Low-power-consumption wireless sensor gateway system and data transmission optimization method

Abstract

The invention discloses a low-power-consumption wireless sensor gateway system and a data transmission optimization method, and relates to the technical field of Internet of Things edge computing and low-power-consumption wireless sensor networks, and the system comprises a heterogeneous computing gateway, a multi-source sensor cluster node and a low-power-consumption communication module. The system realizes that the prediction error is less than or equal to 0.93 and the daily average communication traffic is reduced by 63% under the condition of high data loss through spatial-temporal feature perception and a dynamic threshold decision mechanism. According to the specific scheme, a high-computing-power cluster head is selected based on a clustering network, time-varying features are extracted through a robust matrix decomposition model, a transmission threshold value is dynamically adjusted in combination with the wind speed grade and the equipment temperature state, model parameters are updated through an incremental learning mechanism, the verified parameters are solidified to a nonvolatile storage medium, and long-term operation stability is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things edge computing and low-power wireless sensor networks, and in particular to a low-power wireless sensor gateway system and a data transmission optimization method. Background Art

[0002] Wireless sensor networks, a core technology in the Internet of Things (IoT), are widely used in smart grids, smart transportation, and environmental monitoring. In industrial scenarios like base station monitoring, traditional gateway devices struggle to run multidimensional matrix decomposition algorithms due to insufficient computing power and limited memory. This leads to an imbalance between prediction accuracy and energy consumption. Field measurements show that traditional methods have a mean absolute error exceeding 1.2, while reducing daily transmission volume by less than 30%. Furthermore, the need for high-frequency data collection increases node energy consumption, reducing battery life to less than three months.

[0003] Existing technologies suffer from significant drawbacks: general-purpose microcontrollers lack dedicated acceleration units, making them incapable of real-time computing; existing solutions fail to adequately account for dynamic risk characteristics in industrial environments, such as sudden changes in wind speed, structural tilt, and equipment temperature rise; and static threshold strategies lead to redundant communications, such as base stations continuously transmitting non-critical data during low-risk periods. These issues severely limit the effectiveness of wireless sensor networks in monitoring critical industrial facilities.

[0004] Energy constraints are a core challenge in wireless sensor networks. While model-driven data acquisition can reduce communication traffic, existing solutions rely on cloud computing for dynamic threshold control, resulting in insufficient real-time performance. Furthermore, matrix factorization models are large in number of parameters and require at least 8GB of memory, making deployment impossible on traditional embedded devices due to limited storage resources. Furthermore, frequent erasure of storage media significantly shortens device lifespan, a problem that existing hierarchical storage mechanisms fail to effectively address.

[0005] Therefore, how to achieve dynamic threshold control and real-time prediction in resource-constrained embedded gateways while balancing data accuracy and transmission energy consumption has become an urgent problem to be solved. Summary of the Invention

[0006] This paper addresses two major technical issues in base station monitoring networks: low prediction accuracy due to high rates of time-series data loss and insufficient real-time performance of embedded gateways. By proposing a low-power sensor gateway and a multi-dimensional time-series data optimization method, the gateway is deployed in base station monitoring equipment within a heterogeneous computing architecture. Through dynamic threshold control and incremental learning, the gateway maintains a MAE (Maintaining Error) of 0.93 or less even when the data loss rate is 30% or higher, while reducing average daily traffic by 63%.

[0007] The core innovation of the method is to suppress missing values ​​and noise interference through a robust matrix decomposition model, and to build a The norm loss function, combined with an autoregressive regularization term, captures temporal correlations. The basis matrix U and weight matrix W are calculated in real time by the embedded processor's dedicated acceleration unit, and the coefficient matrix X is dynamically updated based on sensor data. To address the varying wind speeds and device temperature rises in base station environments, dynamically adjusted thresholds based on wind speed and temperature are set. When the sensor data residual exceeds the threshold, a low-power communication module is triggered to transmit data, significantly reducing redundant data backhaul.

