Intelligent concentrator communication path selection method based on mode recognition

Abstract

The invention provides a concentrator communication path intelligent selection method based on mode recognition, and the key point of the technical scheme is that the method comprises the steps: configuring a data collection function module, and obtaining the multi-source index data of each communication path; performing time window alignment and sliding statistical preprocessing on the collected original index data, and normalizing the index data respectively; obtaining a real-time link quality score and a load attenuation mapping function value based on the normalized index data, and calculating a first optimization feature of each communication path; obtaining historical stability of link quality and path switching smoothness, and calculating a second optimization feature of each communication path; calculating a final score of each communication path based on the first optimization feature and the second optimization feature; a communication path is selected to issue a routing strategy, and model parameters are fed back and updated online; according to the method, all-directional balance of path performance, node load and historical stability can be realized, and the timeliness and stability of concentrator communication path selection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of concentrator communication, and in particular to a method for intelligently selecting a concentrator communication path based on pattern recognition. Background Art

[0002] With the continuous advancement of digitalization and intelligentization, the Internet of Things (IoT) has been widely applied in fields such as energy, power generation, urban management, and industrial automation. Numerous sensors, smart meters, environmental monitoring nodes, and video surveillance equipment are connected to the network, forming a massive distributed data acquisition system. As a key node for edge computing and data aggregation, concentrators must possess high-reliability, low-latency, and energy-efficient communication capabilities. However, in actual deployments, concentrators face the burden of wireless channel interference, frequent network topology changes, and high concurrent data traffic. Consequently, the timeliness and stability of communication path selection become bottlenecks in system performance.

[0003] However, traditional path selection methods mostly rely on static routing protocols such as RIP and OSPF, or manually configured routing tables, which cannot reflect link quality and node status in real time. Some adaptive routing protocols, such as RPL, support decision-making based on objective functions, but their objective functions are usually fixed and lack online learning capabilities. Routing strategies based on a single indicator, such as packet loss rate, RSSI, or latency, cannot balance multi-dimensional performance requirements and are prone to problems such as packet loss surge, latency jitter, and uneven energy consumption. Summary of the Invention

[0004] To address the shortcomings of traditional concentrator path selection methods and achieve a comprehensive balance between path performance, node load, and historical stability, the present invention provides a concentrator communication path intelligent selection method based on pattern recognition. The technical solution is as follows:

[0005] A method for intelligently selecting communication paths for a concentrator based on pattern recognition comprises the following steps: configuring a data acquisition function module to acquire multi-source indicator data for each communication path; performing time window alignment and sliding statistical preprocessing on the acquired raw indicator data, and normalizing the indicator data respectively; acquiring a real-time link quality score and a load attenuation mapping function value based on the normalized indicator data, and calculating a first optimization feature for each communication path; acquiring historical stability of link quality and path switching smoothness, and calculating a second optimization feature for each communication path; calculating a final score for each communication path based on the first optimization feature and the second optimization feature; selecting a communication path and issuing a routing strategy based on the final score, and updating model parameters based on online feedback of the transmission result.

[0006] Preferably, a link detection module, a topology discovery module, a load monitoring module, and a transmission log module are configured in the concentrator; the link detection module sends a detection packet to the neighbor node every 5 seconds, collects the indicator data of the packet loss rate PLR, the round-trip delay RTT, the signal-to-noise ratio SNR, and the received signal strength RSSI, and records the timestamp mark and stores it in the local cache queue; the topology discovery module broadcasts the neighbor discovery request every 10 minutes, collects the neighbor list, the link quality indication LQI and the hop count information, and constructs a timing topology map; the load monitoring module collects the indicator data of the load node every minute, and the indicator data is the load value Load; the transmission log module records the sending or receiving timestamp, the number of retransmissions, and the status information of each uplink packet, and pushes it to the log platform.

[0007] Preferably, all data are unified into UTC timestamps and aggregated according to a fixed time window, with the default time window being 5 seconds; outliers in the original indicator data PLR, RTT, and load values ​​are processed, data points exceeding 3 times the standard deviation are removed, and missing values ​​are filled in by linear interpolation; the maximum and minimum values ​​corresponding to each type of indicator data are obtained, a sliding window is configured, and the mean and standard deviation of each type of indicator data are statistically calculated within the window.

