Low-power communication method and system based on internet of things hardware device

Abstract

The application discloses a kind of low-power communication method and system based on internet of things hardware equipment, and the application relates to internet of things technical field, and the original data is preliminarily compressed by first-order forward difference and adaptive segmented activation function, and then compressed again by information entropy screening;Distributed clustering network is constructed using edge gateway, and dynamic feature codebook is generated by local sensitive hashing (LSH) algorithm;High-entropy quantity proportion, remaining power, average codebook index length are normalized to obtain three-dimensional state vector, and the optimal communication protocol is selected by calculating protocol score;The identification module is integrated in the radio frequency front end of the device, and the wake-up control is realized by hamming distance calculation.The application solves the problems of high communication power consumption, low efficiency and poor adaptability of internet of things hardware devices in the prior art, and improves the overall performance of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things, in particular to a low-power communication method and system based on Internet of Things hardware devices. BACKGROUND

[0002] With the popularization of Internet of Things technology, the communication system based on low-power hardware devices is widely used in industrial monitoring, smart agriculture, etc. However, the existing technology has the following defects when realizing low-power communication of devices:

[0003] The original data collected by Internet of Things devices often contains a large amount of redundant information, and traditional compression methods are difficult to realize efficient dimension reduction while preserving key features, resulting in excessive data transmission, which not only occupies a large amount of communication bandwidth, but also increases device power consumption and shortens battery life;

[0004] The existing clustering algorithm mostly uses a static codebook generation mechanism, which cannot adapt to the dynamic changes of device states, resulting in a decrease in feature matching accuracy and difficulty in meeting real-time requirements. At the same time, there is a lack of codebook synchronization mechanism between edge gateways, resulting in data processing redundancy and consistency problems;

[0005] In terms of communication protocol selection, the existing scheme mostly relies on preset rules, which cannot dynamically adjust to multi-dimensional states such as device remaining power and data real-time requirements, often resulting in resource mismatching problems and low communication efficiency.

[0006] Therefore, there is an urgent need for a communication method and system that can balance low power consumption, high real-time performance and strong adaptability. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application provides a low-power communication method and system based on Internet of Things hardware devices, which solves the problems of high power consumption, low efficiency and poor adaptability of existing Internet of Things device communication.

[0008] To achieve the above purpose, the present application realizes the following technical scheme: a low-power communication method based on Internet of Things hardware devices, comprising:

[0009] Step 1, collecting original data using Internet of Things hardware devices, using first-order forward difference combined with adaptive piecewise activation function to preliminarily compress the original data, and then performing secondary compression through information entropy screening;

[0010] Step 2, constructing an edge gateway into a distributed clustering network, generating a dynamic feature codebook based on the received compressed feature vector using the Local Sensitivity Hashing (LSH) algorithm;

[0011] Step 3, the high-entropy number proportion of the edge gateway, the remaining power, and the average codebook index length are combined into a three-dimensional state vector, the protocol score is calculated according to the cosine similarity and the Euclidean distance of the three-dimensional state vector and each protocol preference vector, and the optimal communication protocol is selected according to the protocol score;

[0012] Step 4, an identification module is integrated in the radio frequency front end of the Internet of Things hardware device, wireless signals are collected in real time, frequency domain features are extracted to generate a vector, the vector is binarized, and a Hamming distance is calculated with a preset template, when the distance is less than a threshold, the wake-up is triggered, otherwise the device is put into sleep.

[0013] As a further scheme of the application, the specific steps of using first-order forward difference combined with an adaptive segmented activation function for preliminary compression are as follows:

[0014] The first-order forward difference is performed on the multi-dimensional data collected by the Internet of Things hardware device to obtain data gradients;

[0015] The average value avg and the standard deviation piu of the gradient sequence are calculated;

[0016] The upper threshold is calculated according to the formula a_high=avg+k1×piu;

[0017] The lower threshold is calculated according to the formula a_low=avg-k1×piu;

[0018] The compression coefficient is calculated according to the formula b=1 / (1+exp(-k2×piu)), wherein k1 and k2 are configurable parameters;

[0019] When the gradient value is greater than or equal to a_high, the original value is retained;

[0020] When a_low is less than or equal to the gradient value and less than a_high, the gradient value is compressed by b, wherein b is an adjustment factor;

[0021] When the gradient value is less than a_low, it is compressed to 0.

[0022] As a further scheme of the application, the information entropy of the feature vector after preliminary compression is calculated, the entropy threshold En is set, the feature points with entropy values higher than En are retained, the rest are set to zero, and the proportion of the number of zeros to the total number is calculated, if it is greater than 0.5, the feature vector is marked as high-entropy.