[0008] The gateway hardware architecture integrates a dynamic threshold decision module and an incremental learning unit. Matrix decomposition parameters are stored in real time via high-speed memory and then stored in non-volatile storage media after verification. When the residual exceeds the dynamic threshold, the incremental learning unit triggers an update of the matrix X. A single prediction takes ≤30ms, meeting the requirements for high-frequency data collection at the 5-minute level. This solution optimizes data transmission through a clustering algorithm. Cluster head nodes must maintain a main frequency of ≥2.4GHz and a residual energy of ≥20%, ensuring a balanced computing power and energy consumption.

[0009] To achieve the above object, the technical solution adopted by the present invention is: In a first aspect of the present invention, a low-power wireless sensor gateway system is provided, characterized in that it includes: a main control chip; The main control chip is connected to the communication module, sensor module, power module, storage module and power supply and heat dissipation module respectively; The main control chip uses a Rockchip RK3588 processor equipped with LPDDR48GB memory to execute the multi-dimensional time series prediction model and dual prediction scheme; The communication module includes a SX1262 LoRa communication submodule and an IPEX-4 antenna interface; the sensor module includes a cluster node device of a Witmu RS-WS-N01-8 wind speed sensor, a Wogan Electronics TILT-01H structural tilt sensor, and a Hengge HSTS-01 temperature sensor; The power module includes an 18650 lithium battery pack and a TPS63020 power chip; the storage module includes an LPDDR4 8GB memory and a UFS3.1 128GB storage submodule; the power supply and heat dissipation module includes a power supply interface and an aluminum alloy heat sink; The cluster node device establishes a data link with the gateway through the SX1262LoRa communication submodule and uploads sensor data at intervals of less than or equal to 5 minutes. The main control chip dynamically adjusts the transmission threshold to 0.3 or 0.5 according to the real-time wind speed value input by the wind speed sensor and the real-time temperature value input by the temperature sensor.

[0010] Preferably, when the data residual exceeds the transmission threshold, the coefficient matrix is ​​updated through the LPDDR48GB memory and the model parameters are solidified to the UFS3.1128GB storage submodule.

[0011] Preferably, the main control chip further includes an NPU acceleration unit, and the NPU acceleration unit is used to accelerate the calculation of the autoregressive weight matrix.

[0012] Preferably, the parameters smoothed by the timing prediction algorithm are synchronized to the cluster node device through the SX1262LoRa communication submodule.

[0013] Preferably, the system realizes wide voltage input and regulated power supply through 18650 lithium battery pack and TPS63020 power chip.

[0014] As an advantage, the multi-dimensional time series prediction model is This measures the quality of non-negative matrix factorization and is used to improve the robustness of matrix factorization. Y is the matrix into which the node organizes the collected data. The training process of the multi-dimensional time series prediction model is as follows: Input low-rank matrix rank , time-delay set , learning rate , the regularization parameter , maximum number of iterations , randomly initialize all parameters, including the basis matrix , coefficient matrix and the time regularization parameter matrix ; Calculate the objective function matrix The partial derivatives of , and the basis matrix is ​​updated by the following formula and Adam optimizer, which is in the form of: ; in is the learning rate; Represents a projection operator that forces all elements to be projected onto the non-negative semi-axis. Represents the partial derivative of the loss function with respect to U; Calculate the objective function matrix The partial derivative of , and the coefficient matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ; Calculate the objective function matrix The partial derivative of , and the time regularization parameter matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ; Finally, the model parameters with validation set MAE ≤ 0.9 are stored in the UFS3.1 128GB storage module.

