Computer data storage method and system based on internet of things

Abstract

The present application relates to a kind of computer data storage method and system based on Internet of Things, comprising the following steps: step S1, receive the data package uploaded by Internet of Things terminal, data package includes perception data and the metadata of the description this perception data, metadata at least includes equipment identification, data generation timestamp and geographic position label;Step S2, the time label of extraction data package, content semantic feature, and the credit value of the Internet of Things terminal of generation this data package is combined, the storage priority index of this data package is calculated;Step S3, based on storage priority index, the average reliability score of the multiple heterogeneous edge storage nodes of combination available;The present application realizes the accurate balance of cloud storage cost and data reliability by multidimensional data value evaluation and dynamic redundancy adjustment.

Description

Technical Field

[0001] This invention relates to the field of data storage, and more specifically to a computer data storage method and system based on the Internet of Things. Background Technology

[0002] With the rapid development and widespread application of IoT technology, massive amounts of IoT terminal devices continuously generate sensing data, which is then aggregated through edge gateways and uploaded to the cloud for storage and processing. In typical cloud computing application scenarios such as smart cities, public utility management, and industrial operations and maintenance, cloud data centers need to receive, store, and manage heterogeneous data streams reported by tens of thousands of sensors, posing a severe challenge to the resource scheduling capabilities and data management efficiency of cloud storage systems.

[0003] Currently, the storage and management of IoT data in the cloud mainly faces the following problems: Sensing data from different sources and with different content varies significantly in business value. For example, the importance of pipeline pressure anomaly alarm data to the system is drastically different from that of regular water meter reading data. However, existing storage systems typically employ a uniform redundancy strategy, making it difficult to differentiate data based on its intrinsic value. This results in insufficient protection of critical data while low-value data consumes a large amount of expensive cloud storage resources. IoT data has a distinct time-sensitivity characteristic; the business value of the data often gradually diminishes over time. Traditional storage methods lack quantitative assessment and dynamic response mechanisms for time-sensitivity factors. The reliability parameters of IoT terminals... The quality varies greatly. Some low-quality devices may frequently report erroneous data or interrupt connections. If the data generated by these devices is treated the same, it will not only waste storage space but may also interfere with subsequent data analysis decisions. Fourth, there are often implicit relationships between IoT data items driven by business processes. For example, an abnormal alarm of a certain monitoring indicator will usually trigger the operation and maintenance personnel to query the relevant indicator data synchronously. However, most existing storage management methods only evaluate the storage value based on the independent access frequency of a single data item, ignoring the collaborative needs between data. This can easily lead to important data that is not frequently accessed but is highly correlated with hot data being misjudged as low-value data and prematurely cleaned up.

[0004] In summary, existing technologies for IoT cloud storage scenarios lack a comprehensive storage value assessment and dynamic redundancy adjustment mechanism that can simultaneously consider the importance of data semantic content, time-related degradation characteristics, device reliability, and business relationships between data. This makes it difficult to achieve efficient utilization of cloud storage resources while ensuring data availability and access performance. To address the problems of existing technologies, a computer data storage method based on IoT is proposed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a computer data storage method based on the Internet of Things, comprising the following steps: Step S1: Receive the data packet uploaded by the IoT terminal. The data packet contains sensing data and metadata describing the sensing data. The metadata includes at least the device identifier, data generation timestamp, and geographic location tag. Step S2: Extract the time tag and content semantic features of the data packet, and calculate the storage priority index of the data packet by combining it with the reputation value of the IoT terminal that generated the data packet. Step S3: Based on the storage priority index and combined with the average reliability score of multiple available heterogeneous edge storage nodes, dynamically determine the initial redundancy storage strategy for the data packet. The initial redundancy storage strategy includes the number of replicas R and erasure coding parameters (K,M). Step S4: The perceived data is fragmented and redundantly encoded according to the initial redundant storage strategy to generate multiple encoding blocks. Each encoding block is then distributed and stored on different edge storage nodes according to node load, geographical location and consistent hash mapping rules. Step S5: During the data storage cycle, continuously collect data access logs on each storage node, construct a data semantic association graph with data items as nodes and collaborative access relationships between data items as edges, and calculate the probability of each data item being requested in the active query set through a random walk algorithm. Step S6: Based on the probability of being requested, combined with the topological features of the data item in the data semantic association graph and the duration of data existence, calculate the long-term retention value of each data item, and dynamically adjust the number of replicas and erasure coding parameters of each data item according to the long-term retention value and the current storage resource occupancy status, and reclaim the storage space occupied by data items with a long-term retention value lower than the threshold.

[0006] Furthermore, the perceived data content in the data packet is segmented into words to extract keywords. A pre-trained IoT domain word embedding model is used to map each keyword to a corresponding word vector. This word embedding model uses historical operation and maintenance text corpora from the IoT application domain as training data and is trained using a continuous bag-of-words model or a skip-word model. It is used to map keywords within the domain to fixed-dimensional real-valued vectors. When multiple keywords are extracted, the arithmetic mean of the word vectors of multiple keywords is taken to obtain a comprehensive semantic vector representing the content of the data packet. The comprehensive semantic vector is then multiplied by a preset domain importance weight vector. The domain importance weight vector is obtained by normalizing the statistical frequency of each keyword in historical fault or anomaly records corresponding to the device type. The dot product result is normalized to the 0-1 range using the Sigmoid function to obtain a semantic importance score. The calculation process is expressed as follows: ; In the formula, n is the number of extracted keywords. Let p be the word vector of the p-th keyword. This is the domain importance weight vector. The Sigmoid function, operator Represents the dot product of vectors; Read data to generate a timestamp, and calculate the difference between the current time and the timestamp as the duration the data has existed. The maximum effective duration reference value preset for this type of IoT data. As a normalization benchmark, for existing durations After logarithmic decay processing, the aging sensitivity score is obtained. The specific calculation process is as follows: ; Obtain the reputation value of the IoT terminal that generated the data packet. The calculation process is as follows: The accuracy rate of the data reported by the terminal within a preset historical time window is calculated, which is the proportion of data entries that pass both data integrity verification and numerical range reasonableness verification out of the total number of reported data entries. This proportion is used as the accuracy rate component. ; The proportion of the terminal's online time within a historical time window to the total window duration is used as a stability component. The reputation score is obtained by weighted summation and normalization of the accuracy and stability components. The calculation process is expressed as follows: ; in The preset accuracy weighting coefficients are obtained from a preset level-weight mapping table based on the device importance level of the corresponding terminal. The higher the device importance level, the higher the accuracy weighting coefficients. The larger the value, the range of which is between 0 and 1; Monitor the average round-trip latency and packet loss rate of the current data upload link, and dynamically determine the semantic weight α, timeliness weight β, and reputation weight γ based on the remaining capacity ratio of the edge storage pool, while satisfying α+β+γ=1. semantic importance score Time sensitivity score and equipment reputation value The storage priority index is obtained by multiplying each index by its corresponding weight and then summing the results. The calculation process is expressed as follows: .

[0007] Furthermore, the rules for dynamically determining the semantic weight α, the timeliness weight β, and the reputation weight γ are as follows: When the average round-trip time exceeds the first preset threshold and the packet loss rate increases, the timeliness weight β is increased and the semantic weight α is decreased accordingly, while the reputation weight γ remains unchanged, so as to prioritize the redundant processing of time-sensitive data. When the remaining capacity of the edge storage pool is lower than the second preset threshold, the reputation weight γ is increased and the semantic weight α and the timeliness weight β are decreased accordingly to ensure the storage reliability of data generated by high-reputation devices.

