Machine learning based iot communication link quality prediction and adaptive adjustment method

Abstract

The present application belongs to the technical field of communication quality regulation, and specifically relates to a machine learning-based Internet of Things communication link quality prediction and adaptive adjustment method, aiming to solve the problems of existing technology adjustment lag, low prediction accuracy and poor adaptability. The method comprises: collecting multi-dimensional link original data and preprocessing to obtain a standardized feature data set; constructing a link quality correlation analysis rule base, mining correlation rules through mutual information entropy and an improved Apriori algorithm, obtaining an optimal rule base through verification and outputting a quality grade; constructing a multi-objective optimization decision logic, matching a strategy library to select an optimal adjustment strategy and executing; collecting adjusted data, repeating the preprocessing and prediction process to obtain actual and predicted values; calculating an index change, readjusting the strategy if it does not meet the standard, and simultaneously supplementing data to iteratively update the rule base and strategy library. The present application realizes accurate prediction and multi-objective optimization adjustment of link quality, dynamically adapts to complex link environments, ensures stable operation of Internet of Things services, and has strong applicability.

Description

Technical Field

[0001] This invention belongs to the field of communication quality control technology, and in particular relates to a machine learning-based method for predicting and adaptively adjusting the quality of IoT communication links. Background Technology

[0002] With the rapid development of IoT technology, IoT terminals have been widely used in various fields such as industrial control, smart security, and environmental monitoring. The quality of link communication directly determines the data transmission reliability, real-time performance, and operational stability of the IoT system. IoT communication links are susceptible to interference from complex environments (including electromagnetic interference, temperature and humidity changes, and obstructions), terminal movement, and fluctuations in service load, leading to problems such as increased packet loss rate, increased transmission latency, and unbalanced bandwidth utilization. In severe cases, this can cause data transmission interruptions, affecting the normal operation of IoT services.

[0003] To ensure link communication quality, existing technologies mostly employ link quality monitoring and passive adjustment strategies. This involves real-time collection of link parameters and triggering an adjustment mechanism upon detecting link quality deterioration. However, this approach suffers from adjustment lag and struggles to proactively mitigate the risk of link quality degradation. Some technologies attempt to introduce correlation rules or simple predictive models for early adjustment, but these still have several shortcomings.

[0004] First, the construction of the link quality correlation analysis rule base relies on manual experience to divide feature intervals, which has poor adaptability to complex link features and is difficult to cover multi-dimensional parameter combination scenarios, resulting in insufficient prediction accuracy.

[0005] Second, the lack of quantitative methods in association strength mining makes it easy to retain weak association features, which interferes with the reliability of rule generation.

[0006] Third, the adaptive adjustment strategy lacks multi-objective optimization logic, making it difficult to balance core indicators such as packet loss rate, latency, and bandwidth utilization. Furthermore, it does not consider the strategy execution cost, resulting in an imbalance between adjustment effect and economy. Fourth, the link quality correlation analysis rule base and strategy base update mechanism are rigid, unable to dynamically adapt to the adjusted link state, and the prediction and adjustment capabilities degrade after long-term operation.

[0007] Furthermore, the existing technology lacks sufficient coordination between link quality prediction and adaptive adjustment. The prediction results cannot accurately guide the selection of adjustment strategies, and there is a lack of reverse verification and iterative optimization of the prediction model or link quality correlation analysis rule base after adjustment, forming a disconnected closed loop of "prediction-adjustment", which makes it difficult to meet the high precision and high reliability requirements of IoT services for link quality.

[0008] Therefore, this invention aims to provide a link quality control method that can accurately mine the correlation patterns of link features, achieve early prediction and multi-objective optimization and adjustment, and dynamically iterate and optimize, thereby solving the technical problems of existing technologies such as lag, poor adaptability and insufficient coordination. Summary of the Invention

[0009] The purpose of this invention is to provide a machine learning-based method for predicting and adaptively adjusting the quality of IoT communication links, in order to solve the technical problems of lag, poor adaptability and insufficient coordination in existing technologies.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A machine learning-based method for predicting and adaptively adjusting IoT communication link quality includes the following steps:

[0012] S1: Real-time acquisition of raw data during the link communication process, including physical layer parameters, link layer parameters and environmental interference parameters, and preprocessing of the acquired raw data to obtain a standardized link feature dataset;

[0013] S2: Construct a link quality correlation analysis rule base, which is generated by mining the historical correlation between core features and link quality evaluation indicators; input the standardized link feature dataset into the link quality correlation analysis rule base, and output the link quality level within a future preset time window;

[0014] S3: Construct multi-objective optimization decision logic with the goals of minimizing link packet loss rate, reducing transmission latency and maximizing bandwidth utilization. Match the preset adjustment strategy library, select the optimal adaptive adjustment strategy that is suitable for the current link state, and execute the link parameter adjustment.

[0015] S4: Perform real-time data collection on the adjusted link, repeat steps S1 and S2, and obtain the actual and predicted values ​​of the adjusted link quality.

[0016] S5: Calculate the change in link quality evaluation indicators before and after adjustment. If the change does not reach the preset optimization threshold, return to step S3 and rematch the adjustment strategy based on the new link status. At the same time, supplement the adjusted link data into the original dataset and regularly update the link quality correlation analysis rule base and adjustment strategy base.

[0017] Preferably, the specific process of constructing the link quality correlation analysis rule base in step S2 is as follows:

[0018] S21: Select a standardized link feature dataset and supplement it with historical operation data of IoT communication links within the corresponding time period. The historical operation data includes packet loss rate, transmission delay, bandwidth utilization, and actual signal-to-noise ratio values ​​that correspond one-to-one with the core features, and construct a complete historical sample set; perform time axis alignment processing on the complete historical sample set to obtain the aligned historical sample set.

[0019] S22: Based on the aligned historical sample set, an adaptive threshold partitioning algorithm is used to partition each core feature obtained by screening into intervals. Each core feature is divided into three feature intervals: high, medium, and low according to the numerical distribution, forming a feature interval set. According to the preset link quality level classification standard, the actual values ​​of packet loss rate, transmission delay, bandwidth utilization, and signal-to-noise ratio are mapped to four link quality levels: excellent, good, average, and poor, respectively, and a mapping relationship table between link quality evaluation indicators and quality levels is established.

[0020] S23: Based on the mapping table between link quality evaluation indicators and quality levels, the correlation strength between each core feature interval and the corresponding quality level of each link quality evaluation indicator is calculated by mutual information entropy. Strongly correlated feature intervals with correlation strength greater than a preset threshold are retained to obtain a strongly correlated sample subset. An improved Apriori association rule algorithm is adopted, using the combination of multiple core feature intervals in the strongly correlated sample subset as the antecedent and the quality level and change trend of the corresponding link quality evaluation indicator as the consequent. Minimum support and minimum confidence thresholds are set to mine high-frequency and high-reliability association rules that meet the threshold requirements to form a candidate association rule set.

[0021] S24: Perform rule conflict detection and deduplication on the candidate association rule set; for the rules to be optimized with confidence levels within the preset threshold range, supplement environmental interference parameter constraints based on the actual operating scenarios of IoT communication links to optimize the accuracy of rule description; classify and archive the optimized rules according to the link quality evaluation index type to form an initial link quality association analysis rule library.

[0022] S25: Select a subset of validation samples from the aligned historical sample set, input the core feature data of the validation samples into the initial link quality association analysis rule base, obtain the link quality prediction level through rule matching, compare it with the actual quality level corresponding to the validation sample, and calculate the matching accuracy of each rule; remove invalid rules with accuracy lower than the preset qualified threshold, and for rules with qualified accuracy but insufficient coverage, supplement the corresponding feature combination samples and re-mine and optimize them, finally obtaining the optimal link quality association analysis rule base that meets the link quality prediction requirements.