[0015] In a second aspect of the present invention, a method for reducing data transmission in a wireless sensor network is provided, comprising the following steps: Step 1: Clustering is performed according to the clustering algorithm and cluster heads are elected. Nodes with strong computing power and sufficient remaining energy are selected as cluster heads. After clustering is completed, the cluster head starts to collect cluster data, including wind speed data collected by the wind speed sensor, tilt data collected by the tilt sensor, and temperature data collected by the temperature sensor, through the independent chip STM32L071K8U6 control, and sends it to the base station monitoring gateway of the SX1262LoRa communication submodule; Step 2: The base station monitoring gateway organizes the collected data into a matrix , is the number of cluster members, The amount of data collected by each node, the time series prediction model parameters obtained by training the objective function, that is, the basis matrix , coefficient matrix and the time regularization parameter matrix , the objective function is as follows: Step 3: The gateway sends the matrices U, X, and W to the cluster head through the SX1262LoRa communication submodule. The cluster head and the gateway confirm the synchronization prediction model. Step 4: The cluster head uses the model to predict the current data , the data sending node will compare the predicted value with the actual observed value Compare and calculate residuals , comparing the residual r with the threshold value dynamically set according to the wind speed value input by the wind speed sensor and the temperature value input by the temperature sensor in the base station environment; Step 5: All data at the current moment do not meet the threshold condition, the data is transmitted and the base matrix U and coefficient matrix X are updated through the NPU acceleration unit of the main control chip; Use the model to predict current data , the data sending node will compare the predicted value with the actual observed value Compare and calculate residuals , compared with the residual with predefined error data and without updating the model; Step 6: Synchronize the prediction model parameters through the incremental learning unit of the gateway, where matrices U and W are solidified to the UFS3.1128GB storage submodule, jump to step 4, and perform the next iteration.

[0016] Preferably, in step 4, the method further comprises: Step S41: When the residual When the prediction value is less than the dynamically set threshold, the system considers the prediction reasonable, the cluster head does not send data, and the corresponding sink node obtains the prediction value obtained by the prediction model in step 2. ,The cluster head sends a beacon signal consisting of a small data packet to inform the data sender and receiver to take the model prediction value as the valid prediction value; Step S42: When the residual When the value is greater than or equal to the dynamically set threshold, the system considers the prediction unreasonable and triggers the SX1262 LoRa communication submodule to transmit sensor data and update the matrix X in the LPDDR48GB memory.

[0017] The beneficial effects of the present invention are: 1. Highly robust prediction capability: When the missing rate of wind speed, tilt angle, and temperature data is ≥30%, the prediction error is reduced to 0.82-0.93 by extracting time-varying features through low-rank matrix decomposition, which is 41% higher than the traditional solution. It is suitable for wireless sensing scenarios with high signal-to-noise ratio and severe data loss. 2. Real-time energy efficiency optimization: Dynamic threshold strategies combined with dedicated acceleration units achieve residual calculation times of ≤30ms, meeting the 5-minute high-frequency data collection requirements, reducing average daily communication traffic by 63%, and extending node battery life to over 12 months. 3. Improved storage life: The hierarchical storage architecture increases the model parameter erase and write cycle to 1 million times, which is five times longer than the traditional solution. After verification, the parameters are solidified to non-volatile storage media to ensure long-term stable operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a block diagram of the low-power wireless sensor gateway system for base station environment monitoring and a schematic diagram of the gateway and node equipment structure of the present invention.

[0019] Figure 2 It is a matrix decomposition diagram of the present invention and a dynamic threshold and MAE-energy consumption relationship curve diagram.

[0020] Figure 3 This is a block diagram of a low-power wireless sensor gateway system for base station environment monitoring according to the present invention.

[0021] Figure 4 This is a flow chart of a method for reducing data transmission in a wireless sensor network according to the present invention.

