Edible mushroom growth environment regulation and control method and system based on multi-source sensing data

Abstract

The invention relates to the technical field of agricultural Internet of Things, and discloses an edible mushroom growth environment regulation and control method and system based on multi-source sensing data, and the method comprises the steps: deploying a multi-source sensor network, unifying time synchronization, and carrying out event-driven space-time alignment and three-dimensional mapping. Carrying out multi-dimensional reliability evaluation and environment correction, and obtaining an environment parameter fusion data set by adopting a fuzzy logic reasoning method; performing space-time Kriging interpolation, and obtaining a heterogeneous microenvironment feature database by adopting a multi-source data collaborative correction method; carrying out cross-modal attention fusion; performing multi-scale convolution and time sequence attention classification; carrying out deep reinforcement learning and multi-objective optimization; carrying out online incremental learning optimization to obtain a self-adaptive optimized partition collaborative regulation and control strategy; according to the method, the technical problems of insufficient multi-source data fusion precision, lack of environment spatial and temporal distribution modeling, inaccurate parameter coupling prediction, lack of self-adaptability of regulation and control strategies and the like are solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural Internet of Things (IoT) technology, and more specifically, to a method and system for regulating the growth environment of edible fungi based on multi-source sensor data. Background Technology

[0002] Factory cultivation of edible fungi is an important part of modern agriculture. Its growth process has strict requirements on environmental parameters such as temperature, humidity, light, carbon dioxide concentration, and oxygen concentration. In factory production, different growth stages of fungi often coexist within the same workshop, including mycelial culture, primordia differentiation, and fruiting body growth, forming a complex microenvironmental gradient distribution. There are strong coupling relationships between these environmental parameters; increased temperature accelerates water evaporation, leading to decreased humidity, while mycelial respiration and metabolism alter CO2 and O2 concentrations. The interaction of these parameters forms a complex nonlinear dynamic system.

[0003] Current environmental monitoring for edible fungi mainly relies on point sensors for discrete sampling, and control methods often employ fixed threshold control or simple PID control. Existing technologies suffer from the following shortcomings: misaligned timestamps, inconsistent sampling frequencies, and varying data quality from multi-source heterogeneous sensors; simple data aggregation cannot accurately reflect the true environmental state; point monitoring cannot construct spatial distribution models of environmental parameters, making it difficult to perceive local microenvironmental differences; independent prediction of single parameters ignores the coupling relationship between parameters such as temperature, humidity, and CO2, making it impossible to accurately predict the actual effect of control actions; fixed threshold control cannot adapt to the differentiated needs of different growth stages and lacks multi-objective optimization capabilities; and there is a lack of real-time assessment of growth status, resulting in unclear control objectives and difficulty in quantifying the effects.

[0004] These technical challenges make it difficult for existing methods to address the complex control requirements of multi-regional, multi-stage, and multi-parameter coupling in industrialized production, thus affecting the yield and quality of edible fungi. Therefore, there is a need to develop an intelligent environmental control technology that can integrate multi-source sensor data, construct spatiotemporal distribution models, predict parameter coupling relationships, and generate adaptive control strategies to meet the precision and intelligence requirements of modern industrialized edible fungi production. Summary of the Invention

[0005] This invention provides a method and system for regulating the growth environment of edible fungi based on multi-source sensor data, which solves the technical problems in related technologies such as insufficient accuracy of multi-source data fusion, lack of spatiotemporal distribution modeling of environment, inaccurate parameter coupling prediction, and lack of adaptive regulation strategies.

[0006] This invention provides a method for regulating the growth environment of edible fungi based on multi-source sensor data, comprising: Deploy a multi-source sensor network, unify time synchronization, perform event-driven spatiotemporal alignment and 3D mapping, and obtain a complete spatiotemporally labeled multi-source sensor dataset; Based on a multi-source sensor dataset with complete spatiotemporal annotation, multidimensional reliability assessment and environmental correction are performed, and a fuzzy logic reasoning method is used to obtain a fusion dataset of environmental parameters. Based on the environmental parameter fusion dataset, spatiotemporal kriging interpolation is performed, and a multi-source data collaborative correction method is used to obtain a heterogeneous microenvironment feature database. Based on a heterogeneous microenvironment feature database, cross-modal attention fusion is performed to obtain environmental change prediction results; Based on the environmental change prediction results, multi-scale convolution and temporal attention classification are performed to obtain the growth stage identification results. Based on the growth stage identification results, deep reinforcement learning and multi-objective optimization are performed to obtain a set of instructions for the regional collaborative regulation strategy. Based on the instruction set of the partitioned collaborative control strategy, online incremental learning optimization is performed to obtain an adaptive optimized partitioned collaborative control strategy.

[0007] In a preferred embodiment, the event-driven spatiotemporal alignment and 3D mapping include: Deploy the Network Time Protocol service on the edge computing server to periodically synchronize with the time server of the time service center via the Internet. All sensor nodes perform clock calibration with the Network Time Protocol server at startup to obtain a unified standard time as the time reference. Extract the time series of measurement values ​​from each sensor from the time series database, perform first-order difference calculation on the time series to obtain the rate of change at each moment, and mark the event as an environmental mutation event when the absolute value of the rate of change exceeds the preset rate of change threshold. Extract the characteristic moments of the same event in the response time series of different sensors. The least squares optimization algorithm is adopted to minimize the time deviation of the same event after time delay correction at different sensor response times, and the time delay correction parameters of each sensor are iteratively optimized. Subtract the inherent response delay and dynamic correction delay from the original timestamp to obtain the corrected synchronization timestamp.

[0008] In a preferred embodiment, the event-driven spatiotemporal alignment and 3D mapping further include: The time axis is divided into preset time periods. The first and second derivatives of the measured values ​​are calculated in each time period. When the environmental parameters change steadily, the linear interpolation method is used to interpolate the low-frequency sampling data. When the environmental parameters change drastically, the cubic spline interpolation method is used to interpolate all low-frequency sampled sensor data so that the interpolation target frequency reaches the highest sampling frequency among all sensors. The three-dimensional spatial coordinates of each sensor are read from the sensor network configuration database, and the three-dimensional space of the cultivation workshop is divided into a regular cubic grid. For point sensors, the spatial grid cell to which they belong is determined based on the installation coordinates. For area sensors, a perspective projection transformation model is used to convert the pixel coordinates of the image plane into spatial coordinates based on the camera calibration parameters. Add a spatial labeling field to each sensor data record to obtain a multi-source sensor dataset with complete spatiotemporal labeling.

[0009] In a preferred embodiment, the multidimensional reliability assessment includes: Extract historical measurement data of each sensor within a preset time window from the time series database, and calculate the mean, standard deviation, and coefficient of variation of the measured values; The trend curve is obtained by eliminating short-term fluctuations using moving average filtering, and the drift rate is obtained by calculating the deviation between the original measurement sequence and the trend curve. Detect abnormal jumps in historical measurement sequences and count the number of jumps. Use the Z-score normalization method to calculate the relative position of the sensor in the population. The historical stability reliability score was calculated using a weighted scoring method that combines the coefficient of variation, drift rate, jump frequency, and Z-score for population comparison. Retrieve other similar sensors within the spatial neighborhood of the target sensor from a fully spatiotemporally labeled multi-source sensor dataset, and calculate the median and interquartile range of the neighborhood measurements. A Gaussian function is used to map the bias to a confidence score, thus obtaining the spatial consistency confidence score.

[0010] In a preferred embodiment, obtaining the environmental parameter fusion dataset using fuzzy logic reasoning includes: Read the device file information of each sensor, calculate the ratio of the current usage time of the sensor to its designed lifespan, and use a multi-factor weighted model to calculate the device health score; Historical stability confidence score, spatial consistency confidence score, and equipment health score are used as input variables for fuzzy logic inference. Fuzzy sets and membership functions are defined for each input variable. A fuzzy inference rule base is established. The activation intensity of each rule is calculated by traversing the inference rule base. The fuzzy sets of the consequents of all activated rules are weighted and aggregated. The output fuzzy set is defuzzified to obtain the comprehensive confidence weight value. A threshold screening method is applied to the comprehensive reliability weight. For reliable sensor data with a reliability weight higher than the threshold, a weighted average fusion algorithm is used to calculate the fused environmental parameter values.

[0011] In a preferred embodiment, the spatiotemporal kriging interpolation includes: The fusion values ​​of various environmental parameters at all times and in all spatial grid cells are extracted from the environmental parameter fusion dataset to form a discrete set of spatiotemporal sampling points; Calculate the spatiotemporal distance and the square of the difference in measurements between all sampling point pairs, and use anisotropy measures to calculate the spatiotemporal distance; Half of the squared difference in the measured values ​​is taken as the variability function value, and empirical variability functions are obtained by grouping and statistically analyzing the spatiotemporal distance. The theoretical variogram model is fitted to the empirical variogram, and the Kriging equations are constructed and solved for the target spatiotemporal location to be interpolated to obtain the weight coefficients of each sampling point.