[0008] Furthermore, in step S3, the number of copies R and the erasure coding parameters (K, M) are determined as follows: Get the system's preset minimum number of replicas. and upper limit Store priority index Mapping to the range of replica counts, calculate the base replica count. ; Obtain the average reliability score of all current candidate edge storage nodes. The reliability score is obtained by normalizing the historical online time and failure interval of a node; the higher the score, the more reliable the node. The final number of replicas R is obtained by dividing the basic number of replicas by the average reliability score and rounding up, and then taking the larger value between this and the lower limit of the number of replicas. The calculation process is as follows: ; The number of erasure coding data fragments K is determined based on the target read / write throughput and the size constraints of a single fragment. The number of parity fragments M is calculated by multiplying the number of data fragments K by a storage priority index. With the preset redundancy reliability margin factor The result is obtained by rounding up. The calculation process is as follows: ; in The default value and >1.

[0009] Furthermore, in step S4, the fragmentation and redundancy coding are performed according to the erasure coding parameters (K,M), specifically including: The raw sensing data is divided into K data segments, and M verification segments are generated using erasure coding parameters, for a total of N=K+M coding blocks; Obtain the current available storage capacity, CPU load rate, and geographic location coordinates of each edge storage node. Combined with the node reliability score, calculate and construct a weighted consistent hash ring for each node, and map N coded blocks sequentially and with minimal collisions to the node locations on the ring. During the mapping process, mandatory verification ensures that the physical nodes containing any K coded blocks are distributed in at least two different geographical regions.

[0010] Further, the process of calculating the demanded probability of each data item under the active query set through the random walk algorithm in step S5 is as follows: Each stored data item is taken as a node of the graph. If within the observation time window, the same query request or two consecutive adjacent query requests simultaneously involve data item i and data item j, a directed edge is established between node i and node j, and the weight of the edge is proportional to the co-access frequency; When a query set Q composed of a group of active query tasks is received, the data item nodes contained in Q are used as restart nodes to perform iterative calculation of random walk with restart. The initial value of the demanded probability of all nodes is set to 0, and only the initial value of each node in the query set Q is set to 1 / |Q|. In each iteration, the demanded probability of each node is composed of two parts by weighting: one part is obtained by random walk transfer according to the probability value of the in-edge neighbor nodes of the current node and the edge weight, and the other part is to return the probability mass to the restart node according to the preset restart probability. The iteration is repeated until the variation of the demanded probability value of all nodes between two consecutive iterations is less than the preset convergence threshold , or the number of iterations reaches the preset maximum number of iterations, the iteration is terminated, and the demanded probability of each data item i is obtained , and its iterative calculation process is expressed as: ; where d is a preset restart probability, the value range of which is 0<d<1, representing the probability of returning to the restart node in the query set Q at each step in the random walk process; is the component of the restart vector at the i-th node, and when data item i belongs to query set Q , otherwise , where |Q| represents the number of data items contained in the query set Q; is the weight of the directed edge from node j to node i, whose value is equal to the frequency of co-access of data item j and data item i within the observation time window; In(i) is the set of all in-edge neighbor nodes pointing to node i; Out(j) is the set of all out-edge neighbor nodes of node j; h is an index variable traversing all out-edge neighbor nodes of node j; is the weight of the directed edge from node j to its out-edge neighbor node h; is the demanded probability value of node j at the t-th iteration.

[0011] Further, the manner of calculating the long-term retention value of each data item in step S6 is as follows: For data item i, obtain its probability of being needed using the random walk algorithm. ; Get the duration of data item i from its generation time to the current time. With a preset time decay coefficient The exponential decay process is applied to obtain the aging decay factor. ; Extract the set N(i) of all neighbor nodes of data item i in the data semantic association graph, and calculate the demand probability of each neighbor node j in this set. The weighted average, where the weights are the edge weights between nodes i and j. ; Multiplying the weighted average by the preset correlation enhancement coefficient η and then adding one yields the correlation thermal enhancement factor. The calculation process is expressed as follows: ; The probability of data item i being demanded. The long-term retention value V(i) of this data item is obtained by multiplying the aging decay factor and the associated thermal enhancement factor together. The calculation process is expressed as follows: .

[0012] Furthermore, the specific rules for dynamically adjusting the number of copies and erasure coding parameters for each data item in step S6 are as follows: Set a high retention threshold and low retention threshold And monitor the current space occupancy rate U of the entire edge storage pool in real time; When the long-term retention value V(i) of data item i is lower than the low retention threshold When the space occupancy rate U exceeds the preset storage pressure threshold, a redundancy shrinkage operation is triggered: the number of replicas of data item i is reduced to the preset minimum number of replicas, and its erasure coding configuration is switched to a parameter that increases the proportion of parity shards, so as to reduce the total storage overhead while maintaining basic availability. When the long-term retention value V(i) of data item i is higher than the high retention threshold When the current number of replicas is lower than the preset maximum number of replicas, the number of replicas is increased and the proportion of erasure coding check shards is reduced to improve data reading performance. For long-term retention value between and The data items between them will maintain their current redundancy configuration.

[0013] An Internet of Things (IoT) based computer data storage system includes: The data receiving module is used to receive data packets uploaded by IoT terminals. The data packets contain sensing data and metadata describing the sensing data. The metadata includes at least the device identifier, data generation timestamp, and geographic location tag. The priority calculation module is used to extract the time tag and semantic features of the data packet, and combine them with the reputation value of the IoT terminal that generated the data packet to calculate the storage priority index of the data packet. The initial redundancy configuration module is used to dynamically determine the initial redundancy storage strategy for the data packet based on the storage priority index and the average reliability score of multiple available heterogeneous edge storage nodes. The initial redundancy storage strategy includes the number of replicas R and erasure coding parameters (K,M). The sharding and distribution module is used to shard and redundantly encode the sensing data according to the initial redundant storage strategy, generate multiple coding blocks, and distribute and store each coding block to different edge storage nodes according to node load, geographical location and consistent hash mapping rules. The association analysis module is used to continuously collect data access logs on each storage node during the data storage cycle. It constructs a data semantic association graph with data items as nodes and collaborative access relationships between data items as edges, and calculates the probability of each data item being requested in the active query set through a random walk algorithm. The dynamic adjustment module is used to calculate the long-term retention value of each data item based on the probability of being needed, combined with the topological features of the data item in the data semantic association graph and the duration of data existence. Based on the long-term retention value and the current storage resource occupancy status, it dynamically adjusts the number of replicas and erasure coding parameters of each data item and reclaims the storage space occupied by data items with a long-term retention value below the threshold.