[0008] Preferably, the difference between the value of the packet loss rate PLR ​​at the current moment and the minimum value of the PLR, as well as the total difference between the maximum value of the PLR ​​and the minimum value of the PLR, are calculated, and the ratio of the difference to the total difference is used as the normalized value of the packet loss rate PLR ​​to obtain the normalized index PLR_norm of the packet loss rate; the difference between the reciprocal value of the round-trip delay RTT at the current moment and the minimum value of the reciprocal of the RTT, as well as the total difference between the maximum value of the reciprocal of the RTT and the minimum value of the reciprocal of the RTT, are calculated, and the ratio of the difference to the total difference is used as the normalized value of the round-trip delay RTT to obtain the normalized index RTT_norm of the round-trip delay; the difference between the load value Load of the load node and the minimum load value, as well as the total difference between the maximum load value and the minimum load value, are calculated, and the ratio of the difference to the total difference is used as the normalized value of the corresponding node load to obtain the normalized index Load_norm of the load value.

[0009] Preferably, obtain the real-time link quality score and the load attenuation mapping function value, and calculate the first optimization feature of each communication path, including: calculating the link quality score SLQ(t) at the t-th moment, and the calculation formula is: SLQ(t) = α(t)×(1 - PLR_norm) + β(t)×RTT_norm, where α(t) and β(t) represent the dynamic weights at the t-th moment, and the constraint conditions for α(t) and β(t) are α(t) + β(t) = 1; define the load attenuation mapping function as g(L), when 0 ≤ Load_norm < 0.3, it is the low load area, g(L) = 1 - k1×Load_norm; when 0.3 ≤ Load-norm < 0.7, it is the medium load area, g(L) = 1 - (0.3×k1 + (Load-norm - 0.3)×k2); when 0.7 ≤ Load-norm ≤ 1, it is the high load area, g(L) = 1 - (0.3×k1 + 0.4×k2 + (Load_norm - 0.7)×k3); k1, k2, k3 represent slope parameters, and k1 < k2 < k3; among them, the load attenuation mapping function value at the t-th moment is g(L(t)); calculate the first optimization feature F1(t) of each communication path at the t-th moment, and the calculation formula is: F1(t) = SLQ(t)×g(L(t)).

[0010] Preferably, obtain the historical stability of the link quality and the path switching smoothness, including: configuring a short window T s and a long window T l two kinds of sliding windows, and calculate the standard deviations σ s and σ l of the first optimization feature of each communication path within the short window T s and the long window T l respectively, and map them to the stability D s and D l respectively, where Adopt the method of time series clustering to map each time point t i of the first optimization feature sequence of each communication path to the clustering label c(t i ), and the same category represents a similar fluctuation pattern, so as to divide the continuous numerical sequence into pattern segments, and count the number of times of pattern switching of each communication path. The statistical calculation formula is:

[0011]

[0012] In the formula, N represents the number of sample points;

[0013] Calculate the path switching smoothness H of each communication path, and the calculation formula is:

[0014]

[0015] Preferably, calculating the second optimization feature of each communication path includes: defining the data aging time as t age , t age =tt last , where t represents the timestamp of the current sample, t last Represents the timestamp of the most recent sample and calculates the exponential decay weight ω s 、ω l , the calculation formula is:

[0016]

[0017] Where λ represents the attenuation rate;

[0018] Calculate the second optimization feature F2(t) of each communication path at the tth moment. The calculation formula is: F2(t) = (ω s ×D s +ω l ×D l )×H.

[0019] Preferably, calculating the final score of each communication path includes: taking the current time t as the end point and the length T n In the time window, the values ​​of the first optimization feature F1 and the second optimization feature F2 at the nearest N time points in each communication path are collected respectively, and the mean value μ of F1 and F2 in each communication path is calculated. i and standard deviation σ i , the calculation formula is:

[0020]

[0021] In the formula, the value of i is 1 or 2. The values ​​of F1 and F2 in each communication path at time t are normalized by Sigmoid function to obtain the normalized eigenvalues ​​HF of F1 and F2 in each communication path at time t. i (t), the calculation formula is:

[0022]

[0023] In the formula, the value of i is 1 or 2.

