An internet of things laboratory data acquisition optimization processing method

Abstract

The application relates to the technical field of Internet of Things data acquisition and processing, and discloses a kind of Internet of Things laboratory data acquisition optimization processing method, method includes: extracting the steady-state phase offset constant and DC baseline drift of non-trigger period background noise for cloud template correction;When detecting the reference parameter difference gradient overrun by continuous sampling, extract data to generate event original data matrix;Extract the residual curve by reconstructing the trial sequence of basic relative time difference characteristic vector;Establish inverse cause-effect time window to perform physical energy closed-loop verification, and generate compensation characteristic vector by extracting the curvature inflection point of residual that passes the verification;Two types of vectors are spliced and sent by NB-IoT;Cloud reconstructs the basic waveform using the corrected standard time domain response template, fits the compensation characteristic vector to generate a compensation curve, and outputs a high-frequency data sequence by superposition. The application solves the problem of narrowband communication limitation and baseline drift leading to reconstruction distortion, effectively distinguishes and retains the real physical distortion characteristics.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) data acquisition and processing technology, specifically to an IoT laboratory data acquisition and optimization processing method. Background Technology

[0002] In IoT lab testing environments, multi-dimensional sensor arrays are typically deployed to capture high-frequency transient physical response data in real time. Due to the high sampling rate and numerous monitoring dimensions, the amount of data generated is substantial. However, in real-world applications, narrowband IoT communication technologies such as NB-IoT are commonly used. Their limited communication payload makes it difficult to support the direct transmission of multi-channel, high-frequency raw data; forcing transmission can easily lead to network congestion and data loss. Existing data acquisition solutions often reduce the sampling rate or use simple threshold filtering to decrease the amount of communication data, but this directly results in the loss of crucial physical information during the transient response process.

[0003] Meanwhile, during long-term operation, sensor nodes deployed in the field inevitably experience DC baseline drift and phase synchronization errors between multiple channels due to environmental background noise and the aging of their own components. When the acquisition system directly processes and uploads this data with accompanying steady-state errors, the amount of error accumulates over time in subsequent calculations and waveform restoration, leading to physical reference distortion in the final data solution.

[0004] Furthermore, when capturing complex physical events in the laboratory (such as structural yielding, impact, etc.), the waveforms acquired by sensors often contain nonlinear distortion information. Due to the complex electromagnetic interference and measurement noise present in the field, existing edge processing programs often struggle to accurately identify the excitation sources causing waveform abrupt changes. This leads to conventional smoothing filtering algorithms easily filtering out physically valuable true distortion features, or misclassifying spurious residuals generated by random noise as real physical events without processing, resulting in a lack of physical authenticity and reliability in the final collected data sequence. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an optimized processing method for IoT laboratory data acquisition, which solves the problems of limited high-frequency multidimensional data transmission in narrowband communication environments, waveform reconstruction distortion caused by baseline drift, and difficulty in effectively identifying and preserving real physical distortion features due to complex measurement noise.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing and processing data acquisition in an IoT laboratory, comprising the following steps:

[0007] The discrete environmental background noise sequence under non-triggering period is obtained, the steady-state phase offset constant and DC baseline drift are extracted and reported, and used by the cloud reconstruction server to generate the corrected standard time domain response template.

[0008] Continuous sampling is performed on the multi-dimensional sensor array. When the absolute value of the discrete first-order differential gradient of the reference parameter is detected to be greater than the preset transient trigger threshold, the data segment is locked and extracted based on the trigger time to generate the original event data matrix.

[0009] The basic relative time difference feature vector is extracted from the original event data matrix. The trial sequence is reconstructed using the internally preset low-order simplified time domain response template and the basic relative time difference feature vector. The difference is then calculated with the actual sample to obtain the positive residual curve.

[0010] A reverse causal time window is established for the positive residual curve to perform physical energy closed-loop verification. Curvature inflection point features are extracted from the residual waveform that passes the verification, and a compensation feature vector is generated.

[0011] The basic relative time difference feature vector and the compensation feature vector are concatenated to generate a transmission message, which is then sent to the cloud reconstruction server through the NB-IoT communication module.

[0012] The cloud-based reconstruction server uses the corrected standard time-domain response template and the basic relative time difference feature vector to reconstruct the basic follow-up parameter waveform, fits the compensation feature vector to generate a continuous residual compensation curve, performs numerical superposition and maps it to the global time axis, and outputs a synchronously reconstructed multivariate high-frequency data sequence.

[0013] Furthermore, the specific steps for extracting the steady-state phase offset constant and the DC baseline drift are as follows:

[0014] Within the set steady-state sliding calculation window, perform discrete cross-correlation numerical calculations on the corresponding follower parameter sequence and the reference parameter sequence respectively, and output the discrete cross-correlation function;

[0015] The discrete sampling delay point value corresponding to the maximum value of the discrete cross-correlation function is retrieved and mapped to the deviation value on the time scale to define the steady-state phase offset constant.

[0016] Calculate the arithmetic mean of the follower parameter sequence within the steady-state sliding calculation window, and calculate the difference between the arithmetic mean and the preset zero-point baseline value, the historical calibration baseline value, or the baseline reference value sent from the cloud to generate the DC baseline drift.

[0017] Furthermore, the specific steps for locking and extracting data segments based on the trigger time to generate the original event data matrix are as follows:

[0018] A circular buffer is configured in the internal random access memory. The discrete digital signals output from each channel are written into the circular buffer, and historical data is cyclically overwritten when the memory write address is incremented to the end of the physical boundary.

[0019] When the absolute value of the discrete first-order differential gradient of the reference parameter is detected to be greater than the preset transient trigger threshold, the memory offset address corresponding to the trigger moment in the circular buffer is recorded, and the global logical time index corresponding to the current trigger moment is recorded synchronously.

[0020] Continue writing data to the circular buffer until the accumulated duration reaches the specified duration, trigger the internal interrupt latch mechanism to lock the circular buffer, use the trigger time point address as a reference to determine the start data boundary and end data boundary, extract the data segments within the specified duration before and after the trigger point, and read them out in segments and splice them to generate the original event data matrix.

[0021] In a preferred embodiment of the present invention, the specific steps for extracting the basic relative time difference feature vector from the original event data matrix are as follows:

[0022] In the data sequence of the reference parameters of the original event data matrix, the discrete sampling point with the largest absolute amplitude is retrieved, and the timestamp corresponding to the discrete sampling point with the largest absolute amplitude is designated as the local zero time.

[0023] Traverse the data sequence of each follower parameter to find the extreme response time, calculate the time difference between the extreme response time and the local zero time to generate the time difference delay parameter, and extract the corresponding amplitude to generate the extreme amplitude parameter.

[0024] All time delay parameters and extreme amplitude parameters corresponding to the follow-up parameters are concatenated into an array to generate the basic relative time difference feature vector.

[0025] Furthermore, the specific steps for establishing a reverse causal time window for the positive residual curve and performing physical energy closed-loop verification are as follows:

[0026] When it is detected that the absolute value of consecutive sampling points in the positive residual curve is greater than the distortion threshold, the extreme point with the largest absolute value of the residual is recorded as the suspected distortion moment point.

[0027] Based on the suspected distortion time point, and combined with the steady-state phase offset constant and the set effective causal observation duration, the reverse causal time window is established by tracing back along the historical direction of the time axis.

[0028] Within the reverse causal time window, the sampled values ​​of the reference parameter are subjected to square integral operation to calculate the integral of the reference parameter fluctuation energy.

[0029] When the integral of the energy fluctuation of the reference parameter is greater than the preset excitation energy threshold, it is determined that the corresponding residual has a definite preceding physical excitation source and belongs to a real physical distortion that conforms to causal logic, and the corresponding residual waveform verification passes; when the integral of the energy fluctuation of the reference parameter is not greater than the preset excitation energy threshold, it is determined that the corresponding residual is a pseudo residual that lacks physical excitation support and is filtered out.

[0030] Furthermore, the specific steps for extracting curvature inflection point features from the verified residual waveform and generating a compensation feature vector are as follows:

[0031] For the effective positive residual curves that are determined to be real physical distortions, calculate the first derivative sequence and the second derivative sequence in sequence.

[0032] Traverse and search for sampling points that satisfy the zero-crossing feature, that is, the product of the second-order difference value corresponding to the current sampling point and the second-order difference value corresponding to the previous sampling point is not greater than zero, and the absolute value of the corresponding residual is greater than the distortion threshold, and extract them as the curvature inflection points.

[0033] Record the time offset of the curvature inflection point relative to the local zero moment and the residual amplitude of the curvature inflection point in the positive residual curve. Combine the time offset and the residual amplitude to generate the compensation feature vector containing the corresponding time offset and residual amplitude.

[0034] Furthermore, the step of generating the modified standard time-domain response template is as follows:

[0035] The cloud-based reconstruction server extracts the received steady-state phase offset constant and DC baseline drift, and retrieves the standard time-domain response template of the corresponding follower parameter preset in the time-series database;

[0036] The steady-state phase offset constant is substituted as a phase compensation term into the time independent variable of the standard time-domain response template to perform time axis alignment. The DC baseline drift is updated to the metadata tag associated with the standard time-domain response template as a basic DC bias term to generate the corrected standard time-domain response template.