[0023] As a further scheme of the application, the specific steps of using the local sensitive hashing (LSH) algorithm to generate a dynamic feature codebook are as follows:

[0024] The compressed feature vector is sent to the corresponding edge gateway from the Internet of Things hardware device end, and the feature vector is subjected to min-max normalization processing;

[0025] Configure LSH parameters, setting the number of hash functions to [128, 256] and the number of buckets to [1000, 10000].

[0026] The hash function is designed according to the formula h(v)=sign(v×r), where v is the compressed feature vector and r is the random projection vector;

[0027] Multiple hash functions are combined to generate a hash signature, which is then mapped to the corresponding hash bucket.

[0028] Feature vectors falling into the same hash bucket are clustered using K-means to generate local codebook entries, and a unique index is assigned to each codebook entry.

[0029] As a further aspect of the present invention, each codebook entry is associated with a dynamic weight value w, with an initial value of 1.0. The weights of all codebook entries decay according to the formula w_new=w_old×(T_total / T). When the weight is lower than the threshold, the entry is marked as to be deleted, where T represents the update cycle, T_total represents the time since the codebook entry, and λ is the forgetting factor.

[0030] As a further aspect of the present invention, the specific steps for calculating the protocol score are as follows:

[0031] After normalizing the high-entropy quantity ratio H, the remaining power E, and the average codebook index length L, a three-dimensional state vector SS=[H,E,L] is formed.

[0032] Based on the characteristics of the protocol, a preference vector Pi is predefined for each protocol, where i∈[1,3] and i is an integer. The specific preference vectors are as follows: LPWAN, P1 can be set to [0.2,0.8,0.3]; BLE, P2 can be set to [0.5,0.5,0.5]; 5G, P3 can be set to [0.8,0.2,0.7].

[0033] For each protocol, calculate the cosine similarity (Pi) between its preference vector Pi and its state vector SS;

[0034] The Euclidean distance penalty is calculated using the formula Distance(Pi)=1 / (1+||SS-Pi||).

[0035] The protocol score is calculated using the formula Score(Pi) = Similarity(Pi) × Distance(Pi);

[0036] The protocol with the highest score is selected as the corresponding transmission protocol.

[0037] As a further aspect of the present invention, the historical transmission success rate R(Pi,SS) of each protocol under a specific state is recorded, according to the formula... updating the protocol preference vector, wherein P i,new is the updated protocol preference vector, P i,old is the protocol preference vector before updating, i is the number of protocols, j is the number of transmissions, is the learning rate.

[0038] As a further scheme of the application, the difference between the highest score and the second highest score is calculated according to the formula DecisionMargin=max(Score(Pi))-Secondmax(Score(Pi)), and if DecisionMargin< threshold Scoth, the elastic decision is entered, and the specific rules are as follows: if H>0.7, 5G is preferentially selected; if E<0.3, LPWAN is preferentially selected; otherwise, the protocol with the lowest historical energy consumption under the current power state is selected, wherein max(Score(Pi)) is the highest score, and Secondmax(Score(Pi)) is the second highest score.

[0039] As a further scheme of the application, the abnormality degree A is calculated according to the formula A=Hx(1-E)xL, and if A>threshold Ath, 5G is forced to be selected and the following measures are started: bandwidth maximization, retransmission interval minimization, and enabling a backup communication channel.

[0040] The low-power communication system based on the Internet of Things hardware device comprises:

[0041] The data compression module collects original data by using the Internet of Things hardware device, and the original data is preliminarily compressed by using a first-order forward difference combined with an adaptive segmented activation function, and then is secondarily compressed by information entropy screening.

[0042] The codebook generation module constructs the edge gateway into a distributed clustering network, generates a dynamic feature codebook by using a local sensitive hashing (LSH) algorithm based on the received compressed feature vector.

[0043] The protocol decision module combines the high-entropy proportion of the edge gateway, the remaining power and the average codebook index length into a three-dimensional state vector, calculates the protocol score according to the cosine similarity and the Euclidean distance between the three-dimensional state vector and each protocol preference vector, and selects the optimal communication protocol according to the protocol score.

[0044] The wake-up control module integrates an identification module in the radio frequency front end of the Internet of Things hardware device, collects wireless signals in real time and extracts frequency domain features to generate a vector, binarizes the vector and calculates the Hamming distance with a preset template, and when the distance is less than a threshold, the wake-up is triggered, otherwise the device is controlled to sleep.