[0012] In a preferred embodiment, the method of obtaining a heterogeneous microenvironment feature database using multi-source data collaborative correction includes: The water vapor sensitive band image is extracted from the spectral image cube acquired by the hyperspectral sensor, the spectral reflectance value of each pixel in the sensitive band is extracted, and the normalized humidity spectral index is calculated. Spatial distribution snapshots are extracted from the spatiotemporal distribution fields of temperature, humidity, and gas concentration to form multi-parameter feature vectors for each grid point in three-dimensional space, and then standardized. A density-based clustering algorithm is used for spatial clustering. During the clustering process, if the absolute value of the gradient of the environmental parameter between adjacent points exceeds the preset gradient threshold, the point is not considered to be density reachable. Cluster boundaries are formed at locations that meet the gradient threshold condition.

[0013] In a preferred embodiment, the step of performing cross-modal attention fusion to obtain environmental change prediction results includes: Historical data of a preset duration are extracted from the spatiotemporal distribution field of environmental parameters to form a four-dimensional tensor. Construct a three-dimensional temporal convolutional network model to extract spatiotemporal local patterns in the temporal and spatial dimensions; After convolution, activation and batch normalization are performed, and three-dimensional max pooling is used to extract higher-level spatiotemporal features layer by layer to obtain the spatiotemporal evolution feature tensor. The adjacency relationships of each partition are extracted from the micro-environment partitioning graph, a partition adjacency graph is constructed, and a graph convolutional network is used to aggregate the features of neighbor nodes. Different environmental parameters are treated as different modalities, and multi-head self-attention is used to calculate the cross-attention between parameters to construct the parameter coupling strength matrix.

[0014] In a preferred embodiment, the obtained partitioned collaborative control strategy instruction set includes: Construct a knowledge graph of edible fungi cultivation and query environmental requirements based on the currently cultivated varieties and identified growth stages; The measured values ​​of environmental parameters, the predicted values ​​of environmental parameters, the knowledge set of environmental requirements, and the growth and health status scores of each zone are encoded into state vectors. The deep deterministic policy gradient algorithm is adopted. The Actor network outputs actions based on the state, and the Critic network evaluates the value of the state-action pair. The Actor network inputs the state vector and outputs the control action parameters of each partition through a fully connected neural network. Verify whether the initial control actions meet the physical constraints, correct the action parameters that exceed the constraints, and obtain a feasible control scheme; For each feasible control scheme, the objective function values ​​are calculated to form a multi-objective vector. The Pareto optimal solution screening algorithm is used to screen out the Pareto front solution set, and the final scheme is selected from the Pareto front solution set.

[0015] This invention provides a system for regulating the growth environment of edible fungi based on multi-source sensor data, used to execute the aforementioned method for regulating the growth environment of edible fungi based on multi-source sensor data, comprising: The multi-source sensor network module is used to deploy multi-source sensor networks, unify time synchronization, perform event-driven spatiotemporal alignment and 3D mapping, and obtain a complete spatiotemporally labeled multi-source sensor dataset. The evaluation and data fusion module, based on a multi-source sensor dataset with complete spatiotemporal annotation, performs multi-dimensional reliability evaluation and environmental correction, and uses fuzzy logic reasoning to obtain an environmental parameter fusion dataset. The spatiotemporal distribution modeling and feature extraction module performs spatiotemporal kriging interpolation based on the environmental parameter fusion dataset and uses a multi-source data collaborative correction method to obtain a heterogeneous microenvironment feature database. The multi-parameter coupled prediction module performs cross-modal attention fusion based on a heterogeneous microenvironment feature database to obtain environmental change prediction results; The growth stage identification and evaluation module performs multi-scale convolution and temporal attention classification based on the environmental change prediction results to obtain the growth stage identification results; The regulation strategy generation module, based on the growth stage identification results, performs deep reinforcement learning and multi-objective optimization to obtain a set of regional collaborative regulation strategy instructions; The adaptive optimization module performs online incremental learning and optimization based on the partitioned collaborative control strategy instruction set to obtain an adaptively optimized partitioned collaborative control strategy.

[0016] The beneficial effects of this invention are as follows: An event-driven spatiotemporal alignment algorithm addresses the issues of misaligned timestamps and inconsistent sampling frequencies among different sensors. A multi-dimensional reliability assessment method comprehensively considers sensor historical stability, spatial consistency, environmental interference, and equipment health, dynamically calculating fusion weights and removing outlier data to improve data fusion quality. A spatiotemporal kriging interpolation algorithm constructs a four-dimensional continuous distribution field, combining the high spatial resolution of infrared thermal imaging with the high measurement accuracy of point sensors for collaborative correction, resulting in a spatiotemporal distribution model of environmental parameters with both high precision and high resolution. Adaptive spatial clustering based on density and gradient enables dynamic partitioning of the microenvironment. Computer vision technology is used to extract information on mushroom density and growth status, constructing a heterogeneous microenvironment feature database containing both environmental and growth information. This provides a comprehensive and accurate data foundation for precise regulation, improving environmental perception accuracy and spatial resolution compared to traditional methods.

[0017] This study employs a spatiotemporal graph neural network to model the regional environmental propagation effect. A cross-modal multi-head self-attention mechanism dynamically identifies the coupling strength between parameters such as temperature-humidity and temperature-CO2, constructing an encoder-decoder prediction network to achieve multi-regional, multi-parameter future time-series predictions with higher accuracy than traditional single-parameter independent prediction methods. Based on transfer learning and multi-scale visual features, accurate identification of growth stages and quantitative assessment of health status are achieved, establishing a quantitative correlation between environmental parameters and growth effects. A deep reinforcement learning strategy network intelligently generates regulatory actions based on environmental conditions, predicted trends, and growth requirements. Through multi-objective optimization and Pareto optimal selection, a comprehensive balance is achieved between growth rate, energy efficiency, environmental stability, and regional synergy. Transfer reinforcement learning solves the cold-start problem, and online incremental learning enables continuous model optimization. The system's regulatory intelligence level continuously improves with accumulated operational experience, enhancing adaptability and autonomy, providing effective technical support for the intelligent and high-quality development of the edible fungi industry. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for regulating the growth environment of edible fungi based on multi-source sensor data according to the present invention; Figure 2 This is a block diagram of an edible fungus growth environment control system based on multi-source sensor data according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for regulating the growth environment of edible fungi based on multi-source sensor data, such as... Figure 1 As shown, it includes: Specifically, the following steps are included: Step 1: Deploy a multi-source sensor network, unify time synchronization, perform event-driven spatiotemporal alignment and 3D mapping, and obtain a complete spatiotemporally labeled multi-source sensor dataset; Step 1.1, Layered deployment and network configuration of multi-source sensors; 120 integrated temperature and humidity sensor nodes are deployed in the workshop with a grid spacing of 3m×3m. The temperature measurement range is -10℃ to 50℃, and the humidity range is 0% to 100%RH. Four infrared thermal imagers with a resolution of 640×480 pixels and a field of view of 90 degrees are deployed on the top of the workshop to realize two-dimensional spatial distribution imaging of the temperature field.

[0021] CO2 and O2 concentration sensors were deployed near the ventilation openings in the mycelium cultivation and fruiting management areas, employing the principle of non-dispersive infrared absorption. The CO2 measurement range was 0 to 5000 ppm, and the O2 range was 0% to 25%. Sixteen light intensity sensors were evenly distributed, with a measurement range of 0 to 100,000 lux. Twelve high-definition network cameras with a resolution of 1920×1080 pixels, equipped with automatic zoom and night vision functions, were deployed. Eight hyperspectral imaging sensors were deployed, covering the visible to near-infrared bands, to invert the spatial distribution of air humidity through spectral reflectance characteristics.

[0022] All sensor nodes are connected to the data aggregation node via a Zigbee wireless self-organizing network at a frequency of 2.4 GHz. The data aggregation node is connected to the edge computing server via Ethernet. A unique identification coding system is established for each sensor, assigning it a unique device ID that includes its type, location, and installation time. Spatial attribute parameters such as 3D spatial coordinates, measurement direction, and effective coverage area are recorded, resulting in a multi-source sensor network configuration database containing spatial topology relationships and device parameters.

[0023] Step 1.2: Unify time base synchronization and delay calibration; A network time protocol service is deployed on the edge computing server to periodically synchronize with the national time service center's time server via the internet, with a synchronization cycle of 10 minutes and a time synchronization accuracy down to the millisecond level. All sensor nodes perform clock calibration with the network time protocol server upon startup to obtain a unified UTC standard time as their time reference.

[0024] Various sensors collect data according to preset sampling strategies: temperature and humidity sensors collect data once every 30 seconds, gas concentration sensors collect data once every 60 seconds, light sensors collect data once every 60 seconds, infrared thermal imagers collect data once every 120 seconds, high-definition cameras collect data continuously at a frame rate of 25 frames / second, and hyperspectral sensors collect data once every 300 seconds.