[0037] Furthermore, the specific steps for concatenating the basic relative time difference feature vector and the compensation feature vector to generate the transmission message are as follows:

[0038] The NB-IoT communication module obtains the reference network absolute timestamp issued by the cellular network, and the determined local zero time is numerically bound to the reference network absolute timestamp to generate a global event timestamp.

[0039] Write the global event timestamp and device hardware identifier into the message header;

[0040] For the basic relative time difference feature vector, it is directly transferred into fixed-length data blocks according to the channel index;

[0041] For the compensation feature vector, when the total byte length of the generated transmission message is calculated to be no greater than the preset maximum transmission load threshold, the compensation feature data with the length flag indicating the total number of compensation feature points of the corresponding channel is appended to the end of the data frame in a preset byte order.

[0042] In a preferred embodiment of the present invention, the specific steps for reconstructing the base follow-up parameter waveform using the modified standard time-domain response template and the base relative time difference feature vector are as follows:

[0043] Amplitude normalization is performed on the retrieved modified standard time-domain response template;

[0044] The amplitude of the normalized, corrected standard time-domain response template is scaled proportionally using the obtained extreme amplitude parameters.

[0045] Phase shift is performed on the relative time axis using the obtained time difference delay parameters;

[0046] The retrieved DC baseline drift is used to linearly superimpose constant bias terms to restore the true static physical baseline, complete the equivalent convolution operation, and generate the basic follower parameter waveform.

[0047] Furthermore, the specific steps for fitting the compensation feature vector to generate a continuous residual compensation curve are as follows:

[0048] The cubic spline interpolation function is called to fit and calculate the coordinates of the discrete curvature inflection points in the compensated feature vector;

[0049] Before performing the interpolation calculation, time coordinates corresponding to the start and end points of the event reconstruction time window are added to the first and last ends of the set of discrete curvature inflection point coordinates, respectively, and the residual amplitude at the start and end points is forcibly set to zero, thereby forming a complete set of interpolation constraints.

[0050] In the process of solving the cubic spline interpolation equation system, natural boundary conditions are adopted, that is, the second derivative values ​​at the two endpoints of the fitted curve are forced to be zero, and the continuous residual compensation curve is calculated and output.

[0051] When the total number of compensation feature points in the corresponding channel is equal to 2, the two discrete points are connected by a first-order linear interpolation function; when the total number of compensation feature points in the corresponding channel is less than 2, the zero value sequence is directly output.

[0052] This invention provides a method for optimizing data acquisition and processing in an IoT laboratory. It offers the following advantages:

[0053] 1. This invention extracts the basic relative time difference feature vector and residual compensation feature vector at the edge node, encapsulates the dimensionality-reduced feature data into a transmission message for reporting, and then uses the cloud server to perform continuous waveform reconstruction by combining the standard time domain response template and cubic spline interpolation algorithm. This reduces the byte load of a single data report and solves the problem of network congestion and packet loss caused by directly transmitting high-frequency multidimensional physical sampling sequences in narrowband IoT networks. It achieves data restoration of multivariable high-frequency original waveforms under limited communication bandwidth.

[0054] 2. This invention extracts the steady-state phase offset constant and DC baseline drift by calculating the difference between the extreme value and the arithmetic mean of the discrete cross-correlation function under non-triggering period. The above parameters are then used to correct the standard time-domain response template on the cloud reconstruction server. This can eliminate the static baseline drift and phase synchronization error between channels caused by environmental interference and long-term operation of the field multi-dimensional sensor array, prevent the basic steady-state error from being amplified during the numerical superposition process of the cloud waveform reconstruction, and improve the accuracy of data solution.

[0055] 3. This invention establishes a reverse causal time window for physical energy closed-loop verification based on the positive residual curve generated by edge trial calculations, and extracts curvature inflection point features only for waveform intervals that pass verification using the second-order difference zero-crossing characteristic. By calculating the fluctuation energy integral of the reference parameter within the preceding time window as the judgment criterion, it can identify and retain key nonlinear real physical distortion information such as structural yielding, while filtering out pseudo residual features caused by random electromagnetic interference or measurement noise, ensuring the physical authenticity of the final output high-frequency data sequence. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0057] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0058] Figure 3 This is a schematic diagram of the non-triggering period steady-state baseline calibration principle of the present invention;

[0059] Figure 4 This is a flowchart of the high-frequency synchronous sampling and transient event data capture process of the present invention;

[0060] Figure 5 This is a schematic diagram illustrating the principle of relative time difference feature extraction and causal verification compensation in this invention.

[0061] Figure 6 This is a flowchart of the feature message splicing and network reporting process of the present invention;

[0062] Figure 7 This is a flowchart of the cloud-based multi-parameter waveform synchronous reconstruction process of the present invention;

[0063] Figure 8 This is a comparison chart of the network transmission performance of different data reporting schemes of the present invention;

[0064] Among them, 10 is the edge acquisition node; 101 is the multi-dimensional sensor array; 102 is the microcontroller unit; 103 is the NB-IoT communication module; 20 is the cloud reconstruction server; 201 is the time series database; and 202 is the waveform calculation engine. Detailed Implementation

[0065] The technical solutions in 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.

[0066] See attached document Figure 1 The present invention provides an IoT laboratory data acquisition and optimization processing system, which may include an edge acquisition node 10 and a cloud reconstruction server 20.

[0067] Edge acquisition node 10 is deployed in an IoT laboratory. Edge acquisition node 10 includes a multi-dimensional sensor array 101, a microcontroller unit 102, and an NB-IoT communication module 103. The multi-dimensional sensor array 101 is electrically connected to the microcontroller unit 102.

[0068] The multi-dimensional sensor array 101 includes N+1 synchronous acquisition physical channels, which are configured with one reference parameter sensor and N follower parameter sensors. The reference parameter sensor is the physical quantity measurement channel with the earliest response time and highest signal-to-noise ratio to transient physical events, or the one installed closest to the transient excitation source, used to characterize the main excitation time of the transient physical event. The follower parameter sensors are physical quantity measurement channels that are affected by the same transient physical event as the reference parameter sensor and have a definite timing response relationship with the reference parameter. The microcontroller unit 102 is connected to the NB-IoT communication module 103.

[0069] The microcontroller unit 102 integrates an analog-to-digital converter (ADC) and random access memory (RAM) to perform hardware synchronous sampling, buffer management, and feature data extraction logic. The ADC is either a multi-channel synchronous sampling ADC or the microcontroller unit 102 is connected to a multi-channel sample-and-hold circuit and ADC channels triggered by the same hardware timer. When using a multiplexed ADC for time-division scanning sampling, the fixed scanning delay between each physical channel is pre-calibrated and used as a channel sampling phase compensation factor in subsequent time difference delay parameter calculations. The NB-IoT communication module 103 establishes a communication link with the cloud reconstruction server 20 via a cellular communication network.

[0070] The cloud-based reconstruction server 20 is deployed in an external network environment. The cloud-based reconstruction server 20 includes a time-series database 201 and a waveform calculation engine 202. The time-series database 201 stores standard time-domain response templates corresponding to each physical parameter. The waveform calculation engine 202 receives data packets uploaded by the edge acquisition node 10 and performs multi-channel waveform reconstruction operations with servo parameters aligned with the reference parameter time.

[0071] A standard time-domain response template is specifically used to characterize the response shape data of a physical parameter channel as it changes over time after being subjected to a standard transient excitation. This standard time-domain response template is a set of time-domain response reference data stored in association with the device hardware identifier, sensor channel number, physical parameter type, sampling frequency, standard excitation type, environmental conditions, and template version number.

[0072] Specifically, a standard time-domain response template includes at least the template start time, template end time, template sampling interval, normalized response amplitude sequence, time location of template extrema, amplitude direction of template extrema, applicable physical parameter type, and applicable experimental scenario identifier. For physical parameter channels requiring baseline compensation, the standard time-domain response template also stores the zero-point baseline reference value, steady-state noise reference range, and the time information of the template's most recent update.

[0073] The standard time-domain response template can be obtained through factory calibration. During factory calibration, the edge acquisition node 10 is connected to the corresponding multi-dimensional sensor array 101, and a standard transient excitation is applied to the system under known experimental conditions. This standard transient excitation can be a standard impact, a standard pressure step, a standard temperature pulse, a standard vibration excitation, a standard light pulse, or other known excitation signals that can be stably sensed by the corresponding physical parameter channels. The microcontroller unit 102 acquires complete raw response data of the reference parameter channel and each follower parameter channel at the same sampling frequency as the actual working state, and uploads the complete raw response data to the cloud reconstruction server 20. It should be noted that the upload of complete raw response data mainly occurs during the factory calibration stage, the laboratory debugging stage, or under temporary high-bandwidth communication conditions; during the actual normal operation stage, the edge acquisition node 10 uploads a low-data-volume transmission message composed of the basic relative time difference feature vector and the compensation feature vector through the NB-IoT communication module 103.

[0074] After receiving the complete raw response data, the waveform calculation engine 202 first determines the time reference point of the template based on the extreme points or preset trigger points of the reference parameter channels. Then, it truncates the raw response data of each follower parameter channel, ensuring that each template includes the steady-state segment before the excitation, the rising or falling segment during the excitation, the main response segment, and the recovery segment where the response decays to near the steady state. Subsequently, the waveform calculation engine 202 performs outlier removal, smoothing, baseline correction, and amplitude normalization on the truncated response data, ensuring that the template primarily retains the time-domain response shape of the channel to the standard transient excitation.