[0014] The present invention has the following advantages over the prior art: This invention solves the problem of redundant resource misallocation caused by traditional IoT cloud storage relying solely on a single dimension for data grading by constructing a storage priority evaluation system that integrates semantic importance, time sensitivity, and device reputation. The semantic importance score is obtained by segmenting the perceived data into words, averaging the word vectors, and then normalizing the result with a dot product of the domain importance weight vectors using sigmoid. The time sensitivity score is obtained by logarithmic decay processing based on the data's existing duration. The device reputation score is obtained by weighted summation of the historical reporting accuracy and online stability of the terminal. The calculation process for each component is fully disclosed, ensuring the reproducibility of obtaining the storage priority index. Building upon this foundation, the storage priority index is linked to the average reliability score of edge nodes, ensuring that the number of replicas and erasure coding verification shards are positively correlated with data value. High-value data automatically receives multiple replicas and strong erasure coding fault tolerance, while low-value data is stored with streamlined redundancy, achieving a fine-grained dynamic balance between storage reliability and overhead. By introducing real-time monitoring of network link status and remaining storage pool capacity, the allocation ratio of the three-dimensional weights is dynamically adjusted. When link congestion occurs, the timeliness weight is increased to prioritize the redundant processing of latency-sensitive data; when storage resources are scarce, the reputation weight is increased to ensure that data on high-reliability devices receives priority protection, giving the storage strategy adaptive adjustment capabilities. The system effectively captures implicit relationships between data items formed through business collaboration by collecting data access logs and constructing a data semantic relationship graph. It employs a random walk algorithm with restart capability to calculate the demand probability of data items within the active query set, ensuring that the demand probability balances direct access frequency and indirect relationship demand. In long-term retention value assessment, the system multiplies the demand probability, a time-degradation factor based on duration of existence, and a relationship heat enhancement factor based on a weighted average of the demand probabilities of neighboring nodes. This ensures that data items with decreased access frequency but still within a high-frequency relationship network can still receive a reasonable retention score, preventing the accidental deletion of critical related data based solely on access frequency. The risk is mitigated by a joint decision based on long-term retention value and storage pool space utilization. For low-value data items with limited storage, redundancy contraction is triggered, reducing the number of replicas to a minimum and switching to high-parity erasure coding to free up space. For high-value data items, redundancy enhancement is triggered, increasing the number of replicas and reducing the parity ratio to improve read performance. For intermediate-value data, a stable configuration is maintained, forming a closed-loop management system for the entire lifecycle from data generation to storage and dynamic adjustment. This maximizes the utilization efficiency of cloud storage space while ensuring data availability and access performance. It is especially suitable for long-term reliable storage and flexible access to massive heterogeneous data in IoT cloud-edge collaborative environments with limited resources. Attached Figure Description

[0015] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0017] like Figure 1 As shown, a computer data storage method based on the Internet of Things includes the following steps: Step S1: Receive the data packet uploaded by the IoT terminal. The data packet contains sensing data and metadata describing the sensing data. The metadata includes at least the device identifier, data generation timestamp, and geographic location tag. Step S2: Extract the time tag and content semantic features of the data packet, and calculate the storage priority index of the data packet by combining it with the reputation value of the IoT terminal that generated the data packet. Step S3: Based on the storage priority index and combined with the average reliability score of multiple available heterogeneous edge storage nodes, dynamically determine the initial redundancy storage strategy for the data packet. The initial redundancy storage strategy includes the number of replicas R and erasure coding parameters (K,M). Step S4: The perceived data is fragmented and redundantly encoded according to the initial redundant storage strategy to generate multiple encoding blocks. Each encoding block is then distributed and stored on different edge storage nodes according to node load, geographical location and consistent hash mapping rules. Step S5: During the data storage cycle, continuously collect data access logs on each storage node, construct a data semantic association graph with data items as nodes and collaborative access relationships between data items as edges, and calculate the probability of each data item being requested in the active query set through a random walk algorithm. Step S6: Based on the probability of being requested, combined with the topological features of the data item in the data semantic association graph and the duration of data existence, calculate the long-term retention value of each data item, and dynamically adjust the number of replicas and erasure coding parameters of each data item according to the long-term retention value and the current storage resource occupancy status, and reclaim the storage space occupied by data items with a long-term retention value lower than the threshold.

[0018] The perceived data content in the data packet is segmented into words to extract keywords. Each keyword is mapped to a corresponding word vector using a pre-trained IoT domain word embedding model. This model is trained using historical operational text corpora from the IoT application domain, employing either a continuous bag-of-words model or a skip-word model. It maps domain-specific keywords to fixed-dimensional real-valued vectors. When multiple keywords are extracted, their word vectors are averaged to obtain a comprehensive semantic vector representing the data packet content. This comprehensive semantic vector is then multiplied by a preset domain importance weight vector. This domain importance weight vector is normalized based on the statistical frequency of each keyword in historical fault or anomaly records corresponding to the device type. The dot product result is then normalized to the 0-1 range using the Sigmoid function to obtain a semantic importance score. The calculation process is expressed as follows: ; In the formula, n is the number of extracted keywords. Let p be the word vector of the p-th keyword. This is the domain importance weight vector. The Sigmoid function, operator Represents the dot product of vectors; Read data to generate a timestamp, and calculate the difference between the current time and the timestamp as the duration the data has existed. The maximum effective duration reference value preset for this type of IoT data. As a normalization benchmark, for existing durations After logarithmic decay processing, the aging sensitivity score is obtained. The specific calculation process is as follows: ; Obtain the reputation value of the IoT terminal that generated the data packet. The calculation process is as follows: The accuracy rate of the data reported by the terminal within a preset historical time window is calculated, which is the proportion of data entries that pass both data integrity verification and numerical range reasonableness verification out of the total number of reported data entries. This proportion is used as the accuracy rate component. ; The proportion of the terminal's online time within a historical time window to the total window duration is used as a stability component. The reputation score is obtained by weighted summation and normalization of the accuracy and stability components. The calculation process is expressed as follows: ; in The preset accuracy weighting coefficients are obtained from a preset level-weight mapping table based on the device importance level of the corresponding terminal. The higher the device importance level, the higher the accuracy weighting coefficients. The larger the value, the range of which is between 0 and 1; Monitor the average round-trip latency and packet loss rate of the current data upload link, and dynamically determine the semantic weight α, timeliness weight β, and reputation weight γ based on the remaining capacity ratio of the edge storage pool, while satisfying α+β+γ=1. semantic importance score Time sensitivity score and equipment reputation value The storage priority index is obtained by multiplying each index by its corresponding weight and then summing the results. The calculation process is expressed as follows: ; In cloud computing environments, massive amounts of data generated by IoT terminals need to be aggregated at edge nodes and then uploaded to cloud storage. While cloud storage resources can be elastically expanded, they are costly. This claim addresses this by constructing a storage priority index that integrates semantic importance, time sensitivity, and device reputation value in a three-dimensional weighted manner. This enables cloud storage systems to automatically quantify the storage value the moment data is entered into the database, providing a precise basis for subsequent hierarchical storage, redundancy strategy selection, and lifecycle management. This effectively solves the problem of insufficient protection of critical business data or the consumption of large amounts of expensive cloud storage resources by irrelevant data due to traditional cloud storage relying solely on simple rules.

[0019] In the smart city public utility management scenario supported by cloud computing platform, a large number of pressure sensors, flow sensors and water quality turbidity sensors are deployed along the urban water supply network. Each sensor uploads the collected data to the municipal cloud data center through the NB-IoT network, and the cloud storage system needs to automatically determine the storage strategy based on the data content.