[0023] Preferably, the specific process for obtaining the strongly correlated sample subset in step S23 is as follows:

[0024] S231: Extract different intervals of a single core feature from the feature interval set, denoted as the feature interval variable X, where X = {x1, x2, ..., x...} n}, where n is the number of intervals for a single core feature, here n=3, corresponding to high, medium, and low intervals; extract the quality level corresponding to a certain type of link quality evaluation index from the mapping relationship table between link quality evaluation index and quality level, denoted as quality level variable Y, Y={y1,y2,y3,y4}, corresponding to excellent, good, average, and poor levels respectively; using the aligned historical sample set as the data source, statistically obtain the joint probability distribution P(X,Y) and marginal probability distributions P(X) and P(Y) of feature interval variable X and quality level variable Y;

[0025] S232: Using mutual information entropy to quantify the correlation strength between the feature interval variable X and the quality grade variable Y:

[0026] ;

[0027] in, x For characteristic interval variables X Any value of , that is, a certain interval of a single core feature. y Quality grade variable Y Any value of P(i, j) represents a certain link quality level; x,y P(x) represents the joint probability that the feature interval x and the quality level y occur simultaneously. x P(x) represents the marginal probability of the feature interval x occurring. y () represents the quality level y The marginal probability of occurrence;

[0028] S233: Strong correlation feature interval screening and sample subset construction: A preset correlation strength threshold T is set, and the calculated correlation strength value is compared with the threshold T; core feature intervals with correlation strength values ​​> T are retained, and weak correlation feature intervals with correlation strength values ​​≤ T are removed; at the same time, samples containing weak correlation feature intervals are removed from the aligned historical sample set, and the remaining samples are all composed of combinations of strong correlation feature intervals and corresponding quality levels to form a strong correlation sample subset.

[0029] Preferably, the specific process for obtaining the candidate association rule set in step S23 is as follows:

[0030] S234: Using strongly correlated sample subsets as the data source, define itemsets as single or combined values ​​of core feature intervals, where itemsets containing only a single core feature interval are 1-itemsets; set a minimum support threshold. S min and minimum confidence threshold C min ;

[0031] S235: Traverse the strongly associated sample subsets, count the frequency of all 1-itemsets, calculate the support of each 1-itemet, and retain items with support ≥ S min 1-itemsets that form frequent 1-itemsets are removed, and those with support < S min Infrequent 1-itemsets;

[0032] Self-joining and pruning operations are performed on frequent 1-itemsets to generate candidate k-itemsets; during the pruning stage, subsets of candidate k-itemsets containing non-frequent (k-1)-itemsets are removed, and step S235 is repeated until no new frequent k-itemsets can be generated. All frequent itemsets are then aggregated to form a high-frequency itemset set.

[0033] S236: Association Rule Generation and Filtering: For each frequent itemset in the high-frequency itemset set, split it to generate all possible rule forms, calculate the confidence score of each rule, and retain rules with a confidence score ≥ C min Association rules, removing rules with confidence levels < C min Low reliability rules are identified; redundancy detection is performed on the generated rules, redundant rules are removed, and the remaining rules are classified and organized according to the link quality evaluation index type to form a candidate association rule set.

[0034] Preferably, the specific process of step S25 is as follows:

[0035] S251: Randomly select samples from the aligned historical sample set according to a preset ratio to form a validation sample subset, and retain the remaining samples as a supplementary data source for subsequent rule optimization.

[0036] S252: For each validation sample in the validation sample subset, extract core feature data, repeat the standardization process to obtain standardized validation feature data; input the standardized validation feature data into the generated initial link quality correlation analysis rule base, use the feature interval precise matching algorithm to traverse the correlation rules of the corresponding link quality evaluation index in the link quality correlation analysis rule base, determine the target rule that completely matches the core feature interval combination of the validation sample, and output the link quality prediction level corresponding to the target rule;

[0037] S253: Extract the actual link quality level corresponding to each verification sample from the verification sample subset, compare the predicted link quality level with the actual quality level one by one, and record the matching result of each target rule, which is divided into two categories: successful matching and failed matching.

[0038] S254: Based on the comparison results, the number of successful matches for each association rule and the total number of matches are counted separately, and the accuracy index is used to quantify the effectiveness of a single rule;

[0039] S255: Preset accuracy pass threshold A min The accuracy of each rule is compared with A min Comparison; Elimination accuracy < A min Invalid rules, retain those with an accuracy ≥ A min Effective rules; for accuracy ≥ A min However, rules with a total number of matches less than the preset threshold are marked as rules to be optimized, providing a basis for subsequent supplementary sample mining, and finally forming a set of rules that have been validated for effectiveness.

[0040] Preferably, the specific process of step S3 is as follows:

[0041] S31: Extract the link quality prediction results, including the predicted packet loss rate. L pred Transmission delay prediction value D pred Bandwidth utilization forecast U pred and signal-to-noise ratio prediction value S pred Based on the established link quality level classification standards, the current predicted link quality level and core bottleneck indicators are clearly defined; current link communication requirements are collected, requirement parameters are quantified according to service type, and communication requirement constraint thresholds are set, including transmission delay constraint thresholds for real-time control services. D max Packet loss rate constraint threshold L max Bandwidth utilization constraint threshold for data acquisition services U min This forms a quantified set of communication requirement constraints.

[0042] S32: Using the analytical prediction results and quantified communication requirements as input, a multi-objective optimization decision logic is constructed. The core optimization objectives are to minimize the link packet loss rate, reduce transmission latency, and maximize bandwidth utilization. The predicted signal-to-noise ratio is then used as the input. S pred As a constraint. S pred ≥ S 1 min , S 1 min To determine the minimum signal-to-noise ratio threshold for stable link communication, a weighted approach is used to transform the multi-objective optimization function into a single-objective optimization function.

[0043] S33: Call the preset adjustment strategy library, which includes strategies such as modulation and demodulation mode switching, transmission power adjustment, communication frequency band switching, and data packet strategy adjustment;

[0044] S34: Substitute each strategy in the candidate adjustment strategy set into the optimization objective function F, and calculate the F value corresponding to each strategy; introduce the strategy execution cost coefficient C, including energy consumption cost and handover delay cost, and modify the optimization objective function to F'=F×(1+C), where the energy consumption cost is quantified according to the transmission power adjustment amplitude, and the handover delay cost is quantified according to the time consumption of modulation mode / frequency band switching. Select the strategy with the smallest modified F' value as the optimal adaptive adjustment strategy. If there are multiple strategies with the same F' value, the strategy with the smallest execution cost coefficient C is selected first.

[0045] S35: The IoT gateway parses the optimal adaptive adjustment strategy into standardized control commands and sends the control commands to the corresponding IoT terminal through a secure communication channel. After receiving the commands, the terminal starts the parameter configuration module and completes the link parameter adjustment according to the preset timing.

[0046] Preferably, the specific process of step S4 is as follows:

[0047] S41: Real-time data acquisition is performed on the adjusted IoT communication link to form the adjusted raw dataset, and after preprocessing, the adjusted standardized link feature dataset is obtained.

[0048] S42: Extract the actual values ​​of link quality evaluation indicators from the adjusted standardized link feature data, including the adjusted packet loss rate. L act Adjusted transmission delay D act Adjusted bandwidth utilization U act and the adjusted signal-to-noise ratio S act ;

[0049] S43: Input the standardized link feature data into the optimal link quality correlation analysis rule base, repeat the link quality analysis and prediction process, and obtain the adjusted link quality prediction value. L act_pred , D act_pred , U act_pred , S act_pred This is used to verify the prediction accuracy of the link quality correlation analysis rule base under the adjusted link status.