[0075] To avoid the impact of random noise during a single calibration on the template accuracy, the waveform calculation engine 202 can repeatedly acquire multiple sets of response data obtained under the same device, channel, sampling frequency, and standard excitation conditions. For the multiple sets of response data obtained through repeated acquisition, the waveform calculation engine 202 first removes data sets with significantly deviated extreme value times, significantly abnormal amplitude fluctuations, or noise levels exceeding a preset range. Then, it performs time alignment and amplitude normalization processing on the remaining data sets, ultimately generating a standard time-domain response template for that channel. The generated standard time-domain response template is written into the time-series database 201 and indexed according to the device hardware identifier, channel number, physical parameter type, sampling frequency, and experimental scenario type.

[0076] In actual operation, the standard time-domain response template can also be updated using historical complete event data. Specifically, when the edge acquisition node 10 uploads a complete event raw data matrix during the debugging phase or under high-bandwidth communication conditions, the waveform calculation engine 202 can use this complete event raw data matrix as a template update sample. The waveform calculation engine 202 determines whether the event data meets the signal-to-noise ratio requirements, sampling integrity requirements, and response attenuation integrity requirements. If it does, it performs time alignment, baseline correction, and amplitude normalization on the servo parametric response waveform in the event data, and merges and updates it with the currently valid template to generate a new template version. If it does not meet the requirements, the template is not updated, and the event data is only used as ordinary historical data storage.

[0077] When multiple edge acquisition nodes 10 of the same model, installation structure, and sensor configuration are used in the same experimental scenario, the standard time-domain response template can also be obtained by migrating calibration data from the same model of equipment. During migration, the waveform calculation engine 202 first calls the standard time-domain response template of the same model of equipment as the initial template, and then corrects the initial template based on the steady-state phase offset constant and DC baseline drift obtained by this edge acquisition node 10 during the non-trigger period, thereby obtaining a standard time-domain response template suitable for the current equipment.

[0078] Therefore, the methods for obtaining the standard time-domain response template include at least one or more of the following: factory standard excitation calibration, known excitation experimental calibration, fitting of historical complete event data, and template migration from the same model of equipment. Through these methods, the standard time-domain response template stored in the time-series database 201 can provide a clear time-domain response benchmark for subsequent waveform reconstruction in the cloud using the basic relative time difference feature vector and the compensation feature vector.

[0079] For example, in one specific embodiment, the multi-dimensional sensor array 101 is applied to a structural component impact response monitoring scenario in an IoT laboratory. The reference parameter sensor is an impact force sensor, the first follower parameter sensor is a vibration acceleration sensor, and the second follower parameter sensor is a strain sensor. During factory calibration, the experimenter applies a transient impact of known amplitude and fixed direction to a designated location on the structural component using a standard impact hammer. The microcontroller unit 102 synchronously acquires complete response data from the impact force channel, vibration acceleration channel, and strain channel at a fixed sampling frequency and uploads the complete response data to the cloud reconstruction server 20.

[0080] The waveform calculation engine 202 uses the maximum absolute amplitude point of the impact force channel as the time reference point to extract response data from the vibration acceleration channel and strain channel respectively. For the vibration acceleration channel, the extracted data includes a static vibration baseline before impact, a rapidly rising main peak segment after impact, a subsequent rebound segment, and a tail segment gradually decaying to a stationary state. After removing the zero-point offset, the waveform calculation engine 202 normalizes the vibration acceleration response data according to the maximum absolute amplitude to obtain a standard time-domain response template for the vibration acceleration channel. This template records the complete time-varying shape of the vibration acceleration channel under standard impact, from the static baseline, rapid rise, reaching the main peak, rebound, to damping decay.

[0081] For the strain channel, the extracted data includes the steady-state strain baseline before impact, the relatively delayed strain rise segment after impact, the strain peak holding segment, and the slow recovery segment. The waveform calculation engine 202 performs baseline correction and amplitude normalization on this strain response data to obtain a standard time-domain response template for the strain channel. This template records the response characteristics of the strain channel relative to the impact force channel, which exhibits a certain time lag, a slower rise rate, and a longer recovery process.

[0082] In subsequent actual detection of transient impact events, edge acquisition node 10 no longer uploads the complete high-frequency raw data of vibration acceleration channel and strain channel, but uploads the time difference delay parameter, extreme amplitude parameter and compensation feature vector of each follower parameter channel relative to the reference parameter channel.

[0083] The waveform calculation engine 202 retrieves the standard time-domain response templates of the vibration acceleration channel and strain channel from the time-series database 201, performs time translation on the templates according to the uploaded time difference delay parameters, scales the templates according to the extreme amplitude parameters, performs baseline offset superposition according to the DC baseline drift of the corresponding channel, and generates residual compensation curves by combining the compensation feature vectors, thereby reconstructing the high-frequency response waveforms of the vibration acceleration channel and strain channel in this actual transient event.

[0084] See attached document Figure 2 This invention provides a method for optimizing data acquisition in an IoT laboratory, which may include:

[0085] Step S1: Non-triggering period steady-state baseline calibration. Edge acquisition node 10 acquires the discrete environmental background noise sequence output by multi-dimensional sensor array 101. Microcontroller unit 102 calculates the cross-correlation function of each follower parameter relative to the reference parameter, extracts the steady-state phase offset constant, and extracts the DC baseline drift based on the difference between the arithmetic mean of each follower parameter within the steady-state window and the corresponding channel baseline reference value. NB-IoT communication module 103 uploads the steady-state phase offset constant and DC baseline drift to cloud reconstruction server 20 using application layer periodic keep-alive messages, periodic status reporting messages, or NB-IoT small data transmission messages. Waveform calculation engine 202 corrects the corresponding standard time-domain response template in timing database 201 based on the received parameters.

[0086] Step S2: High-frequency synchronous sampling and transient event data capture. The microcontroller unit 102 configures a circular buffer with a sliding time window in the random access memory. The microcontroller unit 102 controls the multi-dimensional sensor array 101 to perform multi-channel synchronous continuous sampling at a fixed sampling frequency, continuously refreshing the discrete high-frequency data matrix within the circular buffer. The microcontroller unit 102 calculates the discrete first-order difference gradient of the reference parameters in the discrete high-frequency data matrix in real time.

[0087] When the absolute value of the discrete first-order differential gradient is detected to be greater than the preset transient trigger threshold, the microcontroller unit 102 records the logical time index and buffer address corresponding to the trigger point, and locks the circular buffer after continuing to collect subsequent specified durations, extracts the data segments within the specified duration before and after the trigger point, and generates the event original data matrix.

[0088] Step S3: Relative time difference feature extraction and causal verification compensation. Microcontroller unit 102 retrieves the extreme points of the reference parameters in the original event data matrix as local zero moments, calculates the time difference delay parameters and extreme amplitude parameters of each follower parameter's extreme points relative to the local zero moments, and aggregates them to generate a basic relative time difference feature vector. Microcontroller unit 102 reconstructs the trial sequence using an internally preset low-order simplified time-domain response template combined with the basic relative time difference feature vector, and performs difference calculation with the original event data matrix to obtain the positive residual curve. For distortion moments in the positive residual curve that exceed the distortion threshold, microcontroller unit 102 establishes a reverse causal time window based on the steady-state phase offset constant, and calculates the fluctuation energy integral of the reference parameter within the reverse causal time window. When the fluctuation energy integral is greater than the excitation energy threshold, microcontroller unit 102 performs discrete second-order derivative operations on the positive residual curve to extract curvature inflection point features, generating a compensation feature vector containing the corresponding time offset and residual amplitude.

[0089] Step S4: Feature Message Concatenation and Network Reporting. Microcontroller unit 102 concatenates the basic relative time difference feature vector and the compensation feature vector using data bits to generate a transmission message. NB-IoT communication module 103 sends the transmission message to the cloud reconstruction server 20. Microcontroller unit 102 releases the storage space occupied by the original data matrix of the current event.

[0090] Step S5: Synchronous reconstruction of multi-parameter waveforms in the cloud. The waveform calculation engine 202 performs convolution operations using the corrected standard time-domain response template and the basic relative time difference feature vector to reconstruct the basic follower parameter waveform. The waveform calculation engine 202 calls the cubic spline interpolation function to fit the compensation feature vector and generate a continuous residual compensation curve. The waveform calculation engine 202 performs numerical superposition of the basic follower parameter waveform and the continuous residual compensation curve, and maps the superposition result to the global time axis according to the global event timestamp containing the absolute timestamp of the baseline network, outputting a synchronously reconstructed multivariate high-frequency data sequence.

[0091] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0092] See attached document Figure 3 In step S1 of the method provided by this invention, the edge acquisition node 10 and the cloud reconstruction server 20 cooperate to complete the system background error measurement and parameter update operation under non-transient event triggering state. Specifically, it includes the following sub-steps:

[0093] Step S101: Obtain the environmental background noise sequence.