[0024] Preferably, calculating the final score of each communication path further includes: calculating the length T n The variance of the normalized eigenvalue of each communication path in the time window Var i , the calculation formula is:

[0025]

[0026] In the formula, the value of i is 1 or 2. Represents the average value of F1 or F2 normalized feature value;

[0027] Calculate the fusion factor ω of the first optimization feature and the second optimization feature in each communication path. The calculation formula is:

[0028]

[0029] Where, the fusion factor ω represents the automatic weighting of the relative fluctuations of the two characteristics, and 1-ω represents the weighting of the complementary historical stability.

[0030] The two normalized features are linearly combined using the fusion factor ω to obtain the final score of each communication path, where the final score of the p-th communication path at the t-th time is F opt (p, t), the calculation formula is:

[0031] F opt (p,t)=ω×HF1(p,t)+(1-ω)×HF2(p,t),

[0032] Wherein, HF1(p,t) represents the normalized eigenvalue of the first optimized feature of the p-th communication path at the t-th time, and HF2(p,t) represents the normalized eigenvalue of the second optimized feature of the p-th communication path at the t-th time.

[0033] Preferably, the final scores of all communication paths are compared, and the communication path with the highest final score is pushed to the concentrator to update its data forwarding configuration; the actual PLR, RTT and number of retransmissions after each selection of a communication path are recorded and written into the log platform, and the actual transmission results are matched with the first optimization feature and the second optimization feature at the time of decision-making. According to the actual transmission results of the current communication path, the statistics and parameters of each feature are dynamically adjusted and corrected. In the next decision cycle, the new first optimization feature and the second optimization feature are recalculated, so that the model can continuously learn and evolve in actual operation.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention discloses a method for intelligently selecting a concentrator communication path based on pattern recognition. Through high-frequency multi-source data acquisition, two rounds of progressive feature optimization, and adaptive fusion decision-making, the method achieves an all-round balance among path performance, node load, and historical stability based on pattern recognition and online learning, thereby overcoming the defects of traditional concentrator path selection methods and effectively improving the timeliness and stability of concentrator communication path selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an illustrative and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0037] Figure 1 It is a flow chart of the implementation steps of the concentrator communication path intelligent selection method; DETAILED DESCRIPTION

[0038] The technical features of the present invention are further described in detail below with reference to the accompanying drawings so that those skilled in the art can understand them.

[0039] A concentrator communication path intelligent selection method based on pattern recognition involves adaptive path selection for edge computing nodes, which is particularly suitable for the communication path optimization decision of the concentrator in a highly dynamic network environment. The implementation process is as follows Figure 1 As shown in the figure, the complete process of a decision cycle includes: detection and collection of indicator data, indicator normalization, F1 calculation, F1 time series sliding window, F2 calculation, normalization mapping, fusion factor, final score calculation, path selection, route distribution, transmission result feedback, statistics update, and the next decision cycle. The specific implementation steps are as follows:

[0040] Step S1: Configure the data acquisition function module to obtain multi-source indicator data of each communication path;

[0041] Specifically, a link detection module, a topology discovery module, a load monitoring module, and a transmission log module are configured in the concentrator; the link detection module sends a detection packet to the neighbor node every 5 seconds, collects the indicator data of packet loss rate PLR, round-trip delay RTT, signal-to-noise ratio SNR, and received signal strength RSSI, and records the timestamp mark and stores it in the local cache queue; the topology discovery module broadcasts the neighbor discovery request every 10 minutes, collects the neighbor list, link quality indication LQI and hop count information, and constructs a timing topology map; the load monitoring module collects the indicator data of the load node every minute, and the load indicators include CPU utilization, queue length, and bandwidth occupancy. The corresponding load value Load can be calculated by combining the corresponding indicators; the transmission log module records the sending or receiving timestamp, number of retransmissions, and status information of each uplink packet, and pushes it to the log platform.

[0042] In addition, the operating system of the concentrator communication path intelligent selection method is divided into five layers, including: data acquisition layer, preprocessing layer, feature optimization layer, fusion decision layer, and feedback self-correction layer; the data acquisition function module configured in the concentrator needs to collect four types of data in real time: link detection, network topology, node load, and transmission log, and synchronize them accurately.

