A system and method for dynamically monitoring salt content in saline-alkali soil based on narrowband internet of things

Abstract

The application relates to the field of monitoring, in particular to a saline-alkali soil salt dynamic monitoring system and method based on narrowband Internet of Things, which comprises an LSTM model, a drift trend is predicted through the LSTM model, and data is dynamically corrected; linear interpolation weights are dynamically generated according to timestamp difference values, non-uniform sampling data is aligned to fixed time intervals, and the problems of asynchronous sampling time and time interval jitter of original data are solved; a drift sensitive factor is added to an input gate of the LSTM model, a drift rate is statistically adjusted through historical data, and the input gate is dynamically adjusted; a salt diffusion prediction module constructs a salt diffusion model of a finite difference equation according to meteorological data, predicts the space-time distribution of salt in a future period, dynamically adjusts the mechanism of the drift sensitive factor of the input gate, changes the input gate value when the drift rate of the sensor changes, allows more real-time data to participate in correction, and establishes a cross-time-step virtual gradient linkage mechanism of a deep LSTM.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of monitoring, and particularly relates to a saline-alkali soil salt content dynamic monitoring system and method based on narrowband Internet of Things. BACKGROUND

[0002] Narrowband Internet of Things (NB-IoT) has significant technical adaptability and practical application value in the field of saline-alkali soil salt content dynamic monitoring. Saline-alkali soils are mostly distributed in arid and semi-arid regions (such as the saline-alkali regions in northwest and north China), where the terrain is open but sparsely populated, and traditional wired networks are difficult to cover, so wireless communication technology is needed to achieve remote monitoring.

[0003] Monitoring sensors (such as soil salt content sensors and water content sensors) are usually deployed in the wild, and it is difficult to frequently replace batteries or external power sources, so the devices need to support ultra-long standby (several years).

[0004] Saline-alkali soil treatment requires grid deployment of a large number of sensors (possibly tens of nodes per square kilometer), and a single network is required to support a large number of device connections.

[0005] NB-IoT has strong signal penetration (20 dB stronger than GSM), can cover remote areas such as deserts and farmland, and even shallow underground layers (such as buried soil sensors), does not require additional relay equipment, and solves the problem of "last mile" networking in saline-alkali soils.

[0006] Relying on existing operator base stations (such as 700 MHz low frequency bands), in areas without fiber or Wi-Fi coverage, data can still be transmitted back to the cloud platform through cellular networks.

[0007] Currently, narrowband Internet of Things (NB-IoT) has practical applications in the field of saline-alkali soil salt content dynamic monitoring. For example, the smart agriculture project of Huawei and Qingdao Seawater Rice Research and Development Center: soil salt content, water content, and pH value sensors are deployed in saline-alkali soils, data is transmitted in real time to the "Agricultural Soil Cloud Platform" through NB-IoT modules, and AI analysis is combined to achieve precise irrigation, fertilization, and pest control; the 5G smart agriculture project of China Mobile and ZTE in Jilin saline-alkali soil: NB-IoT sensors are deployed to monitor soil pH value and acid-base balance (ESR), combined with 5G network to achieve high-speed data backhaul, driving intelligent irrigation systems and unmanned agricultural machinery operation; the monitoring platform of the National Salt-Tolerant Rice Technology Innovation Center Northeast Center: a "cold saline-alkali soil monitoring platform" is built, integrating an NB-IoT sensor network to collect soil salt content and weather data in real time, and combining them with rice phenotype data to form a multidisciplinary data chain, etc.

[0008] However, distributed sensors need to synchronize sampling time to ensure data consistency, but in the eDRX (extended discontinuous reception) mode of NB-IoT, the sleep cycles of each node are not synchronized, which may cause the data collection time difference in the same area to exceed 10 minutes, affecting the accuracy of the salt space-time distribution analysis. After long-term use of the sensor in a high-salt environment, the zero point drifts due to the polarization effect of the electrode (e.g., an offset of 0.05 dS / m per year), and the NB-IoT system lacks an automatic calibration mechanism (e.g., manual calibration requires on-site operation), and the cumulative error may exceed 10%, affecting the salt trend analysis. SUMMARY

[0009] The purpose of the present application is to provide a salt and alkali soil salt dynamic monitoring system and method based on narrowband Internet of Things, to solve the problems raised in the background art.

[0010] To solve the above technical problems, the present application provides the following technical solutions:

[0011] A salt and alkali soil salt dynamic monitoring system based on narrowband Internet of Things, comprising

[0012] A satellite time module for synchronizing all node clocks and setting a unified sampling period;

[0013] A time synchronization server for sending a synchronization broadcast instruction once every certain period, and the node corrects the local clock after receiving it;

[0014] A timestamp error management module for requesting a timestamp from the base station if the node misses the synchronization instruction due to sleep, and completing the data timestamp error after waking up;

[0015] An LSTM model management module for establishing an LSTM model, predicting the drift trend through the LSTM model, and dynamically correcting the data;

[0016] And before inputting the LSTM model, linear interpolation weights are dynamically generated according to the timestamp difference to align non-uniformly sampled data to fixed time intervals, to solve the problem of non-synchronized sampling time and time interval jitter of original data; and a drift sensitive factor is added to the LSTM model input gate to dynamically adjust the opening and closing of the input gate by statistically analyzing the drift rate of historical data;

[0017] A salt diffusion prediction module for constructing a salt diffusion model of finite difference equation based on meteorological data to predict the future salt space-time distribution for a certain period.