[0094] The microcontroller unit 102 monitors the signal fluctuation amplitude output by the multi-dimensional sensor array 101. When the fluctuation amplitude of the judgment benchmark parameter within a continuously set observation window is lower than a preset minimum value judgment threshold, the system confirms that it is currently in a steady state where no transient physical events have occurred. The value of the minimum value judgment threshold is set according to the normal noise level of the system's environment; for example, it can be set to a fixed value within the range of 0.1% to 0.5% of the corresponding sensor range.

[0095] During the steady-state period, the microcontroller unit 102 controls the multi-dimensional sensor array 101 to synchronously acquire the ambient background noise sequences of the reference parameter and each follower parameter. This sequence data includes a common disturbance component formed by the same environmental perturbation source or a low-amplitude calibration excitation, as well as uncorrelated noise components generated by each channel itself. The common disturbance component provides the correlation basis for steady-state phase shift estimation between different physical channels, while the uncorrelated noise components are suppressed during cross-correlation calculations. The ambient background noise sequence of the reference parameter output by the multi-dimensional sensor array 101 is set as... The output of the first The environmental background noise sequence of each follower parameter is as follows: .

[0096] Step S102: Calculate the discrete cross-correlation function.

[0097] The microcontroller unit 102 allocates a length of [length missing] in its internal memory area. The steady-state sliding calculation window is defined. For each follower parameter measurement channel in the multi-dimensional sensor array 101, the microcontroller unit 102 performs discrete cross-correlation numerical calculations between the corresponding follower parameter sequence and the reference parameter sequence, and outputs the discrete cross-correlation function. .

[0098] By calculating the discrete cross-correlation function, the interference of random white noise can be removed, and the inherent physical delay and amplitude difference characteristics between different measurement channels under steady state can be extracted.

[0099] The specific calculation formula is as follows:

[0100] ;

[0101] In the formula, This represents the total number of discrete sampling points contained within the steady-state sliding calculation window; This indicates the index number of the discrete sampling point in the current summation iteration; This represents the number of discrete sampling delay points, and its value range is... Maximum search range It is determined by the product of the maximum possible transmission delay time between hardware system channels and the system sampling frequency; Indicates the first reference parameter within the window. The absolute value of each discrete noise sample; This indicates that the following parameter within the window has passed through The absolute value of discrete noise samples after each delay point offset.

[0102] Step S103: Extract steady-state calibration parameters.

[0103] Microcontroller unit 102 for discrete cross-correlation function The calculation result array is traversed and compared to retrieve the discrete sampling delay point value corresponding to the maximum value of the function within the preset search range, denoted as . The microcontroller unit 102, in conjunction with the fixed sampling frequency of the multi-dimensional sensor array 101, maps and converts the discrete sampling delay point values ​​into deviation values ​​on a time scale, which are defined as the steady-state phase offset constant. This parameter characterizes the inherent zero-point time misalignment between channels under current physical environmental constraints and sensor aging conditions.

[0104] Based on this, the microcontroller unit 102 calculates the first... The arithmetic mean of the servo parameter sequence within the steady-state window is calculated, and the difference between this arithmetic mean and the preset zero-point baseline value, historical calibration baseline value, or baseline reference value sent from the cloud for the servo parameter channel is calculated to generate the DC baseline drift. This is used to characterize the degree of deviation of the static amplitude baseline of the follower parameter channel itself. For follower parameters with the same dimensions as the reference parameter and which have undergone amplitude normalization, the normalized steady-state mean of the reference parameter can also be used as a comparison benchmark.

[0105] Step S104: Parameter encapsulation and network reporting.

[0106] The microcontroller unit 102 calculates the steady-state phase offset constant. DC baseline drift The calibration data packet is generated by assembling according to the preset byte positions. When the NB-IoT communication module 103 maintains a connection with the cellular network base station, there is a periodic signaling interaction mechanism to maintain the active state of the communication link.

[0107] The microcontroller unit 102 encapsulates the calibration data packet in the data payload area of ​​an application-layer keep-alive message, a periodic status reporting message, or an NB-IoT small data transmission message. The NB-IoT communication module 103 sends a periodic message containing calibration parameters to the cloud reconstruction server 20 according to the preset wake-up period of the communication module.

[0108] Therefore, the system utilizes periodic communication connection maintenance or periodic status reporting processes accompanied by calibration data transmission to reduce the energy consumption overhead of radio frequency wake-up caused by independently initiating service data reporting. Regarding the specific communication implementation of the center-hopping mechanism and paging extension cycle in the narrowband IoT communication protocol, those skilled in the art can perform the underlying configuration according to current cellular IoT standards and specifications. The signaling interaction establishment process is well-known in the field and will not be elaborated upon here.

[0109] Step S105: Correct the standard time domain response template in the cloud.

[0110] The cloud-based reconstruction server 20 receives and parses periodic messages or calibration data packets, extracting the steady-state phase offset constant. DC baseline drift The waveform calculation engine 202 connects to the time series database 201 and retrieves the preset standard time domain response template corresponding to the servo parameters. .

[0111] Waveform calculation engine 202 will calculate the steady-state phase shift constant. Substituting the phase compensation term into the time independent variable of this template to perform time axis alignment, the specific correction formula is as follows:

[0112] ;

[0113] At the same time, the waveform calculation engine 202 will receive the DC baseline drift amount The metadata tag associated with this template is updated as the currently valid base DC bias term for this follower parameter. After time-domain compensation calculation, the waveform calculation engine 202 generates a corrected standard time-domain response template. The waveform calculation engine 202 writes the updated template and associated parameters into the time series database 201 according to the device hardware identifier, channel number, update time, and template version number. It then marks it as the currently valid template, providing a calculation benchmark for eliminating drift errors during subsequent waveform reconstruction in the event of transient events.

[0114] See attached document Figure 4 In step S2 of the method provided by this invention, the edge acquisition node 10 performs continuous acquisition of multiple physical quantities and data fragment locking and extraction operations based on abrupt changes in the local environment. Specifically, it includes the following sub-steps:

[0115] Step S201: Multi-channel synchronous sampling and ring buffer management.

[0116] The microcontroller unit 102 configures a circular buffer in its internal random access memory. The total physical capacity of this circular buffer is set to be greater than the sum of the data occupied by the specified duration before and after the event. Specifically, the capacity of the circular buffer is not less than the total amount of data generated by the reference parameter channel and all follower parameter channels within the specified duration before and after the event. The microcontroller unit 102 uses the underlying timer mechanism to generate a globally consistent hardware trigger signal.

[0117] The multi-dimensional sensor array 101 contains all physical measurement channels, including reference parameters and follow-up parameters, which simultaneously initiate analog-to-digital conversion operations based on this hardware trigger signal, providing a foundation for the synchronous acquisition of multi-dimensional physical quantities along the time axis. The hardware configuration logic for implementing multi-channel synchronous triggering of the analog-to-digital converter based on the microcontroller's underlying timer is well-known in the field and will not be elaborated upon here. Those skilled in the art can configure the registers according to general chip manuals.

[0118] The microcontroller unit 102 writes the discrete digital signals output from each channel into a circular buffer at a set fixed sampling frequency, generating a continuously refreshed discrete high-frequency data matrix. When the memory write address increments to the end of the physical boundary of the circular buffer, the microcontroller unit 102 controls the address pointer to return to the starting position and cyclically overwrites the historical data.

[0119] To ensure the logical continuity of the time series during data truncation, the microcontroller unit 102 maintains a global logical time index that continuously accumulates as data is written, which is used to implement the mapping between physical storage addresses and logical time axes during data truncation.

[0120] Step S202, Transient event trigger determination. To avoid out-of-bounds reads due to missing historical data during system startup, the microcontroller unit 102 activates the transient event trigger determination logic after the cumulative data writing time in the circular buffer reaches the specified preceding time. During real-time data writing, the microcontroller unit 102 synchronously acquires the latest data sequence of the reference parameters and performs discrete first-order difference operations to obtain the discrete first-order difference gradient. .

[0121] Discrete first-order difference operations are mathematically and physically equivalent to approximating the derivative of a physical quantity in the time domain. The calculated results effectively filter out slow, low-frequency background drift and highlight abrupt, high-frequency signals. This gradient parameter numerically reflects the intensity of change in the system's physical excitation source within a unit sampling period.

[0122] The specific calculation formula is as follows:

[0123] ;

[0124] In the formula, This indicates the index of the sampling point at the current moment; This indicates the sampling amplitude corresponding to the index of the reference parameter at the current time. This indicates the sampling amplitude corresponding to the index of the reference parameter at the previous adjacent time.

[0125] The microcontroller unit 102 calculates the discrete first-order difference gradient. With the internally preset transient trigger threshold Numerical comparisons were performed to determine the transient triggering threshold in experimental scenarios involving monitoring minute changes in physical stress. Configured to 1% to 5% of the full scale of the reference parametric sensor; transient trigger threshold in experimental scenarios for capturing destructive physical impact signals. Configured to 10% to 20% of the full scale of the reference parameter sensor.

[0126] When the discrete first-order difference gradient The absolute value is not greater than the transient trigger threshold. When the microcontroller unit 102 determines that no transient physical event has occurred, it maintains the normal sliding refresh of the circular buffer. When the discrete first-order differential gradient... The absolute value is greater than the transient trigger threshold. When the microcontroller unit 102 determines that the system has detected the occurrence of a transient physical event, the transient trigger threshold in this embodiment is... The value is between 1% and 20% of the full scale of the reference parameter sensor.