[0050] Preferably, the process of verifying the prediction accuracy of the link quality correlation analysis rule base under the adjusted link state in step S43 is as follows:

[0051] S431: Calculation accuracy evaluation indicators: The mean absolute error (MAE), root mean square error (RMSE), and prediction accuracy (Acc) are selected as the core accuracy evaluation indicators to quantify the prediction deviation of single-type link quality indicators and the overall prediction reliability.

[0052] S432: Preset accuracy evaluation threshold standards, including MAE threshold ≤ 5% of preset indicator benchmark value, RMSE threshold ≤ 8% of preset indicator benchmark value, and prediction accuracy Acc ≥ 85%;

[0053] S433: When all three accuracy indicators meet the corresponding threshold requirements, it is determined that the prediction accuracy of the link quality correlation analysis rule base under the adjusted link state is up to standard, and it will continue to be used for subsequent link quality prediction; when any accuracy indicator does not meet the threshold requirements, it is marked as insufficient adaptability of the link quality correlation analysis rule base. In the iterative optimization process, the data after this adjustment is given priority to be used as the core sample to supplement the dataset, and the feature association logic and interval division standard of the link quality correlation analysis rule base are optimized in a targeted manner.

[0054] Preferably, the specific process of step S5 is as follows:

[0055] S51: Extract the link quality prediction value before adjustment ( L pred , D pred , U pred , S pred Using this as a benchmark, the changes in each core link quality evaluation index are calculated, including the change in packet loss rate Δ. L Transmission delay change Δ D Change in bandwidth utilization Δ U、 Signal-to-noise ratio change ΔS; ΔL and ΔD are positive optimization changes, ΔU and ΔS are positive optimization changes; preset optimization threshold set. T ={ T L , T D , T U , T S} are the minimum optimization thresholds for packet loss rate, transmission delay, bandwidth utilization, and signal-to-noise ratio, respectively;

[0056] S52: If the positive change of all core indicators is greater than or equal to the corresponding optimization threshold, the adjustment effect is deemed to be satisfactory; if the positive change of any core indicator is less than the corresponding optimization threshold, or if a negative change occurs, the adjustment effect is deemed to be unsatisfactory, and the process immediately returns to step S3. Based on the adjusted link status data, steps S31-S35 are re-executed to match and execute a new optimal adaptive adjustment strategy until the adjustment effect is satisfactory.

[0057] S53: Add the adjusted original dataset and the actual and predicted values ​​of the adjusted quality indicators to the original dataset of step S1 to form the expanded original dataset, and update the standardized link feature dataset at the same time; start the iterative optimization process regularly.

[0058] The beneficial effects of this invention include:

[0059] 1. Improved accuracy and foresight in link quality prediction, mitigating the risk of quality degradation: This invention quantifies feature intervals using an adaptive threshold partitioning algorithm, combines mutual information entropy to accurately mine the correlation strength between core feature intervals and link quality levels, eliminates weak correlation interference, and then generates a high-frequency, high-reliability link quality correlation analysis rule base using an improved Apriori algorithm. Compared to manually constructing a link quality correlation analysis rule base, this significantly improves the scientific rigor and adaptability of the prediction logic. Simultaneously, the accuracy of the link quality correlation analysis rule base is verified using a subset of validation samples, further ensuring prediction accuracy and achieving precise prediction of link quality levels within a preset time window. This breaks away from the traditional passive adjustment model and proactively mitigates the risk of link quality degradation.

[0060] 2. Achieving multi-objective collaborative optimization and adjustment, balancing adjustment effectiveness and economy: This invention constructs a multi-objective optimization decision logic with minimizing packet loss rate, reducing transmission latency, and maximizing bandwidth utilization as core objectives. It dynamically allocates indicator weights based on service type and introduces strategy execution cost coefficients (energy consumption, handover latency) to correct the optimization objectives, thus selecting the optimal strategy suitable for the current link state. Compared to single-objective adjustment strategies, this method can balance multi-dimensional link indicators, adapt to different service needs (real-time control, data acquisition), and reduce energy consumption and time costs during the adjustment process, thereby improving the overall efficiency of link adjustment.

[0061] 3. Strong dynamic iterative optimization capability, adaptable to complex link environment changes: After adjustment, this invention collects link data, repeats the preprocessing and prediction process, obtains actual and predicted values, and verifies the accuracy of the link quality correlation analysis rule base. Simultaneously, the adjusted data is added to the dataset, and the correlation link quality analysis rule base and adjustment strategy base are updated periodically. This iterative mechanism can dynamically optimize feature association logic, interval division standards, and strategy adaptation range, enabling the system to adapt to complex environmental interference, business load fluctuations, and link status changes. Even under long-term operation, it maintains high prediction and adjustment reliability, solving the problems of rigid databases and degraded adaptability in existing technologies.

[0062] 4. High versatility and practicality, adaptable to various IoT application scenarios: The collected link parameters cover multiple dimensions including the physical layer, link layer, and environmental interference. The adjustment strategies cover core methods such as modulation and demodulation, power adjustment, frequency band switching, and packet sub-strategies, making it suitable for different IoT application scenarios such as industrial control, environmental monitoring, and smart security. Furthermore, various threshold parameters (association strength threshold, support / confidence threshold, optimization threshold, etc.) can be dynamically set according to specific scenarios without requiring large-scale hardware modifications, effectively reducing costs. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the machine learning-based IoT communication link quality prediction and adaptive adjustment method of the present invention.

[0064] Figure 2 This is a flowchart illustrating the prediction accuracy of the verification link quality correlation analysis rule base of the present invention under the adjusted link state. Detailed Implementation

[0065] The following is in conjunction with the appendix Figures 1-2 The present invention will be further described in detail below:

[0066] Example 1

[0067] See appendix Figure 1 As shown, the machine learning-based method for predicting and adaptively adjusting IoT communication link quality includes the following steps:

[0068] S1: IoT Communication Link Data Acquisition and Preprocessing: A multi-dimensional data acquisition module deployed between IoT terminals and gateways collects raw data in real time during the link communication process. The raw data includes physical layer parameters, link layer parameters, and environmental interference parameters. The collected raw data is processed sequentially by outlier removal, missing value completion, data standardization, and feature filtering to obtain a standardized link feature dataset. The feature filtering adopts a joint filtering strategy of mutual information entropy and variance threshold, retaining core features that are more correlated with the link quality evaluation index than a preset threshold, providing data support for subsequent link quality analysis.

[0069] S2: Real-time Link Quality Analysis and Prediction: Based on the standardized link feature dataset, a link quality correlation analysis rule base is constructed. The link quality correlation analysis rule base is generated by mining the historical correlation between core features and link quality evaluation indicators (packet loss rate, transmission delay, bandwidth utilization, signal-to-noise ratio), covering the link quality change patterns corresponding to different feature combinations. The preprocessed real-time link core feature data is input into the link quality correlation analysis rule base. Through feature matching and logical reasoning, the link quality prediction results within a preset time window are output. Combined with the preset link quality level classification standard, the link quality level corresponding to the prediction results is determined.