[0018] Further, it further comprises a sleep cycle module for turning off the sensor circuit during the eDRX sleep cycle and only maintaining the RTC clock running; after waking up, collecting data and transmitting it through NB-IoT single uplink transmission;

[0019] It also includes a data processing module for removing outliers based on the Grubbs test and filling in missing data using spatiotemporal neighborhood interpolation.

[0020] Furthermore, the LSTM model management module is also used to superimpose a sinusoidal modulation term into the forget gate activation function of the LSTM model to enhance the memory retention ability of seasonal drift patterns, targeting the annual or seasonal periodicity of electrode drift.

[0021] Furthermore, the LSTM model management module is also used to add virtual gradient links across time steps in the deep LSTM of the LSTM model, forcing the model to retain long-term drift characteristics in order to capture drift patterns across the device lifecycle.

[0022] Furthermore, linear interpolation weights are dynamically generated based on the timestamp differences to align non-uniformly sampled data to a fixed time interval. Specifically, this includes determining the target time point T. k Actual sampling points before and after (t) j ,S j ) and (t j+1 ,S j+1 ), calculate the time interval Δt1=T k -t j and Δt2=t j+1 -T k The weights are inversely proportional to the time percentage: w j =Δt2 / (Δt1+Δt2),w j+1 =Δt1 / (Δt1+Δt2), the aligned value is S k =w j ⋅S j +w j+1 ⋅S j+1 .

[0023] Furthermore, a drift-sensitive factor is added to the input gate of the LSTM model, and the opening and closing of the input gate are dynamically adjusted by statistically analyzing the drift rate using historical data, including:

[0024] A sliding time window is used to extract historical drift data within the window and calculate the average drift rate within the window: drift rate = (drift amount at the end of the window - drift amount at the beginning of the window) / window duration; the drift rate is converted into a dimensionless factor of 0.1-0.9, i.e., the drift sensitivity factor β, through normalization mapping;

[0025] By introducing a drift sensitivity factor β after a linear combination of traditional input gates, the output value of the input gate at time t is: i t =σ((W i ⋅[h t−1 ,x t ]+b i )×(1+β)), W iis the weight matrix of the input gate, σ is the Sigmoid activation function, h t−1 is the hidden state of the previous time, x t is the input vector of the current time, i.e., the sensor measurement value, b i is the bias vector of the LSTM input gate.

[0026] The input gate value is dynamically adjusted by the drift sensitivity factor β.

[0027] Further, a salt diffusion model based on finite difference equation is constructed according to the meteorological data to predict the spatial and temporal distribution of salt in the future period, including:

[0028] Based on the water and salt transport theory, the soil salt dynamic is regarded as a convection and diffusion process on a two-dimensional vertical section. The initial salt value is the sensor data corrected by the LSTM model. The model is initialized at 0 o'clock every day according to the future 72-hour hourly meteorological forecast data. The data is accessed through the NB-IoT meteorological API or satellite inversion;

[0029] Rainfall is the water input of the top layer of soil, and evaporation is the output of the top layer. When the rainfall exceeds the saturated water holding capacity of the soil, the excess water is discharged in the form of runoff. When the evaporation exceeds the current water content, the current water content value is taken to avoid negative water content. The water flux is calculated by Darcy's law. The permeability is preset according to the soil type, and the convection and diffusion terms are discretized into the flux difference of adjacent grids by using the explicit finite difference method. The salt convection flux between adjacent subgrids in the horizontal direction is equal to the product of the water flux and the salt concentration. The diffusion flux is proportional to the concentration gradient and the diffusion coefficient. The bottom layer salt only exchanges with the upper layer by diffusion, without external water input. The migration rate of the bottom layer salt under the action of gravity is set to 15% of the top layer.

[0030] Further, the surface layer salt concentration is determined by the measured value and the model calculation, and is updated every hour.

[0031] The bottom layer is set as a closed boundary, consistent with the monitoring depth of the sensor.

[0032] The salt change of each grid is calculated iteratively every hour, considering the lateral infiltration within 5% of adjacent grids.

[0033] After each round of calculation, the real-time monitoring data corrected by LSTM is used to calibrate the model. If the error is out of limit, the local grid soil porosity and root water absorption rate are automatically adjusted.

[0034] Then a future salt prediction matrix is generated, marking the high-risk areas where the salt value is greater than or equal to the threshold value.

[0035] Further, for the annual or seasonal periodicity of electrode drift, a sinusoidal modulation term is superimposed in the LSTM model forgetting gate activation function to enhance the memory retention capability of seasonal drift rules, including: dividing a year into 12 monthly periods or 4 quarterly periods, taking the month m or the quarter q as the input variable of the sinusoidal function, generating a sinusoidal curve with a period of 12 months or 3 months, and the expression of the modulation term is sin(2πm / 12) or sin(2πq / 4), the amplitude of which is automatically adjusted according to the seasonal fluctuation amplitude of the historical drift data, and the output of the forgetting gate is obtained by multiplying the output value of the traditional sigmoid function and the sinusoidal modulation term, when it is in the summer high temperature period, the value of the sinusoidal term is close to 1, the forgetting gate control value is reduced, and the forgetting of the high drift data in this period is reduced; when it is in winter, the value of the sinusoidal term is close to 0, the forgetting gate control value is increased, and the filtering of the low drift data is enhanced.

[0036] Further, a cross-time step virtual gradient link is added to the deep LSTM of the LSTM model to force the model to retain long-term drift features to capture the drift mode across the life cycle of the device, including: when the sensor node runs for more than a threshold, the hidden state of the key time node in the whole life cycle of the device is automatically extracted, and these historical states are spliced with the current time data through virtual connection and then input into the deep LSTM unit, and the link weight is dynamically adjusted according to the running time of the device, the recent state adopts linear weight, and the early state adopts exponential decay weight.