[0045] The application provides a low-power communication method and system based on an Internet of Things hardware device, which has the following beneficial effects compared with the prior art:

[0046] (1) The present application realizes two-stage compression of original data by combining first-order forward difference with adaptive segmented activation function and information entropy screening, reduces data transmission volume, reduces communication bandwidth occupation and device power consumption, and prolongs battery life;

[0047] (2) The present application generates a dynamic feature codebook using an edge distributed clustering network and an LSH algorithm, and combines a weight decay mechanism to dynamically adapt to changes in different states of the device, improve feature matching accuracy, and ensure real-time requirements;

[0048] (3) The present application selects the optimal protocol based on the similarity of three-dimensional state vectors and protocol preference vectors, and balances communication efficiency and power consumption by combining a Hamming distance matching wake-up mechanism for radio frequency fingerprint recognition to improve system adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The present application is a step flowchart;

[0050] Figure 2 The present application is a system framework diagram. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] As Figure 1 , the present application provides a low-power communication method based on Internet of Things hardware devices, comprising:

[0053] Step 1, collect original data using the Internet of Things hardware device end, and compress the original data by first-order forward difference and adaptive segmented activation function, and then compress it again by information entropy screening;

[0054] The Internet of Things hardware device end has the following characteristics:

[0055] It is usually an embedded sensor, terminal node, etc., such as an STM32L series MCU sensor-equipped device, a LoRa wireless sensor, etc.

[0056] It adopts low-power design and relies on battery or energy harvesting technology for power supply, such as farmland soil humidity sensors and intelligent building door and window magnetic sensors.

[0057] The original data collected by the Internet of Things hardware device end usually has the following defects:

[0058] There is a lot of static redundancy and dynamic redundancy, for example, the three-axis acceleration data collected by the vibration sensor when the industrial equipment is running normally, more than 85% of which belongs to invalid fluctuations of smooth operation, and only 5% contains fault characteristics;

[0059] The original data collected by multiple sensors usually has more than 10 dimensions and is mixed with environmental noise, for example, the smart home environment monitoring device simultaneously collects temperature, humidity, light, and human infrared signals. The temperature data may suddenly change due to the air flow from the air conditioner outlet, which may mask the true human activity characteristics. Therefore, filtering and feature extraction are needed to identify effective information.

[0060] The amount of original data exceeds the carrying capacity of low-power communication protocols. For example, the typical bandwidth of LoRa protocol is 125 kHz, and it takes about 600 ms to transmit 10 KB of original data. If an agricultural soil moisture sensor collects data every 30 minutes, the daily data transmission volume reaches 4.8 MB, which is far beyond the ideal transmission load of LoRa, easily leading to network congestion and packet loss.

[0061] Real-time analysis of original data requires a large amount of computing power, which is difficult for low-power MCUs to support. For example, PCA dimensionality reduction of 100-dimensional original vibration data requires about 100 ms on an STM32L0 chip. However, the device usually requires fault detection delay to be less than 50 ms. Processing of uncompressed original data may lead to ineffective early warning.

[0062] The specific steps of preliminary compression of original data through first-order forward difference and adaptive segmented activation function are as follows:

[0063] First-order forward difference is calculated for multi-dimensional data collected by Internet of Things hardware devices (such as temperature and humidity, vibration acceleration) to approximate the data change rate and extract dynamic characteristics of the data.

[0064] For example, the gradient calculation of the temperature data sequence [25, 25.1, 26, 26.5, 27] gives [0.1, 0.9, 0.5, 0.5], highlighting the mutation characteristics around 26°C.

[0065] Calculate the average value avg and the standard deviation piu of the gradient sequence.

[0066] Calculate the upper threshold according to the formula a_high=avg+k1×piu.

[0067] Calculate the lower threshold according to the formula a_low=avg-k1×piu.

[0068] Calculate the compression coefficient according to the formula b=1 / (1+exp(-k2×piu)), where k1 and k2 are configurable parameters, and the default values are k1=1.5 and k2=0.5.

[0069] When the gradient value is ≥ a_high, the original value is retained;

[0070] High gradient usually reflects important features such as data mutation, and compression will cause gradient distortion;

[0071] When a_low≤gradient value<a_high, it is compressed as gradient value×b, where b is the adjustment factor;

[0072] Medium gradient contains certain effective information, but direct transmission of full precision will increase redundancy, and through linear compression, the direction and relative size relationship can be retained while reducing data volume;

[0073] When the gradient value is <a_low, it is compressed to 0;

[0074] Low gradient is usually caused by noise or non-critical features of input data, and zero processing can achieve gradient sparsification, greatly reducing the amount of transmitted data;

[0075] For example, the original gradient sequence is [-0.3, -0.1, 0.2, 0.5, 0.8, 1.2, 1.5, 2.0], the avg of the gradient sequence is 0.6, the piu is 0.8, the adaptive parameter a_low is-0.6, a_high is 1.8, b is 0.377, and the processed gradient sequence is [0, -0.038, 0.075, 0.189, 0.302, 0.452, 0.566, 2.0];