[0022] Legend: 1. Main control chip RK3588; 2. LoRa module SX1262; 3. Independent control chip STM32L071K8U6; 4. Vitom RS-WS-N01-8 wind speed sensor; 5. Hengge HSTS-01 temperature sensor; 6. Wogan Electronics TILT-01H structural tilt sensor; 7. UFS3.1 128GB storage module; 8. 18650 lithium battery pack; 9. TPS63020 power chip; 10. Intel AX200 WiFi module; 11. LPDDR4 8GB memory; 12. 12V / 5A power supply interface; 13. 40×40mm aluminum alloy heat sink; 14. LoRa antenna interface; 15. WiFi antenna interface; 16. IPEX-4 antenna interface. DETAILED DESCRIPTION

[0023] See also Figure 1-3 As shown, in a first aspect of the present invention, a low-power wireless sensor gateway system is provided, characterized in that it includes: a main control chip 1; The main control chip 1 is connected to the communication module, sensor module, power module, storage module and power supply and heat dissipation module respectively; The main control chip uses a Rockchip RK3588 processor configured with LPDDR48GB memory 11, which is used to execute the multi-dimensional time series prediction model and the dual prediction scheme; The communication module includes a SX1262 LoRa communication submodule 2 and an IPEX-4 antenna interface 16; the sensor module includes a cluster node device of a Wittm RS-WS-N01-8 wind speed sensor 4, a Wogan Electronics TILT-01H structural tilt sensor 6, and a Hengge HSTS-01 temperature sensor 5; The power module includes an 18650 lithium battery pack 8 and a TPS63020 power chip 9; the storage module includes an LPDDR4 8GB memory 11 and a UFS3.1 128GB storage submodule 7; the power supply and heat dissipation module includes a power supply interface 12 and an aluminum alloy heat sink 13; The cluster node device establishes a data link with the gateway through the SX1262LoRa communication sub-module 2 and uploads sensor data at intervals of less than or equal to 5 minutes. The main control chip 1 dynamically adjusts the transmission threshold to 0.3 or 0.5 according to the real-time wind speed value input by the wind speed sensor 3 and the real-time temperature value input by the temperature sensor 5.

[0024] Preferably, when the data residual exceeds the transmission threshold, the coefficient matrix is ​​updated through the LPDDR48GB memory 11 and the model parameters are solidified to the UFS3.1128GB storage submodule 7.

[0025] Preferably, the main control chip 2 further includes an NPU acceleration unit, and the NPU acceleration unit is used to accelerate the calculation of the autoregressive weight matrix.

[0026] Preferably, the parameters smoothed by the timing prediction algorithm are synchronized to the cluster node device through the SX1262LoRa communication submodule 3.

[0027] Preferably, the system realizes wide voltage input and regulated power supply through 18650 lithium battery pack 8 and TPS63020 power chip 9.

[0028] As an advantage, the multi-dimensional time series prediction model is This measures the quality of non-negative matrix factorization and is used to improve the robustness of matrix factorization. Y is the matrix into which the node organizes the collected data. The training process of the multi-dimensional time series prediction model is as follows: Input low-rank matrix rank , time-delay set , learning rate , the regularization parameter , maximum number of iterations , randomly initialize all parameters, including the basis matrix , coefficient matrix and the time regularization parameter matrix ; Calculate the objective function matrix The partial derivatives of , and the basis matrix is ​​updated by the following formula and Adam optimizer, which is in the form of: ; in is the learning rate; Represents a projection operator that forces all elements to be projected onto the non-negative semi-axis. Represents the partial derivative of the loss function with respect to U; Calculate the objective function matrix The partial derivative of , and the coefficient matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ; Calculate the objective function matrix The partial derivative of , and the time regularization parameter matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ; Finally, the model parameters with validation set MAE ≤ 0.9 are stored in the UFS3.1128GB storage module 7.

[0029] In a second aspect of the present invention, a method for reducing data transmission in a wireless sensor network is provided, comprising the following steps: Step 1: Clustering is performed according to the clustering algorithm and cluster heads are elected. Nodes with strong computing power and sufficient remaining energy are selected as cluster heads. After clustering is completed, the cluster head starts to collect cluster data, including wind speed data collected by the wind speed sensor, tilt data collected by the tilt sensor, and temperature data collected by the temperature sensor, through the independent chip STM32L071K8U6 control, and sends it to the base station monitoring gateway of the SX1262LoRa communication submodule; Step 2: The base station monitoring gateway organizes the collected data into a matrix , is the number of cluster members, The amount of data collected by each node, the time series prediction model parameters obtained by training the objective function, that is, the basis matrix , coefficient matrix and the time regularization parameter matrix , the objective function is as follows: , The data will first perform maximum-minimum (Min-Max) normalization, mapping the values ​​to the range [0, 1]. Then, it will be divided into training set and validation set (8:2 ratio) according to the characteristics of the base station environment monitoring scenario data, and then input into the model for training.

