation result as the local anomaly confidence.

5. The wireless sensor data transmission and storage method based on distributed nodes according to claim 1, characterized in that, Based on the quality score and local anomaly confidence level of the local field reconstruction report, a decision is made to either store or transmit the report or package the raw data to generate a high-fidelity diagnostic data package, specifically including: A first threshold and a second threshold are preset; the sensor node compares the quality score with the first threshold in real time, and compares the local anomaly confidence level with the second threshold; When the quality score is greater than the first threshold and the local anomaly confidence is less than the second threshold, it is determined that the local structure response state is stable and the local field reconstruction report is reliable and can be effectively reconstructed. Then, the local field reconstruction report and key metadata are compressed and stored in the local memory and transmitted at an opportune time. When the quality score is less than or equal to the first threshold or the local anomaly confidence is greater than or equal to the second threshold, it is determined that the local structure response state is abnormal or difficult to reconstruct reliably. Then, the original time-series data segment that triggered the determination, the corresponding fused data stream, and the local field reconstruction report are packaged together to generate a high-fidelity diagnostic data packet, which is preferentially written to the local memory and wireless transmission is started immediately or with high priority.

6. The wireless sensor data transmission and storage method based on distributed nodes according to claim 1, characterized in that, The periodic analysis of time-series data in the long-term cognitive database calls the built-in lightweight analysis routines to perform time-series analysis on the local feature vectors, quality scores, and local anomaly confidence levels in the long-term cognitive database; the evolution trend of the structural state of the local area under the responsibility of the sensor node and the change pattern of the local anomaly confidence level are derived using moving average, trend fitting, or autoregressive models.

7. The wireless sensor data transmission and storage method based on distributed nodes according to claim 6, characterized in that, The autonomous generation of an adaptive monitoring strategy for a future period includes: predicting the structural response level in the future period based on the evolution trend, and dynamically adjusting the data collection frequency accordingly; updating the neighbor node selection priority list based on the historical communication quality and collaborative reconstruction contribution between nodes; and adaptively adjusting the first and second thresholds used for judgment based on the change pattern and statistical distribution of local anomaly confidence.

Images