[0025] When the sensor completes a measurement, the acquisition module reads the local clock to generate a timestamp, using the ISO standard format accurate to milliseconds. Response delay calibration is performed on different types of sensors. A step-change standard stimulus signal is applied in a controlled environment, and the time required for the sensor output to reach a steady-state value is recorded. This measurement is repeated 10 times, and the average value is taken. The inherent delay is 5s for the temperature and humidity sensor, 30s for the gas sensor, and 0.1s for the thermal imager.

[0026] When a data aggregation node receives a data packet, it records the reception time and calculates the network transmission delay by comparing the sending and receiving timestamps. Time information such as sensor measurement timestamps, inherent response delays, and network transmission delays are encapsulated with the measurement data into a standardized data packet containing fields such as sensor ID, data type, measurement value, original timestamp, and delay parameters. This yields a raw multi-source sensor data stream with complete time stamps, which is stored in a time-series database.

[0027] Step 1.3, Event-driven spatiotemporal alignment and adaptive interpolation; The time series of measurements from each sensor were extracted from a time-series database, and the rate of change was calculated using first-order difference. A threshold was applied to the rate of temperature change; when the absolute value exceeded 0.5℃ / minute, it was marked as a temperature abrupt change event, and the time of occurrence and magnitude of the change were recorded. The same method was applied to humidity (threshold 2% / minute) and CO2 concentration (threshold 100ppm / minute) to detect environmental abrupt change events.

[0028] Environmental abrupt events are typically caused by external control actions and are detected by different sensors with varying time response characteristics. This study extracts characteristic moments of the same event from the response time series of different sensors and dynamically corrects the time delay parameters by comparing the differences in response moments. An event propagation time delay model is established based on sensor spatial distance and the propagation speed of environmental parameters. A least squares optimization algorithm is used to iteratively optimize the time delay correction parameters, aiming to minimize the time deviation of the same event after time delay correction at different sensor response moments. The optimized dynamic time delay correction parameters are obtained, and the original timestamp is subtracted from the inherent response delay and the dynamic correction delay to obtain the corrected synchronization timestamp.

[0029] To address the issue of inconsistent sampling frequencies among different sensors, a piecewise adaptive interpolation strategy is adopted. The time axis is divided into several time periods to analyze the characteristics of environmental parameter changes and calculate the first and second derivatives. When environmental parameters change steadily, the second derivative is close to zero, and linear interpolation is used; when changes are drastic, the second derivative is significantly non-zero, and cubic spline interpolation is used to ensure the continuity of the interpolation function and its first and second derivatives, accurately characterizing the nonlinear trend of parameter changes.

[0030] Interpolation is performed on all low-frequency sampled sensor data, with the interpolation target frequency set to the highest sampling frequency, i.e., once every 30 seconds. Image data retains its original sampling time, and is matched with the most recent image data when needed. After time delay correction and adaptive interpolation, an aligned sensor dataset with optimized time synchronization accuracy and unified sampling frequency is obtained.

[0031] Step 1.4, 3D spatial coordinate mapping and mesh annotation; The three-dimensional spatial coordinates of each sensor are read from the sensor network configuration database. The coordinate system adopts a right-handed coordinate system with the southwest corner of the workshop as the origin, east as the X-axis, north as the Y-axis, and vertically upward as the Z-axis. The spatial position of point sensors is represented by the three-dimensional coordinates of the installation point, while that of area sensors is described by the equipment installation coordinates, field of view angle, and field of view direction vector.

[0032] The three-dimensional space of the cultivation workshop was divided into a regular cubic grid with a side length of 0.5m. The entire workshop space was discretized into several spatial grid units, each identified by its center point coordinates. For point-type sensors, the spatial grid unit to which they belong was determined based on their installation coordinates, and the measured values ​​were labeled as the corresponding environmental parameter values ​​for that grid unit. The effective measurement radius was set based on the diffusion characteristics of the environmental parameters and the sensor sensitivity. Within the effective radius, all spatial grid units could use the sensor's measured values ​​as a reference.

[0033] For surface-mount sensors such as infrared thermal imagers, the output is a two-dimensional image, with each pixel corresponding to a region in space. Based on the thermal imager's installation location, field of view, and imaging plane geometry, a mapping relationship between image pixel coordinates and three-dimensional spatial coordinates is established. Using a perspective projection transformation model, the image pixel coordinates are converted into spatial coordinates according to the camera calibration parameters (intrinsic and extrinsic matrices). For each pixel in the thermal image, the corresponding spatial ray is calculated through inverse perspective projection transformation; the intersection of this ray and a plane at a specific height represents the spatial position of that pixel.

[0034] Thermal imaging image pixels are mapped to spatial grid cells, with each grid cell potentially corresponding to one or more pixels in the image. The average of these pixel temperature measurements is taken as the thermal imaging temperature value for that grid cell. The spatial mapping method for hyperspectral sensors is similar, establishing a mapping between spectral image pixels and spatial grid cells based on imaging geometry. Visible light images acquired by high-definition cameras are primarily used for growth state identification, associating the image with the shooting time, location, and field of view to establish a correspondence between the image and a spatial region.

[0035] A spatial annotation field is added to each sensor data record, including spatial attributes such as the 3D coordinates of the measurement point, the ID of the spatial grid cell to which it belongs, and the effective measurement radius. For cases where multiple sensor measurements cover the same spatial grid cell, all sensor measurement records are retained and their data sources are labeled, providing raw data support for subsequent data fusion. This results in a fully spatiotemporally annotated multi-source sensor dataset, with each record containing a precise timestamp and spatial location information, achieving a unified representation of multi-source heterogeneous sensor data in both time and space dimensions.

[0036] Step 2: Based on the multi-source sensor dataset with complete spatiotemporal annotation, perform multi-dimensional reliability assessment and environmental correction, and use fuzzy logic reasoning to obtain the environmental parameter fusion dataset; Specifically, the following steps are included: Step 2.1, Multidimensional Reliability Assessment; Historical measurement data for each sensor within a 30-day time window are extracted from the time-series database to form a historical measurement time series. Statistical characteristics of the historical measurement time series are calculated, including the mean, standard deviation, and coefficient of variation. The coefficient of variation is defined as the standard deviation divided by the mean, reflecting the relative dispersion of the measured values.

[0037] Analyzing the long-term trend of the measurement time series, a 24-hour window-length moving average filter is used to eliminate short-term fluctuations, resulting in a smooth trend curve. The deviation between the original measurement series and the trend curve is calculated. If the deviation gradually increases over time, it indicates sensor drift. The drift rate is defined as the amount of deviation increase per unit time. A drift rate less than 0.05℃ / day (adjustable within the range of 0.02℃ / day to 0.1℃ / day) indicates sensor stability.

[0038] Abnormal jump points are detected in historical measurement sequences. A jump is defined as the absolute value of the difference between adjacent measurements exceeding three times the standard deviation of the normal fluctuation range. The number of jumps within a historical time window is counted; a higher number of jumps indicates lower sensor reliability. Statistical characteristics of similar sensor groups are extracted, and the stability index of an individual sensor is compared with the group's statistical characteristics. The Z-score standardization method is used to calculate the sensor's relative position within the group.

[0039] A weighted scoring method was used to calculate the historical stability reliability score, taking into account the coefficient of variation, drift rate, jump frequency, and population comparison Z-score. The weights were: coefficient of variation 0.3, drift rate 0.3, jump frequency 0.2, and Z-score 0.2. A historical stability reliability score was obtained for each sensor, ranging from 0 to 1, with a higher score indicating more stable and reliable historical performance.

[0040] For the target sensor to be evaluated at the current moment, other sensors of the same type within its spatial neighborhood are retrieved from the complete spatiotemporally labeled multi-source sensor dataset. The spatial neighborhood is defined as all sensors of the same type within a spherical region with a radius of 5m centered on the target sensor's location. If the number of sensors of the same type within the neighborhood is less than 3, the neighborhood radius is expanded to 10m.

[0041] Extract the current measurement values ​​of all sensors within the neighborhood to form a set of neighborhood measurement values. Calculate the statistical characteristics of the neighborhood measurement values, including the mean, median, standard deviation, and interquartile range. Determine whether the target sensor measurement value falls within a reasonable range, defined as the interval between the median and 1.5 times the interquartile range. If it falls within the reasonable range, use a Gaussian function to map the deviation to a confidence score; if it falls outside the range, set an appropriate confidence score based on the degree of deviation. Considering the spatial gradient characteristics of environmental parameters, a spatial gradient correction method is used to perform gradient compensation on the neighborhood measurement values ​​before consistency verification. Obtain the spatial consistency confidence score for each sensor measurement value at the current time.

[0042] A knowledge base linking sensor types and environmental interference factors is established, including the working principles of various sensors, susceptible environmental factors, and quantitative relationships of interference effects. Environmental state parameters are extracted from the fused data at the current moment. Interference impact models are queried from the knowledge base based on sensor type, and the current environmental state parameters are substituted into the model to calculate the interference impact. Considering the influence of various environmental interference factors, a multiplicative correction model is used to calculate the comprehensive environmental interference correction coefficient. The sensor measurement value is multiplied by the environmental interference correction coefficient to obtain the corrected measurement value.