[0020] Each keyword is mapped to a corresponding word vector using a pre-trained IoT domain-specific word embedding model. This model is trained on historical operation and maintenance text corpora from the water supply network monitoring domain. A pre-defined domain importance weight vector is also used. Based on the statistical frequency of each keyword in the historical fault or anomaly records within the water supply network monitoring field, and after normalization processing, characteristic keywords reflecting network leakage or pollution events, such as "sudden pressure drop," "abnormal flow," and "excessive turbidity," were assigned high weights of 0.93, 0.88, and 0.85, respectively.

[0021] At a certain moment, the cloud receives a data packet from a pressure sensor in section A of the water supply network. After word segmentation, the keyword "sudden pressure drop" is extracted. Let the word vector of "sudden pressure drop" after being mapped by the word embedding model be... Since only one keyword was extracted, the mean vector is [0.82, 0.10, 0.08].

[0022] This vector and Performing the dot product operation, we get .

[0023] After Sigmoid normalization, the semantic importance score .

[0024] The timestamp for this data generation is 0.5 hours old. The maximum valid duration for this type of water supply monitoring data is... Set to 24 hours and substitute into the time sensitivity calculation formula. .

[0025] In the past 30 days, the pressure sensor terminal reported a total of 86,400 data entries, of which 85,536 entries passed cyclic redundancy check and numerical range check. The accuracy component... ; The terminal's online time accounts for 0.996% of the total window time, and the stability component... .

[0026] This pressure sensor is a critical monitoring device in the water supply network, classified as Level I in importance. The corresponding accuracy weight coefficient is obtained by looking up the data in a pre-defined level-weight mapping table. Reputation value .

[0027] The current cloud storage gateway's uplink average round-trip latency is 10 milliseconds and the packet loss rate is 0.01%, both within the normal range. The remaining capacity of the storage pool is 62%, and the conditions for dynamic weight adjustment have not been triggered. The weights remain at their default values ​​of α=0.4, β=0.3, and γ=0.3.

[0028] Final storage priority index .

[0029] This high score will trigger the cloud to allocate a higher number of copies to the data packet and use low redundancy erasure coding parameters to ensure that the key monitoring data of suspected pipeline leaks are adequately protected by redundancy in cloud storage, supporting the subsequent training of pipe burst early warning models and emergency dispatch decision analysis.

[0030] In contrast, the total water meter readings from section B of the water supply network at the same time, after word segmentation, have the keyword "cumulative flow," and the dot product of its word vector and domain weight vector is only 0.18. This water meter terminal is a conventional monitoring device with an equipment importance level of III. The information can be obtained from the level-weight mapping table. Furthermore, this terminal has a historical reporting accuracy rate of 0.88, online stability of 0.93, and a reputation score of [missing information]. ,exist Under the same circumstances, The cloud automatically adopts a low-copy, high-compression storage strategy, effectively saving cloud storage space costs.

[0031] The rules for dynamically determining semantic weight α, timeliness weight β, and reputation weight γ are as follows: When the average round-trip time exceeds the first preset threshold and the packet loss rate increases, the timeliness weight β is increased and the semantic weight α is decreased accordingly, while the reputation weight γ remains unchanged, so as to prioritize the redundant processing of time-sensitive data. When the remaining capacity of the edge storage pool is lower than the second preset threshold, the reputation weight γ is increased and the semantic weight α and the timeliness weight β are decreased accordingly to ensure the storage reliability of data generated by high-reputation devices. The first preset threshold is determined based on the maximum response latency allowed for this type of IoT data service, and the second preset threshold is determined based on the historical capacity warning water level of the edge storage pool. There is no necessary relationship between the first preset threshold and the second preset threshold. By introducing real-time monitoring of network link status and remaining capacity of edge storage pools during the storage priority index calculation process, the semantic weight α, timeliness weight β, and reputation weight γ are dynamically adjusted, enabling the three-dimensional weighted system to have adaptive adjustment capabilities. When the link is congested, the timeliness sensitivity weight is increased to ensure that real-time data receives more timely redundancy protection. When storage resources are scarce, the device reputation weight is increased to ensure that data from high-reliability devices receives priority protection. Thus, in the cloud computing environment, storage resource allocation is coordinated with network conditions and device reliability for optimization, avoiding the drawbacks of static weight configuration, such as insufficient protection of critical data or excessive resource consumption of low-quality data when the network deteriorates or storage is scarce.

[0032] This scenario involves the management of urban water supply networks under a cloud computing platform. Under normal conditions, the average round-trip latency of the cloud storage gateway link is 10 milliseconds, the packet loss rate is 0.01%, the remaining capacity of the storage pool is 62%, no adjustment conditions are triggered, and the semantic weight α = 0.4, the timeliness weight β = 0.3, and the reputation weight γ = 0.3.

[0033] Scenario 1: Link congestion triggers an increase in timeliness weight. During a certain period, network congestion occurs due to concurrent uploads of municipal video surveillance data, causing the average round-trip latency to surge to 65 milliseconds (exceeding the first preset threshold of 50 milliseconds), and the packet loss rate to rise to 1.8%. At this time, the first dynamic adjustment rule is triggered: the timeliness weight β is increased and the semantic weight α is decreased accordingly, while the reputation weight γ remains unchanged.

[0034] The adjusted weights are α=0.2, β=0.5, and γ=0.3. For the same "pressure drop" data packet mentioned above, its semantic importance score... Time sensitivity score Equipment reputation value Storage priority index before adjustment ; The adjusted storage priority index is calculated as follows: ; The adjusted priority index rose from 0.846 to 0.878, indicating that the system automatically strengthened the redundancy protection of this time-sensitive data in a network congestion environment, enabling it to obtain more copies or a lower erasure code verification ratio, ensuring that important pipeline abnormal information can be stored in the cloud in a timely and reliable manner.

[0035] Scenario 2: Storage resource shortage triggers reputation weight increase. During another period, due to a large amount of historical monitoring data not being cleaned up in time, the remaining capacity of the edge storage pool drops to 18% (below the second preset threshold of 30%), triggering the second dynamic adjustment rule: increasing the reputation weight γ and correspondingly decreasing the semantic weight α and the timeliness weight β. The adjusted weight configuration is α=0.2, β=0.2, and γ=0.6. For the same "sudden pressure drop" data packet, the adjusted storage priority index is calculated as follows: ; The adjusted priority index rose further to 0.913, as the data comes from high-reputation terminals. When storage space is limited, the system prioritizes critical data from high-reputation devices. Data from low-reputation devices, even if similar in content, receives a lower priority due to its higher reputation weight, thus being allocated more economical storage strategies and effectively freeing up storage space. If, at this time, similar data is reported from a device with a reputation value of only 0.85, its... It will drop to about 0.77, automatically triggering high compression ratio storage, thus tilting storage resources toward highly reliable data.