[0127] Step S203: Locking and extracting the original event data matrix.

[0128] When the transient physical event triggering condition is met, the microcontroller unit 102 records the memory offset address corresponding to the current triggering time point in the circular buffer, and simultaneously records the global logical time index corresponding to the current triggering time point. The circular buffer is not locked at this time. The microcontroller unit 102 does not immediately interrupt the sampling task of the multi-dimensional sensor array 101, but controls it to continue writing data to the circular buffer to record the physical response process after being stimulated.

[0129] When the cumulative duration of newly written data since the trigger point reaches a subsequent specified duration, the microcontroller unit 102 triggers its internal interrupt latch mechanism, pausing the write operation of the multi-dimensional sensor array 101 to the circular buffer. At this time, the circular buffer is locked to prevent the data corresponding to the event from being overwritten by subsequent sampling. The microcontroller unit 102 uses the recorded trigger point address as a reference, backtracks along the time axis in the historical record direction to determine the preceding specified duration to determine the starting data boundary, and continues backward to the end of the subsequent specified duration to determine the ending data boundary.

[0130] To obtain the complete physical evolution of the system, the initial specified duration is configured to cover the observation time window during the early stage of the system's evolution from steady state to transient state, and the subsequent specified duration is configured to cover the damping decay period of the dominant response of the physical system's transfer function. The microcontroller unit 102 reads the data sequence of all channels within the time period defined by the determined start and end data boundaries from the circular buffer.

[0131] If the calculated start data boundary to end data boundary crosses the end of the circular buffer in the physical storage address, the microcontroller unit 102 performs segmented reading, first reading the data segment before the address jump, then reading the data segment after the jump, and concatenating the sequence in chronological order to finally generate the event original data matrix.

[0132] The microcontroller unit 102 transfers the extracted raw event data matrix to a separate memory workspace for subsequent data dimensionality reduction and causal verification calculations. Furthermore, to avoid system resource exhaustion or repeated extraction caused by consecutive high-frequency events, after completing the extraction and transfer of the raw event data matrix, the microcontroller unit 102 pauses the trigger determination of the discrete first-order differential gradient for a set blocking time. Monitoring resumes after the blocking time ends. This blocking time is set to be no less than the longest computation time required to process a single event.

[0133] See attached document Figure 5 In step S3 of the method provided by this invention, the edge acquisition node 10 performs feature dimensionality reduction calculation of the high-frequency data matrix and anomaly detection and filtering operations based on physical energy logic locally. Specifically, it includes the following sub-steps:

[0134] Step S301: Establish local zero time and extract the basic relative time difference feature vector.

[0135] The microcontroller unit 102 traverses the raw event data matrix stored in the memory working area. In specific implementations, to unify the time reference system of multi-dimensional physical quantities, the microcontroller unit 102 retrieves the discrete sampling point with the largest absolute amplitude in the data sequence of the reference parameter, and designates the timestamp corresponding to this point in the data matrix as the local zero moment. .

[0136] If multiple maxima with equal absolute amplitudes are found during the search process, the microcontroller unit 102 selects the first point with the smallest time index as the local zero time. To eliminate positioning ambiguity, the point with the maximum absolute amplitude is selected to accommodate both forward and reverse abrupt changes in physical excitation states, avoiding local zero-time positioning failure due to polarity changes.

[0137] Subsequently, the microcontroller unit 102 traverses each... For each data sequence of a follower parameter, the same absolute maximum retrieval logic is used to find the extreme response time in each follower parameter sequence. The microcontroller unit 102 calculates the extreme response time. With local zero time The time difference between them generates the time difference delay parameter. Simultaneously extract the extreme response time. The corresponding amplitude is used as the extreme amplitude parameter. .

[0138] The microcontroller unit 102 concatenates the time delay parameters and extreme amplitude parameters corresponding to all follow-up parameters into an array to generate a basic relative time difference feature vector, expressed by the formula:

[0139] ;

[0140] This operation reduces the dimensionality of the original high-frequency sampling sequence to a feature set containing only key time and amplitude scalars.

[0141] Step S302: Local dimensionality reduction trial calculation and positive residual curve generation.

[0142] Due to the limited computing power of the edge acquisition node 10, the microcontroller unit 102 internally contains a low-order simplified time-domain response template. Specifically, this low-order simplified time-domain response template can be represented as a first-order inertial delay model, and the trial sequence... The equivalent formula for reconstruction is:

[0143] ;

[0144] In the formula, The system sampling period set for the hardware. The time constant of the corresponding physical channel is determined in advance, with a value ranging from 10ms to 50ms. This segmented constraint eliminates invalid exponential calculations before the physical response occurs.

[0145] Microcontroller unit 102 will use the basic relative time difference feature vector Substituting the parameters into the aforementioned low-order simplified time-domain response template, trial sequences of each follower parameter are reconstructed on the local time axis. .

[0146] The microcontroller unit 102 will input the actual sampling sequence corresponding to the original event data matrix. With trial sequence Index by Time Perform point-by-point subtraction to obtain the positive residual curve. ;

[0147] The formula is:

[0148] ;

[0149] The microcontroller unit 102 will display the positive residual curve. The absolute value and the preset distortion threshold Compare the results. Distortion threshold. The value of takes into account both the background noise floor and the system baseline drift, and is set to the corresponding background noise sequence obtained in step S1. The standard deviation is 3 to 5 times the DC baseline drift. The absolute value is the sum of the two values ​​to prevent threshold failure when the drift is zero.

[0150] When a positive residual curve is detected to contain no less than The absolute value of consecutive sampling points Greater than the distortion threshold At that time, these consecutive sampling points are classified into a distorted data segment, and the microcontroller unit 102 records the extreme point with the largest absolute value of the residual in the distorted data segment as the suspected distortion time point. And indicate the corresponding residual deviation direction. The preset threshold for the number of consecutive out-of-limit points can be 2 to 5, or configured according to the sampling frequency to the number of sampling points corresponding to the preset minimum duration.

[0151] Specifically, the low-order simplified time-domain response template is a low-parameter mathematical function pre-set in the non-volatile memory of the microcontroller unit at the edge acquisition node, used to characterize the basic profile of the main response segment of the corresponding follower parameter to the transient excitation.

[0152] The low-order simplified time-domain response template retains only the monotonically rising segment and the main peak falling segment of the physical response waveform, actively stripping away complex tail-damped oscillations, multi-order harmonic components, and steady-state recovery details. In actual deployment, its data structure is reduced to a small number of key control point coordinates or low-order polynomial (such as first- or second-order) coefficients. The aim is to quickly fit the macroscopic energy envelope of the basic response waveform with extremely low microcontroller memory usage and computational overhead, thereby providing a benchmark for subsequent stripping of conventional waveforms and exposure of nonlinear distortion residuals.

[0153] For example, when a multidimensional sensor array is applied to a structural impact experiment, and the follow-up parameter is the vibration acceleration, the low-order simplified time-domain response template corresponding to this channel inside the microcontroller unit is not a complex decaying oscillation waveform sequence containing thousands of sampling points, but specifically a standardized half-sine wave or a piecewise quadratic parabola mathematical expression.

[0154] During edge-end calculations, the microcontroller unit directly uses the extracted extreme amplitude parameters as the amplitude coefficients of the half-sine wave, aligns the extracted extreme response times (local zero time plus time delay parameters) to the peak coordinates of the half-sine wave, and, combined with the preset baseline response pulse width, quickly generates a smooth macroscopic reference response curve (i.e., the calculation sequence). Subsequently, the microcontroller unit subtracts the original data matrix containing high-frequency structural yield signals from the smooth half-sine wave calculation sequence point by point. This allows for the filtering out of the conventional macroscopic impact envelope with extremely low computing power, highlighting the true nonlinear physical responses that deviate from the conventional contours, thereby quickly obtaining the positive residual curve for subsequent closed-loop verification.

[0155] Step S303: Construction of the reverse causal time window and verification of physical energy closed loop.

[0156] To eliminate non-physical signal abrupt changes caused by spatial electromagnetic interference or common-mode noise, the microcontroller unit 102 initiates a verification mechanism based on the causality of physical excitation for each suspected distortion point. Based on the principle that the response lags behind the excitation in a conventional physical system, the microcontroller unit 102 uses the suspected distortion point... Based on this, a reverse causal time window is established by tracing back along the historical timeline. The time interval of the reverse causal time window is defined as follows:

[0157] ;

[0158] In the formula, The corresponding channel steady-state phase offset constant extracted in step S1 is used to compensate for the inherent transmission delay; The effective causal observation duration is determined by the resonant frequency period of the multidimensional sensor array 101, and in this embodiment, the value is between 20ms and 100ms.

[0159] If the calculated starting point of the reverse causal time window is earlier than the starting point of the original event data matrix, the microcontroller unit 102 truncates and aligns the starting point of the reverse causal time window to the starting point of the original event data matrix to prevent memory reads from exceeding the bounds. Subsequently, the microcontroller unit 102 performs square integral operations on the sampled values ​​of the reference parameter within the reverse causal time window to calculate the integral of the reference parameter fluctuation energy. The calculation formula is:

[0160] ;

[0161] The summation interval is the index of all sampling points falling within the inverse causal time window. , The baseline parameter in the original event data matrix is ​​located at the corresponding index. The actual sampling amplitude at that location.