[0070] S3: Generation and execution of link adaptive adjustment strategy based on prediction results: Combining the link quality prediction results and current link communication requirements, a multi-objective optimization decision logic is constructed to minimize the link packet loss rate, reduce transmission latency, and maximize bandwidth utilization. A preset adjustment strategy library is matched, which includes applicable scenarios and parameter ranges for modulation / demodulation mode switching, transmission power adjustment, communication frequency band switching, and data packet splitting strategy adjustment. The optimal adaptive adjustment strategy that adapts to the current link state is selected through logical decision-making. The IoT gateway issues adjustment instructions to the IoT terminal according to the adjustment strategy to execute the link parameter adjustment.

[0071] S4: Perform real-time data collection on the adjusted link from step S3, and repeat steps S1 and S2 to obtain the actual and predicted values ​​of the adjusted link quality.

[0072] S5: Calculate the change in link quality evaluation indicators before and after adjustment. If the change does not reach the preset optimization threshold, return to step S3 and rematch the adjustment strategy based on the new link status. At the same time, supplement the adjusted link data into the original dataset, regularly update the link quality correlation analysis rule base and adjustment strategy base, optimize the feature correlation logic and strategy adaptation standard, and realize the dynamic iteration of link adjustment capability.

[0073] Example 2

[0074] Based on Example 1, the specific process of constructing the link quality correlation analysis rule base in step S2 is as follows:

[0075] S21: Historical Dataset Expansion and Alignment Processing: Select the standardized link feature dataset from step S1, and supplement it with historical operating data of IoT communication links within the corresponding time period. The historical operating data includes packet loss rate, transmission delay, bandwidth utilization, and actual values ​​of signal-to-noise ratio that correspond one-to-one with the core features, thus constructing a complete historical sample set. Perform time axis alignment processing on the complete historical sample set to ensure that each core feature combination data and the corresponding actual value of the link quality evaluation index are accurately matched in the time dimension. Eliminate invalid samples with time misalignment to obtain the aligned historical sample set.

[0076] S22: Core Feature Interval Division and Quality Level Mapping: Based on the aligned historical sample set, an adaptive threshold division algorithm is used to divide the core features (physical layer, link layer, and environmental interference core features) obtained in step S1 into intervals. Each core feature is divided into three feature intervals—high, medium, and low—according to its numerical distribution, forming a feature interval set. Simultaneously, according to the preset link quality level classification standard, the actual values ​​of packet loss rate, transmission delay, bandwidth utilization, and signal-to-noise ratio are mapped to four link quality levels: excellent, good, average, and poor, respectively, establishing a mapping relationship table between link quality evaluation indicators and quality levels.

[0077] S23: Two-Dimensional Association Mining: Based on the feature interval set obtained in S22 and the mapping table between link quality evaluation indicators and quality levels, the association strength between each core feature interval and the corresponding quality level of each link quality evaluation indicator is first calculated using mutual information entropy. Strongly associated feature intervals with association strength greater than a preset threshold are retained, while weakly associated feature intervals and their corresponding samples are removed, resulting in a strongly associated sample subset. Subsequently, an improved Apriori association rule algorithm is used, taking the combination of multiple core feature intervals in the strongly associated sample subset as the antecedent and the quality level and trend of the corresponding link quality evaluation indicator as the consequent. Minimum support and minimum confidence thresholds are set to mine high-frequency, high-reliability association rules that meet the threshold requirements, forming a candidate association rule set.

[0078] S24: Deduplication and Optimization of Association Rules: The candidate association rule set undergoes rule conflict detection and deduplication, eliminating contradictory and duplicate rules. For rules to be optimized with confidence levels within a preset threshold range, environmental interference parameter constraints are added based on the actual operating scenarios of IoT communication links, including temperature ranges and electromagnetic interference intensity ranges, to optimize the accuracy of rule descriptions. Simultaneously, the optimized rules are categorized and archived according to link quality evaluation index types (packet loss rate, transmission delay, bandwidth utilization, signal-to-noise ratio) to form an initial link quality association analysis rule base.

[0079] S25: Validation and Correction of Link Quality Correlation Analysis Rule Base: A subset of validation samples is selected from the aligned historical sample set. The core feature data of the validation samples are input into the initial link quality correlation analysis rule base. The link quality prediction level is obtained through rule matching and compared with the actual quality level corresponding to the validation samples to calculate the matching accuracy of each rule. Invalid rules with accuracy below the preset qualified threshold are removed. For rules with qualified accuracy but insufficient coverage, corresponding feature combination samples are added and re-mined and optimized. Finally, the optimal link quality correlation analysis rule base that meets the requirements of link quality prediction is obtained, providing support for feature matching and logical reasoning for subsequent real-time link quality prediction.

[0080] In this embodiment, the specific process of obtaining the strongly correlated sample subset in step S23 is as follows:

[0081] S231: Extract different intervals of a single core feature from the feature interval set, denoted as the feature interval variable X, where X = {x1, x2, ..., x...} n}, where n is the number of intervals for a single core feature, here n=3, corresponding to high, medium, and low intervals. The quality level corresponding to a certain type of link quality evaluation indicator is extracted from the mapping table between link quality evaluation indicators and quality levels, denoted as the quality level variable Y, Y={y1,y2,y3,y4}, corresponding to excellent, good, average, and poor levels respectively. Using the corresponding aligned historical sample set as the data source, the joint probability distribution P(X,Y) and marginal probability distributions P(X) and P(Y) of the feature interval variable X and the quality level variable Y are statistically obtained.

[0082] S232: Mutual Information Entropy Calculation of Association Strength: Mutual information entropy (MI) is used to quantify the association strength between feature interval variable X and quality level variable Y. The larger the mutual information entropy, the higher the association between the core feature interval and the link quality level.

[0083] ;

[0084] in, x For characteristic interval variables X Any value of , that is, a certain interval of a single core feature. y For quality grade variables Y Any value of P(i, j) represents a certain link quality level; x,y P(x) represents the joint probability that the feature interval x and the quality level y occur simultaneously. x ) is the characteristic interval x The marginal probability of occurrence, P( y () represents the quality level yThe probability of occurrence of marginal features. The mutual information entropy value between each core feature interval and the corresponding quality level of each link quality evaluation index is calculated using the above formula, and this value is used as the correlation strength value between the two.

[0085] S233: Strong Correlation Feature Interval Filtering and Sample Subset Construction: A preset correlation strength threshold T is set according to the requirements of IoT link communication scenarios and historical data characteristics, with a value range of [0.1, 0.3]. The calculated correlation strength values ​​are compared with the threshold T; core feature intervals with correlation strength values ​​> T are retained, i.e., strong correlation feature intervals, while weak correlation feature intervals with correlation strength values ​​≤ T are removed. At the same time, samples containing weak correlation feature intervals are removed from the aligned historical sample set. The remaining samples are all composed of combinations of strong correlation feature intervals and corresponding quality levels, forming a strong correlation sample subset.

[0086] The specific process of obtaining the candidate association rule set in step S23 is as follows:

[0087] S234: Algorithm Initialization and Parameter Setting: Using a subset of strongly correlated samples as the data source, define itemsets as single or combined values ​​of core feature intervals, where itemsets containing only a single core feature interval are 1-itemsets; set the minimum support threshold. S min The values ​​are set based on the data volume of the strongly correlated sample subset and the requirements of the link scenario, with a range of [2%, 5%], and a minimum confidence threshold. C min The value range is [70%, 85%]. S min Used for filtering high-frequency itemsets C min Used to filter highly reliable association rules.

[0088] S235: High-frequency itemset mining: Traverse strongly associated sample subsets, count the frequency of occurrence of all 1-itemsets, and calculate the support of each 1-itemset. The support calculation formula is as follows:

[0089] Support (I)= Count (I) / Total ;

[0090] Where I is a certain 1-itemset, Count (I) represents the number of samples in a strongly correlated subset that contain itemset I. Total The total number of samples in the strongly correlated subset; retaining support ≥ S min 1-itemsets that form frequent 1-itemsets are removed, and those with support < S min Infrequent 1-itemsets.