[0036] In step S3, the number of copies R and the erasure coding parameters (K, M) are determined as follows: Get the system's preset minimum number of replicas. and upper limit Store priority index Mapping to the range of replica counts, calculate the base replica count. ; Obtain the average reliability score of all current candidate edge storage nodes. The reliability score is obtained by normalizing the historical online time and failure interval of a node; the higher the score, the more reliable the node. The final number of replicas R is obtained by dividing the basic number of replicas by the average reliability score and rounding up, and then taking the larger value between this and the lower limit of the number of replicas. The calculation process is as follows: ; The number of erasure coding data fragments K is determined based on the target read / write throughput and the size constraints of a single fragment. The number of parity fragments M is calculated by multiplying the number of data fragments K by a storage priority index. With the preset redundancy reliability margin factor The result is obtained by rounding up. The calculation process is as follows: ; in The default value and >1; In a cloud computing environment, by applying the storage priority index to both the number of replicas and the number of erasure coding verification shards, high-value data automatically receives more replicas and stronger erasure coding fault tolerance (a larger M), ensuring complete data recovery even when edge node reliability fluctuates. Low-value data, on the other hand, has fewer replicas and a smaller proportion of erasure coding verification, maintaining basic availability with streamlined storage overhead. This achieves a positive correlation between storage resource consumption and data importance, avoiding the problems of low-value data excessively occupying cloud storage space and high-value data insufficient protection in traditional static configurations, and significantly improving storage cost-effectiveness.

[0037] A scenario for public utility management of urban water supply networks supported by a cloud computing platform. The cloud storage system has a preset minimum number of replicas. Upper limit Redundancy tolerance strength factor The average reliability score of all current candidate edge storage nodes. .

[0038] For the "sudden pressure drop" data packet in section A of the water supply network, its storage priority index is... Basic replica count calculation: ; After node reliability adjustments: ; Selected data shard number K=6, verification shard number: ; Finally, the erasure coding parameters (6,5) were adopted, the total number of coding blocks was 11, each coding block was stored in 5 copies, and the total number of storage units was 5×11=55.

[0039] For the data packet containing "cumulative flow" water meter readings in section B of the water supply network, Base number of replicas: ; After reliability adjustments: ; Verify the number of fragments: ; The erasure code is (6,2), the total number of encoded blocks is 8, and the total number of storage units is 4×8=32.

[0040] The total storage units (55) of high-value abnormal data are significantly higher than the total storage units (32) of regular readings. The allocation of storage resources is strictly positively correlated with the value of data, and the cloud space is used efficiently and reasonably.

[0041] When the average reliability of nodes further decreases (e.g.) At that time, the number of high-value data copies will increase to Unchanged, but if Adjustments will be made accordingly to changes; the number of copies of low-value data will be increased. The solution maintains a reasonable gap. It ensures that critical water supply monitoring data receives strong redundancy protection in dynamically changing edge environments, while ordinary meter reading data only occupies necessary storage resources, achieving a refined balance between storage cost and reliability.

[0042] In step S4, fragmentation and redundancy coding are performed according to the erasure coding parameters (K,M), specifically including: The raw sensing data is divided into K data segments, and M verification segments are generated using erasure coding parameters, for a total of N=K+M coding blocks; Obtain the current available storage capacity, CPU load rate, and geographic location coordinates of each edge storage node, and calculate a comprehensive hash weight for each node based on the node reliability score; Construct a weighted consistent hash ring to map N encoded blocks sequentially and with minimal collisions to the node positions on the ring; During the mapping process, mandatory verification ensures that the physical nodes containing any K coded blocks are distributed in at least two different geographical regions; In a cloud computing environment, edge storage nodes are characterized by strong heterogeneity, geographical dispersion, and dynamic changes. This claim obtains the current available storage capacity, CPU load rate, and geographical coordinates of each node, and calculates a comprehensive hash weight by combining node reliability scores. It then constructs a weighted consistent hash ring, maps multiple coded blocks generated by erasure coding to nodes on the ring according to the principle of minimizing collisions, and forces verification after mapping that the physical nodes where any K coded blocks are located are distributed in at least two different geographical regions. This achieves cross-regional fault isolation and load balancing during the data writing stage, avoiding the risk of data irrecoverability due to the failure of a single node or a single region, and effectively utilizing the real-time processing capabilities of each node, significantly improving the data persistence and access efficiency of the cloud edge storage system.

[0043] In the scenario of urban water supply network public utility management, using defined redundancy parameters, for high-value data packets experiencing a sudden drop in pressure, the erasure coding parameters are K=6, M=5, the total number of coding blocks is N=11, each coding block stores R=5 replicas, and the total number of storage units is 55. There are currently 6 candidate edge storage nodes, located in two geographical regions, A and B, with 3 nodes in each region. The real-time status of each node is as follows:

[0044] The node's overall hash weight is obtained by weighted summation of available storage capacity percentage, CPU idle rate, and reliability score, with weights of 0.4, 0.3, and 0.3 respectively. The calculation results are as follows: ; ; ; ; ; ; Each node is assigned a corresponding number of virtual nodes on the consistent hash ring according to its comprehensive hash weight. The higher the weight, the more virtual nodes there are, thus increasing the probability that a coded block will be mapped to that node. The 11 coded blocks are mapped onto the ring sequentially and with minimal collisions to obtain the initial distribution. Then, a geographic region check is performed: the set of physical nodes containing any 6 (K=6) coded blocks is extracted, and it is checked whether they cover at least two regions. If all 6 blocks fall in region A, some blocks are moved to nodes in region B. The final distribution ensures that any 6 coded blocks meet the cross-region condition, meaning that the data can still be recovered even if any single region fails completely. For example, the coded blocks can be distributed in N1 (2 copies), N2 (1 copy), N3 (2 copies), N4 (2 copies), N5 (1 copy), and N6 (3 copies). The total number of replicas meets the storage capacity of each node, and nodes in region A hold 5 different coded blocks, while nodes in region B hold 6. Any K=6 blocks must cross regions. This mapping process fully utilizes the large capacity and high reliability of the high-weight node N6 while avoiding overload on the low-load node N5. It also enhances data resilience through multi-regional distribution. For low-value "cumulative traffic" data, with 8 encoded blocks and 4 replicas, the mapping and verification are performed similarly, resulting in fewer overall storage units and reflecting differentiated distribution.

[0045] The process described in step S5, which uses a random walk algorithm to calculate the demand probability of each data item in the active query set, is as follows: Taking each stored data item as a node of the graph, if within the observation time window, the same query request or sequentially adjacent query requests involve both data item i and data item j, a directed edge is established between node i and node j, and the weight of the edge is proportional to the co-access frequency; When receiving a query set Q composed of a group of active query tasks, the data item nodes contained in Q are taken as restart nodes to perform iterative calculation of random walk with restart. The initial value of the demanded probability of all nodes is set to 0, and only the initial value of each node in the query set Q is set to 1 / |Q|. In each iteration, the demanded probability of each node is composed of two weighted parts: one part is obtained by random walk transition according to the probability values of the in-edge neighbor nodes of the current node and the edge weights, and the other part returns the probability mass to the restart nodes according to a preset restart probability. The iteration is repeated until the variation of the demanded probability value of all nodes between two consecutive iterations is less than the preset convergence threshold or the number of iterations reaches the preset maximum number of iterations, the iteration is terminated to obtain the demanded probability of each data item i , and its iterative calculation process is expressed as: ; where d is the preset restart probability, with a value range of 0 < d < 1, representing the probability of returning to the restart node in the query set Q at each step in the random walk process; is the component of the restart vector at the i-th node, and when data item i belongs to the query set Q , otherwise , where |Q| represents the number of data items contained in the query set Q; is the weight of the directed edge pointing from node j to node i, and its value is equal to the frequency of co-access of data item j and data item i within the observation time window; In(i) is the set of all in-edge neighbor nodes pointing to node i; Out(j) is the set of all out-edge neighbor nodes of node j; h is an index variable that traverses all out-edge neighbor nodes of node j; is the weight of the directed edge pointing from node j to its out-edge neighbor node h; is the demanded probability value of node j at the t-th iteration; IoT data items often exhibit implicit relationships driven by business logic. Traditional methods assess their value solely based on the independent access frequency of individual data items, easily overlooking important data that, despite low access volume, is closely linked to hot data. This invention constructs a data semantic association graph and employs a random walk algorithm with restart to calculate the probability of being requested. This ensures that the probability of each data item being requested depends not only on whether it is directly queried but also on the propagation of probability quality along the association edges from the active query set. This accurately reflects the comprehensive demand for data items within the business collaboration network, providing a more reliable quantitative basis for subsequent long-term retention value calculations and differentiated redundancy adjustments.