[0037] The application discloses an electronic device, comprising:

[0038] at least one processor; and

[0039] a memory connected in communication with the at least one processor; wherein

[0040] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the module functions of the above-mentioned salt and alkali soil salt dynamic monitoring system based on narrowband Internet of Things.

[0041] Beneficial effect: through the drift sensitive factor dynamic adjustment mechanism of the input gate, when the sensor drift rate changes, the input gate control value changes, and more real-time data is allowed to participate in correction.

[0042] The cross-time step virtual gradient link mechanism of the deep LSTM captures the drift trend of the whole life cycle of the device, and the sinusoidal modulation term strengthens the memory of the seasonal rules, so that the annual cumulative drift error is reduced.

[0043] Based on the physical model of the finite difference equation, the real-time data corrected by the LSTM are fused to realize the prediction of the future salt distribution, identify the high-risk area of the salt value in advance, and provide decision basis in advance for precise irrigation and soil improvement. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A module schematic diagram of a salt and alkali ground salt dynamic monitoring system based on narrowband Internet of Things. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] The present application discloses a salt and alkali ground salt dynamic monitoring system based on narrowband Internet of Things, which can be realized in the form of hardware and / or software. The present application discloses a salt and alkali ground salt dynamic monitoring system based on narrowband Internet of Things, preferably a software application, which can be installed in terminal or server electronic devices.

[0047] As Figure 1 A salt and alkali ground salt dynamic monitoring system based on narrowband Internet of Things includes:

[0048] The satellite time service module 100 is used for synchronizing all node clocks and setting a unified sampling period (such as once every 10 minutes).

[0049] The time synchronization server 101 is used for sending a synchronization broadcast instruction once every certain period (24 hours), and the node receives and calibrates the local clock after the synchronization broadcast instruction (accuracy ±1ms).

[0050] The timestamp error management module 102 is used for if the node misses the synchronization instruction due to hibernation, actively requesting a timestamp from the base station after waking up, and completing the data timestamp error.

[0051] The LSTM model management module 103 is used for establishing an LSTM model, predicting the drift trend through the LSTM model, and dynamically correcting the data.

[0052] And the LSTM model management module is used for dynamically generating a linear interpolation weight according to the timestamp difference before the LSTM model input, aligning the non-uniform sampling data to a fixed time interval, so as to solve the problem of asynchronous sampling time and time interval jitter of original data.

[0053] And the LSTM model management module is used for increasing a drift sensitive factor in the LSTM model input gate, dynamically adjusting the opening and closing of the input gate through the historical data statistics drift rate.

[0054] The hibernation period module 104 is used for closing the sensor circuit of the sensor and maintaining only the RTC clock during the eDRX hibernation period. After waking up, data is collected and transmitted through a single NB-IoT uplink transmission (data packet ≤1KB).

[0055] The data processing module 105 is used for removing abnormal values based on Grubbs test and filling missing data through spatial and temporal neighborhood interpolation method.

[0056] The salt diffusion prediction module 106 is used for constructing a salt diffusion model of finite difference equation according to meteorological data (rainfall, evaporation) to predict the spatial and temporal distribution of salt in a certain period in the future.

[0057] Preferably, the LSTM model management module is used for superimposing a sinusoidal modulation term in the LSTM model forgetting gate activation function for the annual or seasonal periodicity of electrode drift, and enhancing the memory retention capability for seasonal drift rules.

[0058] Preferably, the LSTM model management module is used for adding a cross-time step virtual gradient link in the deep LSTM of the LSTM model, forcing the model to retain long-term drift features to capture drift patterns across the life cycle of the device.

[0059] In the implementation, for the satellite timekeeping module, a unified sampling period (such as every 10 minutes) is set to synchronize all node clocks. When the sensor node is powered on for the first time after deployment, the built-in GPS / Beidou satellite timekeeping module (hardware-level integration) is automatically activated, the current UTC time (accuracy ≤100ns) is calculated by receiving signals from at least 4 satellites, and the local time zone time (such as Beijing time) is converted.

[0060] The node takes the calculated time as the initial clock reference, sends an initialization signal (including node ID, latitude and longitude coordinates, and initial timestamp) to the cloud platform through the NB-IoT module, and the cloud records the initial time state of the node.

[0061] The cloud platform presets a global sampling period (such as 10 minutes / time), broadcasts configuration instructions to all nodes through the NB-IoT core network. After receiving the instructions, the node writes the sampling period parameters into the timing register of the local RTC (real-time clock) module, and activates the periodic wake-up mechanism (such as triggering a data collection task every 10 minutes).

[0062] The time synchronization server deployed in the cloud is synchronized with the atomic clock of the national time center (such as the National Time Service Center of the Chinese Academy of Sciences) to generate a high-precision time reference (error ≤1μs) as the source of clock calibration for the entire network.

[0063] In implementation, for time synchronization server, send synchronization broadcast instruction once every certain period (24 hours), node receives and calibrates local clock (accuracy ±1 ms). Server generates synchronization broadcast instruction (including current timestamp, check code) once every 24 hours from UTC time 0, sends to all base stations in monitoring area through NB-IoT core network, and base station covers sensor nodes in coverage area in broadcast form.

[0064] The node in wake-up state receives the synchronization instruction, extracts the timestamp (accurate to ms) in the instruction, compares with the local RTC clock, and calculates the time deviation (such as local clock fast / slow X ms).