[0091] Self-joining and pruning operations are performed on frequent 1-itemsets to generate candidate k-itemsets, where k ≥ 2, and k is the number of core feature intervals in the itemsets. In the self-joining stage, frequent (k-1)-itemsets are paired according to preset rules to generate candidate k-itemsets. In the pruning stage, subsets of candidate k-itemsets containing non-frequent (k-1)-itemsets are removed to reduce unnecessary computation. Steps S235 are repeated until no new frequent k-itemsets can be generated, and all frequent itemsets are aggregated to form a high-frequency itemset set.

[0092] S236: Association Rule Generation and Filtering: For each frequent itemset in the high-frequency itemset set, split it to generate all possible rule forms, that is, split the antecedent A (multi-core feature interval combination) and consequent B (quality level and trend of link quality evaluation indicators), and A∩B=∅, A≠∅, B≠∅; calculate the confidence of each rule, and the confidence calculation formula is as follows:

[0093] ;

[0094] Where A→B represents the association rule "if the combination of multiple core feature intervals of the antecedent A is satisfied, then the link quality state of the consequent B will appear", Support(A∪B) is the support of the joint occurrence of itemset A and itemset B, and Support(A) is the support of the antecedent A.

[0095] Retain association rules with confidence ≥ C_min and remove low-reliability rules with confidence < C_min; at the same time, perform redundancy detection on the generated rules and remove redundant rules that can be derived from other rules. The remaining rules are classified and organized according to the link quality evaluation index type to form a candidate association rule set, which provides a data source for rule deduplication and optimization in S2.4.

[0096] The specific process of step S25 is as follows:

[0097] S251: From the aligned historical sample set, randomly select samples according to a preset ratio (range 20%-30%) to form a verification sample subset, and retain the remaining samples as a supplementary data source for subsequent rule optimization; the verification sample subset must maintain the same data distribution as the aligned historical sample set to ensure that each core feature interval combination and corresponding link quality level are covered, and avoid verification deviation.

[0098] S252: Validation Sample Preprocessing and Rule Matching: For each validation sample in the validation sample subset, extract core feature data, repeat the standardization process in S1, and obtain standardized validation feature data; input the standardized validation feature data into the initial link quality correlation analysis rule base, use the feature interval precise matching algorithm to traverse the correlation rules of the corresponding link quality evaluation indicators in the link quality correlation analysis rule base, determine the target rule that completely matches the core feature interval combination of the validation sample, and output the link quality prediction level corresponding to the target rule.

[0099] S253: Comparison of prediction results with actual results: Extract the actual link quality level corresponding to each verification sample from the verification sample subset (determined based on the mapping relationship table between link quality evaluation indicators and quality levels), compare the predicted link quality level with the actual quality level one by one, record the matching result of each target rule, and divide it into two categories: successful matching (predicted level is consistent with actual level) and failed matching (predicted level is inconsistent with actual level).

[0100] S254: Single Rule Matching Accuracy Calculation: Based on the comparison results, the number of successful matches for each associated rule and the total number of matches are counted separately. The accuracy index is used to quantify the effectiveness of a single rule. The accuracy calculation formula is as follows:

[0101] Accuracy ( R i )=S success ( R i ) / Total Match ( R i );

[0102] in, R i For the first rule in the initial link quality correlation analysis rule base i Association rules, S success ( R i The number of successful matches for this rule in the subset of validation samples is denoted as . Total Match ( R i ) represents the total number of matches for this rule in the subset of validation samples, i.e., how many validation samples trigger this rule; if a rule has no matching samples ( Total Match ( R i If )=0), then it is marked as a sample validation rule to be supplemented.

[0103] S255: Rule Filtering and Classification: Preset Accuracy Threshold Amin The value range is [75%, 85%]. The accuracy of each rule is compared with... A min Comparison; Elimination accuracy < A min Invalid rules, retain those with an accuracy ≥ A min Effective rules; for accuracy ≥ A min However, rules with a total number of matches less than the preset threshold (i.e., insufficient coverage) are marked as rules to be optimized, providing a basis for subsequent supplementary sample re-mining, and finally forming a set of rules with validated effectiveness to support the subsequent link quality correlation analysis rule base optimization process.

[0104] Example 3

[0105] Based on Example 1 or Example 2, the specific process of step S3 is as follows:

[0106] S31: Prediction Result Analysis and Communication Demand Quantification: Extract the link quality prediction results output from step S2, including the predicted packet loss rate. L pred Transmission delay prediction value D pred Bandwidth utilization forecast U pred and signal-to-noise ratio prediction value S pred Based on the established link quality level classification standards, the current predicted link quality level and core bottleneck indicators are identified, such as excessive packet loss rate and high latency. Simultaneously, current link communication requirements are collected, and requirement parameters are quantified according to service type (real-time control, data acquisition), setting communication requirement constraint thresholds. Specifically, the transmission latency constraint threshold corresponds to the real-time control service. D max Packet loss rate constraint threshold L max Bandwidth utilization constraint threshold for data acquisition services U min This forms a quantified set of communication requirement constraints.

[0107] S32: Construction of Multi-Objective Optimization Decision Logic: Using the analytical prediction results and quantified communication requirements as input, a multi-objective optimization decision logic is constructed. The core optimization objectives are to minimize the link packet loss rate, reduce transmission latency, and maximize bandwidth utilization, while also incorporating the predicted signal-to-noise ratio. S pred As a constraint ( S pred ≥ S 1 min , S 1min (This refers to the minimum signal-to-noise ratio threshold for stable link communication). To achieve multi-objective collaborative optimization, a weighted method is used to transform the multi-objectives into a single-objective optimization function. The formula for calculating the objective function is as follows:

[0108] F= w 1· L pred / L max + w 2. D pred / D max + w 3· U pred / U max ;

[0109] Among them, among them, w 1. w 2. w 3 are the weighting coefficients for packet loss rate, transmission delay, and bandwidth utilization, respectively, satisfying... w 1+ w 2+ w 3=1, and the weighting coefficients are dynamically allocated according to the business type. For real-time control businesses, the weighting ratio of ω2 is ≥40%, and for data acquisition businesses... w 3. Weighting percentage ≥ 40%; L max This is the packet loss rate constraint threshold for the corresponding service. D max This is the transmission delay constraint threshold for the corresponding service. U max The maximum bandwidth utilization of the link is determined by the gateway hardware performance. The smaller the objective function F value, the better the adaptability of the corresponding adjustment strategy.

[0110] S33: Preset Adjustment Strategy Library Invocation and Matching: Invokes the preset adjustment strategy library, which is built based on the communication characteristics of IoT links and covers four core strategies: modulation and demodulation mode switching, transmission power adjustment, communication frequency band switching, and data packet strategy adjustment. Each strategy corresponds to a clear applicable scenario, parameter adjustment range, and impact coefficient on link quality indicators, as detailed below:

[0111] Modulation / demodulation mode switching: Applicable scenario for signal-to-noise ratio prediction. S pred In [ S 1 min , S midThe adjustment range includes three modulation methods: QPSK, 16QAM, and 64QAM. The corresponding packet loss rate reduction coefficient, latency change coefficient, and bandwidth utilization improvement coefficient are preset in the strategy library.