[0030] Step S21: Input low-rank matrix rank , time-delay set , learning rate , the regularization parameter The time lag set defaults to the existence of short-term trends and seasonality in the time series. For example, you can set To represent the short-term dependencies of hourly monitoring data. Longer lag sets yield better algorithm performance.

[0031] The regularization parameters and learning rates are selected from the validation set by the NPU acceleration unit of the RK3588 processor (1) through the hyperparameter grid search. The hyperparameters of each model that obtain the minimum RMSE on the validation set of the original dataset are set as the final hyperparameters. Set the maximum number of iterations . Randomly initialize all matrices , , , The initialization method is to randomly generate a matrix of (0.01,0.09).

[0032] Step S22: Calculate the target function matrix The partial derivatives of , and the basis matrix is ​​updated by the following formula and Adam optimizer, which is in the form of: ; in is the learning rate; Represents a projection operator that forces all elements to be projected onto the non-negative semi-axis. Step S23: Calculate the target function pair matrix The partial derivative of , and the coefficient matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ; Step S24: Calculate the target function matrix The partial derivative of , and the coefficient matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ; The model with the smallest RMSE in the validation set is selected as the final model and stored in the UFS3.1128GB storage module 7.

[0033] Step 3: The gateway sends the matrices U, X, and W to the cluster head through the SX1262 LoRa communication submodule that supports a transmission interval of ≤5 minutes. The cluster head and the gateway confirm the synchronization prediction model. During the model parameter synchronization phase, the ACK confirmation mechanism of the SX1262 LoRa module (2) ensures that the transmission success rate is ≥99.9%, ensuring that the matrix U, X, and W parameters of the cluster head and the gateway are strictly consistent. In view of the dynamic response differences between the wind speed and structural inclination data in the base station environment, a dynamic gradient adaptation strategy for error calculation is predefined: the wind speed data fluctuates significantly due to sudden fluctuations, and η=0.3 calculated based on Formula 5 is used to capture instantaneous changes; the structural inclination is affected by continuous loads, and η=0.5 is set to balance trend tracking and energy consumption. The high-precision mode is triggered by the real-time wind speed value input by the wind speed sensor (4). The specific threshold settings are as shown in the attached figure. Figure 1 The dynamic threshold curve on the right is shown.

[0034] Figure 2 The verification curve on the right demonstrates the effect of a dynamic threshold of 0.3 ≤ Threshold ≤ 0.5 on MAE and energy consumption. Actual measured data shows that when wind speed is ≥ 10.8 m / s (force 6) and the threshold is 0.3, MAE is 0.87, and the gateway transmits an average of 41 times per day. Under normal operating conditions (wind speed < force 6), when the threshold is 0.5, MAE is 0.92, and the number of transmissions is reduced to 15, a 63% reduction compared to the traditional solution. Figure 4 The energy-saving decision module of step S04 enables the gateway to automatically switch to the energy-saving mode according to the remaining energy of the node (when the remaining energy is less than 30%), and dynamically adjusts the threshold value in combination with the wind speed sensor (4) input to achieve adaptive optimization in the base station scenario.

[0035] The model update process relies on the incremental learning unit of the RK3588 processor (main control chip 1). When the residual r ≥ the threshold, the Figure 4In the data transmission instruction of step S05, the LPDDR48GB memory 11 updates the matrix X_{t+1} within ≤50ms, and Figure 1 The UFS3.1 128GB storage submodule 7 solidification parameters are shown in . This mechanism can quickly respond to the structural vibration caused by the gust load and maintain the prediction accuracy MAE ≤ 0.9. The specific verification data is as follows Figure 2 The curve on the right is marked.