[0162] Microcontroller unit 102 determines the integral of the fluctuation energy of the reference parameter. Is it greater than the preset excitation energy threshold? Excitation energy threshold The linear function representing the absolute value of the residual amplitude at the suspected distortion point is defined by the following formula:

[0163] ;

[0164] In the formula, For the positive residual curve at the suspected distortion time point The absolute value of the residual amplitude at that point, The energy conversion coefficient is obtained through experimental calibration, and its dimensions are the same as those of the reference parameter fluctuation energy integral. The conversion relationship between the residual amplitude and the residual amplitude is matched, and the value can range from 0.5 to 2.0. If Greater than The microcontroller unit 102 determines that the residual has a definite preceding physical excitation source and belongs to a real physical distortion that conforms to causal logic; if Not greater than If the residual is determined to be a pseudo-residual lacking physical excitation support, it will be filtered out and no subsequent compensation feature extraction will be performed on it.

[0165] Step S304: Generate compensation feature vectors based on second derivatives.

[0166] Microcontroller unit 102 extracts waveform compensation features only for the real physical distortion residual segments that have passed the causal closed-loop verification. Microcontroller unit 102 performs discrete second-order derivative operations on the retained valid positive residual curves to locate the curvature inflection points where the waveform curvature fundamentally changes. The specific calculation logic is as follows: first, calculate the first-order derivative sequence:

[0167] ;

[0168] Then calculate the second derivative sequence:

[0169] ;

[0170] The microcontroller unit 102 iterates through the data to find the zero-crossing characteristic, that is:

[0171] ;

[0172] In the formula, This is the residual amplitude corresponding to the previous sampling point; This represents the residual amplitude corresponding to the current sampling point; This is the first-order difference value corresponding to the current sampling point; This is the second-order difference value corresponding to the current sampling point; This is the first-order difference value corresponding to the next sampling point; This is the second-order difference value corresponding to the previous sampling point;

[0173] To avoid a large number of invalid false inflection points caused by small fluctuations in residual noise, the microcontroller unit 102 further requires that the absolute value of the residual corresponding to the zero-crossing point must be greater than the aforementioned distortion threshold. The microcontroller unit 102 extracts the curvature inflection points that meet the above two conditions and records them relative to the local zero time. time offset And the residual magnitude of that point in the positive residual curve. The microcontroller unit 102 combines the above parameters to generate a compensation feature vector:

[0174] ;

[0175] In the formula, For the first The compensation feature vector corresponding to each follower parameter channel For the extracted first The time offset of each curvature inflection point relative to the local zero moment and the residual amplitude of that point in the positive residual curve.

[0176] By extracting the aforementioned feature points, the microcontroller unit 102 further reduces the dimension of a continuous high-frequency residual waveform into multiple discrete coordinate points that can characterize the boundary of the distorted contour, providing core data support for the subsequent reliable restoration of the waveform under limited cellular network bandwidth.

[0177] See attached document Figure 6 In step S4 of the method provided by this invention, the edge acquisition node 10 performs protocol encapsulation of local feature data, wireless network transmission, and reclamation of underlying hardware storage resources. Specifically, it includes the following sub-steps:

[0178] Step S401, global absolute time alignment.

[0179] The microcontroller unit 102 obtains the reference network absolute timestamp issued by the cellular network through the NB-IoT communication module 103. The microcontroller unit 102 numerically binds the local zero time determined in step S3 with the reference network absolute timestamp to generate a global event timestamp. .

[0180] The specific conversion logic is as follows: Record the absolute timestamp of the obtained baseline network. Local system clock tick value Calculate the local zero time. The corresponding system clock tick value and The tick difference is calculated, and then converted into a time difference based on the tick period of the local system clock, and the time difference is then compensated. In this way, the global event timestamp is calculated. .

[0181] The conversion formula is as follows:

[0182] ;

[0183] This alignment operation effectively eliminates the accumulated deviation of the local clock caused by the long-term operation of the device, so that discrete physical events have globally unified time coordinates in the cloud time series database 201.

[0184] Step S402: Adaptive data bit splicing and transmission message generation.

[0185] Microcontroller unit 102 initializes the underlying message structure in the memory transmit buffer. Microcontroller unit 102 sets the global event timestamp. The device hardware identifier is written into the message header. For the basic relative time difference feature vector, since the number of time difference delay parameters and extreme amplitude parameters of each channel contained therein is fixed, the microcontroller unit 102 directly transfers them into fixed-length data blocks in ascending order of channel index.

[0186] For the compensation feature vector, since the number of curvature inflection points caused by different physical events in different channels varies dynamically, the microcontroller unit 102 first inserts a length flag indicating the total number of compensation feature points in that channel before writing the compensation feature vector for each channel. .

[0187] The microcontroller unit 102 appends the compensation feature data, including the time offset and residual amplitude, to the end of the data frame in a preset byte order. To prevent the number of extracted compensation feature points from exceeding the maximum transmission unit limit of the narrowband IoT underlying protocol, the microcontroller unit 102 simultaneously calculates the total byte length of the generated transmission message during the data bit appending process. The specific calculation formula is as follows:

[0188] ;

[0189] In the formula, Indicates the fixed number of bytes occupied by the message header; This indicates the total number of follow-up parameter channels configured in the multi-dimensional sensor array 101; This represents the fixed number of bytes occupied by the basic relative time difference feature vector of a single channel; Indicates the first The total number of compensation feature points extracted from each channel; This indicates the fixed number of bytes occupied by a single compensated feature point parameter.

[0190] The microcontroller unit 102 calculates the total byte length. With the preset maximum transmission load threshold Compare them. The maximum data packet capacity is determined according to the NB-IoT protocol configuration, and in this embodiment, the value is 512 bytes.

[0191] Adding the current compensation feature point will result in Greater than At that time, the microcontroller unit 102 terminates the addition of subsequent compensation feature data and sets a truncation flag in the message header.

[0192] Through the above data bit splicing logic, the microcontroller unit 102 integrates the fixed-length basic features and the variable-length compensation features to generate an adaptive transmission message that adapts to the bandwidth of narrowband networks.

[0193] Step S403: Reporting on the NB-IoT narrowband network.

[0194] The microcontroller unit 102 sends the generated transmission message to the NB-IoT communication module 103 via a serial communication interface. The NB-IoT communication module 103 establishes a radio frequency connection with the cellular communication network base station and sends the small data transmission message to the cloud reconstruction server 20.

[0195] For the underlying interaction process of establishing wireless radio frequency bearer, control plane data transmission state transition and unstructured supplementary service data encapsulation of NB-IoT communication module 103, those skilled in the art can configure the software and hardware through relevant communication standards. Its baseband modulation and wireless transmission process are well known technologies in the field and will not be described in detail here.

[0196] Step S404: Release local storage and reset state.

[0197] When the microcontroller unit 102 receives a message transmission success confirmation signal returned by the NB-IoT communication module 103, or when a transmission failure triggers a retransmission at the underlying communication layer and the number of retransmissions reaches the preset maximum retry threshold, the microcontroller unit 102 executes resource reclamation logic locally.

[0198] By setting a timeout-based forced memory reclamation mechanism, the device can avoid running out of memory due to continuous waiting for network responses in weak signal environments. The microcontroller unit 102 issues a memory unit release command, clearing all storage space occupied by the current event raw data matrix in the independent memory working area and resetting the relevant memory addressing pointers.

[0199] Simultaneously, the microcontroller unit 102 releases the write interrupt latch state set for the circular buffer in step S2, restoring the hardware permission of the multi-dimensional sensor array 101 to cyclically refresh discrete high-frequency data to the circular buffer. The microcontroller unit 102 resets the system's internal state machine to the non-triggering steady-state monitoring mode and releases the decision block on the discrete first-order differential gradient, enabling the system to have the normalized monitoring capability to capture the next transient physical event at any time.

[0200] See attached document Figure 7 In step S5 of the method provided by this invention, the cloud reconstruction server 20 and its internal waveform calculation engine 202 perform parsing, high-order numerical fitting, and time-domain restoration operations on the received edge feature messages. Specifically, this includes the following sub-steps:

[0201] Step S501: Restore the basic follow-up parameter waveform based on the modified template.

[0202] The waveform decoding engine 202 parses and extracts the basic relative time difference feature vectors of each channel from the received underlying transmission messages, and obtains the extreme amplitude parameters of the corresponding channels. With time difference delay parameter The waveform calculation engine 202 connects to the time series database 201 and retrieves the corrected standard time domain response template. .

[0203] Waveform solving engine 202 retrieves the corrected standard time-domain response template Perform amplitude normalization to make the amplitude scale corresponding to the extreme points equal to 1, so as to ensure that the extreme amplitude parameters can be used directly in the future. Numerical accuracy during scaling calculations. Since edge acquisition node 10 extracts feature parameters based on extreme value triggering, the dominant response of the physical system to the excitation can be equivalently regarded as the linear time-invariant system response to this extreme value feature.

[0204] The waveform decoding engine 202 substitutes the aforementioned basic feature parameters into the standard time-domain response template to perform equivalent convolution operations, generating a reconstructed waveform based on the basic follower parameters. To reduce computational complexity, the equivalent convolution operation is transformed into amplitude scaling and time axis translation operations on the normalized template waveform in the cloud-based numerical implementation.