[0065] The node corrects the local clock through the hardware level clock adjustment interface (such as the frequency control register of the RTC module) to make the deviation converge within ±1 ms, and records the calibration time and deviation value, and returns to the cloud log system through NB-IoT.

[0066] In implementation, for timestamp error management module, if the node misses the synchronization instruction due to sleep, actively request timestamp from base station after wake-up to complete data timestamp error. The node only maintains RTC (real-time clock) module running during eDRX sleep period, RTC uses independent low-power power supply (such as button cell) for power supply, clock accuracy ≤±5ppm (i.e. daily error ≤0.432 seconds).

[0067] During sleep, the node closes the sensor circuit, NB-IoT communication module and satellite timing module, and only keeps RTC timer for wake-up trigger (such as wake-up every 10 minutes to collect data).

[0068] After the node wakes up according to the preset sampling period (such as 10 minutes), it first queries the "synchronization status flag bit" of the local RTC clock (the flag bit is automatically set by hardware after successfully receiving the time synchronization instruction).

[0069] If the flag bit is "unsynchronized" (i.e. no synchronization broadcast instruction in the last 24 hours is received), it is determined that the synchronization instruction is missed, and the active timestamp request process is triggered.

[0070] The node activates the NB-IoT communication module, sends "timestamp request frame" to the base station through random access channel (RACH), including: node ID (unique identification), local RTC current timestamp (as reference benchmark); request type is time synchronization completion. After the base station receives the request, it is forwarded to the cloud time synchronization server through NB-IoT core network, and the server extracts its high-precision timestamp (error ≤1 μs), and encapsulates it as "timestamp response frame" to return to the node.

[0071] After the node receives the response frame, the server timestamp (T_server) and the local RTC timestamp (T_local) are parsed, and the time deviation ΔT=T_server-T_local is calculated.

[0072] If the absolute value of ΔT>10ms (exceeding the maximum time error allowed by the system), calibration is performed through the following steps:

[0073] Through the frequency control register of the RTC module, the load capacitance of the clock crystal oscillator (or the digital calibration word) is adjusted to align the RTC frequency to the server time, ensuring that the subsequent timing accuracy is ≤±1ms / day; the local clock is directly forced to update to T_server, and the calibration time and deviation value are recorded (e.g., "2025-05-2010:00:00, deviation +50ms").

[0074] In implementation, for the LSTM model management module, an LSTM model is established, and drift trends are predicted through the LSTM model to dynamically correct data. LSTM realizes selective processing of information through three gates (forget gate, input gate, and output gate), and each gate is composed of a sigmoid layer and a point multiplication operation (the sigmoid output is a value between 0 and 1, indicating the proportion of information allowed to pass). The forget gate (ForgetGate) input is the current input and the previous hidden state, which is output after the sigmoid layer. The input gate (InputGate) includes two parts:

[0075] The sigmoid layer (input gate) determines which new information needs to be updated to the cell state.

[0076] The tanh layer generates a candidate update value for updating the cell state.

[0077] The final update of the cell state is determined by the forget gate and the input gate.

[0078] The output gate (OutputGate) is output after the sigmoid layer, and then multiplied by the cell state processed by the tanh to obtain the current hidden state, which determines which information in the cell state will be used as the current output.

[0079] In implementation, for the LSTM model management module, it is also used to dynamically generate linear interpolation weights according to the timestamp difference before the LSTM model input, aligning non-uniformly sampled data to fixed time intervals to solve the problem of asynchronous sampling time and time interval jitter in original data. Specifically, it includes:

[0080] First, all sensor nodes achieve initial clock synchronization (accuracy ≤100ns) through satellite timing modules (GPS / BeiDou), and are calibrated daily by the time synchronization server (error ≤1ms) to ensure that the timestamp error of the entire network is controlled within 10ms.

[0081] Based on the hourly Beijing time, a fixed sampling interval of 10 minutes is set to form a standardized time series (such as 09:00, 09:10, 09:20, etc.), requiring all node data to align to this time grid.

[0082] Then, for non-uniformly sampled data at any node {(t) i ,S i (e.g., if the actual sampling times are 09:05:30 and 09:17:15), they need to be mapped to the nearest fixed time point (e.g., 09:10:00). The specific steps are as follows:

[0083] Adjacent time point matching: Determine the target time point T k (e.g., 09:10:00) actual sampling points (t) before and after j ,S j ) and (t j+1 ,S j+1 (e.g., 09:05:30 and 09:17:15).

[0084] Time difference weighting: Calculate the time interval Δt1=T k -t j (4.5 minutes) and Δt2=t j+1 -T k (7.25 minutes), weights are inversely distributed according to time percentage: w j =Δt2 / (Δt1+Δt2),w j+1 =Δt1 / (Δt1+Δt2;Example: In the above scenario, w j =7.25 / (4.5+7.25)≈0.62, w j+1 ≈0.38, the aligned value is S k =0.62⋅S j +0.38⋅S j+1 .

[0085] In implementation, the LSTM model management module also incorporates a drift-sensitive factor into the LSTM model input gates, dynamically adjusting the gate opening and closing based on historical data and statistical drift rate analysis. Specifically:

[0086] The sliding time window (window length of 12 sampling periods, such as 12 days) is used to extract the historical drift data (i.e. the electrode zero point offset predicted by the LSTM model) within the window. The average drift rate within the window is calculated: drift rate = (drift at the end of the window - drift at the beginning of the window) / window duration; for example, if the drift increases from 0.02 dS / m to 0.05 dS / m in 12 days, the rate is 0.0025 dS / m / day.