[0036] Data reduction benefits from the model's predictions, based on the matrix obtained through training 、 and the time regularization parameter matrix , predicting the future moment in a single step using autoregressive method The potential time factor vector of : ; Step 4: The cluster head uses the model to predict the current data , the data sending node will compare the predicted value with the actual observed value Compare and calculate residuals , comparing the residual r with the threshold value dynamically set according to the wind speed value input by the wind speed sensor and the temperature value input by the temperature sensor in the base station environment; Step S41: When the residual When the prediction value is less than the dynamically set threshold, the system considers the prediction reasonable, the cluster head does not send data, and the corresponding sink node obtains the prediction value obtained by the prediction model in step 2. ,The cluster head sends a beacon signal consisting of a small data packet to inform the data sender and receiver to take the model prediction value as the valid prediction value; Step S42: When the residual When the value is greater than or equal to the dynamically set threshold, the system considers the prediction to be unreasonable, and the triggered SX1262 LoRa communication submodule (2) transmits the sensor data and updates the matrix X in the LPDDR48GB memory (11).

[0037] Step 5: All data at the current moment do not meet the threshold condition, the data is transmitted and the base matrix U and coefficient matrix X are updated through the NPU acceleration unit of the main control chip; Use the model to predict current data , the data sending node will compare the predicted value with the actual observed value Compare and calculate residuals , compared with the residual with predefined error data and without updating the model; Step S51: The first case is that all data at the current moment do not meet the threshold condition. The data sending node needs to transmit data with corresponding errors exceeding the threshold, and needs to update the model bilaterally.

[0038] The incremental update method is as follows. Assume The newly collected data at the moment is , fixed-dimensional correlation feature matrix in incremental learning , historical time feature vector and the parameters of the temporal regularizer , only update Potential time factor of the moment .

[0039] The incremental learning problem is as follows, and the loss function of incremental learning is: ; in It is a newly collected sample; is the latent dimension factor matrix learned on the old samples; It is to be learned The potential time factor vector corresponding to the moment; is the time regularization term The predicted value of the potential time factor at time .

[0040] Use the projected gradient descent method to solve the equation, where The gradient is as follows: ; Compare and , select the one with the smallest RMSE The learned latent temporal factor vector The future will be carried out through formula (6) Prediction of temporal characteristics of moments.

[0041] Step S52: The second situation is that only part of the data at the current moment does not meet the threshold condition. The data sending node only needs to transmit the data with the corresponding error exceeding the threshold, and the model is not updated in real time.

[0042] Step 6: Synchronize the prediction model parameters through the incremental learning unit of the gateway, where matrices U and W are solidified to the UFS3.1128GB storage submodule, jump to step 4, and perform the next iteration.

[0043] Example: S01: Clustering and cluster head election: Execute the clustering algorithm and preferentially select the node equipped with Rockchip RK3588 processor (main control chip 1) as the cluster head. The cluster head collects wind speed data collected by the Wittm RS-WS-N01-8 wind speed sensor 4, tilt data collected by the Wogan Electronics TILT-01H structural tilt sensor 6, and temperature data collected by the Hengge HSTS-01 temperature sensor 5 through the IPEX-4 antenna interface (16), and uploads it to the gateway via the SX1262 LoRa communication submodule 2.

[0044] S02: Matrix Construction and Model Training: The gateway organizes data into matrix Y and stores the basis matrix U, coefficient matrix X, and autoregressive weights W in 8GB of LPDDR4 memory 11. A robust NMF model is used, with an L2,1 norm as the loss function and an integrated autoregressive regularization term. The Adam optimizer is accelerated by the NPU to update parameters. The goal is to minimize the validation set MAE ≤ 0.9 before being stored in UFS 3.1 128GB storage 7.