[0112] Transmission power adjustment: Applicable scenarios include excessive packet loss rate related to signal strength. L pred > L max The adjustment range is 30%-100% of the terminal's rated transmission power, with a power adjustment step size of 5%. The correlation coefficient between the corresponding power and the signal-to-noise ratio is pre-calibrated and archived.

[0113] Communication frequency band switching: Applicable scenarios where high electromagnetic interference intensity leads to excessive latency. The switching range includes IoT licensed frequency bands (including 868MHz and 915MHz) and unlicensed frequency bands (including 2.4GHz). The anti-interference capabilities and bandwidth characteristics of different frequency bands are preset as the matching criteria.

[0114] Data packet splitting strategy adjustment: Applicable scenarios are those with low bandwidth utilization. U pred <U min If the latency exceeds the standard, the adjustment range includes packet size (128B-1024B) and packet retransmission mechanism. The correlation model between the corresponding packet parameters and link utilization and latency is pre-embedded in the policy library.

[0115] S34: Optimal Adjustment Strategy Selection: Substitute each strategy in the candidate adjustment strategy set into the constructed optimization objective function F to calculate the F value for each strategy. Simultaneously, introduce a strategy execution cost coefficient C, including energy consumption cost and handover delay cost, and modify the optimization objective function to F' = F × (1 + C), where energy consumption cost is quantified based on the transmission power adjustment magnitude, and handover delay cost is quantified based on the modulation scheme / frequency band switching time. Select the strategy with the smallest modified F' value as the optimal adaptive adjustment strategy. If multiple strategies have the same F' value, prioritize the strategy with the smallest execution cost coefficient C to ensure a balance between adjustment effectiveness and economy.

[0116] S35: Adjustment Command Issuance and Parameter Execution: The IoT gateway parses the selected optimal adaptive adjustment strategy into standardized control commands. These commands include the strategy type, specific adjustment parameters (including modulation scheme code, transmission power value, target frequency band, and packet size), and execution timing requirements. The gateway issues control commands to the corresponding IoT terminal via a secure communication channel. Upon receiving the commands, the terminal activates the parameter configuration module and completes the link parameter adjustment according to a preset timing sequence. During the adjustment process, the execution status (success / failure) is fed back to the gateway in real time. If execution fails, the gateway automatically calls the second-best strategy from the candidate adjustment strategy set and reissues the commands until the adjustment is successful, ensuring that the link parameters are configured according to the optimal strategy.

[0117] Example 4

[0118] Based on Example 1, Example 2, or Example 3, the specific process of step S4 is as follows:

[0119] S41: Adjusted Link Data Acquisition and Preprocessing: After the link parameter adjustment is successfully executed and runs stably for a preset time (range 30s-5min, ensuring link stability), the multi-dimensional data acquisition module is started to collect real-time data from the adjusted IoT communication link. The data types collected are completely consistent with those in step S1, including physical layer parameters, link layer parameters, and environmental interference parameters, forming the adjusted original dataset. The preprocessing process of step S1 is repeated, sequentially performing outlier removal (combination of 3σ criterion and isolated forest algorithm), missing value completion (KNN weighted completion), data standardization (Z-Score method), and feature selection (joint selection of mutual information entropy and variance threshold) to obtain the adjusted standardized link feature dataset, ensuring that the data processing standards are consistent with the previous steps.

[0120] S42: Obtaining Actual and Predicted Values ​​of Link Quality After Adjustment: Extract the actual values ​​of link quality evaluation indicators from the adjusted and standardized link feature data after preprocessing in S41, including the adjusted packet loss rate ( L act ), adjusted transmission delay ( D act ), adjusted bandwidth utilization ( U act ) and adjusted signal-to-noise ratio ( S act (This serves as the baseline data for effect verification).

[0121] S43: Input the standardized link feature data into the optimal link quality correlation analysis rule base constructed in step S2, repeat the link quality analysis and prediction process of S2, and obtain the adjusted link quality prediction value. L act_pred , Dact_pred , U act_pred , S act_pred This is used to verify the prediction accuracy of the link quality correlation analysis rule base under the adjusted link status. See [link to relevant documentation]. Figure 2 As shown, the verification process is as follows:

[0122] S431: Calculation of accuracy evaluation indicators: The mean absolute error (MAE), root mean square error (RMSE), and prediction accuracy (Acc) are selected as the core accuracy evaluation indicators to quantify the prediction deviation of individual link quality indicators and the overall prediction reliability. The calculation formulas are as follows:

[0123] ;

[0124] ;

[0125] Where, P1= L act_pred P2= D act_pred P3= U act_pred P4= S act_pred This corresponds to the predicted values ​​of four types of indicators, A1= L act A2= D act A3= U act A4= S act These correspond to the actual values ​​of the four categories of indicators; the smaller the MAE and RMSE values, the smaller the prediction deviation and the higher the accuracy.

[0126] ;

[0127] Where Acc represents the overall prediction accuracy, with a value ranging from [0% to 100%]. A higher accuracy indicates a stronger adaptability of the link quality correlation analysis rule base to the adjusted link status; regarding the signal-to-noise ratio (SNR) S act In extreme scenarios where the denominator is 0, the absolute deviation threshold is used instead of the relative deviation calculation to avoid the anomaly of the denominator being 0.

[0128] S432: Accuracy threshold determination and link quality correlation analysis rule base adaptability assessment: preset accuracy evaluation threshold standards, where MAE threshold ≤ 5% of preset index benchmark value, RMSE threshold ≤ 8% of preset index benchmark value, and prediction accuracy Acc ≥ 85%. The preset index benchmark value is determined based on the historical average value when the link is running stably, which is consistent with the quality level mapping logic constructed in step S2 link quality correlation analysis rule base.

[0129] S433: When all three accuracy indicators meet the corresponding threshold requirements, it is determined that the prediction accuracy of the link quality correlation analysis rule base under the adjusted link state is up to standard, and it will continue to be used for subsequent link quality prediction; if any accuracy indicator does not meet the threshold requirements, it is marked as insufficient adaptability of the link quality correlation analysis rule base. In the iterative optimization process, the data after this adjustment is given priority to be used as the core sample to supplement the dataset, and the feature association logic and interval division standard of the link quality correlation analysis rule base are optimized in a targeted manner to improve the prediction ability of the link quality correlation analysis rule base for the adjusted link state.

[0130] The specific process of step S5 is as follows:

[0131] S51: Calculation of changes in indicators before and after adjustment and determination of optimization effect: Extract the predicted link quality value before adjustment ( L pred , D pred , U pred , S pred Using this as a benchmark, the changes in each core link quality evaluation index are calculated. The formula for calculating the changes is as follows:

[0132] Δ L = L pred -L act ;

[0133] Δ D = D pred - D act ;

[0134] Δ U = U act - U pred ;

[0135] ΔS= S act - S pred ;

[0136] Wherein, ΔL and ΔD are positive optimization change quantities, with values ​​≥0 indicating effective optimization, i.e., reduced packet loss rate and latency; ΔU and ΔS are positive optimization change quantities, with values ​​≥0 indicating effective optimization, i.e., improved bandwidth utilization and signal-to-noise ratio; a preset set of optimization thresholds. T ={ T L , T D , T U , T S} are the minimum optimization thresholds for packet loss rate, transmission delay, bandwidth utilization, and signal-to-noise ratio, respectively (set by link communication requirements and hardware performance).

[0137] S52: Construct a comprehensive judgment logic for optimization effect: If the positive change of all core indicators is ≥ the corresponding optimization threshold (ΔL≥ T L , ΔD≥ T D , ΔU≥ T U ΔS≥ T If the positive change of any core indicator is less than the corresponding optimization threshold, or if a negative change occurs (indicator deterioration), the adjustment effect is determined to be unsatisfactory. The process immediately returns to step S3. Based on the adjusted link status data obtained in S41, steps S31-S35 are re-executed to match a new optimal adaptive adjustment strategy and execute it until the adjustment effect is satisfactory.