[0043] The system retrieves file information for each sensor device from the sensor device management system, including installation time, cumulative runtime, historical calibration records, historical fault records, and maintenance records. It queries the design life and performance degradation curves based on sensor type and models sensor reliability over time using a Weibull distribution function. The system calculates the ratio of current usage time to design life, analyzes calibration intervals, and statistically analyzes historical fault frequencies. Finally, it calculates a device health score using a multi-factor weighted model, considering usage time, calibration status, and fault frequency.

[0044] Step 2.2, Fuzzy logic comprehensive reliability reasoning and weighted fusion; Historical stability confidence score, spatial consistency confidence score, and equipment health score are used as input variables for fuzzy logic inference, with environmental interference correction coefficients as correction factors. A fuzzy set and membership function are defined for each input variable, mapping continuous score values ​​to fuzzy linguistic variables. For example, historical stability scores are divided into three fuzzy sets: low (0 to 0.4), medium (0.4 to 0.7), and high (0.7 to 1).

[0045] A fuzzy inference rule base is established, containing multiple IF-THEN inference rules. The antecedent of each rule is a combination of fuzzy sets of input variables, and the consequent of each rule is a fuzzy set of output confidence weights. Based on the current sensor scores, the membership degree of each rule to each fuzzy set is calculated. The activation strength of each rule is calculated by traversing the inference rule base, and all activated rule consequent fuzzy sets are aggregated by weighted activation strength.

[0046] The output fuzzy set is defuzzified, and the centroid method is used to calculate the centroid position of the membership function of the output fuzzy set as the output value, yielding the preliminary comprehensive reliability weight. An environmental interference correction coefficient is applied to the comprehensive reliability weight, and the final reliability weight is equal to the preliminary comprehensive reliability weight multiplied by the environmental interference confidence factor. The comprehensive reliability weights of all sensors are normalized so that the sum of the reliability weights of all sensors for the same environmental parameter at the same time equals 1. This yields the real-time comprehensive reliability weight for each sensor at the current time.

[0047] A threshold screening process is applied to the overall reliability weights, setting a minimum reliability threshold (typically 0.3). Sensor data with a weight below this threshold are considered unreliable. Sensor data with a reliability weight below the threshold are marked as anomalous and removed from the fusion calculation. The anomalous data sensor ID, timestamp, measurement value, and reason for the anomalous data are recorded and stored in the anomalous data log. A sensor health check prompt is triggered, notifying maintenance personnel to inspect, calibrate, or replace the corresponding sensor.

[0048] For reliable sensor data with a confidence weight higher than a threshold, a weighted average fusion algorithm is used to calculate the fused environmental parameter value. For a given environmental parameter in a specific spatial grid cell, if N reliable sensors cover that location, the fused value is calculated as the sum of the products of each sensor measurement and its corresponding confidence weight, divided by the sum of the weights. Since the weights have been normalized to make the weight sum equal to 1, the fused value is simply the weighted sum of the measurements.

[0049] A spatial weighted fusion method is employed for the spatial distribution data provided by infrared thermal imaging and hyperspectral sensors. The thermal imaging data is calibrated using precise measurements from point sensors. The deviation between the thermal imaging value and the point measurement value is calculated at the point sensor location. A spatial interpolation method is used to extend the deviation distribution across the entire space, resulting in a global correction of the thermal imaging field. The corrected thermal imaging field exhibits both high spatial resolution and high measurement accuracy.

[0050] The fused environmental parameter data undergoes a rationality check, which includes verifying whether the values ​​are within physically possible ranges, whether the changes in values ​​between adjacent time points conform to physical laws, and whether the values ​​at spatially adjacent locations maintain continuity. The fused environmental parameter values, along with corresponding timestamps, spatial grid cell IDs, a list of participating sensors and their weights, and a fusion quality score, are stored together. A comprehensive data fusion process is then performed, iterating through all time points, all spatial locations, and all environmental parameter types, to obtain a high-quality, reliable fused environmental parameter dataset.

[0051] Step 3: Based on the environmental parameter fusion dataset, perform spatiotemporal kriging interpolation and use a multi-source data collaborative correction method to obtain a heterogeneous microenvironment feature database; Specifically, the following steps are included: Step 3.1, Spatiotemporal Kriging interpolation and multi-scale data co-correction; The fused values ​​of various environmental parameters, such as temperature, humidity, CO2 concentration, and O2 concentration, are extracted from the environmental parameter fusion dataset at all times and in all spatial grid cells to form a discrete spatiotemporal sampling point set. Each sampling point contains four-dimensional coordinate information (three-dimensional spatial coordinates X, Y, Z and one-dimensional time coordinate T) and the corresponding environmental parameter measurement value.

[0052] Spatiotemporal Kriging interpolation is an extension of the classic Kriging spatial interpolation method into the time dimension. Based on the known spatial and temporal locations of sampling points and measured values, it performs optimal linear unbiased estimation of parameter values ​​at unknown locations and times. First, the spatiotemporal autocorrelation characteristics of environmental parameters are analyzed, and the spatiotemporal variogram is calculated. The spatiotemporal variogram describes the relationship between the difference in measured values ​​between two sampling points and their spatiotemporal distance. The spatiotemporal distance comprehensively considers both spatial and temporal distances and is normalized using spatially and temporally correlated scale parameters, respectively.

[0053] Calculate the spatiotemporal distance and the squared difference of measured values ​​between all sampling point pairs. Use half of the squared difference of measured values ​​as the variogram value, and group them by spatiotemporal distance to obtain the empirical variogram. Fit the empirical variogram to a theoretical variogram model (spherical model, exponential model, Gaussian model), and use the least squares method to estimate the variogram model parameters (nuclear value, sill value, and range) to obtain the fitted spatiotemporal variogram model.

[0054] To determine the spatiotemporal location of the interpolation target, the corresponding variogram value is obtained by querying the variogram model based on its spatiotemporal distance to all known sampling points. A Kriging equation system is constructed, and the weighting coefficients of each sampling point are solved using the Lagrange multiplier method. The interpolation result is equal to the weighted sum of the measured values ​​of each sampling point and their corresponding weighting coefficients. Kriging interpolation is performed on all grid points in the workshop's three-dimensional space at all times to obtain the estimated environmental parameters for each grid point at each time, forming a complete four-dimensional spatiotemporal distribution field.

[0055] Infrared thermal imagers output thermal images with a resolution of 640×480 pixels, capable of sensing detailed features of spatial temperature distribution, but their absolute temperature measurement accuracy is relatively lower than that of point-type temperature sensors. Point-type temperature sensors can achieve a temperature measurement accuracy of ±0.2℃, but only provide the temperature value at the installation point, resulting in low spatial resolution. This method extracts the measured value and spatial coordinates of the point-type temperature sensor at a specific moment. In the corresponding infrared thermal image at that moment, the position of the corresponding image pixel is calculated using perspective projection transformation based on the sensor's spatial coordinates. The temperature values ​​of that pixel and its neighboring pixels (a 5×5 pixel area) are extracted from the thermal image, and the average of the neighboring temperature values ​​is calculated as the thermal imaging temperature measurement result at that point.

[0056] The deviation between the precise measurement values ​​of point sensors and the thermal imaging temperature measurement results is calculated. For all point temperature sensor locations, the corresponding thermal imaging deviation is calculated, resulting in a set of discrete spatial position deviation sampling points. Spatial interpolation (radial basis function interpolation or Kriging interpolation) is performed on these deviation sampling points to expand the discrete deviation values ​​into a continuous spatial deviation distribution field. After obtaining the thermal imaging deviation distribution field for the entire workshop, the temperature value of each pixel in the original thermal imaging image is corrected. The correction method is that the pixel-corrected temperature equals the original pixel temperature plus the deviation distribution field value corresponding to the pixel's spatial position. After correction, the absolute temperature measurement accuracy of thermal imaging is significantly improved, reaching a level close to that of point sensors, while maintaining the high spatial resolution advantage of thermal imaging.

[0057] In the case of multiple infrared thermal imagers within the workshop, their fields of view may partially overlap. In the overlapping areas, the corrected temperature values ​​from multiple imagers are fused. The fusion method calculates fusion weights based on the temperature measurement accuracy and viewing angle relationship of each imager; images with a closer viewing angle to the front have higher weights, while those with a tilted viewing angle have lower weights. The corrected and fused thermal imaging temperature field is compared and verified with the temperature distribution field obtained through spatiotemporal Kriging interpolation. If differences exist between the two at non-sensor locations, the high-resolution detail features of the thermal imaging are prioritized, and the Kriging interpolation results are locally corrected using the thermal imaging data. This results in a temperature spatial distribution model that combines high accuracy and high spatial resolution.