[0076] Calculate the information entropy of the compressed feature vector, set the entropy value threshold En, and only retain the feature points with entropy value higher than En, and set the rest to zero. If the proportion of the number of zeros to the total number is greater than 0.5, mark the feature vector as high entropy;

[0077] Entropy value represents feature uncertainty, and retaining high-entropy features can maximize the retention of original data information;

[0078] Secondary compression based on information entropy screening shifts from data amplitude to data information value evaluation, avoiding feature omission caused by relying solely on gradient threshold, for example, in the early stage of device failure, there may be abnormal signals with no obvious gradient change but high information entropy, which can be accurately retained by this step.

[0079] Step 2, build a distributed clustering network based on edge gateway, receive the compressed feature vector and generate a dynamic feature codebook using the Local Sensitivity Hashing (LSH) algorithm;

[0080] Divide the edge gateway into multiple distributed clusters according to geographical location (such as workshops in a factory, city blocks) or functional domain (such as temperature sensor clusters, vibration sensor clusters), each cluster contains 3-5 gateways, forming a ring / star hybrid topology;

[0081] Each gateway maintains the state information of adjacent gateways, such as load rate, connection state, and monitors in real time through a heartbeat mechanism;

[0082] When a gateway fails, the device feature vectors it is responsible for are automatically routed to a standby gateway in the cluster, and the routing table is quickly synchronized and updated through the Paxos algorithm;

[0083] The device end sends the compressed feature vector to the corresponding edge gateway, and the edge gateway normalizes the vector, which can be normalized to the [0, 1] interval using min-max normalization to reduce the dimension effect;

[0084] Configure LSH parameters for each edge node:

[0085] The number of hash functions is [128, 256], and the dimension of the feature vector in the edge scenario is usually 512-2048. The number of hash functions is D / 16 to D / 8, which can retain enough feature information and avoid similarity loss due to excessive compression. Where D is the dimension of the compressed feature vector;

[0086] The number of buckets is [1000, 10000], and the data volume processed by the edge gateway is usually 10^4 to 10^6 feature vectors. The number of buckets is N / 100 to N / 1000, which can make each bucket store an average of 10 to 100 vectors, avoiding too dense or too sparse data in the bucket;

[0087] According to the formula h(v)=sign(v×r), a hash function is designed, where v is the compressed feature vector and r is a random projection vector. Different r directly leads to different hash functions;

[0088] Multiple hash functions are combined to generate a hash signature, such as 1010110, and mapped to the corresponding hash bucket;

[0089] The edge gateway clusters the feature vectors falling into the same hash bucket, which can use K-means;

[0090] LSH maps similar features into the same bucket through hash functions, and the semantic similarity of feature vectors in the bucket is high. At this time, clustering features in the same bucket can directly extract representative vectors from the local similar feature set, reducing the clustering calculation amount;

[0091] Generate local codebook entries and assign each codebook entry a unique index;

[0092] The local codebook is a set of representative feature vectors generated by the edge gateway. By clustering features within the same hash bucket, the center vector of each cluster is taken as an entry. For example, if the features in a hash bucket are clustered into 5 clusters, 5 codebook entries are generated. Each entry is essentially a low-dimensional feature vector with the same dimensions as the original features.

[0093] The codebook entries serve as semantic labels for the features within the cluster, allowing the original features to be mapped to the codebook space. At the same time, the local codebook can be directly used for real-time inference. For example, an industrial robot can determine the current workpiece type by matching the codebook entries without uploading the original sensor data to the cloud.

[0094] For example, 10 feature vectors from different devices are collected in hash bucket B32. The edge gateway N1 performs K-means clustering on B32 and generates codebook entry C1-001 = [0.75, 0.18, 0, 0.05]. Subsequent device features are directly indexed by C1-001 instead of the original vectors if they are mapped to B32.

[0095] The edge gateway needs to exchange local codebooks regularly.

[0096] In the edge scenario, the device state changes over time, causing the feature distribution to drift. For example, the initial features of an industrial sensor may be clustered into 5 clusters, corresponding to the same number of codebook entries. However, after running for 3 days, some features may move to new clusters due to device aging. Regularly exchanging codebooks allows the edge gateway to timely perceive global feature changes and avoid outdated local codebooks.

[0097] A single edge gateway covers a limited number of devices, and the local codebook it generates may not be complete. Through exchange, different gateways can obtain the codebooks of other gateways to supplement their missing patterns.

[0098] If duplicate codebook entries are found, the better version is retained through a voting mechanism.