[0043] Step S2: perform time window alignment and sliding statistical preprocessing on the collected original indicator data, and normalize the indicator data respectively;

[0044] Specifically, all data are unified into UTC timestamps and aggregated according to a fixed time window, with the default time window being 5 seconds. Outliers in the original indicator data, such as PLR, RTT, and load values, are processed, data points exceeding three times the standard deviation are removed, and missing values ​​are filled in through linear interpolation. The maximum and minimum values ​​corresponding to each type of indicator data are obtained, a sliding window is configured, and the mean and standard deviation of each type of indicator data are statistically calculated within the window to generate real-time and historical views.

[0045] In addition, the difference between the current packet loss rate PLR ​​and the minimum PLR value, as well as the total difference between the maximum PLR and the minimum PLR value, are calculated. The ratio of the difference to the total difference is used as the normalized value of the packet loss rate PLR ​​to obtain the normalized indicator PLR_norm of the packet loss rate. The value range is [0,1]. The normalization calculation formula is:

[0046]

[0047] Calculate the difference between the reciprocal value of the current round-trip time delay (RTT) and the minimum reciprocal value of the RTT, as well as the total difference between the maximum reciprocal value of the RTT and the minimum reciprocal value of the RTT. Use the ratio of the difference to the total difference as the normalized value of the round-trip time delay (RTT) to obtain the normalized round-trip time delay indicator RTT_norm. The normalization calculation formula is:

[0048]

[0049] Calculate the difference between the load value Load and the minimum load value of the load node, as well as the total difference between the maximum load value and the minimum load value. Take the ratio of the difference to the total difference as the normalized value of the corresponding node load to obtain the normalized index Load_norm of the load value. The normalization calculation formula is:

[0050]

[0051] The value range is [0,1].

[0052] Step S3: Based on the normalized indicator data, obtain the real-time link quality score and the load attenuation mapping function value, and calculate the first optimization feature of each communication path;

[0053] Specifically, when evaluating communication paths solely based on link quality, such as low packet loss rate and high throughput, it is easy to overlook the processing pressure of the concentrator or relay node itself. When the node's CPU utilization, queue length, or bandwidth occupancy is close to saturation, even if the link conditions appear to be excellent, data will still experience queuing delays or be dropped at that node, resulting in overall transmission delay jitter and increased packet loss rate. Node load and link performance are not independent of each other. High load is often accompanied by buffer backlogs and reduced processing capacity, which directly affects forwarding delay. Therefore, path scores based solely on link quality cannot reflect node load conditions. High-load nodes can easily become transmission bottlenecks, leading to increased delay jitter and packet loss. It is necessary to incorporate node load into the link score attenuation mapping and introduce the load factor into the link score model. This will enable high-load nodes to automatically reduce their priority in path selection, thereby dynamically avoiding possible bottlenecks.

[0054] Obtaining the real-time link quality score and load attenuation mapping function value includes: calculating the link quality score SLQ(t) at time t, using the following formula: SLQ(t) = α(t) × (1-PLR_norm) + β(t) × RTT_norm, where α(t) and β(t) represent the dynamic weights at time t. When the packet loss rate (PLR) fluctuates greatly, increasing α(t) indicates increasing the weight of PLR_norm. When the round-trip delay (RTT) fluctuates greatly, increasing β(t) indicates increasing the weight of PLR_norm. Indicates increasing the weight of RTT_norm. The constraint condition of α(t) and β(t) is α(t) + β(t) = 1. α(t) and β(t) can be adaptively adjusted based on historical fluctuations to balance the importance of PLR_norm and RTT_norm while ensuring adaptability to different network environments. SLQ(t) can reflect the reliability of the link at the current moment. The value range of SLQ(t) is (0, 1). The closer the value is to 1, the more reliable the link is at the current moment.