[0058] Step 3.2, Optimize the coupling of spectral inversion humidity distribution with temperature and humidity; Hyperspectral sensors acquire multiple narrow-band spectral images in the visible to near-infrared range. The spectral reflectance or absorption characteristics of different bands are related to the water vapor content in the air. Water vapor has absorption peaks near specific wavelengths (such as 940 nm and 1140 nm), and the spectral intensity of these bands is negatively correlated with air humidity.

[0059] Water vapor-sensitive band images were extracted from the spectral image cube acquired by the hyperspectral sensor. These sensitive bands were determined through spectral library lookup and experimental calibration. The spectral reflectance value of each pixel in the sensitive band was extracted to form a spectral feature vector. Simultaneously, the spectral reflectance of the reference band (bands with less water vapor absorption influence) was extracted for normalization to eliminate the influence of light intensity variations.

[0060] The normalized humidity spectral index is calculated as follows: Subtract the reflectance of the reference band from the reflectance of the humidity-sensitive band to obtain the reflectance difference; add the reflectance of the reference band to the reflectance of the humidity-sensitive band to obtain the reflectance sum; and divide the reflectance difference by the reflectance sum to obtain the normalized humidity spectral index. The advantage of the normalized index is that it compensates for interference factors such as changes in illumination and differences in the reflectance of object surfaces, highlighting the spectral characteristics of water vapor absorption.

[0061] A quantitative inversion model between spectral indices and air humidity was established, and the model was calibrated experimentally. The calibration experiment was conducted in a controlled environment chamber, with different humidity gradients set, and hyperspectral data and high-precision humidity sensor data were collected simultaneously. The regression relationship between the spectral indices and actual humidity was then fitted. The regression model can be linear regression, multinomial regression, or a machine learning regression model such as support vector regression.

[0062] A regression model was applied to hyperspectral images acquired in the field. A normalized humidity spectral index was calculated for each pixel, and then substituted into the inversion model to calculate the estimated air humidity at the corresponding spatial location. High-precision measurements from a point humidity sensor were used to correct the hyperspectral inverted humidity field. The deviation between the hyperspectral humidity inversion value and the sensor's measured value at the point humidity sensor location was calculated. A humidity deviation distribution field was generated through spatial interpolation, and the hyperspectral humidity field was globally corrected to obtain a high-precision humidity spatial distribution model.

[0063] Temperature and humidity are physically coupled. In high-temperature areas, rapid evaporation may decrease humidity, but the diffusion of evaporated water vapor to other areas will increase humidity in those areas. A coupled physical constraint model of temperature and humidity is established. Based on thermodynamic and fluid dynamics principles, humidity distribution should be coordinated with temperature distribution, ventilation conditions, and evaporation source distribution. A coupled constraint optimization method is employed, simultaneously considering both temperature and humidity distribution models. Physical consistency constraints are used to jointly optimize and adjust both models, ensuring that temperature and humidity distributions satisfy their respective observational data while conforming to the physical coupling laws. This yields a spatial distribution model of air humidity that is consistent with the temperature distribution.

[0064] Step 3.3, Adaptive spatial clustering partitioning and target detection feature extraction; A snapshot of the spatial distribution at a specific moment is extracted from the spatiotemporal distribution field of temperature, humidity, and CO2 concentration to form a multi-parameter feature vector for each grid point in three-dimensional space, containing the temperature, humidity, and CO2 concentration values ​​at that point. The feature vectors are then standardized to eliminate the influence of different parameter units and numerical ranges.

[0065] Calculate the feature similarity between spatially adjacent grid points, using Euclidean distance as the similarity metric. Construct a spatial adjacency graph, where the graph nodes are spatial grid points, and adjacent grid points are connected by edges, with the edge weight being the reciprocal of the feature similarity. Perform spatial clustering using a density-based spatial clustering algorithm (DBSCAN algorithm or its improved version), with a neighborhood radius preferably of 1.5m (adjustable within the range of 0.5m to 3m).

[0066] The clustering process considers not only feature similarity but also the spatial gradient of environmental parameters. The gradient of environmental parameters for each grid point is calculated using the central difference method. The density reachability criteria for the density clustering algorithm are modified: when the absolute value of the environmental parameter gradient between two adjacent points exceeds a preset gradient threshold (preferably 1℃ / m for temperature gradient, adjustable within the range of 0.5℃ / m to 2℃ / m), even if the features are similar, the point is not considered density reachable, thus forming cluster boundaries at locations with large gradients.

[0067] An improved DBSCAN clustering algorithm is executed, traversing all grid points and clustering contiguous regions with similar features and small gradients into a single cluster, each cluster representing a microenvironmental partition. Post-processing optimizations are performed on the clustering results, including small cluster merging (with a preferred threshold of 10 m³), ​​boundary smoothing, and outlier handling. This results in a dynamically partitioned microenvironmental map, dividing the workshop space into several microenvironmental regions. Environmental parameters are relatively uniform within each region, but significant environmental differences or gradients exist between regions. The optimal update cycle for the partitioned map is 2 hours (adjustable within the range of 1 to 4 hours).

[0068] For each zone in the microenvironment zoning map, extract the environmental characteristic statistics of that zone, including the mean, standard deviation, maximum, and minimum temperatures, the mean and standard deviation of humidity, the mean CO2 concentration, and the mean O2 concentration. Calculate the spatial gradient characteristics of the zone, including the average gradient magnitude and principal gradient direction of the environmental parameters within the zone, reflecting the uniformity and changing trends of the environment within the zone.

[0069] The temporal variation characteristics of each partition are calculated by extracting the time series of environmental parameters for that partition over the past 24 hours, and calculating the rate of change, fluctuation amplitude, and periodicity. The rate of change is the average change in parameter values ​​per unit time; the fluctuation amplitude is the difference between the maximum and minimum parameter values; and the periodicity characteristics are identified through Fourier transform or autocorrelation analysis to determine if there are regular fluctuations such as daily cycles or regulatory cycles. The environmental characteristics of each partition are constructed into a multi-dimensional feature vector containing dozens of feature components, comprehensively describing the environmental state and dynamic characteristics of that partition.

[0070] Visual information for each zone is extracted from images captured by a high-definition camera. First, image-zone spatial registration is performed, determining which areas in the image correspond to which zones based on the camera position, viewing angle, and zone spatial range. Then, deep learning-based object detection algorithms (such as YOLO and Faster R-CNN) are used to identify mushroom stick objects in the image. The detection algorithm outputs the bounding box coordinates and confidence score for each mushroom stick.

[0071] The number of detected mycelium logs within each zone was counted, and the mycelium log distribution density was calculated by dividing the number of logs by the zone area. Areas with higher mycelium log density have a more crowded growth environment, with increased mutual shading and competition among logs, resulting in a greater impact on environmental parameters such as CO2 concentration and humidity. The detected mycelium log images were preliminarily determined to be in their growth stage. Based on visual characteristics such as color, texture, and the presence of fruiting bodies, the logs were classified into stages such as mycelial growth, primordium formation, fruiting body growth, and maturity. The classification method employed was an image feature-based classifier or a convolutional neural network classification model. The number and proportion of logs at different growth stages within each zone were counted to generate a statistical distribution of growth stages for each zone.

[0072] Mycelial coverage features were extracted, and image segmentation was performed on the surface image of the mushroom logs to separate mycelial-covered and uncovered areas. The percentage of mycelial-covered area to the total surface area of ​​the mushroom logs was calculated. Fruiting body quantity and size features were extracted. For mushroom logs that had entered the fruiting stage, the number of fruiting bodies was detected and counted, and the diameter or height of the fruiting bodies was measured. These growth indicators were quantified as numerical features and incorporated into the partitioned feature vector.

[0073] By integrating environmental and growth characteristics, a comprehensive database of heterogeneous microenvironment features is constructed. The database organizes data by partition, with each partition corresponding to one record. Each record contains dozens of fields, including partition ID, timestamp, spatial extent, environmental statistical characteristics, environmental dynamic characteristics, mycelial density, growth stage distribution, mycelial coverage, and fruiting body statistics. The database supports spatiotemporal queries, feature retrieval, and statistical analysis, providing rich feature inputs for subsequent environmental prediction and control decisions.

[0074] Step 4: Based on the heterogeneous microenvironment feature database, perform cross-modal attention fusion to obtain the environmental change prediction results; Specifically, the following steps are included: Step 4.1, 3D Temporal Convolution and Graph Neural Network Modeling; Supported by the partition features of the heterogeneous microenvironment feature database, and combined with the spatiotemporal distribution field of environmental parameters obtained in step 3, historical data from the past 12 hours are extracted. The historical data constitutes a 5-dimensional tensor, with dimensions of time, spatial X, spatial Y, spatial Z, and parameter channel.