[0099] Adjacent edge gateways may process data from some of the same devices. For example, edge gateways N1 and N2 monitor machine M1 on a production line at different sampling times. The local codebooks they generate may contain similar normal operation entries. On the other hand, the feature distribution of similar devices has similarities, which may also lead to duplicate codebook entries in different edge gateways.

[0100] The quality of duplicate entries generated by different edge gateways may differ. Voting can retain the more representative version, improving the accuracy of global feature matching.

[0101] If all duplicate entries are directly merged, it will cause codebook inflation, increasing the storage burden and matching delay of edge gateways. The voting mechanism can simplify the codebook and maintain computational efficiency.

[0102] For example, the codebook of edge gateway N1 contains entries {C1-001, C1-002}, and the codebook of N2 contains {C2-001, C2-003}, where C1-001 is similar to C2-003, N1 finds that C1-001 is similar to C2-003, and by voting, the better C1-001 is retained, and N2 retains C1-001 in the same way, finally, the codebook of N1 is updated to {C1-001, C1-002}, and the codebook of N2 is updated to {C1-001, C2-001}, both of which share C1-001, but still retain their respective entries;

[0103] Each codebook entry is associated with a dynamic weight value w, the initial value is 1.0, which represents the representativeness of the entry to the current feature distribution, and the weight decays over time, reflecting that the importance of historical data gradually decreases, and the specific operation is as follows:

[0104] The update period T is set, and all codebook entry weights decay according to the formula w_new = w_old x (T_total / T), where T_total represents the time of the codebook entry from now on, and λ is a forgetting factor, with a value range of [0.95, 0.99];

[0105] When the weight is lower than the threshold, the entry is marked for deletion, for example, if the codebook entry C1-005 of edge gateway N1 is not matched for 50 consecutive periods, the weight decreases to below the specified threshold, and is cleaned up to release storage space;

[0106] When the similarity between the new feature vector v_new and all existing codebook entries C_i is lower than the threshold value Si, a new entry is created, and the specific operation is as follows:

[0107] The similarity between the new feature v_new and all entries is calculated, and if max(sim(v_new, C_i)) < Si, the creation is triggered;

[0108] v_new is taken as the new cluster center to generate entry C_j;

[0109] The initial weight w = 1.0, and the metadata, such as the creation time and the associated device ID, is associated;

[0110] If the similarity between the subsequent feature v and C_j is sim(v, C_j) > Si, then v is included in the cluster corresponding to C_j, and the cluster center is updated;

[0111] The device compares the new feature vector with the codebook, finds the most similar entry, and only transmits the codebook index;

[0112] The original feature may contain sensitive information, and the use of the transmission index can avoid data leakage, even if the attacker obtains the index, it is also difficult to restore the original feature from the index, reducing the risk of data tampering;

[0113] For example, the edge gateway D5 new feature [0.92, 0, 0.81, 0] has a similarity of 0.99 with the codebook index C1-002, and D5 sends the index C1-002 instead of the original vector, not only effectively compressing the data volume, but also protecting the data;

[0114] The edge gateway uploads the local code to the cloud regularly, and the cloud fuses the codebooks of each edge node based on the idea of federated learning to generate a global model;

[0115] For example, the cloud collects 100 codebook entries of edge gateway N1 and 80 codebook entries of N2, merges similar entries to generate a global codebook G-001, and pushes the update to each edge gateway to optimize the quality of the local codebook.

[0116] Step 3: Normalize the proportion of high-entropy quantity, remaining power, and average codebook index length of the edge gateway to obtain a three-dimensional state vector, calculate the cosine similarity and Euclidean distance between the three-dimensional state vector and each protocol preference vector, and comprehensively obtain the protocol score. According to the protocol score, the optimal communication protocol is selected;

[0117] In the above step 1, it has been marked whether each feature vector is high-entropy. By calculating the proportion of high-entropy quantity in the total quantity in each edge gateway, the proportion of high real-time data is reflected, directly reflecting the real-time of data.

[0118] The average codebook index length can directly reflect the data size. For example, 10 groups of feature vectors are obtained at the same time, and 3 codebooks are generated after clustering, with indexes as follows: Codebook one, 2 bytes (corresponding to 5 groups of feature vectors); Codebook two, 3 bytes (corresponding to 3 groups of feature vectors); Codebook three, 1 byte (corresponding to 2 groups of feature vectors). Then the total codebook index length Ltotal = 2 × 5 + 3 × 3 + 1 × 2 = 21 bytes, and the average length Lavg = 21 / 10 = 2.1 bytes.

[0119] After normalizing the proportion of high-entropy quantity H, the remaining power E, and the average codebook index length L, a three-dimensional state vector SS = [H, E, L] is formed, H ∈ [0, 1], E is normalized by dividing by 100, and L is normalized by dividing by the preset maximum length threshold.