[0055] The load attenuation mapping function is defined as g(L). When 0≤Load_norm<0.3, it is the low load area, g(L)=1-k1×Load_norm;

[0056] When 0.3≤Load-norm<0.7, it is the medium load region, g(L)=1-(0.3×k1+(Load_norm-0.3)×k2);

[0057] When 0.7≤Load_norm≤1, it is the high load area, g(L)=1-(0.3×k1+0.4×k2+(Load_norm-0.7)×k3);

[0058] In the formula, k1, k2, and k3 represent slope parameters, where k1 < k2 < k3; among them, the value of the load attenuation mapping function at the t-th moment is g(L(t)); the slope parameters k1, k2, and k3 can be adaptively adjusted and artificial parameter tuning is suppressed. The formula for their adaptive adjustment is:

[0059]

[0060] In the formula, the value of i is 1, 2, 3, and k 0,i represents the set reference slope. The reference slope k 0,i is generally set to an empirical value, such as 0.5, 1, 1.5, and σ L represents the load standard deviation, and μ L represents the average load. γ represents the set adaptive factor used to control the sensitivity to load fluctuations. The default value of γ is 0.5. ε represents a small constant used to prevent the denominator from being zero, and its value is a constant approaching 0, such as 0.01; when the load fluctuates greatly, the slope is automatically amplified to further suppress possible bottlenecks. When the load is stable, the attenuation mapping is milder to avoid over-punishing occasional high-load situations.

[0061] In addition, calculate the first optimization feature of each communication path, including: calculate the first optimization feature F1(t) of each communication path at the t-th moment. The calculation formula is: F1(t) = SLQ(t) × g(L(t)); the first optimization feature F1 focuses on the real-time quality of the current link and the load situation of the node, and reflects the instant performance and resource utilization rate of the link.

[0062] Step S4: Obtain the historical stability of the link quality and the path switching smoothness, and calculate the second optimization feature of each communication path;

[0063] Since the first optimization feature F1 only reflects the link quality and load attenuation at the current moment, if the link jitters violently within a very short time, such as jittering within a few seconds, it may only show a short-term fluctuation in F1, and it is difficult to distinguish between occasional noise and trend deterioration; in a real network, jitters often show clusters or periodicity, and there is an essential difference between short-term spike perturbations and continuous degradation fault mechanisms. Only analyzing the single first optimization feature F1 cannot determine whether this fluctuation is short-term noise or the link enters a high-risk state; therefore, it is necessary to introduce two perspectives of historical stability and switching smoothness. On the one hand, quantify the fluctuation strength of the first optimization feature F1 on different time scales, and on the other hand, quantify the switching frequency of the path on the time series clustering labels to make up for the deficiencies of the first optimization feature F1 in jitter recognition and trend perception.

[0064] Specifically, obtaining the historical stability of the link quality and the path switching smoothness includes: configuring a short window T s and a long window Tl Two sliding windows, the short window is used for real-time fluctuation analysis and is set to 1 minute, and the long window is used for trend analysis and is set to 10 minutes; s With long window T l The standard deviation σ of the first optimization feature of each communication path is calculated s , σ l , and are mapped to stability D s 、D l ,in, σ s The bigger, D s The smaller the value, the more severe the link jitter in the last minute. l The bigger, D l The smaller the value, the greater the overall link fluctuation in the last 10 minutes.

[0065] The time series clustering method is used to cluster each time point t of the first optimized feature sequence of each communication path. i Mapped to cluster label c(t i ), the same category represents similar fluctuation patterns. Preferably, the present invention uses the DTW+k-Shape algorithm to divide the F1 sequence into K categories, with K=3 by default; thereby dividing the continuous numerical sequence into pattern segments, and counting the number of mode switches for each communication path. The statistical calculation formula is:

[0066]

[0067] Where N represents the number of sample points;

[0068] Calculate the path switching smoothness H of each communication path. The calculation formula is:

[0069]

[0070] In addition, the second optimization feature of each communication path is calculated, including: defining the data aging time as t age , t age =tt last , if the aging time t age A larger value indicates that the data is no longer the latest dynamics, where t represents the timestamp of the current sample, t last Represents the timestamp of the most recent sample and calculates the exponential decay weight ω s 、ω l , the calculation formula is:

[0071]