[0205] The specific calculation formula is as follows:

[0206] ;

[0207] In the formula, This represents the relative time independent variable in a cloud-based reconstruction environment, and is set as follows: The time is absolutely aligned with the local zero time established in step S301; This represents the extreme amplitude parameter obtained from message parsing, used to scale the amplitude of the standard time-domain response template proportionally. This represents the time difference delay parameter, used to phase-shift the standard time-domain response template on the relative time axis; The current effective DC baseline drift of this channel, retrieved from the time-series database, is used as a constant bias term for linear superposition to restore the true static physical baseline.

[0208] Through this numerical substitution operation, the system reconstructs the basic low-frequency response profile of the system in this physical event using discrete scalar characteristics.

[0209] Step S502: Generating a smooth fit of the continuous residual compensation curve.

[0210] Waveform solving engine 202 calculates the coordinates of discrete curvature inflection points in the analytically obtained variable-length compensation eigenvector. Perform high-order numerical fitting to generate continuous residual compensation curves.

[0211] To ensure the continuity of the second derivative of the fitted curve throughout the entire reconstruction interval and avoid the Runge phenomenon often seen in high-order polynomial fitting, the waveform solving engine 202 calls the cubic spline interpolation function to perform piecewise fitting calculations on the discrete curvature inflection points. To prevent numerical divergence of the cubic spline interpolation in intervals outside the feature point coverage area, before performing the interpolation calculation, the waveform solving engine 202 appends the time coordinates corresponding to the start and end points of the event reconstruction time window to the beginning and end of the analyzed discrete curvature inflection point set, respectively, and forces the residual amplitude at these two points to be zero, thereby forming a complete set of interpolation constraints.

[0212] During the solution of the cubic spline interpolation equations, the waveform calculation engine 202 employs natural boundary conditions, forcibly setting the second derivative values ​​at both endpoints of the fitted curve to zero. This ensures that the reconstructed compensation curve naturally decays at both ends and smoothly transitions to the background baseline. Furthermore, to guarantee the algorithm's completeness under extreme conditions, the waveform calculation engine 202 detects the total number of compensation feature points for the corresponding channel before fitting. .

[0213] when At that time, perform the above cubic spline interpolation calculation;

[0214] when When the degrees of freedom for solving cubic splines are not satisfied, the waveform solving engine 202 downgrades to use a first-order linear interpolation function to connect the two discrete points.

[0215] when When it is determined that no high-frequency waveform compensation is needed, the zero-value sequence is directly output. For the construction of the piecewise polynomial coefficient matrix and the catch-up method solution process in cubic spline interpolation calculation, those skilled in the art can write the program based on fundamental numerical analysis theory. The specific matrix solution derivation is a well-known technique in this field and will not be elaborated here.

[0216] After fitting calculations, the waveform solving engine 202 calculates and outputs the continuous residual compensation curves for each channel. .

[0217] Step S503: Waveform values ​​are superimposed and precisely mapped to the global time axis.

[0218] The waveform calculation engine 202 reconstructs the waveform from the basic servo parameters generated in step S501 under a unified cloud time resolution. Compensation curve of continuous residual generated in step S502 Perform point-by-point superposition of values ​​in the time domain. The calculation formula is:

[0219] ;

[0220] This addition process fuses the low-order dominant response with high-frequency nonlinear distortion details, resulting in a complete waveform sequence in a relative time reference frame. Subsequently, the waveform decoding engine 202 reads the global event timestamp parsed from the message header. Perform a substitution mapping from the relative time system to the global absolute time system.

[0221] Waveform solving engine 202 will use relative time as the independent variable With global event timestamps Perform numerical addition to establish a global timeline variable. The mapping relationship. Because constraints were already enforced in the previous computing system. Corresponding to the local zeros of the maxima of physical events, this mapping operation ensures the synchronous anchoring of the physical excitation origin of the reconstructed waveform with the absolute timestamp of the reference network.

[0222] Waveform Solving Engine 202 Translate the entire timeline The final multi-parameter reconstructed waveform is generated. The waveform decoding engine 202 will generate a final multi-parameter reconstructed waveform that includes global real-time coordinates. The data is written back to the time series database 201 for persistent storage, completing the waveform data restoration of this transient event in the cloud and providing a standardized data source for subsequent upper-layer business logic analysis and visualization terminal retrieval.

[0223] To enable those skilled in the art to more clearly understand the purpose, technical solution, and advantages of this invention, the present invention will be further described in detail below with reference to specific application embodiments, real experimental test data, and corresponding drawings. It should be noted that the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0224] This embodiment uses a destructive impact test scenario of a steel beam structure in an IoT laboratory as an example. In this scenario, edge acquisition nodes are deployed on the steel beam test bench, using a system sampling frequency of 1000Hz (i.e., sampling period). (1ms).

[0225] Reference parameter sensor (channel 0): Impact force sensor (range 0-10000N).

[0226] Follow-up parametric sensor (channel 1): Vibration acceleration sensor (range -50g to 50g). During the non-triggered period, the system extracts the DC baseline drift of the vibration acceleration channel caused by sensor temperature drift. g, due to the steady-state phase shift constant caused by the acquisition circuit and physical transmission. ms.

[0227] exist At 10:00:00, the hammer strikes the steel beam in free fall.

[0228] The microcontroller calculates the discrete first-order difference gradient of the impact force channel in real time. Within a single millisecond, the impact force increases from 100N to 2100N, a gradient... The value is 2000N, far exceeding the preset trigger threshold. (500N).

[0229] The system determines that a trigger has been triggered, locks the circular buffer, and extracts the original event data matrix from 50ms before the trigger point to 150ms after the trigger point (a total of 200 discrete sampling points, with a duration of 200ms).

[0230] In the original data matrix, the impact force channels are at relative timestamps. The absolute maximum value of 8000N is reached at 0ms after triggering; this point is set as the local zero moment. ms.

[0231] By traversing channel 1 (vibration acceleration), it was found that it reaches a maximum value of 30g 5ms after the relative local zero moment. From this, the time delay parameter of channel 1 can be extracted. ms, extreme amplitude parameter It weighs 30g. Basic feature vector. It lasts for 5ms and weighs 30g.

[0232] Preset vibration channel time constant The delay is 20ms. The microcontroller calculates this using a first-order inertial delay model. ms (i.e., index) Trial amplitude at ms):

[0233] g;

[0234] At this time, the actual data originally sampled by the sensor g. Calculate the positive residual. g. This value is greater than the distortion threshold. (Set to 2g), marked as a suspected distortion point.

[0235] The microcontroller constructs a reverse causal time window (−15ms, −5ms) by backtracking to the past, and calculations show that the impact force has a significant early energy integral release within this time window. The residual was determined to be a secondary vibration caused by the nonlinear yielding of the actual physical structure.

[0236] By detecting the zero-crossing of the second derivative, the coordinates of three curvature inflection points in the distorted segment of the waveform are extracted: (10ms, 4.0g), (12ms, 8.0g), and (15ms, -1.5g). A compensation feature vector is then generated. .

[0237] The microcontroller combines extreme value features with compensation features to generate an NB-IoT message with a total length of only 46 bytes and uploads it. The cloud-based waveform processing engine receives the message and then:

[0238] The first step is to retrieve the corrected standard time-domain response template. It was shifted for 5ms, magnified 30 times, and superimposed with a steady-state baseline of 0.5g;

[0239] The second step is to use the cubic spline interpolation function to smoothly fit the three discrete points (10, 4.0), (12, 8.0), and (15, −1.5) to generate a continuous residual compensation curve that naturally converges to 0 at both ends.

[0240] The third step is to overlay the basic reconstructed waveform with the compensation curve and align it with the network absolute timestamp (10:00:00) to finally restore the nonlinear vibration waveform with structural yielding characteristics of the impact.

[0241] To verify the data transmission performance of this invention in a narrowband IoT communication environment, a destructive impact testing platform for NB-IoT structural components was built. This platform includes 100 test nodes, each configured with an edge acquisition node, a multi-dimensional sensor array, and an NB-IoT communication module. During the experiment, 1000 standard drop-weight impact tests were uniformly performed on the test nodes, and the performance of each scheme in terms of single-transmission data volume and network reporting success rate was statistically analyzed.

[0242] The following control group was set up in this experiment:

[0243] Experimental group: The data acquisition optimization processing scheme provided by this invention is adopted, namely, extracting discrete features through steady-state calibration, basic extreme value dimensionality reduction and causal residual compensation, and then concatenating the discrete features into an adaptive feature message before reporting.

[0244] Control Group A: This group used a raw data pass-through method, meaning that after a transient event was triggered, all high-frequency raw data sequences were directly packaged and uploaded. The data for each event included 200 sampling points per channel, and the sampling data was processed using a 16-bit ADC resolution.

[0245] Control group B: Threshold-triggered downsampling was used, that is, after the transient event was triggered, the original high-frequency data sequence was extracted at equal intervals in a ratio of 5:1, and the extracted sampled data was uploaded.