[0120] Based on the characteristics of the protocol, a preference vector Pi (i ∈ [1, 3], i is an integer) is predefined for each protocol. The specific preference vector is as follows:

[0121] LPWAN, characterized by low real-time requirement, high power sensitivity, and low data volume sensitivity, can be set to P1 = [0.2, 0.8, 0.3];

[0122] BLE, characterized by balanced indicators, can be set to P2 = [0.5, 0.5, 0.5],

[0123] 5G, characterized by high real-time demand, low power sensitivity, high data processing capacity, P3 can be set as [0.8, 0.2, 0.7];

[0124] For each protocol, calculate the cosine similarity Similarity(Pi) between its preference vector Pi and the state vector SS;

[0125] To avoid the state vector deviating too much from the protocol vector in space, increase the Euclidean distance penalty, the specific formula is Distance(Pi)=1 / (1+||SS-Pi||);

[0126] According to the formula Score(Pi)=Similarity(Pi)×Distance(Pi), calculate the protocol score;

[0127] Select the protocol with the highest score as the corresponding transmission protocol;

[0128] Record the historical transmission success rate R(Pi,SS) of each protocol under a certain state, according to the formula Update the protocol preference vector, where P i,new P is the updated protocol preference vector, P i,old is the protocol preference vector before updating, i is the number of protocols, j is the number of transmissions, is the learning rate;

[0129] According to the formula DecisionMargin=max(Score(Pi))-Secondmax(Score(Pi)), calculate the difference between the highest score and the second highest score, where max(Score(Pi)) is the highest score and Secondmax(Score(Pi)) is the second highest score;

[0130] If DecisionMargin<Threshold Scoth, enter the flexible decision, the specific rules are as follows:

[0131] If H>0.7, corresponding to high real-time, prefer 5G;

[0132] If E<0.3, corresponding to low power, prefer LPWAN;

[0133] Otherwise, select the protocol with the lowest historical energy consumption under the current power state;

[0134] According to the formula A=H×(1-E)×L, calculate the abnormality degree, quantify the degree of system state deviation from the normal level, provide objective basis for triggering emergency communication strategy, and avoid system failure caused by misjudgment;

[0135] If A > threshold Ath, Ath is usually 0.6, forced to choose 5G;

[0136] 5G has high bandwidth, meets the demand of burst data transmission in abnormal situation; low delay, guarantees the rapid delivery of decision instructions; high reliability, reduces the data loss rate, and can effectively cope with the communication quality challenge in abnormal state;

[0137] The specific parameters of the start are as follows:

[0138] Maximize bandwidth, a large amount of alarm data may be generated in abnormal scenarios, maximizing bandwidth can ensure unblocked data transmission, and avoid loss of critical information due to insufficient bandwidth;

[0139] Minimize retransmission interval, which can shorten the data retransmission period, quickly repair the packet loss caused by network interference, reduce communication delay, and ensure the real-time performance of system response;

[0140] Enable standby communication channel, through the redundancy mechanism of main and standby channels, avoid communication interruption caused by single link failure, and improve the communication fault tolerance of the system in abnormal state.

[0141] Step 4, integrate the radio frequency fingerprint identification module in the radio frequency front end of the Internet of Things hardware device, sample the surrounding wireless signals in real time, and extract the frequency domain features to generate a frequency domain feature vector. The frequency domain feature vector is binarized by threshold to obtain a binary vector. The binary vector and the template in the preset mode library are calculated by Hamming distance. When the distance is less than the threshold, the device is triggered to wake up, otherwise the device enters deep sleep;

[0142] Select low-power radio frequency chip to adapt to the endurance demand of the device, avoid the sharp increase of power consumption after integrating the module, and capture the physical layer features of the surrounding wireless signals in real time to provide data basis for wake-up judgment;

[0143] Adopt sliding window sampling, extract signal frequency domain features such as center frequency and harmonic component through fast Fourier transform FFT, generate 32-dimensional frequency domain feature vector, FFT calculation complexity is low, 32-dimensional vector calculation time is low on processor, meet the power limit of low-power device;

[0144] The 32-dimensional floating-point frequency domain feature vector extracted in real time is binarized by threshold, for example, elements greater than 0.5 are set to 1, and elements less than or equal to 0.5 are set to 0, and converted to 32-bit binary vector;

[0145] The binarization process filters the signal amplitude fluctuation, only retains the pattern shape feature, and is suitable for complex electromagnetic environment;

[0146] The preset mode library also stores the binarized template;

[0147] 32-bit binary vector only needs 4 bytes of storage, while floating point vector needs 128 bytes, and the preset mode library capacity can be expanded to more than 20 groups;

[0148] The Hamming distance between the real-time vector and the template is calculated, and if the distance is less than or equal to a threshold, the wake-up is triggered;

[0149] When there is no match, the chip enters deep sleep (CPU is turned off, only RF wake-up circuit is reserved), and the sampling is activated once every 5 seconds through a timer, and after a successful match, the MCU is woken up for data compression, and is transmitted to the corresponding edge gateway through the communication module;

[0150] In the intermittent sampling mode, the average power consumption is greatly reduced compared to the continuous listening case, which can effectively prolong the battery life.