[0045] S03: Model parameter synchronization: The gateway sends U, X, and W to the cluster head through the SX1262LoRa communication sub-module 2, and uses the ACK confirmation mechanism to ensure parameter consistency.

[0046] S04: Dynamic threshold decision: The cluster head calculates the residual r of the comparison between the predicted value Y^t+1 and the actual value Yt+1, and dynamically generates a transmission threshold according to the real-time wind speed value input by the wind speed sensor 4 and the real-time temperature value input by the temperature sensor 5.

[0047] S05: Data transmission control: If r exceeds the threshold, the SX1262 module is triggered to transmit the original data (interval ≤ 5 minutes) and the parameters are refreshed within 50ms of the LPDDR4 memory by incrementally updating Xt+1.

[0048] S06: Incremental Learning and Parameter Solidification: The incremental learning unit fixes U and W and optimizes only Xt+1. The updated X is verified by the NPU and persisted to the UFS 3.1 storage module 7 to ensure that the MAE of structural condition monitoring under gust loads is ≤ 0.9.

[0049] S07: Model prediction iteration: Return to step S04 to make predictions for the next moment, and generate X^t+2 closed-loop optimization.

[0050] S08: Hardware collaboration mechanism: On the gateway side, the RK3588 processor 1 uses the LoRa protocol stack, the business logic module integrates dynamic threshold decision-making, and the LWIP protocol stack supports remote configuration. On the node side, the sensor driver module collects data in real time and compresses and transmits it through the LoRa protocol stack.

[0051] In addition, for technical details not fully described in this embodiment, please refer to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0052] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0053] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0054] Through the above description of the embodiments, those skilled in the art will clearly understand that the methods of the above embodiments can be implemented using software plus the necessary general-purpose hardware platform. Of course, hardware can also be used, but in many cases the former is a more preferred implementation method. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, a magnetic disk, or an optical disk) and includes a number of instructions for enabling a terminal device (such as a mobile phone, computer, server, air conditioner, or network device) to execute the methods described in the various embodiments of the present invention.

[0055] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A low-power wireless sensor gateway system, characterized in that: include: Main control chip (1); The main control chip (1) is respectively connected to the communication module, the sensor module, the power module, the storage module and the power supply and heat dissipation module; The main control chip (1) uses a Rockchip RK3588 processor configured with LPDDR48GB memory (11) to execute a multi-dimensional time series prediction model and a dual prediction scheme; The communication module includes a SX1262 LoRa communication submodule (2) and an IPEX-4 antenna interface (16); the sensor module includes a cluster node device of a Wittm RS-WS-N01-8 wind speed sensor (4), a Wogan Electronics TILT-01H structural tilt sensor (6) and a Hengge HSTS-01 temperature sensor (5); The power module includes an 18650 lithium battery pack (8) and a TPS63020 power chip (9); the storage module includes an LPDDR4 8GB memory (11) and a UFS3.1 128GB storage submodule (7); the power supply and heat dissipation module includes a power supply interface (12) and an aluminum alloy heat sink (13); The cluster node device establishes a data link with the gateway via the SX1262LoRa communication submodule (2) and uploads sensor data at intervals of less than or equal to 5 minutes. The main control chip (1) dynamically adjusts the transmission threshold to 0.3 or 0.5 according to the real-time wind speed value input by the wind speed sensor (4) and the real-time temperature value input by the temperature sensor (5).

2. A low-power wireless sensor gateway system according to claim 1, characterized in that: When the data residual exceeds the transmission threshold, the coefficient matrix is ​​updated through the LPDDR48GB memory (11) and the model parameters are solidified to the UFS3.1128GB storage submodule (7).

3. A low-power wireless sensor gateway system according to claim 1, characterized in that: The main control chip (1) further comprises an NPU acceleration unit, and the NPU acceleration unit is used to accelerate the calculation of the autoregressive weight matrix.

4. A low-power wireless sensor gateway system according to claim 1, characterized in that: The parameters smoothed by the timing prediction algorithm are synchronized to the cluster node devices through the SX1262LoRa communication submodule (2).