[0046] This scenario involves the management of a city's water supply network. The cloud storage system contains the following data items: Data item 1 is "Pressure Drop Alarm Record for Section A," Data item 2 is "Flow Change Record for Section A," Data item 3 is "Historical Maintenance Record for Section A," and Data item 4 is "Regular Temperature Reading for Section A." Analysis of historical access logs reveals that when a "Pressure Drop" alarm is triggered, maintenance personnel typically retrieve the corresponding section's "Flow Change Record" to confirm the severity of the event and occasionally consult the section's "Historical Maintenance Record" for further analysis. However, "Regular Temperature Reading" is almost never accessed in conjunction with alarm events. Based on this, a directed semantic association graph is constructed: within the observation window, the frequency of collaborative access between Data item 1 and Data item 2 is 10. A bidirectional edge is established, with weights... The frequency of collaborative access between data item 1 and data item 3 is 3. A bidirectional edge is established with weights... Data item 4 has no collaborative access edge.

[0047] The currently active query set Q contains only data item 1 (the system detected a sudden drop in pressure and pushed an alarm; the operations and maintenance personnel are currently viewing this record), |Q|=1.

[0048] The restart probability is set to d=0.15. The component corresponding to data item 1 in the restart vector r. The rest are 0.

[0049] Initialize the probability of each node being in demand as follows: Other values ​​are 0. Calculations are performed according to the iterative formula: ; First iteration: ; ; ; ; After multiple iterations, the probability converges, and the steady-state value can be solved using the system of equations. Let the convergence values ​​be respectively... ,satisfy: ; ; ; Substitution Solving for ,but .

[0050] The probability of being requested is as follows: alarm records themselves have a probability of 0.5405, strongly correlated traffic mutation records have a probability of 0.3534, historical maintenance records have a probability of approximately 0.1061, while unrelated regular temperature readings have a probability of 0. This result indicates that although traffic mutation records do not appear in the current active query set, they are still assigned a high probability of being requested due to their close business correlation with alarm records, thus ensuring that they are not prematurely discarded due to being "unpopular" in subsequent storage strategy adjustments. Meanwhile, the truly irrelevant data item 4 has a zero probability of being requested and can be preferentially downgraded. Through this mechanism, the cloud system can more accurately identify business-important related data, achieving efficient allocation of storage resources.

[0051] The method for calculating the long-term retention value of each data item in step S6 is as follows: For data item i, obtain its probability of being needed using the random walk algorithm. ; Get the duration of data item i from its generation time to the current time. With a preset time decay coefficient The exponential decay process is applied to obtain the aging decay factor. ; Extract the set N(i) of all neighbor nodes of data item i in the data semantic association graph, and calculate the demand probability of each neighbor node j in this set. The weighted average, where the weights are the edge weights between nodes i and j. ; Multiplying the weighted average by the preset correlation enhancement coefficient η and then adding one yields the correlation thermal enhancement factor. The calculation process is expressed as follows: ; The probability of data item i being demanded. The long-term retention value V(i) of this data item is obtained by multiplying the aging decay factor and the associated thermal enhancement factor together. The calculation process is expressed as follows: ; In a cloud computing environment, the value of data diminishes over time, and some data, while having low direct demand, are strongly related to high-demand data in business contexts. This claim calculates the long-term retention value by multiplying the probability of demand, the time-related decay factor, and the correlation enhancement factor. This allows the assessment of the storage value of data items to simultaneously consider the intensity of current business demand, data freshness, and its importance in the semantic association network. It can effectively prevent the accidental deletion of data with relevant value due to excessively rapid time-related decay or isolated assessment, and provide accurate quantitative basis for subsequent redundancy adjustments. Thus, it maximizes the long-term utilization efficiency of cloud storage space while ensuring data availability.

[0052] Continuing with the public utility management scenario of urban water supply networks, the probability of demand is calculated.

[0053] The probability of each data item being requested is as follows: Data item 1 (A-section pressure drop alarm record) Data item 2 (Traffic mutation record in section A) Data item 3 (Historical maintenance records for section A) Data item 4 (routine temperature readings for section A) .

[0054] Preset time decay coefficient (per hour), the correlation enhancement coefficient η=0.5.

[0055] The duration of each data item from its generation time to the current time is set as follows: Alarm Records Hourly Traffic Flow Change Records Hours, historical maintenance records Hourly (7-day) routine temperature readings Hours. The time-related decay factor is calculated as follows: ; ; ; ; When calculating the associated thermal enhancement factor, the neighbor set and edge weights are the same as before: The neighbors of data item 1 are 2 and 3, with weights... ; The neighbor of data item 2 is 1, and the weight is... The neighbor of data item 3 is 1, and the weight is... ; Data item 4 has no neighbors, according to regulations .

[0056] The weighted average of the demand probability of the neighbors of data item 1 is: ; ; The weighted average of data item 2 is 0.5405. Data item 3 is also... .

[0057] Long-term retention value calculation: ; ; ; ; The results show that alarm records have the highest retention value due to their high demand and timeliness; traffic mutation records, although not directly queried, have a strong correlation with alarm records, and the correlation enhancement factor increases their retention value, preventing them from being prematurely eliminated; historical maintenance records, due to their long generation time and severe time decay, have extremely low retention value even with enhancement factors and can be prioritized for downgraded storage; routine temperature data has no business relevance and no demand, and its value is zero. If only the probability of being needed is considered, data item 2 scores 0.353, and data item 1 scores 0.541. However, after introducing timeliness and correlation, the relative gap between data item 2 and 1 narrows, demonstrating the case's ability to protect related data.

[0058] The specific rules for dynamically adjusting the number of copies and erasure coding parameters for each data item in step S6 are as follows: Set a high retention threshold and low retention threshold And monitor the current space occupancy rate U of the entire edge storage pool in real time; When the long-term retention value V(i) of data item i is lower than the low retention threshold When the space occupancy rate U exceeds the preset storage pressure threshold, a redundancy shrinkage operation is triggered: the number of replicas of data item i is reduced to the preset minimum number of replicas, and its erasure coding configuration is switched to a parameter that increases the proportion of parity shards, so as to reduce the total storage overhead while maintaining basic availability. When the long-term retention value V(i) of data item i is higher than the high retention threshold When the current number of replicas is lower than the preset maximum number of replicas, the number of replicas is increased and the proportion of erasure coding check shards is reduced to improve data reading performance. For long-term retention value between and The data items between them will maintain their current redundancy configuration. In a cloud computing environment, by jointly deciding on long-term retention value and storage pool space occupancy, the system can automatically perform redundant shrinking of low-value data when storage resources are scarce, reducing its replica count and switching to erasure coding with a high checksum ratio to free up space. At the same time, when the demand for high-value data increases, its replica count is increased and the checksum ratio is reduced to improve access performance, while maintaining stability for intermediate value data. This forms a closed-loop adaptive scheduling mechanism for storage resources, which not only ensures the high availability and read efficiency of critical business data, but also effectively suppresses the continuous occupation of cloud storage space by inefficient data, achieving a dynamic balance between storage cost and service quality.