[0151] As Figure 2 The low-power communication system based on the Internet of Things hardware device comprises:

[0152] The data compression module collects original data using the Internet of Things hardware device, and uses first-order forward difference combined with an adaptive segmented activation function to preliminarily compress the original data, and then performs secondary compression through information entropy screening;

[0153] The codebook generation module constructs the edge gateway into a distributed clustering network, and generates a dynamic feature codebook based on the received compressed feature vector using the local sensitive hashing (LSH) algorithm;

[0154] The protocol decision module combines the high-entropy quantity proportion, the remaining power, and the average codebook index length of the edge gateway into a three-dimensional state vector, calculates the protocol score according to the cosine similarity and Euclidean distance between the three-dimensional state vector and each protocol preference vector, and selects the optimal communication protocol according to the protocol score;

[0155] The wake-up control module integrates the identification module in the RF front end of the Internet of Things hardware device, collects wireless signals in real time and extracts frequency domain features to generate a vector, performs binaryzation on the vector, and calculates the Hamming distance with a preset template, and triggers the wake-up when the distance is less than a threshold, otherwise controls the device to sleep.

[0156] Some data in the above formula are dimensionless for numerical calculation, and the contents not described in detail in the specification all belong to the prior art known to those skilled in the art.

[0157] The above embodiments are only used to illustrate the technical method of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A low-power communication method based on Internet of Things (IoT) hardware devices, characterized in that, include: Step 1: Collect raw data using IoT hardware devices, compress the raw data initially using first-order forward difference combined with an adaptive piecewise activation function, and then compress it a second time by filtering information entropy. The specific operation of the second compression by filtering information entropy is as follows: calculate the information entropy of the feature vector after initial compression, set the entropy threshold to En, retain feature points with entropy values ​​higher than En, set the rest to zero, and calculate the proportion of the number of zeros to the total number. If it is greater than 0.5, then mark the feature vector as high entropy. Step 2: Construct the edge gateway into a distributed clustering network, and generate a dynamic feature codebook based on the received compressed feature vector using the Locality Sensitive Hash (LSH) algorithm. Step 3: Combine the high entropy ratio H of edge gateways, the remaining power E, and the average codebook index length L into a three-dimensional state vector. Calculate the protocol score based on the cosine similarity and Euclidean distance between the three-dimensional state vector and each protocol preference vector. Select the optimal communication protocol based on the protocol score. Step 4: Integrate an identification module into the radio frequency front-end of the IoT hardware device to collect wireless signals in real time and extract frequency domain features to generate vectors. After binarizing the vectors, calculate the Hamming distance with a preset template. When the distance is less than a threshold, trigger wake-up; otherwise, control the IoT hardware device to go into sleep mode. The object that triggers wake-up is the core processing unit MCU and related functional modules of the IoT hardware device.

2. The low-power communication method based on IoT hardware devices according to claim 1, characterized in that, The specific steps for initial compression using first-order forward difference combined with an adaptive piecewise activation function are as follows: The data gradient is obtained by performing a first-order forward difference on the multidimensional data collected by IoT hardware devices. Calculate the mean avg and standard deviation piu of the gradient sequence; The upper threshold is calculated using the formula a_high=avg+k1×piu; The lower threshold is calculated using the formula a_low=avg-k1×piu; The compression coefficient is calculated using the formula b=1 / (1+exp(-k2×piu)), where k1 and k2 are configurable parameters. When the gradient value is greater than or equal to a_high, the original value is retained; When a_low ≤ gradient value < a_high, it is compressed to gradient value × b, where b is the compression coefficient; When the gradient value is less than a_low, it is compressed to 0.

3. The low-power communication method based on IoT hardware devices according to claim 1, characterized in that, The specific steps for generating a dynamic feature codebook using the Locality Sensitive Hashing (LSH) algorithm are as follows: The IoT hardware device sends the compressed feature vector to the corresponding edge gateway, and performs min-max normalization on the feature vector. Configure LSH parameters, setting the number of hash functions to [128, 256] and the number of buckets to [1000, 10000]. The hash function is designed according to the formula h(v)=sign(v×r), where v is the compressed feature vector and r is the random projection vector; Multiple hash functions are combined to generate a hash signature, which is then mapped to the corresponding hash bucket. Feature vectors falling into the same hash bucket are clustered using K-means to generate local codebook entries, and a unique index is assigned to each codebook entry.