5. A low-power wireless sensor gateway system according to claim 1, characterized in that: The system realizes wide voltage input and regulated power supply through an 18650 lithium battery pack (8) and a TPS63020 power chip (9).

6. A low-power wireless sensor gateway system according to claim 1, characterized in that: The multidimensional time series prediction model is constructed by the norm This measures the quality of non-negative matrix factorization and is used to improve the robustness of matrix factorization. Y is the matrix into which the node organizes the collected data. The training process of the multi-dimensional time series prediction model is as follows: Input low-rank matrix rank , time-delay set , learning rate , the regularization parameter , maximum number of iterations , randomly initialize all parameters, including the basis matrix , coefficient matrix and the time regularization parameter matrix ; Calculate the objective function matrix The partial derivatives of , and the basis matrix is ​​updated by the following formula and Adam optimizer, which is in the form of: ； in is the learning rate; Represents a projection operator that forces all elements to be projected onto the non-negative semi-axis. Represents the partial derivative of the loss function with respect to U; Calculate the objective function matrix The partial derivative of , and the coefficient matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ； Calculate the objective function matrix The partial derivative of , and the time regularization parameter matrix is ​​updated through the above formula and Adam optimizer, which is in the form of: ； Finally, the model parameters with validation set MAE ≤ 0.9 are stored in the UFS3.1128GB storage module (7).

7. A method for reducing data transmission in a wireless sensor network, characterized in that: The following steps are involved: Step 1: Clustering is performed according to the clustering algorithm and cluster heads are selected. Nodes with strong computing power and sufficient remaining energy are selected as cluster heads. After clustering is completed, the cluster heads start to collect cluster data including wind speed data collected by the wind speed sensor (4), tilt data collected by the tilt sensor (6) and temperature data collected by the temperature sensor (5) through the independent chip STM32L071K8U6 (3), and send them to the base station monitoring gateway of the SX1262LoRa communication submodule (2); Step 2: The base station monitoring gateway organizes the collected data into a matrix , is the number of cluster members, The amount of data collected by each node, the time series prediction model parameters obtained by training the objective function, that is, the basis matrix , coefficient matrix and the time regularization parameter matrix , The objective function is specifically as follows: ； Step 3: The gateway sends the matrices U, X, and W to the cluster head through the SX1262LoRa communication submodule (2). The cluster head and the gateway confirm the synchronization prediction model. Step 4: The cluster head uses the model to predict the current data , the data sending node will compare the predicted value with the actual observed value Compare and calculate residuals , comparing the residual r with a threshold value dynamically set according to the wind speed value input by the wind speed sensor (4) and the temperature value input by the temperature sensor (5) in the base station environment; Step 5: All data at the current moment do not meet the threshold condition, transmit the data and update the base matrix U and coefficient matrix X through the NPU acceleration unit of the main control chip (1); Use the model to predict current data , the data sending node will compare the predicted value with the actual observed value Compare and calculate residuals , compared with the residual with predefined error data and without updating the model; Step 6: Synchronize the prediction model parameters through the incremental learning unit of the gateway, where the matrices U and W are solidified to the UFS3.1128GB storage submodule (7), jump to step 4, and perform the next iteration.

8. The method for reducing data transmission in a wireless sensor network according to claim 7, characterized in that: In step 4, the method further includes: Step S41: When the residual When the prediction value is less than the dynamically set threshold, the system considers the prediction reasonable, the cluster head does not send data, and the corresponding sink node obtains the prediction value obtained by the prediction model in step 2. ,The cluster head sends a beacon signal consisting of a small data packet to inform the data sender and receiver to take the model prediction value as the valid prediction value; Step S42: When the residual When the value is greater than or equal to the dynamically set threshold, the system considers the prediction to be unreasonable, and the triggered SX1262 LoRa communication submodule (2) transmits the sensor data and updates the matrix X in the LPDDR48GB memory (11).

Images