[0059] In the scenario of public utility management of urban water supply networks, the long-term retention value V is calculated previously.

[0060] Set a high retention threshold Low retention threshold Storage stress threshold .

[0061] The current edge storage pool occupancy rate U=0.88, exceeding the stress threshold. Minimum replica count preset. , the maximum number of copies The current configuration and long-term retention value of each data item are as follows: Data item 1 (Alarm record of sudden pressure drop in section A): , the current number of copies Erasure coding (K=6, M=5), check fragment percentage .

[0062] Data item 2 (Traffic surge record in section A): , between and In between, maintain the current configuration: The erasure code (6,4) accounts for 0.667%.

[0063] Data item 3 (Historical maintenance records for section A): ,current The erasure code (6,3) accounts for 0.5%.

[0064] Data item 4 (routine temperature readings for section A): ,current The erasure code (6,2) accounts for 0.333.

[0065] because Triggering adjustments: Data item 1 meets the enhancement conditions and the current replica count is 5 < 6, so perform redundancy enhancement: increase the replica count to 1. The proportion of erasure code verification fragments is reduced, M is adjusted from 5 to 3, and the new erasure code (6,3) accounts for 0.5%.

[0066] The adjusted number of storage units is 6 × (6 + 3) = 54.

[0067] Data item 3 satisfies the contraction condition ( and ): The number of copies has been reduced to The proportion of erasure code verification fragments has increased, and M has been adjusted from 3 to 8. The new erasure code is (6,8), accounting for 1.333.

[0068] The adjusted number of storage units is 1×(6+8)=14.

[0069] Data item 4 also shrinks: When M increases from 2 to 10, the erasure code is (6,10), and the number of storage units is 1×16=16.

[0070] The following is a comparison of storage resource usage before and after the adjustment: Before adjustment, the total number of storage units was 5×11+4×10+3×9+4×8=55+40+27+32=154; The total number of storage units after adjustment is 6×9+4×10+1×14+1×16=54+40+14+16=124; The total number of storage units was reduced by approximately 19.5%, effectively freeing up cloud storage space. Data item 1, a critical alarm record, saw its replicas increased to 6 while its verification ratio decreased. This allows for concurrent retrieval from more nodes, improving the real-time performance of fault diagnosis. Although the verification ratio for data items 3 and 4 increased, the total overhead was significantly reduced due to the replica count decreasing to 1, and they can still be recovered via erasure coding, meeting basic availability requirements. If the storage pool occupancy rate subsequently falls below the stress threshold, and the long-term retention value of certain low-value data increases (e.g., due to new business connections), the system can be dynamically adjusted again for continuous optimization.

[0071] An Internet of Things (IoT) based computer data storage system includes: The data receiving module is used to receive data packets uploaded by IoT terminals. The data packets contain sensing data and metadata describing the sensing data. The metadata includes at least the device identifier, data generation timestamp, and geographic location tag. The priority calculation module is used to extract the time tag and semantic features of the data packet, and combine them with the reputation value of the IoT terminal that generated the data packet to calculate the storage priority index of the data packet. The initial redundancy configuration module is used to dynamically determine the initial redundancy storage strategy for the data packet based on the storage priority index and the average reliability score of multiple available heterogeneous edge storage nodes. The initial redundancy storage strategy includes the number of replicas R and erasure coding parameters (K,M). The sharding and distribution module is used to shard and redundantly encode the sensing data according to the initial redundant storage strategy, generate multiple coding blocks, and distribute and store each coding block to different edge storage nodes according to node load, geographical location and consistent hash mapping rules. The association analysis module is used to continuously collect data access logs on each storage node during the data storage cycle. It constructs a data semantic association graph with data items as nodes and collaborative access relationships between data items as edges, and calculates the probability of each data item being requested in the active query set through a random walk algorithm. The dynamic adjustment module is used to calculate the long-term retention value of each data item based on the probability of being needed, combined with the topological features of the data item in the data semantic association graph and the duration of data existence. Based on the long-term retention value and the current storage resource occupancy status, it dynamically adjusts the number of replicas and erasure coding parameters of each data item and reclaims the storage space occupied by data items with a long-term retention value below the threshold.

[0072] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A computer data storage method based on the Internet of Things, characterized in that, Includes the following steps: Step S1: Receive the data packet uploaded by the IoT terminal. The data packet contains sensing data and metadata describing the sensing data. The metadata includes at least the device identifier, data generation timestamp, and geographic location tag. Step S2: Extract the time tag and content semantic features of the data packet, and calculate the storage priority index of the data packet by combining it with the reputation value of the IoT terminal that generated the data packet. Step S3: Based on the storage priority index and combined with the average reliability score of multiple available heterogeneous edge storage nodes, dynamically determine the initial redundancy storage strategy for the data packet. The initial redundancy storage strategy includes the number of replicas and erasure coding parameters. Step S4: The perceived data is fragmented and redundantly encoded according to the initial redundant storage strategy to generate multiple encoding blocks. Each encoding block is then distributed and stored on different edge storage nodes according to node load, geographical location and consistent hash mapping rules. Step S5: During the data storage cycle, continuously collect data access logs on each storage node, construct a data semantic association graph with data items as nodes and collaborative access relationships between data items as edges, and calculate the probability of each data item being requested in the active query set through a random walk algorithm. Step S6: Based on the probability of being requested, combined with the topological features of the data item in the data semantic association graph and the duration of data existence, calculate the long-term retention value of each data item, and dynamically adjust the number of replicas and erasure coding parameters of each data item according to the long-term retention value and the current storage resource occupancy status, and reclaim the storage space occupied by data items with a long-term retention value lower than the threshold.

2. The computer data storage method based on the Internet of Things according to claim 1, characterized in that: The perceived data content in the data packet is segmented into words to extract keywords. Each keyword is mapped to a corresponding word vector using a pre-trained IoT domain word embedding model. When multiple keywords are extracted, the word vectors of multiple keywords are averaged to obtain a comprehensive semantic vector representing the content of the data packet. The comprehensive semantic vector is then multiplied by a preset domain importance weight vector. The domain importance weight vector is normalized based on the statistical frequency of each keyword in the domain corresponding to the device type in historical fault or anomaly records. The dot product result is normalized to the 0 to 1 range using the Sigmoid function to obtain the semantic importance score. Read the data to generate a timestamp, calculate the difference between the current time and the timestamp as the data's existing duration, use the maximum effective duration reference value preset for this type of IoT data as the normalization benchmark, and after logarithmic decay processing on the existing duration, obtain the timeliness sensitivity score. The reputation value of the IoT terminal that generates the data packet is obtained. The calculation process is as follows: the accuracy rate of the data reported by the terminal within the preset historical time window is calculated, that is, the proportion of the number of data items that pass the data integrity check and the numerical range reasonableness check to the total number of reported data items, which is used as the accuracy rate component. The proportion of the terminal's online time within a historical time window to the total window time is used as a stability component. The reputation value is obtained by weighted summation and normalization of the accuracy and stability components. Monitor the average round-trip latency and packet loss rate of the current data upload link, and dynamically determine the semantic weight, timeliness weight and reputation weight based on the remaining capacity ratio of the edge storage pool, with the sum of the semantic weight, timeliness weight and reputation weight being 1. The storage priority index is obtained by multiplying the semantic importance score, timeliness sensitivity score, and device reputation score by their respective weights and then summing them.