[0072] Where λ represents the decay rate, which can be calculated based on the standard deviation σ of the first optimized feature. s , σl Perform online dynamic update; short window attenuation weight ω s Focusing on sensitivity to the latest burst jitter, the long window attenuation weight ω l Focusing on the stable grasp of the overall trend, the two are automatically balanced without the need to manually set a fixed ratio;

[0073] Afterwards, the short-term and long-term stability evaluations are integrated, the latest or overall perspective is adaptively emphasized, and the path label switching smoothness H is combined to suppress the jitter path with frequent mode breaks, and the second optimization feature of each communication path is obtained; where the second optimization feature of each communication path at time t is F2(t), and the calculation formula is: F2(t) = (ω s ×D s +ω l ×D l )×H; When the link is stable in both short and long window periods and mode switching is rare, the value of F2 approaches 1. When the link experiences sudden jitter or frequent mode switching in both short and long window periods, the value of F2 approaches 0.

[0074] Step S5: Calculating a final score for each communication path based on the first optimization feature and the second optimization feature;

[0075] The purpose of calculating the final score is to adaptively compress the absolute value domain of the first and second optimized features into the interval (0, 1), while considering the mean and fluctuation range in the current environment, eliminating the dimension and scale differences between different paths or different features, mapping values ​​greater than the mean close to 1, and mapping values ​​less than the mean close to 0, while retaining the difference amplitude information.

[0076] Specifically, the final score of each communication path is calculated, including: n Within the time window, such as T n In a 5-minute time window, the values ​​of the first optimization feature F1 and the second optimization feature F2 at the nearest N time points in each communication path are collected, and the mean value μ of F1 and F2 in each communication path is calculated. i and standard deviation σ i , the calculation formula is:

[0077]

[0078] In the formula, the value of i is 1 or 2. The values ​​of F1 and F2 in each communication path at time t are normalized by Sigmoid function to obtain the normalized eigenvalues ​​HF of F1 and F2 in each communication path at time t. i (t), the calculation formula is:

[0079]

[0080] In the formula, the value of i is 1 or 2.

[0081] In addition, the final score of each communication path is calculated, including: calculating the length of T n The variance of the normalized eigenvalue of each communication path in the time window Var i , the calculation formula is:

[0082]

[0083] In the formula, the value of i is 1 or 2. Represents the average value of F1 or F2 normalized feature value;

[0084] Calculate the fusion factor ω of the first optimization feature and the second optimization feature in each communication path. The calculation formula is:

[0085]

[0086] Where, the fusion factor ω represents the automatic weighting of the relative fluctuations of the two characteristics, and 1-ω represents the weighting of the complementary historical stability.

[0087] The two normalized features are linearly combined using the fusion factor ω to obtain the final score of each communication path, where the final score of the p-th communication path at the t-th time is F opt (p, t), the calculation formula is:

[0088] F opt (p,t)=ω×HF1(p,t)+(1-ω)×HF2(p,t),

[0089] Wherein, HF1(p,t) represents the normalized eigenvalue of the first optimized feature of the p-th communication path at the t-th time, and HF2(p,t) represents the normalized eigenvalue of the second optimized feature of the p-th communication path at the t-th time.

[0090] Step S6: Select a communication path based on the final score and issue a routing strategy, and update the model parameters based on the online feedback of the transmission results;

[0091] Specifically, the final scores of all communication paths are compared, and the communication path with the highest final score is pushed to the concentrator to update its data forwarding configuration; the actual PLR, RTT and number of retransmissions after each communication path is selected are recorded and written to the log platform, and the actual transmission results are matched with the first optimization feature and the second optimization feature at the time of decision-making. The actual PLR and RTT can directly reflect the decision effect. If they deviate from expectations, the distribution of the first optimization feature F1 and the second optimization feature F2 will change; according to the actual transmission results of the current communication path, the statistics and parameters of each feature are dynamically adjusted and corrected, and the internal parameters of the mapping and fusion functions are automatically adjusted through feedback without manual intervention; in the next decision cycle, the new first optimization feature and the second optimization feature are recalculated, so that the model can continuously learn and evolve in actual operation to avoid model aging and inaccurate decision-making; as the number of feedback loops increases, the predicted expected value and fluctuation amplitude gradually converge, which can make the decision more accurate and more robust.