4. The low-power communication method based on IoT hardware devices according to claim 3, characterized in that, Each codebook entry is associated with a dynamic weight value w, initially set to 1.

0. The weights of all codebook entries decay according to the formula w_new=w_old×(T_total / T). When the weight falls below the threshold, the entry is marked for deletion. Here, T represents the update cycle, and T_total represents the time since the codebook entry was created.

5. The low-power communication method based on IoT hardware devices according to claim 1, characterized in that, The specific steps for calculating the protocol score are as follows: After normalizing the high-entropy quantity ratio H, the remaining power E, and the average codebook index length L, a three-dimensional state vector SS=[H,E,L] is formed. Based on the characteristics of the protocols, a preference vector Pi is predefined for each protocol, where i∈[1,3] and i is an integer. The specific preference vectors are as follows: LPWAN, P1 is set to [0.2,0.8,0.3]; BLE, P2 is set to [0.5,0.5,0.5]; 5G, P3 is set to [0.8,0.2,0.7]. For each protocol, calculate the cosine similarity (Pi) between its preference vector Pi and its state vector SS; The Euclidean distance penalty is calculated using the formula Distance(Pi)=1 / (1+||SS-Pi||). The protocol score is calculated using the formula Score(Pi) = Similarity(Pi) × Distance(Pi); The protocol with the highest score is selected as the corresponding transmission protocol.

6. The low-power communication method based on IoT hardware devices according to claim 5, characterized in that, Record the historical transmission success rate R(Pi,SS) of each protocol under a specific state, according to the formula Update the protocol preference vector, where P i,new For the updated protocol preference vector, P i,old This is the protocol preference vector before the update, where i is the number of protocols and j is the number of transmissions. The learning rate is used to define the specific state, which is the state corresponding to the three-dimensional state vector SS composed of the proportion of high-entropy quantities H, the remaining power E, and the average codebook index length L.

7. The low-power communication method based on IoT hardware devices according to claim 1, characterized in that, The difference between the highest and second-highest scores is calculated using the formula DecisionMargin=max(Score(Pi))-Secondmax(Score(Pi)). If DecisionMargin < the threshold Scoth, then a flexible decision is made, with the following specific rules: if H > 0.7, 5G is selected first; if E < 0.3, LPWAN is selected first; otherwise, the protocol with the lowest historical energy consumption under the current power condition is selected, where max(Score(Pi)) is the highest score and Secondmax(Score(Pi)) is the second-highest score.

8. The low-power communication method based on IoT hardware devices according to claim 1, characterized in that, The anomaly degree is calculated according to the formula A=H×(1-E)×L in step 3. If A>th threshold, 5G is forcibly selected and the following three measures are activated at the same time: bandwidth maximization, retransmission interval minimization, and activation of backup communication channel.

9. A low-power communication system based on Internet of Things (IoT) hardware devices, used to execute the low-power communication method based on IoT hardware devices as described in any one of claims 1-8, characterized in that, include: The data compression module uses IoT hardware devices to collect raw data, uses first-order forward difference combined with an adaptive piecewise activation function to initially compress the raw data, and then performs secondary compression through information entropy filtering. The specific operation of the secondary compression through information entropy filtering is as follows: calculate the information entropy of the feature vector after initial compression, set the entropy threshold to En, retain feature points with entropy values ​​higher than En, set the rest to zero, and calculate the proportion of the number of zeros to the total number. If it is greater than 0.5, the feature vector is marked as high entropy. The codebook generation module constructs the edge gateway into a distributed clustering network and generates a dynamic feature codebook based on the received compressed feature vector using the Locality Sensitive Hash (LSH) algorithm. The protocol decision module combines the proportion of high-entropy edge gateways (H), remaining power (E), and average codebook index length (L) into a three-dimensional state vector. Based on the cosine similarity and Euclidean distance between the three-dimensional state vector and each protocol preference vector, the protocol score is calculated, and the optimal communication protocol is selected based on the protocol score. The wake-up control module integrates an identification module in the radio frequency front-end of the IoT hardware device. It collects wireless signals in real time and extracts frequency domain features to generate vectors. After binarizing the vectors, it performs Hamming distance calculation with a preset template. When the distance is less than a threshold, it triggers wake-up; otherwise, it controls the IoT hardware device to go into sleep mode. The object that triggers wake-up is the core processing unit MCU and related functional modules of the IoT hardware device.

Images