3. The computer data storage method based on the Internet of Things according to claim 2, characterized in that: The rules for dynamically determining semantic weight, timeliness weight, and reputation weight are as follows: When the average round-trip time exceeds the first preset threshold and the packet loss rate increases, the timeliness weight is increased and the semantic weight is decreased accordingly, while the reputation weight remains unchanged, so as to prioritize the redundant processing of time-sensitive data. When the remaining capacity of the edge storage pool is lower than the second preset threshold, the reputation weight is increased and the semantic weight and timeliness weight are decreased accordingly to ensure the storage reliability of data generated by high-reputation devices.

4. The computer data storage method based on the Internet of Things according to claim 3, characterized in that: The number of copies and erasure coding parameters in step S3 are determined as follows: Obtain the system's preset lower and upper limits for the number of replicas, map the storage priority index to the replica number range, and calculate the basic number of replicas; Obtain the average reliability score of all current candidate edge storage nodes. The reliability score is obtained by normalizing the node's historical online time and failure interval. The higher the value, the more reliable the node. Divide the basic number of replicas by the average reliability score and round up. Take the larger value between the base number of replicas and the lower limit of the number of replicas to obtain the final number of replicas. The number of erasure coding data fragments is determined based on the target read / write throughput and the size constraints of a single fragment. The number of verification fragments is obtained by multiplying the number of data fragments by the storage priority index and the preset redundancy reliability margin factor, and then rounding up.

5. A computer data storage method based on the Internet of Things according to claim 4, characterized in that: In step S4, fragmentation and redundancy coding are performed according to erasure coding parameters, specifically including: The raw sensing data is divided into K data segments, and M verification segments are generated using erasure coding parameters, for a total of N=K+M coding blocks; Obtain the current available storage capacity, CPU load rate, and geographic location coordinates of each edge storage node, and calculate a comprehensive hash weight for each node based on the node reliability score; Construct a weighted consistent hash ring to map N encoded blocks sequentially and with minimal collisions to the node positions on the ring; During the mapping process, mandatory verification ensures that the physical nodes containing any K coded blocks are distributed in at least two different geographical regions.

6. The computer data storage method based on the Internet of Things according to claim 5, characterized in that: The process described in step S5, which uses a random walk algorithm to calculate the demand probability of each data item in the active query set, is as follows: Using each stored data item as a node in the graph, if within the observation time window there is the same query request or consecutive query requests involving data item i and data item j, then a directed edge is established between node i and node j, and the weight of the edge is proportional to the frequency of collaborative access. When a query set consisting of a group of active query tasks is received, the data item nodes contained in the query set are used as restart nodes. Random walk iteration calculation with restart is performed. The initial value of the demand probability of all nodes is set to 0, and only the initial value of each node in the query set is set to a preset value. In each iteration, the demand probability of each node is composed of two weighted parts: one part is obtained by random walk transfer based on the probability value of the incoming neighbor nodes of the current node and the edge weight, and the other part is the probability quality is pushed back to the restart node according to the preset restart probability. Repeat the iteration until the change in the demand probability of all nodes between two consecutive iterations is less than the preset convergence threshold, or the number of iterations reaches the preset maximum number of iterations, and then terminate the iteration to obtain the demand probability of each data item i.

7. A computer data storage method based on the Internet of Things according to claim 6, characterized in that: The method for calculating the long-term retention value of each data item in step S6 is as follows: For data item i, obtain its probability of being needed using the random walk algorithm; The duration of data item i from its generation time to the current time is obtained, and it is subjected to exponential decay processing with a preset time decay coefficient to obtain the time decay factor. Extract the set of all neighbor nodes of data item i in the data semantic association graph, and calculate the weighted average of the demand probability of each neighbor node j in the set of all neighbor nodes, with the weight being the edge weight between node i and j. Multiply the weighted average by the preset correlation enhancement coefficient and add one to obtain the correlation thermal enhancement factor; The long-term retention value of data item i is obtained by multiplying its demand probability, time decay factor, and associated thermal enhancement factor.

8. A computer data storage method based on the Internet of Things according to claim 7, characterized in that: The specific rules for dynamically adjusting the number of copies and erasure coding parameters for each data item in step S6 are as follows: Set high and low retention thresholds and monitor the current space occupancy of the entire edge storage pool in real time; When the long-term retention value of data item i is lower than the low retention threshold and the space occupancy rate U exceeds the preset storage stress threshold, a redundancy shrinkage operation is triggered: the number of replicas of data item i is reduced to the preset minimum number of replicas, and its erasure coding configuration is switched to a parameter that increases the proportion of parity fragments, so as to reduce the total storage overhead while maintaining basic availability. When the long-term retention value of data item i is higher than the high retention threshold, a redundancy enhancement operation is triggered: if the current number of replicas is lower than the preset maximum number of replicas, the number of replicas is increased and the proportion of erasure coding check fragments is reduced to improve data reading performance. For data items whose long-term retention value falls between the low retention threshold and the high retention threshold, their current redundancy configuration remains unchanged.

9. A computer data storage system based on the Internet of Things, applied in any one of the methods described in claims 1-8, characterized in that: include: The data receiving module is used to receive data packets uploaded by IoT terminals. The data packets contain sensing data and metadata describing the sensing data. The metadata includes at least the device identifier, data generation timestamp, and geographic location tag. The priority calculation module is used to extract the time tag and semantic features of the data packet, and combine them with the reputation value of the IoT terminal that generated the data packet to calculate the storage priority index of the data packet. The initial redundancy configuration module is used to dynamically determine the initial redundancy storage strategy for the data packet based on the storage priority index and the average reliability score of multiple available heterogeneous edge storage nodes. The initial redundancy storage strategy includes the number of replicas R and erasure coding parameters (K,M). The sharding and distribution module is used to shard and redundantly encode the sensing data according to the initial redundant storage strategy, generate multiple coding blocks, and distribute and store each coding block to different edge storage nodes according to node load, geographical location and consistent hash mapping rules. The association analysis module is used to continuously collect data access logs on each storage node during the data storage cycle. It constructs a data semantic association graph with data items as nodes and collaborative access relationships between data items as edges, and calculates the probability of each data item being requested in the active query set through a random walk algorithm. The dynamic adjustment module is used to calculate the long-term retention value of each data item based on the probability of being needed, combined with the topological features of the data item in the data semantic association graph and the duration of data existence. Based on the long-term retention value and the current storage resource occupancy status, it dynamically adjusts the number of replicas and erasure coding parameters of each data item and reclaims the storage space occupied by data items with a long-term retention value below the threshold.

Images