[0087] The rate is converted to a dimensionless factor of 0.1-0.9, i.e. drift sensitivity factor (denoted as β) through normalization mapping, where 0.1 corresponds to low drift rate (such as winter stable period), and 0.9 corresponds to high drift rate (such as summer high temperature polarization period).

[0088] The drift sensitivity factor β is introduced after the linear combination of the traditional input gate, the formula is: i t =σ((W i ⋅[h t−1 ,x t ]+b i )×(1+β)),W i is the weight matrix of the input gate, σ is the Sigmoid activation function, h t−1 is the hidden state at the last time (historical information),

[0089] x t is the input vector at the current time (such as sensor measurement), i t represents the output value of the input gate (Input Gate) at time t, β is the drift sensitivity factor, b i is the bias vector of the LSTM input gate;

[0090] When the drift rate is high (such as β = 0.8), the input gate value is amplified, allowing more current data to participate in cell state update, responding quickly to short-term and severe drift. When the drift rate is low (such as β = 0.2), the input gate value is reduced, inhibiting redundant data flow, and strengthening the memory of long-term trends.

[0091] By embedding the drift sensitivity factor in the input gate, the LSTM model realizes the "rate perception-gate regulation-dynamic correction" closed loop for the sensor drift in saline-alkali soil, so that the electrode drift prediction error is controlled within a very small range, effectively solving the cumulative error problem of long-term monitoring in high salt environment, and providing more reliable data support for saline-alkali soil treatment.

[0092] In implementation, for the sleep cycle module, the sensor is used to turn off the sensor circuit in the eDRX sleep cycle, only maintain the RTC clock running; collect data after waking up and transmit it through NB-IoT single uplink (data packet ≤1 KB). In the eDRX sleep cycle, the power management module cuts off the power supply circuit of the soil salinity sensor, the moisture sensor and the pH sensor, and at the same time, the RF transceiver and the baseband processing unit of the NB-IoT communication module are turned off, only the RTC (real-time clock) module is provided with an independent low-power power supply (such as a button cell), and the clock is maintained running (accuracy ≤±5ppm / day). The RTC module triggers the wake-up mechanism according to the preset sampling period (such as 10 minutes), and when the wake-up time arrives, the sensor node activates the sensor circuit and the NB-IoT module in turn: first, complete the hardware level automatic calibration through the built-in redundant reference electrode, collect soil salinity, moisture and pH value data, and synchronously obtain the current timestamp of the RTC; then, encapsulate the data, node ID, latitude and longitude coordinates and other information into a data packet of ≤1 KB, and transmit it to the cloud platform through the NB-IoT module in a single uplink at a frequency of 700 MHz. After transmission is completed, the node immediately enters the next eDRX sleep cycle, and repeats the above low-power cycle. If the unsynchronized time mark is detected after waking up (i.e. no recent 24-hour synchronization instruction is received), the timestamp error is requested from the base station first, and then the data collection task is performed, to ensure that the time accuracy of the uploaded data is ≤10 ms.

[0093] In implementation, for the data processing module, abnormal values are removed based on Grubbs test, and missing data is filled by spatial and temporal neighborhood interpolation method. When removing abnormal values based on Grubbs test, for the grid-deployed sensor nodes (such as one node deployed every 100 meters in a 1km×1km grid), the time series data (such as the salinity value collected every 10 minutes) of each node is tested point by point: first, calculate the mean value and standard deviation of the node in a certain period (such as 144 data points in 24 hours), set the significance level α=0.05, and determine the corresponding threshold value through the Grubbs test critical value table. Calculate the deviation multiple (Z value) of each data point from the mean value, if the Z value exceeds the critical value (such as ±2.576), it is determined as an abnormal value.

[0094] For example, the salinity value of a certain node at a certain time in summer is 6.1 dS / m, which is much higher than the mean value of 3.5 dS / m of the node at the same period and the standard deviation is 0.8 dS / m, the Z value is calculated as 3.25, which exceeds the critical value, combined with the hardware calibration record, it is determined as an abnormal value caused by temporary interference of the electrode, and is removed.

[0095] The missing data is filled by the space-time neighborhood interpolation method, including, if the data of a node at a certain time (such as 15:30) is missing, the salt content values of the four nearest neighboring nodes (east, south, west and north directions, interval 50-100 meters) in the same time in the grid where the node is located are obtained, and the historical values of the node at the adjacent time (15:20 and 15:40) are obtained. The spatial weight is calculated in inverse proportion to the square of the distance between nodes (such as 50 meters weight 0.4, 80 meters weight 0.2), and the time weight is calculated in inverse proportion to the interval minutes (such as interval 10 minutes weight 0.3). After normalizing the space and time weights, the missing values are generated by weighted average of the neighborhood data.

[0096] For example, the value of the node 50 meters east is 3.6 dS / m (weight 0.35), and the value of the node 80 meters south is 3.4 dS / m (weight 0.2). The value of the node at 15:20 is 3.5 dS / m (weight 0.25), and the value of the node at 15:40 is 3.7 dS / m (weight 0.2). The filling value is 3.6 x 0.35 + 3.4 x 0.2 + 3.5 x 0.25 + 3.7 x 0.2 = 3.54 dS / m. At the same time, the result is corrected in combination with the evaporation amount data of the region on the same day (such as 5 mm), to ensure that the filling value meets the physical law of salt concentration with evaporation.