[0075] A three-dimensional temporal convolutional network model is constructed. The network input is a 5-dimensional tensor of historical environmental parameters, and the output is a spatiotemporal feature tensor. The network structure contains multiple three-dimensional convolutional layers, with the convolutional kernels sliding simultaneously in the three dimensions of time and space to extract spatiotemporal local patterns. The kernel size of the first convolutional layer is set to 3 steps in the time dimension, 3 steps in the X dimension, 3 steps in the Y dimension, and 3 steps in the Z dimension, which can capture the correlation patterns between adjacent time steps and adjacent spatial locations.

[0076] After extracting local features through convolution, nonlinearity is introduced through an activation function (ReLU or LeakyReLU). A batch normalization layer is added after the convolutional layer to normalize the features, accelerating training convergence and improving the model's generalization ability. A three-dimensional max-pooling layer is used to downsample the features, with a pooling window size of 2×2×2 being optimal. This reduces the spatiotemporal resolution of the feature map, decreases computational cost, and enhances feature invariance.

[0077] The network employs a stacked multi-layered approach combining 3D convolutions, activations, normalization, and pooling to extract higher-level spatiotemporal features layer by layer. Shallow convolutional layers extract simple spatiotemporal patterns, while deeper convolutional layers extract complex patterns such as large-scale temperature propagation, periodic environmental fluctuations, and the diffusion of the effects of regulatory actions. The network is optimally configured with five convolutional blocks, each containing convolution, activation, normalization, and pooling operations. No pooling is performed after the final convolutional layer to maintain high feature resolution and preserve spatial detail. The number of feature channels gradually increases with network depth, from four channels for input environmental parameters to 128 channels in deeper layers.

[0078] The network training uses historical data, with training samples consisting of input historical time window environmental parameter tensors and labels representing the actual environmental parameter values ​​at future times. The loss function is mean squared error, which measures the difference between the predicted values ​​obtained from subsequent prediction layers and the actual values ​​of the network's output features. The Adam optimizer is preferably used for gradient descent optimization, with a preferred learning rate of 0.001. Training is iterated for 200 epochs or until the loss on the validation set no longer decreases. The resulting trained 3D temporal convolutional network model can extract feature representations containing spatiotemporal evolution patterns from historical environmental parameters.

[0079] The adjacency relationships between partitions are extracted from the microenvironment partitioning graph. Two partitions are defined as adjacent if they are spatially adjacent and share a common boundary. A partition adjacency graph is constructed, where the graph nodes represent the microenvironment partitions, and edges connect adjacent partition nodes. The edge weights are set based on the contact area or boundary length between adjacent partitions; a larger contact area indicates stronger environmental exchange between partitions and a higher weight.

[0080] Each partition node is associated with a feature vector, extracted from a heterogeneous microenvironment feature database, containing statistical, dynamic, and growth state features of the partition environment. A graph neural network, through an iterative message-passing mechanism, enables each node to aggregate feature information from its neighboring nodes, achieving partition-based environmental impact modeling. A graph convolutional network is preferably used as the basic structure of the graph neural network, with the graph convolution operation defined as weighted aggregation of node features. Through multiple layers of graph convolution (preferably three layers), nodes can aggregate multi-hop neighbor information. Finally, each partition node receives an enhanced feature representation, which not only includes the partition's own environmental information but also incorporates environmental information from neighboring partitions, reflecting the spatial propagation and diffusion effects of environmental parameters.

[0081] Step 4.2, cross-modal attention fusion and encoder-decoder prediction; Different environmental parameters such as temperature, humidity, CO2, and O2 are considered as different modes, with each parameter corresponding to a feature channel. Feature sequences corresponding to each parameter are extracted from the spatiotemporal evolution feature tensor to form a multi-channel feature matrix.

[0082] A multi-head self-attention mechanism is employed to calculate cross-attention between parameters. In cross-modal attention, the feature of one parameter channel is used as the query, and the features of other parameter channels are used as keys and values. The attention distribution of that parameter to the other parameters is calculated. The magnitude of the attention weights reflects the strength of the correlation between parameters; a high weight indicates a strong coupling relationship between the two parameters. The average attention weights between all parameter pairs are collected to construct a parameter coupling strength matrix, where the rows and columns of the matrix correspond to different parameters, and the element values ​​represent the coupling strength of the corresponding parameter pair.

[0083] The multi-head attention mechanism employs multiple attention heads (preferably eight) for parallel computation. Each head learns different parameter association patterns, and the results from the multiple heads are concatenated and fused through a linear transformation. Through attention weighting, each parameter feature is enhanced by other parameter features, forming a cross-modal enhanced feature representation. This representation integrates multi-parameter comprehensive information, providing a foundation for accurately predicting multi-parameter collaborative changes. The resulting dynamic parameter coupling strength matrix and cross-modal enhanced feature representation are then obtained.

[0084] Cross-modal augmentation features and partition-level features output from graph neural networks are concatenated and fused to form comprehensive predictive input features. An encoder-decoder architecture prediction network is constructed, where the encoder encodes the input features into hidden state vectors, compressing historical context information and parameter coupling relationships within the hidden states. The encoder employs a multi-layer LSTM or GRU recurrent neural network, capable of capturing long-term dependencies in time series data.

[0085] The decoder starts from the hidden state and progressively decodes to generate predicted values ​​for environmental parameters at future time steps. The decoder also employs a recurrent neural network structure, where the input for each time step is the predicted output and hidden state of the previous time step. The next time step prediction is generated through iterative updates. The decoder's output layer is a fully connected layer that maps the hidden state to various environmental parameters and predicted values ​​for each partition. The output dimension is the number of partitions multiplied by the number of parameter types multiplied by the number of future prediction time steps.

[0086] During the decoding process, a parameter coupling strength matrix is ​​introduced as a constraint. Parameters with high coupling strength have a stronger mutual influence during prediction. Coupling constraints are implemented through attention or gating mechanisms. During training, historical data and future true values ​​are used as supervision signals. The loss function is the mean square error between the predicted and true values, summed over all partitions, all parameters, and all future time steps. Backpropagation and gradient descent are used to optimize the network parameters, resulting in a well-trained prediction network model.

[0087] The model inference process takes the current moment and historical environmental characteristics as input and outputs predicted time series of parameters such as temperature, humidity, CO2 concentration, and O2 concentration for each region within the next 4 hours. The environmental change prediction results are multi-dimensional time series, including regional, parameter, and time dimensions, comprehensively describing the future spatiotemporal evolution trend of the environment. Obtaining multi-regional, multi-parameter future time series predictions provides a basis for predicting future environmental changes in regulation strategies.

[0088] Step 5: Based on the environmental change prediction results, perform multi-scale convolution and temporal attention classification to obtain the growth stage identification results; Specifically, the following steps are included: Step 5.1, Adaptive image enhancement and fine-tuning of the transfer learning model; Images of mushroom spawn captured by high-definition cameras are affected by lighting conditions at different times and locations, resulting in significant differences in image brightness and contrast. An adaptive histogram equalization algorithm is used to enhance the images. This algorithm dynamically adjusts pixel values ​​based on the local histogram distribution of the image, ensuring uniform brightness and moderate contrast across all regions.

[0089] The image is converted from the RGB color space to the HSV color space. The HSV space separates color and luminance information; the H channel represents hue, the S channel represents saturation, and the V channel represents luminance. Histogram equalization is applied to the V channel to eliminate luminance differences, the S channel is moderately enhanced to highlight color features, and the H channel remains unchanged to preserve true hue information. White balance correction is then performed to eliminate color casts caused by different light source color temperatures. The white balance algorithm identifies white or gray areas in the image and adjusts the gain of the RGB three channels to make the RGB values ​​of white areas equal, thereby correcting the overall color cast. This results in a standard mushroom stick image with uniform brightness and accurate color.

[0090] A deep convolutional neural network model pre-trained on a large-scale image dataset is used as the base model, preferably employing mature network architectures such as ResNet, as the pre-trained model has already learned general visual features. The last few classification layers of the pre-trained model are replaced with new fully connected layers, and the number of output categories of the new classification layers is set to the number of categories of edible fungi growth stages, including six categories: mycelial germination stage, mycelial growth stage, mycelial maturity stage, primordium formation stage, fruiting body growth stage, and mature harvesting stage.

[0091] Fine-tuning was performed using a collected dataset of edible fungus growth images. The dataset contained labeled image samples at each growth stage, with each sample labeled with its respective growth stage category. The fine-tuning strategy involved freezing the first few layers of the pre-trained model to retain its general feature extraction capabilities, training only the later layers and the newly added classification layer, thus adapting the model to the specific visual features of edible fungi. Training employed a cross-entropy loss function and a gradient descent optimizer, with a learning rate of 0.0001 preferred to avoid disrupting the pre-trained weights, and 50 training epochs preferred. Data augmentation techniques such as random cropping, rotation, and flipping were used to expand the training samples and improve the model's generalization ability.

[0092] The environmental change prediction results are aligned by region and shooting time, and used as auxiliary contextual features to be fused with image features in the classification head. A temporal attention mechanism is introduced to weight and aggregate the prediction values ​​at different time steps, highlighting environmental change patterns that contribute significantly to the current stage's judgment, thereby improving classification stability and foresight.