[0246] Test results are as follows Figure 8 As shown:

[0247] Reference Figure 8 It can be seen that, in terms of the average data volume transmitted per transmission, the control group A transmitted 1240 bytes per transmission; after equal-interval downsampling, the control group B reduced the data volume transmitted per transmission to 280 bytes; and the experimental group, using the scheme of this invention, transmitted 52 bytes per transmission. Therefore, this invention, by converting the multi-channel high-frequency raw sequence into basic relative time difference features and compensation features, significantly reduces the data reporting load per event while preserving the main temporal information and distortion compensation information of transient events. Compared with control group A, the experimental group reduced the data volume transmitted per transmission by approximately 95.8%.

[0248] Furthermore, regarding the reporting success rate on the NB-IoT network, control group A, due to its larger single-transmission data volume, required fragmented transmission or multiple retransmissions under weak network conditions, resulting in a reporting success rate of 68.5% according to experimental statistics. Control group B, due to its reduced transmission data volume, achieved a reporting success rate of 92.1%. After adopting the adaptive feature message generated by this invention, the experimental group achieved a single message data volume of 52 bytes, which reduced the probability of fragmented transmission and retransmission, resulting in a reporting success rate of 99.6%.

[0249] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing data acquisition and processing in an IoT laboratory, characterized in that, Includes the following steps: The discrete environmental background noise sequence under non-triggering period is obtained, the steady-state phase offset constant and DC baseline drift are extracted and reported, and used by the cloud reconstruction server to generate the corrected standard time domain response template. Continuous sampling is performed on the multi-dimensional sensor array. When the absolute value of the discrete first-order differential gradient of the reference parameter is detected to be greater than the preset transient trigger threshold, the data segment is locked and extracted based on the trigger time to generate the original event data matrix. The basic relative time difference feature vector is extracted from the original event data matrix. The trial sequence is reconstructed using the internally preset low-order simplified time domain response template and the basic relative time difference feature vector. The difference is then calculated with the actual sample to obtain the positive residual curve. A reverse causal time window is established for the positive residual curve to perform physical energy closed-loop verification. Curvature inflection point features are extracted from the residual waveform that passes the verification, and a compensation feature vector is generated. The basic relative time difference feature vector and the compensation feature vector are concatenated to generate a transmission message, which is then sent to the cloud reconstruction server through the NB-IoT communication module. The cloud-based reconstruction server uses the corrected standard time-domain response template and the basic relative time difference feature vector to reconstruct the basic follow-up parameter waveform, fits the compensation feature vector to generate a continuous residual compensation curve, performs numerical superposition and maps it to the global time axis, and outputs a synchronously reconstructed multivariate high-frequency data sequence.

2. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for extracting the steady-state phase offset constant and DC baseline drift are as follows: Within the set steady-state sliding calculation window, perform discrete cross-correlation numerical calculations on the corresponding follower parameter sequence and the reference parameter sequence respectively, and output the discrete cross-correlation function; The discrete sampling delay point value corresponding to the maximum value of the discrete cross-correlation function is retrieved and mapped to the deviation value on the time scale to define the steady-state phase offset constant. Calculate the arithmetic mean of the follower parameter sequence within the steady-state sliding calculation window, and calculate the difference between the arithmetic mean and the preset zero-point baseline value, the historical calibration baseline value, or the baseline reference value sent from the cloud to generate the DC baseline drift.

3. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for generating the original event data matrix by locking and extracting data segments based on the trigger time are as follows: A circular buffer is configured in the internal random access memory. The discrete digital signals output from each channel are written into the circular buffer, and historical data is cyclically overwritten when the memory write address is incremented to the end of the physical boundary. When the absolute value of the discrete first-order differential gradient of the reference parameter is detected to be greater than the preset transient trigger threshold, the memory offset address corresponding to the trigger moment in the circular buffer is recorded, and the global logical time index corresponding to the current trigger moment is recorded synchronously. Continue writing data to the circular buffer until the accumulated duration reaches the specified duration, trigger the internal interrupt latch mechanism to lock the circular buffer, use the trigger time point address as a reference to determine the start data boundary and end data boundary, extract the data segments within the specified duration before and after the trigger point, and read them out in segments and splice them to generate the original event data matrix.

4. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for extracting the basic relative time difference feature vector from the original event data matrix are as follows: In the data sequence of the reference parameters of the original event data matrix, the discrete sampling point with the largest absolute amplitude is retrieved, and the timestamp corresponding to the discrete sampling point with the largest absolute amplitude is designated as the local zero time. Traverse the data sequence of each follower parameter to find the extreme response time, calculate the time difference between the extreme response time and the local zero time to generate the time difference delay parameter, and extract the corresponding amplitude to generate the extreme amplitude parameter. All time delay parameters and extreme amplitude parameters corresponding to the follow-up parameters are concatenated into an array to generate the basic relative time difference feature vector.

5. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for establishing a reverse causal time window and performing physical energy closed-loop verification for the positive residual curve are as follows: When it is detected that the absolute value of consecutive sampling points in the positive residual curve is greater than the distortion threshold, the extreme point with the largest absolute value of the residual is recorded as the suspected distortion moment point. Based on the suspected distortion time point, and combined with the steady-state phase offset constant and the set effective causal observation duration, the reverse causal time window is established by tracing back along the historical direction of the time axis. Within the reverse causal time window, the sampled values ​​of the reference parameter are subjected to square integral operation to calculate the integral of the reference parameter fluctuation energy. When the integral of the fluctuation energy of the reference parameter is greater than the preset excitation energy threshold, it is determined that the corresponding residual has a definite preceding physical excitation source, which belongs to the real physical distortion that conforms to causal logic, and the corresponding residual waveform verification is passed. When the integral of the fluctuation energy of the reference parameter is not greater than the preset excitation energy threshold, the corresponding residual is determined to be a pseudo residual lacking physical excitation support and is filtered out.

6. The IoT laboratory data acquisition and optimization processing method according to claim 5, characterized in that, The specific steps for extracting curvature inflection point features from the verified residual waveform and generating a compensation feature vector are as follows: For the effective positive residual curves that are determined to be real physical distortions, calculate the first derivative sequence and the second derivative sequence in sequence. Traverse and search for sampling points that satisfy the zero-crossing feature, that is, the product of the second-order difference value corresponding to the current sampling point and the second-order difference value corresponding to the previous sampling point is not greater than zero, and the absolute value of the corresponding residual is greater than the distortion threshold, and extract them as the curvature inflection points. Record the time offset of the curvature inflection point relative to the local zero moment and the residual amplitude of the curvature inflection point in the positive residual curve. Combine the time offset and the residual amplitude to generate the compensation feature vector containing the corresponding time offset and residual amplitude.

7. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The steps for generating the corrected standard time-domain response template are as follows: The cloud-based reconstruction server extracts the received steady-state phase offset constant and DC baseline drift, and retrieves the standard time-domain response template of the corresponding follower parameter preset in the time-series database; The steady-state phase offset constant is substituted as a phase compensation term into the time independent variable of the standard time-domain response template to perform time axis alignment. The DC baseline drift is updated to the metadata tag associated with the standard time-domain response template as a basic DC bias term to generate the corrected standard time-domain response template.

8. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for concatenating the basic relative time difference feature vector with the compensation feature vector to generate the transmission message are as follows: The NB-IoT communication module obtains the reference network absolute timestamp issued by the cellular network, and the determined local zero time is numerically bound to the reference network absolute timestamp to generate a global event timestamp. Write the global event timestamp and device hardware identifier into the message header; For the basic relative time difference feature vector, it is directly transferred into fixed-length data blocks according to the channel index; For the compensation feature vector, when the total byte length of the generated transmission message is calculated to be no greater than the preset maximum transmission load threshold, the compensation feature data with the length flag indicating the total number of compensation feature points of the corresponding channel is appended to the end of the data frame in a preset byte order.

9. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for reconstructing the base follow-up parameter waveform using the modified standard time-domain response template and the base relative time difference feature vector are as follows: Amplitude normalization is performed on the retrieved modified standard time-domain response template; The amplitude of the normalized, corrected standard time-domain response template is scaled proportionally using the obtained extreme amplitude parameters. Phase shift is performed on the relative time axis using the obtained time difference delay parameters; The retrieved DC baseline drift is used to linearly superimpose constant bias terms to restore the true static physical baseline, complete the equivalent convolution operation, and generate the basic follower parameter waveform.

10. The IoT laboratory data acquisition and optimization processing method according to claim 1, characterized in that, The specific steps for fitting the compensation feature vector to generate a continuous residual compensation curve are as follows: The cubic spline interpolation function is called to fit and calculate the coordinates of the discrete curvature inflection points in the compensated feature vector; Before performing the interpolation calculation, time coordinates corresponding to the start and end points of the event reconstruction time window are added to the first and last ends of the set of discrete curvature inflection point coordinates, respectively, and the residual amplitude at the start and end points is forcibly set to zero, thereby forming a complete set of interpolation constraints. In the process of solving the cubic spline interpolation equation system, natural boundary conditions are adopted, that is, the second derivative values ​​at the two endpoints of the fitted curve are forced to be zero, and the continuous residual compensation curve is calculated and output. When the total number of compensation feature points in the corresponding channel is equal to 2, the two discrete points are connected by a first-order linear interpolation function; when the total number of compensation feature points in the corresponding channel is less than 2, the zero value sequence is directly output.

Images