[0097] In the implementation, for the salt diffusion prediction module, a salt diffusion model for constructing a finite difference equation according to meteorological data (rainfall, evaporation) is used to predict the salt space-time distribution in the future period. Specifically:

[0098] When constructing the salt diffusion model according to the meteorological data, on the basis of the 1km x 1km grid deployed by the system, each grid is divided into 10 x 10 sub-grids in the horizontal direction at an interval of 100m, and divided into 0-10cm and 10-20cm layers in the vertical direction according to the sensor depth, to form a spatial grid matching the node distribution.

[0099] The model is based on the theory of water and salt transport, and the dynamic soil salt content is regarded as a convection and diffusion process on a two-dimensional vertical section. The specific implementation is as follows:

[0100] The horizontal step is 100m (corresponding to the node interval), the vertical step is 10cm, and the time step is set to 1 hour (consistent with the meteorological data update frequency). The initial salt content value is the sensor data corrected by the LSTM model (error ≤3%), and the model is initialized at 0 o'clock every day according to the future 72-hour hourly meteorological forecast data (rainfall, evaporation). The data is obtained by accessing the meteorological API or satellite inversion through NB-IoT.

[0101] Rainfall is input as water into the top layer of soil (0-10 cm), and evaporation is output from the top layer. When the rainfall exceeds the saturated water-holding capacity of the soil (preset at 20 mm per layer), the excess water is discharged as runoff; when the evaporation exceeds the current layer's water content, the current water content value is taken to avoid negative water content. The water flux is calculated by Darcy's law, combined with the preset permeability of the soil type (e.g., sandy soil 0.5 cm / h, clay soil 0.1 cm / h).

[0102] The explicit finite difference method is used to discretize the convection term and the diffusion term into the difference of adjacent grid fluxes. The salt convection flux between adjacent subgrids in the horizontal direction is equal to the product of water flux and salt concentration, and the diffusion flux is proportional to the concentration gradient and diffusion coefficient (the diffusion coefficient is automatically increased by 25% in high salt environments to simulate ion migration acceleration). In the vertical direction, the salt in the bottom layer (10-20 cm) only exchanges with the upper layer through diffusion, without external water input, and the salt migration rate in the bottom layer under gravity is set to 15% of the top layer.

[0103] The salt concentration in the surface layer (0-10 cm) is determined by both measured values and model calculations, updated every hour; the bottom layer (10-20 cm) is set as a closed boundary, consistent with the monitoring depth of the sensor. The grid edge uses mirror boundary conditions, which eliminate edge effects by copying the data of adjacent grids.

[0104] The salt content of each grid is calculated iteratively every hour, considering lateral infiltration within 5% of adjacent grids. After each calculation, the real-time monitoring data are used to calibrate the model (root mean square error RMSE≤5%) using LSTM correction, and if the error exceeds the limit, the local grid parameters (such as soil porosity and root water absorption rate) are automatically adjusted.

[0105] A salt prediction matrix is generated for the next 72 hours, with a spatial resolution of 100 m x 10 cm and a time resolution of 1 hour, marking high-risk areas with a salt value ≥3 dS / m.

[0106] Preferably, the LSTM model management module is further configured to superimpose a sinusoidal modulation term in the LSTM model's forget gate activation function for annual or seasonal periodicity of electrode drift, to enhance the memory retention capability for seasonal drift regularity. For annual / seasonal periodicity of electrode drift, a periodic adjustment mechanism is constructed with time (month or season) as a variable when superimposing a sinusoidal modulation term in the LSTM model's forget gate activation function: using the precise timestamp provided by the satellite timing module, divide a year into 12 monthly periods or 4 quarterly periods, use the month m (1-12) or the quarter q (1-4) as the input variable of the sinusoidal function to generate a sinusoidal curve with a period of 12 months or 3 months. The expression of the modulation term is sin(2πm / 12) or sin(2πq / 4), and its amplitude is automatically adjusted according to the seasonal fluctuation amplitude of historical drift data (range 0.2-0.5), for example, in the spring salt return peak period of saline-alkali land, the drift fluctuation in this period is larger through historical data statistics, and the amplitude is automatically set to 0.4, so that the peak of the sinusoidal curve corresponds to the spring drift peak. The output of the forget gate is obtained by multiplying the output value of the traditional sigmoid function and the sinusoidal modulation term, when in the summer high temperature period, the sinusoidal term value is close to 1, the forget gate control value is reduced, and the forgetting of high drift data in this period is reduced; when in winter, the sinusoidal term value is close to 0, the forget gate control value is increased, and the filtering of low drift data is enhanced. Through this mechanism, the memory retention capability of the model for seasonal drift regularity is improved, for example, in the saline-alkali area of northeast cold land, the regularity that the electrode drift rate in the spring thaw period (March-April) is higher than that in winter (December-February) can be significantly captured, the seasonal drift prediction error is reduced, and the accuracy of salt trend analysis in different seasons is ensured.