[0093] Step 5.2, Multi-task learning and growth health status assessment; Based on the growth stage classification model, a multi-task learning branch is extended. In addition to the stage classification task, tasks such as mycelial density regression, mycelial uniformity assessment, color health determination, and fruiting body detection are added. Multiple tasks share the convolutional neural network feature extraction layer, and each task's dedicated output layer is connected to a higher layer of the network.

[0094] The mycelial density regression task outputs a continuous value from 0 to 1 representing mycelial coverage, using a regression loss function such as mean squared error. The mycelial uniformity task assesses the evenness of mycelial distribution on the substrate surface, calculating a uniformity score by analyzing the spatial variance or distribution entropy of mycelial density. The color health task determines whether the mycelial color is normal. Healthy mycelium is white or the normal color for a specific variety; abnormal colors such as yellow or black indicate poor growth or contamination. Color feature analysis and classification methods are used to determine health.

[0095] The sub-entity detection task employs an object detection network to identify the location and size of sub-entities in an image, count the number of sub-entities, measure their diameter, and assess their developmental status. During multi-task joint training, the weighted sum of the losses from each task is used as the total loss, with weights set according to task importance. Multi-task learning leverages the correlation between tasks to mutually promote each other, improving the overall performance of the model.

[0096] Environmental prediction context gating is introduced into the multi-task branch: the outputs of each task are dynamically weighted and corrected based on the predicted trajectories of temperature, humidity, CO2, and O2 for the next 4 hours. Environmental risk factors are added to the comprehensive score of growth health status; if the prediction indicates a risk of exceeding the limits, the score is reduced or the relevant warning weight is increased; when the prediction shows that the environment will stabilize, the score stability item is increased accordingly, so that the score has both real-time and forward-looking characteristics.

[0097] The model inference process takes a standard mushroom stick image as input and outputs multiple growth indicators, including growth stage category, mycelial density value, mycelial uniformity score, color health level, and fruiting body detection results. These multi-dimensional indicators are then integrated, weighted according to the current growth stage, and weighted summed and normalized to a score of 0 to 100 to obtain a comprehensive growth health status score. A higher score indicates that the growth status is closer to the ideal state. This yields accurate growth stage identification results and a quantified growth health status score.

[0098] Step 6: Based on the growth stage identification results, perform deep reinforcement learning and multi-objective optimization to obtain the partitioned collaborative regulation strategy instruction set; Specifically, the following steps are included: Step 6.1, Knowledge Graph Demand Query and Deep Reinforcement Learning Decision; A knowledge graph for edible mushroom cultivation was constructed, containing optimal environmental parameters for different edible mushroom varieties and growth stages. This knowledge is derived from the experience of cultivation experts, scientific literature, and historical best-practice cultivation data. The knowledge graph is organized in a graph structure, with nodes including variety nodes, growth stage nodes, and parameter nodes. Edges represent relationships between nodes, such as the stages included in a variety and the parameters required for each stage.

[0099] Based on the currently cultivated edible fungi varieties (e.g., shiitake mushrooms) and the identified growth stages (e.g., mycelial growth stage), corresponding environmental requirements are queried in a knowledge graph. The query results include the optimal temperature range, optimal humidity range, optimal CO2 concentration range, optimal O2 concentration range, and optimal light intensity range for that stage. The requirements differ significantly across growth stages: the mycelial growth stage requires higher temperatures and CO2 concentrations to promote metabolism; the primordia differentiation stage requires temperature differences and lower CO2 concentrations; and the fruiting body growth stage requires moderate temperatures and increased light. This yields a stage-specific and variety-specific set of environmental requirements knowledge.

[0100] The system encodes multi-dimensional information such as current measured values ​​of environmental parameters in each zone, predicted values ​​of future environmental parameters, ranges of environmental demand parameters, growth health status scores, mushroom density, and historical regulatory actions and effects into a state vector. This state vector has a high dimensionality, containing hundreds of feature components, comprehensively describing the current system state.

[0101] The Deep Deterministic Policy Gradient (DDPG) algorithm is employed, which is suitable for reinforcement learning problems with continuous action spaces. The algorithm comprises an Actor network and a Critic network. The Actor network outputs actions based on the state, while the Critic network evaluates the value of state-action pairs. The Actor network takes a state vector as input, passes it through a multi-layer fully connected neural network, and outputs control action parameters for each zone. These actions include continuous values ​​such as target temperature setpoints, target humidity setpoints, ventilation volume adjustment coefficients, and light intensity adjustment coefficients.

[0102] The Critic network takes a state and an action as input and outputs a Q-value estimate, which represents the expected long-term cumulative reward for performing the action in that state. During training, the Actor network policy is optimized by maximizing the Q-value, and the Critic network is optimized by minimizing the Q-value prediction error. Training data comes from historical regulation experience and simulation data. Through extensive iterative training, the policy network learns the mapping relationship from states to optimal actions. During model inference, the current state vector is input, and the Actor network outputs a preliminary regulation action plan, which theoretically maximizes long-term rewards, i.e., optimizes growth effect and energy efficiency.

[0103] Step 6.2, Multi-objective Pareto optimization and instruction set generation; Verify whether the initial control actions meet physical constraints, including equipment capacity constraints (e.g., heater maximum power), adjustment rate constraints (e.g., maximum temperature rise / fall per hour), and safety range constraints (e.g., temperature must not exceed the safety upper limit). Correct any action parameters that exceed the constraints by either trimming the out-of-limit parameters to the constraint boundaries or breaking down large-amplitude adjustments into multiple smaller adjustments performed step-by-step. This yields a feasible control scheme that meets the physical constraints.

[0104] A multi-objective evaluation system is established, with objectives including maximizing growth rate, minimizing energy consumption cost, maximizing environmental stability, and minimizing inter-regional interference. The growth rate objective is calculated based on the degree of similarity between the post-regulation environment and the optimal environment, and the growth rate predicted by the growth model. The energy consumption objective is calculated based on the power consumption and operating time of the equipment corresponding to the regulation action. The environmental stability objective is calculated based on the fluctuation range of environmental parameters before and after regulation. The inter-regional interference objective is calculated based on the degree of impact of regulation on the environment of non-target zones.

[0105] For each feasible control scheme, the objective function values ​​are calculated, forming a multi-objective vector. A Pareto optimal solution selection algorithm is used to compare the multi-objective vectors of all candidate schemes, selecting the Pareto front solution set. A Pareto optimal solution is one in which no other scheme is inferior to it in all objectives and superior in at least one objective. The front solution set represents the optimal trade-offs among the multiple objectives.

[0106] Based on the current operational mode priority settings, the final solution is selected from the Pareto front solution set. If the current period is a critical growth phase, growth rate is prioritized; if it is an energy consumption control phase, energy efficiency is prioritized. A weighted scoring method is used to convert the multi-objective vector into a comprehensive score, with weights set according to priority. The solution with the highest comprehensive score is selected as the optimal control strategy.

[0107] The optimal control strategy is translated into equipment control commands, including specific parameters such as heater power percentage, fan speed (Hertz), humidifier spray duration (in seconds), and supplementary lighting brightness level. Considering the coordinated timing of multiple devices, basic environmental adjustments such as ventilation and temperature are executed first, followed by fine-tuning adjustments such as humidity and light intensity, to avoid offsetting control actions. This results in a time-sequential, zoned coordinated control strategy command set, which includes the control parameters and execution sequence for each device.

[0108] Step 7: Based on the instruction set of the partitioned collaborative control strategy, perform online incremental learning optimization to obtain an adaptive optimized partitioned collaborative control strategy; Specifically, the following steps are included: Step 7.1, transfer reinforcement learning pre-training; During the initial deployment phase of the system, due to the lack of historical regulatory experience data for the current scenario, directly training the reinforcement learning policy network faces a cold start problem, requiring a large amount of random exploration, which may lead to poor regulatory effects or even damage to the growth of the mushroom sticks. To solve this problem, a transfer reinforcement learning method is adopted to accelerate policy learning by utilizing existing experience from other scenarios.

[0109] We collected historical environmental control experience data from other edible mushroom cultivation scenarios or similar facility agriculture settings. This data included environmental conditions, control actions implemented, resulting environmental changes, and growth effects under different scenarios. Although the specific parameter values ​​differed across scenarios, the basic principles of environmental control were common, such as heating for warming, ventilation for cooling, misting for humidification, and dehumidification for dehumidification.

[0110] Deep reinforcement learning networks are pre-trained on source domain data, ensuring sufficient data volume for adequate training. The pre-trained policy network learns general regulatory knowledge and decision-making rules, including how to select regulatory directions based on environmental deviations, how to balance multi-objective conflicts, and how to avoid over-regulation. While the pre-trained network achieves good performance on source domain tasks, direct application to target scenarios may present adaptation issues.