[0138] S53: Dataset Supplementation and Iterative Optimization: The adjusted original dataset, along with the actual and predicted values ​​of the adjusted quality indicators, are added to the original dataset from step S1 to form an expanded original dataset. Simultaneously, the standardized link feature dataset is updated to ensure data coverage of the entire link state before and after the adjustment. The iterative optimization process is initiated periodically, with the iteration cycle dynamically adjusted based on the link data change rate. When the change rate is >10%, it is shortened to 24 hours / cycle; when the change rate is ≤10%, it is set to 72 hours / cycle. The specific iteration process is as follows:

[0139] Link quality correlation analysis rule base update: Based on the expanded dataset, repeat steps S22-S25, add correlation rules corresponding to the adjusted link feature combinations, and remove high-frequency failure rules (accuracy < A min This involves optimizing the confidence and coverage of existing rules, synchronously updating the feature interval division criteria, and the mapping relationship table between link quality evaluation indicators and quality levels.

[0140] Adjustment strategy library update: Statistically analyze the execution effect of each adjustment strategy, update the applicable scenario boundaries and parameter adjustment range of various strategies, supplement strategy combinations with high performance achievement rate, such as the joint strategy of modulation and demodulation method and frequency band switching, correct the impact coefficient of strategy on link quality indicators, and optimize strategy matching logic; at the same time, update the weight coefficient allocation standard of multi-objective optimization function to adapt to the dynamic changes in link communication requirements.

Claims

1. A machine learning-based method for predicting and adaptively adjusting the quality of IoT communication links, characterized in that, Includes the following steps: S1: Real-time acquisition of raw data during the link communication process, including physical layer parameters, link layer parameters and environmental interference parameters, and preprocessing of the acquired raw data to obtain a standardized link feature dataset; S2: Construct a link quality correlation analysis rule base, which is generated by mining the historical correlation between core features and link quality evaluation indicators; input the standardized link feature dataset into the link quality correlation analysis rule base, and output the link quality level within a future preset time window; S3: Construct multi-objective optimization decision logic with the goals of minimizing link packet loss rate, reducing transmission latency and maximizing bandwidth utilization. Match the preset adjustment strategy library, select the optimal adaptive adjustment strategy that is suitable for the current link state, and execute the link parameter adjustment. S4: Perform real-time data collection on the adjusted link, repeat steps S1 and S2, and obtain the actual and predicted values ​​of the adjusted link quality. S5: Calculate the change in link quality evaluation indicators before and after adjustment. If the change does not reach the preset optimization threshold, return to step S3 and rematch the adjustment strategy based on the new link status. At the same time, supplement the adjusted link data into the original dataset and regularly update the link quality correlation analysis rule base and adjustment strategy base.

2. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 1, characterized in that, The specific process of constructing the link quality correlation analysis rule base in step S2 is as follows: S21: Select a standardized link feature dataset and supplement it with historical operation data of IoT communication links within the corresponding time period. The historical operation data includes packet loss rate, transmission delay, bandwidth utilization, and actual signal-to-noise ratio values ​​that correspond one-to-one with the core features, and construct a complete historical sample set. The complete historical sample set is time-axis aligned to obtain the aligned historical sample set. S22: Based on the aligned historical sample set, an adaptive threshold partitioning algorithm is used to partition each core feature obtained by screening into intervals. Each core feature is divided into three feature intervals: high, medium and low according to the numerical distribution, forming a feature interval set. Based on the preset link quality level classification standard, the actual values ​​of packet loss rate, transmission delay, bandwidth utilization, and signal-to-noise ratio are mapped to four link quality levels: excellent, good, average, and poor, respectively, and a mapping relationship table between link quality evaluation indicators and quality levels is established. S23: Based on the mapping table between link quality evaluation indicators and quality levels, the correlation strength between each core feature interval and the corresponding quality level of each link quality evaluation indicator is calculated by mutual information entropy. Strongly correlated feature intervals with correlation strength greater than a preset threshold are retained to obtain a strongly correlated sample subset. An improved Apriori association rule algorithm is adopted, using the combination of multiple core feature intervals in the strongly correlated sample subset as the antecedent and the quality level and change trend of the corresponding link quality evaluation indicator as the consequent. Minimum support and minimum confidence thresholds are set to mine high-frequency and high-reliability association rules that meet the threshold requirements to form a candidate association rule set. S24: Perform rule conflict detection and deduplication on the candidate association rule set; for the rules to be optimized with confidence levels within the preset threshold range, supplement environmental interference parameter constraints based on the actual operating scenarios of IoT communication links to optimize the accuracy of rule description; classify and archive the optimized rules according to the link quality evaluation index type to form an initial link quality association analysis rule library. S25: Select a subset of validation samples from the aligned historical sample set, input the core feature data of the validation samples into the initial link quality association analysis rule base, obtain the link quality prediction level through rule matching, compare it with the actual quality level corresponding to the validation sample, and calculate the matching accuracy of each rule; remove invalid rules with accuracy lower than the preset qualified threshold, and for rules with qualified accuracy but insufficient coverage, supplement the corresponding feature combination samples and re-mine and optimize them, finally obtaining the optimal link quality association analysis rule base that meets the link quality prediction requirements.

3. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 2, characterized in that, The specific process of obtaining the strongly correlated sample subset in step S23 is as follows: S231: Extract different intervals of a single core feature from the feature interval set, denoted as the feature interval variable X, where X = {x1, x2, ..., x...} n }, where n is the number of intervals for a single core feature, here n=3, corresponding to high, medium, and low intervals; extract the quality level corresponding to a certain type of link quality evaluation index from the mapping relationship table between link quality evaluation index and quality level, denoted as quality level variable Y, Y={y1,y2,y3,y4}, corresponding to excellent, good, average, and poor levels respectively; using the aligned historical sample set as the data source, statistically obtain the joint probability distribution P(X,Y) and marginal probability distributions P(X) and P(Y) of feature interval variable X and quality level variable Y; S232: Using mutual information entropy to quantify the correlation strength between the feature interval variable X and the quality grade variable Y: ； in, x For characteristic interval variables X Any value of , that is, a certain interval of a single core feature. y For quality grade variables Y Any value of P(i, j) represents a certain link quality level; x,y P(x) represents the joint probability that the feature interval x and the quality level y occur simultaneously. x P(x) represents the marginal probability of the feature interval x occurring. y () represents the quality level y The marginal probability of occurrence; S233: Strong correlation feature interval screening and sample subset construction: A preset correlation strength threshold T is set, and the calculated correlation strength value is compared with the threshold T; core feature intervals with correlation strength values ​​> T are retained, and weak correlation feature intervals with correlation strength values ​​≤ T are removed; at the same time, samples containing weak correlation feature intervals are removed from the aligned historical sample set, and the remaining samples are all composed of combinations of strong correlation feature intervals and corresponding quality levels to form a strong correlation sample subset.

4. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 2, characterized in that, The specific process of obtaining the candidate association rule set in step S23 is as follows: S234: Using strongly correlated sample subsets as the data source, define itemsets as single or combined values ​​of core feature intervals, where itemsets containing only a single core feature interval are 1-itemsets; set a minimum support threshold. S min and minimum confidence threshold C min ; S235: Traverse the strongly associated sample subsets, count the frequency of all 1-itemsets, calculate the support of each 1-itemet, and retain items with support ≥ S min 1-itemsets that form frequent 1-itemsets are removed, and those with support < S min Infrequent 1-itemsets; Self-joining and pruning operations are performed on frequent 1-itemsets to generate candidate k-itemsets; during the pruning stage, subsets of candidate k-itemsets containing non-frequent (k-1)-itemsets are removed, and step S235 is repeated until no new frequent k-itemsets can be generated. All frequent itemsets are then aggregated to form a high-frequency itemset set. S236: Association Rule Generation and Filtering: For each frequent itemset in the high-frequency itemset set, split it to generate all possible rule forms, calculate the confidence score of each rule, and retain rules with a confidence score ≥ C min Association rules, removing rules with confidence levels < C min Low reliability rules are identified; redundancy detection is performed on the generated rules, redundant rules are removed, and the remaining rules are classified and organized according to the link quality evaluation index type to form a candidate association rule set.

5. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 2, characterized in that, The specific process of step S25 is as follows: S251: Randomly select samples from the aligned historical sample set according to a preset ratio to form a validation sample subset, and retain the remaining samples as a supplementary data source for subsequent rule optimization. S252: For each validation sample in the validation sample subset, extract core feature data, repeat the standardization process to obtain standardized validation feature data; input the standardized validation feature data into the generated initial link quality correlation analysis rule base, use the feature interval precise matching algorithm to traverse the correlation rules of the corresponding link quality evaluation index in the link quality correlation analysis rule base, determine the target rule that completely matches the core feature interval combination of the validation sample, and output the link quality prediction level corresponding to the target rule; S253: Extract the actual link quality level corresponding to each verification sample from the verification sample subset, compare the predicted link quality level with the actual quality level one by one, and record the matching result of each target rule, which is divided into two categories: successful matching and failed matching. S254: Based on the comparison results, the number of successful matches for each association rule and the total number of matches are counted separately, and the accuracy index is used to quantify the effectiveness of a single rule; S255: Preset accuracy pass threshold A min The accuracy of each rule is compared with A min Comparison; Elimination accuracy < A min Invalid rules, retain those with an accuracy ≥ A min Effective rules; for accuracy ≥ A min However, rules with a total number of matches less than the preset threshold are marked as rules to be optimized, providing a basis for subsequent supplementary sample mining, and finally forming a set of rules that have been validated for effectiveness.

6. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 1, characterized in that, The specific process of step S3 is as follows: S31: Extract the link quality prediction results, including the predicted packet loss rate. L pred Transmission delay prediction value D pred Bandwidth utilization forecast U pred and signal-to-noise ratio prediction value S pred Based on the established link quality level classification standards, the current predicted link quality level and core bottleneck indicators are clearly defined; current link communication requirements are collected, requirement parameters are quantified according to service type, and communication requirement constraint thresholds are set, including transmission delay constraint thresholds for real-time control services. D max Packet loss rate constraint threshold L max Bandwidth utilization constraint threshold for data acquisition services U min This forms a quantified set of communication requirement constraints. S32: Using the analytical prediction results and quantified communication requirements as input, a multi-objective optimization decision logic is constructed. The core optimization objectives are to minimize the link packet loss rate, reduce transmission latency, and maximize bandwidth utilization. The predicted signal-to-noise ratio is then used as the input. S pred As a constraint. S pred ≥ S 1 min , S 1 min To determine the minimum signal-to-noise ratio threshold for stable link communication, a weighted approach is used to transform the multi-objective optimization function into a single-objective optimization function. S33: Call the preset adjustment strategy library, which includes strategies such as modulation and demodulation mode switching, transmission power adjustment, communication frequency band switching, and data packet strategy adjustment; S34: Substitute each strategy in the candidate adjustment strategy set into the optimization objective function F, and calculate the F value corresponding to each strategy; introduce the strategy execution cost coefficient C, including energy consumption cost and handover delay cost, and modify the optimization objective function to F'=F×(1+C), where the energy consumption cost is quantified according to the transmission power adjustment amplitude, and the handover delay cost is quantified according to the time consumption of modulation mode / frequency band switching. Select the strategy with the smallest modified F' value as the optimal adaptive adjustment strategy. If there are multiple strategies with the same F' value, the strategy with the smallest execution cost coefficient C is selected first. S35: The IoT gateway parses the optimal adaptive adjustment strategy into standardized control commands and sends the control commands to the corresponding IoT terminal through a secure communication channel. After receiving the commands, the terminal starts the parameter configuration module and completes the link parameter adjustment according to the preset timing.

7. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 6, characterized in that, The specific process of step S4 is as follows: S41: Real-time data acquisition is performed on the adjusted IoT communication link to form the adjusted raw dataset, and after preprocessing, the adjusted standardized link feature dataset is obtained. S42: Extract the actual values ​​of link quality evaluation indicators from the adjusted standardized link feature data, including adjusted packet loss rate, adjusted transmission delay, adjusted bandwidth utilization and adjusted signal-to-noise ratio; S43: Input the standardized link feature data into the optimal link quality correlation analysis rule base, repeat the link quality analysis and prediction process, and obtain the adjusted link quality prediction value, which is used to verify the prediction accuracy of the link quality correlation analysis rule base under the adjusted link state.

8. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 7, characterized in that, The process of verifying the prediction accuracy of the link quality correlation analysis rule base under the adjusted link status in step S43 is as follows: S431: Calculation accuracy evaluation index: The mean absolute error, root mean square error and prediction accuracy are selected as the core accuracy evaluation indexes to quantify the prediction deviation of single link quality indexes and the overall prediction reliability, respectively. S432: Preset accuracy evaluation threshold standards, including MAE threshold, RMSE threshold, and prediction accuracy threshold; S433: When all three accuracy indicators meet the corresponding threshold requirements, it is determined that the prediction accuracy of the link quality correlation analysis rule base under the adjusted link state is up to standard, and it will continue to be used for subsequent link quality prediction. If any accuracy metric fails to meet the threshold requirement, it is marked as insufficient adaptability of the link quality correlation analysis rule base. In the iterative optimization process, the data adjusted in this instance is prioritized as the core sample to supplement the dataset, and the feature association logic and interval division criteria of the link quality correlation analysis rule base are optimized in a targeted manner.

9. The method for predicting and adaptively adjusting IoT communication link quality based on machine learning according to claim 7, characterized in that, The specific process of step S5 is as follows: S51: Extract the predicted link quality value before adjustment as the comparison benchmark, and calculate the change in each core link quality evaluation indicator, including the change in packet loss rate Δ. L Transmission delay change Δ D Change in bandwidth utilization Δ U、 Signal-to-noise ratio change Δ S ;Δ L Δ D To optimize the change, Δ U ΔS represents the positive optimization change. Preset Optimize threshold set T ={ T L , T D , T U , T S } are the minimum optimization thresholds for packet loss rate, transmission delay, bandwidth utilization, and signal-to-noise ratio, respectively; S52: If the positive change of all core indicators is greater than or equal to the corresponding optimization threshold, the adjustment effect is deemed to be satisfactory; if the positive change of any core indicator is less than the corresponding optimization threshold, or if a negative change occurs, the adjustment effect is deemed to be unsatisfactory, and the process immediately returns to step S3. Based on the adjusted link status data, steps S31-S35 are re-executed to match and execute a new optimal adaptive adjustment strategy until the adjustment effect is satisfactory. S53: Add the adjusted original dataset and the actual and predicted values ​​of the adjusted quality indicators to the original dataset of step S1 to form the expanded original dataset, and update the standardized link feature dataset at the same time; start the iterative optimization process regularly.

Images