[0107] Preferably, the LSTM model management module is used to add a cross-time step virtual gradient link in the deep LSTM of the LSTM model, forcing the model to retain long-term drift features to capture drift patterns across the life cycle of the device. In the deep network of the LSTM model, the cross-time step virtual gradient link is implemented by establishing long-range associations of hidden states at different time stages. Specifically, when the sensor node runs for more than 6 months, the system automatically extracts the hidden states at key time nodes (such as the 1st, 3rd, 6th, and 12th months) in the life cycle of the device, and inputs the spliced data of these historical states and the current time data into the deep LSTM unit through virtual connection. The link weight is dynamically adjusted according to the running time of the device. The recent state (such as the last 12 months) adopts a linear weight (such as the last 6 months with a weight of 0.8 and the last 6 months with a weight of 0.6), and the early state (such as more than 24 months) adopts an exponential decay weight (such as the second year with a weight of 0.3 and the third year with a weight of 0.1), to ensure that the model pays more attention to historical features with high correlation with the current drift pattern. For example, when a node runs for 24 months, the system activates the cross-time step link at the 6th, 18th, and 24th months, and forcibly associates the drift data of the same season in different years (such as the high drift rate in summer every year), so that the model can identify the cumulative drift trend of the electrode as the service life increases (such as the drift rate increasing from 0.03 dS / m / year to 0.06 dS / m / year year by year). In practical applications, this mechanism can make the model predict the drift acceleration caused by device aging in advance, and when the prediction error of multiple consecutive cycles exceeds the threshold, it automatically triggers an "electrode aging warning" in combination with the hardware calibration cycle, prompting the operation and maintenance personnel to replace the device before the error accumulates to the threshold, thereby realizing the error control of the saline-alkali soil monitoring device throughout its life cycle.

[0108] The various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a complex programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0109] The application further discloses a saline-alkali soil salt dynamic monitoring method based on narrowband Internet of Things, which comprises the following steps:

[0110] The sensor nodes are deployed in a 1km×1km grid, with a burial depth of 0-20cm and a node spacing of 100m, about 100 nodes per square kilometer. During initialization, all node clocks are synchronized through the satellite timing module, and a uniform sampling period is set (e.g., every 10 minutes).

[0111] The time synchronization server sends a synchronization broadcast instruction every 24 hours, and the node adjusts the local clock after receiving it (accuracy ±1ms).

[0112] If the node misses the synchronization instruction due to sleep, it actively requests a timestamp from the base station after waking up to complete the data timestamp error.

[0113] Before inputting the LSTM model, linear interpolation weights are dynamically generated according to the timestamp difference to align non-uniformly sampled data to fixed time intervals, solving the problem of asynchronous sampling time and time interval jitter in original data.

[0114] The drift trend is predicted by the LSTM model, and the data is dynamically corrected.

[0115] A drift-sensitive factor is added to the input gate of the LSTM model, and the drift rate is dynamically adjusted by historical data statistics.

[0116] The sensor turns off the sensor circuit during the eDRX sleep period, and only maintains the RTC clock running; after waking up, it collects data and transmits it through NB-IoT single uplink transmission (data packet ≤1KB).

[0117] Based on Grubbs test to remove outliers, and through spatial and temporal neighborhood interpolation method to fill in missing data.

[0118] According to meteorological data (rainfall, evaporation), a salt diffusion model based on finite difference equation is constructed to predict the spatial and temporal distribution of salt in a certain period in the future.

[0119] Preferably, for the annual or seasonal periodicity of electrode drift, a sinusoidal modulation term is added to the activation function of the forget gate of the LSTM model, enhancing the memory retention ability of seasonal drift patterns.

[0120] Preferably, a cross-time step virtual gradient link is added to the deep LSTM of the LSTM model to force the model to retain long-term drift features and capture drift patterns across device life cycles.

[0121] It will be apparent that the method of the present application can be implemented by a computer program, and that a computer program for implementing the method of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the computer program, when executed by the processor, implements the functions / operations specified in the flow diagrams and / or block diagrams. The computer program can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0122] It will be apparent that the method of the present application can be implemented by a computer program, and that a computer program for implementing the method of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the computer program, when executed by the processor, implements the functions / operations specified in the flow diagrams and / or block diagrams. The computer program can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0123] at least one processor; and

[0124] a memory in communication with the at least one processor; wherein

[0125] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned method for dynamically monitoring salinity of saline-alkali soil based on narrowband Internet of Things or the module functions of the system for dynamically monitoring salinity of saline-alkali soil based on narrowband Internet of Things.

[0126] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A saline-alkali soil salt dynamic monitoring system based on narrowband Internet of Things, characterized in that, Comprise: Satellite time module for synchronizing all node clocks, setting a unified sampling period; Time synchronization server for sending a synchronization broadcast instruction once every certain period, and the node corrects the local clock after receiving it; Timestamp error management module for requesting a timestamp from the base station if the node misses the synchronization instruction due to hibernation, and completing the timestamp error of the data; LSTM model management module for establishing an LSTM model, predicting the drift trend through the LSTM model, and dynamically correcting the data; And for generating linear interpolation weights dynamically according to the timestamp difference before inputting the LSTM model, aligning non-uniformly sampled data to fixed time intervals to solve the problem of asynchronous sampling time and time interval jitter of original data; and adding a drift sensitive factor in the input gate of the LSTM model to dynamically adjust the opening and closing of the input gate by statistically analyzing the drift rate from historical data; Salt diffusion prediction module for constructing a salt diffusion model based on finite difference equations according to meteorological data to predict the spatial and temporal distribution of salt in the future; The salt diffusion model is calibrated using real-time monitoring data corrected by LSTM; Adding a drift sensitive factor in the input gate of the LSTM model to dynamically adjust the opening and closing of the input gate by statistically analyzing the drift rate from historical data includes: Using a sliding time window to extract historical drift data within the window and calculate the average drift rate: drift rate = (drift amount at the end of the window - drift amount at the beginning of the window) / window duration; Convert the drift rate to a dimensionless factor of 0.1-0.9, i.e. drift sensitive factor β, through normalization mapping; Introduce the drift sensitive factor β after the linear combination of the traditional input gate, then the output value of the input gate at time t is: it=σ((Wi⋅[ht−1,xt]+bi)×(1+β)), Wi is the weight matrix of the input gate, σ is the Sigmoid activation function, ht−1 is the hidden state at the previous time, xt is the input vector at the current time, i.e. the sensor measurement value, and bi is the bias vector of the LSTM input gate; Dynamically adjust the input gate value through the drift sensitive factor β.