[0092] The embodiments described in the present invention are merely descriptions of preferred implementations of the present invention and are not limited to the precise structures described above and shown in the accompanying drawings. Various modifications and changes can be made without departing from the scope of protection thereof. Without departing from the design concept of the present invention, various variations and improvements made to the technical solutions of the present invention by engineers and technicians in this field should fall within the scope of protection of the present invention.

Claims

1. A method for intelligently selecting a concentrator communication path based on pattern recognition, characterized in that: Configure a data acquisition module to acquire multi-source indicator data for each communication path; perform time window alignment and sliding statistical preprocessing on the collected raw indicator data, and normalize the indicator data separately; based on the normalized indicator data, obtain the real-time link quality score and load attenuation mapping function value, and calculate the first optimization feature of each communication path; Obtaining historical stability of link quality and path switching smoothness, and calculating the second optimization feature of each communication path; Calculating a final score for each communication path based on the first optimization feature and the second optimization feature; Based on the final score, the communication path is selected and the routing strategy is issued, and the model parameters are updated based on the online feedback of the transmission results.

2. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 1, characterized in that: The configuration data acquisition function module obtains multi-source indicator data of each communication path, including: configuring a link detection module, a topology discovery module, a load monitoring module, and a transmission log module in the concentrator; sending a detection packet to the neighbor node every 5 seconds through the link detection module, collecting indicator data of packet loss rate PLR, round-trip delay RTT, signal-to-noise ratio SNR, and received signal strength RSSI, and recording the timestamp mark, and storing it in the local cache queue; broadcasting a neighbor discovery request every 10 minutes through the topology discovery module, collecting neighbor lists, link quality indication LQI and hop count information, and constructing a timing topology map; collecting indicator data of load nodes every minute through the load monitoring module, and the indicator data is the load value Load; recording the sending or receiving timestamp, number of retransmissions, and status information of each uplink packet through the transmission log module, and pushing it to the log platform.

3. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 2, characterized in that: The collected original indicator data is subjected to time window alignment and sliding statistical preprocessing, including: unifying all data into UTC timestamps and aggregating them according to a fixed time window, with the time window defaulting to 5 seconds; processing outliers in the original indicator data PLR, RTT, and load values, removing data points exceeding 3 times the standard deviation, and filling in missing values ​​through linear interpolation; obtaining the maximum and minimum values ​​corresponding to each type of indicator data, configuring a sliding window, and statistically calculating the mean and standard deviation of each type of indicator data within the window.

4. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 3, characterized in that: The normalization of the indicator data includes: calculating the difference between the current packet loss rate PLR ​​value and the minimum PLR value, as well as the total difference between the maximum PLR value and the minimum PLR value, using the ratio of the difference to the total difference as the normalized value of the packet loss rate PLR, and obtaining the normalized indicator PLR_norm of the packet loss rate; calculating the difference between the reciprocal value of the round-trip delay RTT at the current moment and the minimum RTT reciprocal value, as well as the total difference between the maximum RTT reciprocal value and the minimum RTT reciprocal value, using the ratio of the difference to the total difference as the normalized value of the round-trip delay RTT, and obtaining the normalized indicator RTT_norm of the round-trip delay; calculating the difference between the load value Load of the load node and the minimum load value, as well as the total difference between the maximum load value and the minimum load value, using the ratio of the difference to the total difference as the normalized value of the corresponding node load, and obtaining the normalized indicator Load_norm of the load value.

5. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 4, characterized in that: The obtaining of the real-time link quality score and the load attenuation mapping function value, and calculating the first optimization feature of each communication path includes: calculating the link quality score SLQ(t) of each communication path at the t-th moment, and the calculation formula is: SLQ(t) = α(t)×(1 - PLR_norm) + β(t)×RTT_norm, where α(t) and β(t) represent the dynamic weights at the t-th moment, and the constraint conditions of α(t) and β(t) are α(t) + β(t) = 1; define the load attenuation mapping function as g(L), when 0 ≤ Load_norm < 0.3, it is the low load area, g(L) = 1 - k1×Load_norm; when 0.3 ≤ Load_norm < 0.7, it is the medium load area, g(L) = 1 - (0.3×k1 + (Load_norm - 0.3)×k2); when 0.7 ≤ Load_norm ≤ 1, it is the high load area, g(L) = 1 - (0.3×k1 + 0.4×k2 + (Load_norm - 0.7)×k3); k1, k2, k3 represent slope parameters, and k1 < k2 < k3; where, the load attenuation mapping function value at the t-th moment is g(L(t)); calculate the first optimization feature F1(t) of each communication path at the t-th moment, and the calculation formula is: F1(t) = SLQ(t)×g(L(t)).

6. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 1 or 5, characterized in that: The acquisition of the historical stability of the link quality and the path switching smoothness, and the calculation of the second optimization feature of each communication path include: configuring a short window T s and long window T l Two sliding windows, respectively in the short window T s With long window T l The standard deviation σ of the first optimization feature of each communication path is calculated s , σ l , and are mapped to stability D s 、D l ,in, The time series clustering method is used to cluster each time point t of the first optimized feature sequence of each communication path. i Mapped to cluster label c(t i ), the same category represents similar fluctuation patterns, thus dividing the continuous numerical sequence into pattern segments, and counting the number of mode switching of each communication path. The statistical calculation formula is: Where, N represents the number of sample points; Calculate the path switching smoothness H of each communication path. The calculation formula is:

7. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 6, characterized in that: The calculation of the second optimization feature of each communication path further includes: defining the data aging time as t age , t age =tt last , where t represents the timestamp of the current sample, t last Represents the timestamp of the most recent sample and calculates the exponential decay weight ω s 、ω l , the calculation formula is: Where, λ represents the attenuation rate; Calculate the second optimization feature F2(t) of each communication path at the tth moment. The calculation formula is: F2(t) = (ω s ×D s +ω l ×D l )×H.

8. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 7, characterized in that: The calculation of the final score of each communication path includes: taking the current time t as the end point and the length T n In the time window, the values ​​of the first optimization feature F1 and the second optimization feature F2 at the nearest N time points in each communication path are collected respectively, and the mean value μ of F1 and F2 in each communication path is calculated. i and standard deviation σ i , the calculation formula is: In the formula, the value of i is 1 or 2. The values ​​of F1 and F2 in each communication path at time t are normalized by Sigmoid function to obtain the normalized eigenvalues ​​HF of F1 and F2 in each communication path at time t. i (t), the calculation formula is: Where, the value of i is 1, 2.

9. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 8, characterized in that: The calculation of the final score of each communication path also includes: calculating the length T n The variance of the normalized eigenvalue of each communication path in the time window Var i , the calculation formula is: In the formula, the value of i is 1 or 2. Represents the average value of F1 or F2 normalized feature value; Calculate the fusion factor ω of the first optimization feature and the second optimization feature in each communication path, and the calculation formula is: Where, the fusion factor ω represents the relative fluctuation of the two features to automatically allocate weights, and 1 - ω represents the complementary historical stability to allocate weights; The two normalized features are linearly combined using the fusion factor ω to obtain the final score of each communication path, where the final score of the p-th communication path at the t-th time is F opt (p, t), the calculation formula is: F opt (p,t)=ω×HF1(p,t)+(1-ω)×HF2(p,t), Where, HF1(p,t) represents the normalized eigenvalue of the first optimization feature of the p-th communication path at the t-th moment, and HF2(p,t) represents the normalized eigenvalue of the second optimization feature of the p-th communication path at the t-th moment.

10. The method for intelligently selecting a concentrator communication path based on pattern recognition according to claim 2, characterized in that: The selecting of the communication path based on the final score and issuing the routing policy, and online feedback and updating of the model parameters according to the transmission result includes: comparing the final scores of all communication paths, pushing the communication path with the highest final score to the concentrator, and updating its data forwarding configuration; recording the actual PLR, RTT and retransmission times results after each selection of the communication path, and writing them into the log platform, matching the actual transmission result with the first optimization feature and the second optimization feature at the time of decision-making, dynamically adjusting and correcting the statistics and parameters of each feature according to the actual transmission result of the current communication path, and recalculating the new first optimization feature and the second optimization feature in the next decision cycle, so that the model continuously learns and evolves during actual operation.

Images