2. The system according to claim 1, wherein, It also includes a hibernation period module for turning off the sensor circuit during the eDRX hibernation period and only maintaining the RTC clock running; after waking up, collect data and transmit it through a single NB-IoT uplink transmission; It also includes a data processing module for removing outliers based on Grubbs test and filling missing data through spatiotemporal neighborhood interpolation method.

3. The system according to claim 2, wherein, The LSTM model management module is also used to superimpose a sinusoidal modulation term in the LSTM model's forget gate activation function for the annual or seasonal periodicity of electrode drift, enhancing the memory retention ability for seasonal drift patterns.

4. The system according to claim 3, wherein, The LSTM model management module is also used to add cross-time step virtual gradient links in the deep LSTM of the LSTM model to force the model to retain long-term drift features and capture drift patterns across device lifecycles.

5. The saline-alkali soil salt dynamic monitoring system based on narrowband Internet of Things according to claim 4, characterized in that, According to the timestamp difference value, linear interpolation weights are dynamically generated to align the non-uniform sampling data to a fixed time interval, specifically including: determining a target time point T k The actual sampling points before and after (t j ,S j ) and (t j+1 ,S j+1 ), calculate the time interval Δt1=T k −t j and Δt2=t j+1 −T k , the weight is inversely proportional to the proportion of time: w j =Δt2 / (Δt1+Δt2),w j+1 =Δt1 / (Δt1+Δt2),the aligned value is S k =w j ⋅S j +w j+1 ⋅S j+1 .

6. The saline-alkali soil salt dynamic monitoring system based on narrowband Internet of Things according to claim 5, characterized in that, Constructing a salt diffusion model based on finite difference equations according to meteorological data to predict the spatial and temporal distribution of salt in the future includes: Based on the theory of water and salt transport, the soil salt dynamic is regarded as the convection and diffusion process on the two-dimensional vertical profile, the initial salt value is the sensor data corrected by the LSTM model, and the model is initialized at 0 o'clock every day according to the future 72-hour hourly meteorological forecast data, and the data is obtained through the access of NB-IoT to the meteorological API or satellite inversion; Rainfall is the water input of the top layer of soil, and evaporation is the output of the top layer. When the rainfall exceeds the saturated water holding capacity of the soil, the excess water is discharged in the form of runoff. When the evaporation exceeds the current layer water content, the current water content value is taken to avoid negative water content. The water flux is calculated by Darcy's law, the permeability is preset according to the soil type, and the explicit finite difference method is used to discretize the convection term and the diffusion term into the flux difference of adjacent grids. The salt convection flux between adjacent subgrids in the horizontal direction is equal to the product of the water flux and the salt concentration, and the diffusion flux is proportional to the concentration gradient and the diffusion coefficient. The salt in the vertical direction of the bottom layer only exchanges with the upper layer through diffusion, and there is no external water input. The migration rate of the salt in the bottom layer under the action of gravity is set to 15% of that in the top layer. The salt concentration of the surface layer is determined by the measured value and the model calculation, and is updated every hour. The bottom layer is set as a closed boundary, consistent with the monitoring depth of the sensor. The salt content of each grid is calculated iteratively every hour, considering the lateral infiltration of adjacent grids within 5%. After each round of calculation, the real-time monitoring data corrected by the LSTM is used to calibrate the model. If the error is out of limit, the local grid soil porosity and root water absorption rate will be automatically adjusted. Then generate a future salt prediction matrix, mark the high-risk area where the salt value is greater than or equal to the threshold.

7. The system according to claim 6, wherein, In view of the annual or seasonal periodicity of electrode drift, a sinusoidal modulation term is added to the forget gate activation function of the LSTM model to enhance the memory retention ability of seasonal drift rules. The year is divided into 12 months or 4 seasons, and the month m or season q is used as the input variable of the sinusoidal function to generate a sinusoidal curve with a period of 12 months or 3 months. The expression of the modulation term is sin(2πm / 12) or sin(2πq / 4), and the amplitude is automatically adjusted according to the seasonal fluctuation amplitude of the historical drift data. The output of the forget gate is obtained by multiplying the output value of the traditional sigmoid function and the sinusoidal modulation term. When it is in the summer high temperature period, the value of the sinusoidal term is close to 1, the control value of the forget gate is reduced, and the forgetting of the high drift data in this period is reduced. When it is in winter, the value of the sinusoidal term is close to 0, the control value of the forget gate is increased, and the filtering of the low drift data is enhanced.

8. The system according to claim 7, wherein, A cross-time-step virtual gradient link is added to the deep LSTM of the LSTM model to force the model to retain long-term drift characteristics to capture the drift pattern across the life cycle of the device. When the sensor node runs for more than a threshold, the hidden state of the key time node in the whole life cycle of the device is automatically extracted, and these historical states are spliced with the current time data and input into the deep LSTM unit through virtual connection. The link weight is dynamically adjusted according to the running time of the device, the recent state adopts linear weight, and the early state adopts exponential decay weight.

9. An electronic device, comprising: at least one processor; and a memory communicatively connected with the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the module functions of the salt content dynamic monitoring system based on narrowband Internet of Things in claim 1.

Images