[0111] Transfer learning techniques are employed to adapt a pre-trained network to the current cultivation scenario. The method involves freezing the network's bottom-level general feature extraction layers and fine-tuning only the top-level decision output layers. This allows the network to adapt to the specific environmental characteristics of the current scenario, such as workshop layout, equipment configuration, and strain characteristics. Fine-tuning uses a small amount of initial data from the current scenario and converges quickly on a limited number of samples using fast adaptation algorithms such as meta-learning. This results in an initial policy network adapted to the scenario. This network combines general knowledge from the source domain with specific knowledge for fine-tuning in the target domain, demonstrating good controllability even in the initial system operation phase. This addresses the problem of poor policy quality requiring extensive exploration and trial-and-error during the cold start phase of reinforcement learning.

[0112] Step 7.2, online incremental learning and priority experience replay optimization; During actual operation, the environmental prediction model and policy network need to be continuously optimized based on the actual feedback data of the current scenario to adapt to the specific dynamic characteristics of the environment and the response characteristics of the equipment. The environmental prediction model and policy network are optimized online, with both running in parallel and mutually reinforcing each other.

[0113] First, the environmental prediction model is optimized through online incremental learning. After the control strategy is implemented, environmental parameter change data for each zone are continuously collected through a multi-source sensor network to form the actual environmental response time series. The actual response data is compared with the predicted values ​​generated by the prediction model before the control is implemented, and the prediction error of each parameter at each time point is calculated. The prediction error includes statistical indicators such as root mean square error, mean absolute error, and maximum deviation.

[0114] Analyze the distribution characteristics of prediction errors to identify which parameters, regions, and time periods have larger prediction errors. Possible reasons for larger errors include insufficient model learning of the scene's features, the presence of unmodeled special factors in the area such as equipment malfunctions or changes in mushroom stick layout, and sudden environmental changes exceeding the model's training data range. Focus on analyzing high-error areas and extract error patterns to provide a basis for model improvement.

[0115] An online incremental learning method is employed to update the prediction model. Incremental learning uses the latest real-world data as training samples and continuously updates the model parameters through mini-batch gradient descent. Incremental learning does not require retraining the entire model; it only fine-tunes the model parameters, making it computationally efficient and suitable for real-time systems. A relatively small learning rate is set to avoid excessive impact of new data on already learned knowledge, and momentum optimization is used to maintain learning stability.

[0116] Regularly evaluate model performance on the validation set, which includes recent historical data. When model performance metrics, such as prediction error, decrease by more than a preset threshold (preferably 5%), or after a preset update cycle (preferably 7 days), save the updated model parameters as the new version. Use A / B testing to compare the performance of the old and new model versions. If the new version significantly outperforms the old version and passes stability testing, execute the model update deployment, loading the new model parameters into the production system to replace the old model. If performance degrades or anomalies occur after the model update, the system automatically rolls back to the previous stable version and records the anomaly log for analysis.

[0117] While continuously optimizing the environmental prediction model, the reinforcement learning policy network is optimized in parallel, enabling the network to continuously improve its decision-making capabilities based on feedback from actual control effects. A quantitative evaluation value of the control effect is calculated based on indicators such as changes in growth health status scores before and after control, environmental parameter compliance rates, total equipment energy consumption, and control time efficiency.

[0118] A hierarchical reward function is designed to map evaluation metrics to reinforcement learning reward signals. Improved growth status is the primary reward, reduced energy consumption is the secondary reward, environmental parameters exceeding safe ranges are given a large negative reward, and drastic environmental fluctuations are given a moderate negative reward. The total reward value for this regulation cycle is calculated by combining all rewards and penalties.

[0119] The current state, the action performed, the reward obtained, and the new state are combined into an experience quadruple and stored in the experience replay pool. The experience replay pool uses a priority sampling mechanism, determining the sampling priority based on the temporal difference error of the experience. Batch experience data is periodically sampled from the experience replay pool, and these experiences are used to update the Actor-Critic network.

[0120] Through multiple rounds of experience learning, the policy network is continuously improved, learning the optimal control strategy for the current scenario. The performance of the strategy is periodically evaluated on simulation environments or validation sets. When the average cumulative reward of a new strategy is significantly higher than that of the current strategy and it passes stability and security tests, the policy network is updated and deployed. Through continuous optimization, the quality of control strategy decisions has improved, enabling more accurate timing of control measures, more precise determination of control magnitude, and more reasonable coordination of control actions across multiple zones.

[0121] Through transfer learning pre-training, online optimization of the prediction model, and online optimization of the policy network, the system's environmental prediction and control decision-making capabilities are continuously improved. A continuously improved zone-based collaborative control strategy is obtained, which fully adapts to the dynamic characteristics of the current environment, equipment response characteristics, and strain growth characteristics. The control accuracy and efficiency continuously improve with accumulated operational experience, ultimately forming the optimal control scheme for the current cultivation workshop.

[0122] A system for regulating the growth environment of edible fungi based on multi-source sensor data is used to execute the aforementioned method for regulating the growth environment of edible fungi based on multi-source sensor data, such as... Figure 2 As shown, it includes: The multi-source sensor network module is used to deploy multi-source sensor networks, unify time synchronization, perform event-driven spatiotemporal alignment and 3D mapping, and obtain a complete spatiotemporally labeled multi-source sensor dataset. The evaluation and data fusion module, based on a multi-source sensor dataset with complete spatiotemporal annotation, performs multi-dimensional reliability evaluation and environmental correction, and uses fuzzy logic reasoning to obtain an environmental parameter fusion dataset. The spatiotemporal distribution modeling and feature extraction module performs spatiotemporal kriging interpolation based on the environmental parameter fusion dataset and uses a multi-source data collaborative correction method to obtain a heterogeneous microenvironment feature database. The multi-parameter coupled prediction module performs cross-modal attention fusion based on a heterogeneous microenvironment feature database to obtain environmental change prediction results; The growth stage identification and evaluation module performs multi-scale convolution and temporal attention classification based on the environmental change prediction results to obtain the growth stage identification results; The regulation strategy generation module, based on the growth stage identification results, performs deep reinforcement learning and multi-objective optimization to obtain a set of regional collaborative regulation strategy instructions; The adaptive optimization module performs online incremental learning and optimization based on the partitioned collaborative control strategy instruction set to obtain an adaptively optimized partitioned collaborative control strategy.

[0123] In one embodiment of the present invention, a specific example is provided: This invention was applied to 20 standardized cultivation workshops of a large-scale shiitake mushroom factory, each with an area of ​​1000 square meters. A 180-day field test was conducted at a factory cultivation base in a city in Fujian Province. During the test, a complete multi-source sensor network was deployed, covering an area of ​​20,000 square meters and involving approximately 500,000 shiitake mushroom logs. The system ran continuously for 6 months, accumulating massive amounts of environmental monitoring data and growth status data.

[0124] The real-time data collected by the multi-source sensors are shown in Table 1: Table 1: Real-time data acquired by multi-source sensors;

[0125] The environmental parameter fusion and partition feature extraction data are shown in Table 2: Table 2: Environmental parameter fusion and zoning feature extraction data;

[0126] Through the multi-source sensor data fusion, spatiotemporal distribution modeling, deep learning prediction, and reinforcement learning control technologies of this invention, precise and intelligent control of complex factory-style cultivation environments is achieved.

[0127] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for regulating the growth environment of edible fungi based on multi-source sensor data, characterized in that, Includes the following steps: Deploy a multi-source sensor network, unify time synchronization, perform event-driven spatiotemporal alignment and 3D mapping, and obtain a complete spatiotemporally labeled multi-source sensor dataset; Based on a multi-source sensor dataset with complete spatiotemporal annotation, multidimensional reliability assessment and environmental correction are performed, and a fuzzy logic reasoning method is used to obtain a fusion dataset of environmental parameters. Based on the environmental parameter fusion dataset, spatiotemporal kriging interpolation is performed, and a multi-source data collaborative correction method is used to obtain a heterogeneous microenvironment feature database. Based on a heterogeneous microenvironment feature database, cross-modal attention fusion is performed to obtain environmental change prediction results; Based on the environmental change prediction results, multi-scale convolution and temporal attention classification are performed to obtain the growth stage identification results. Based on the growth stage identification results, deep reinforcement learning and multi-objective optimization are performed to obtain a set of instructions for the regional collaborative regulation strategy. Based on the instruction set of the partitioned collaborative control strategy, online incremental learning optimization is performed to obtain an adaptive optimized partitioned collaborative control strategy.

2. The method for regulating the growth environment of edible fungi based on multi-source sensor data according to claim 1, characterized in that, The event-driven spatiotemporal alignment and 3D mapping include: Deploy the Network Time Protocol service on the edge computing server to periodically synchronize with the time server of the time service center via the Internet. All sensor nodes perform clock calibration with the Network Time Protocol server at startup to obtain a unified standard time as the time reference. Extract the time series of measurement values ​​from each sensor from the time series database, perform first-order difference calculation on the time series to obtain the rate of change at each moment, and mark the event as an environmental mutation event when the absolute value of the rate of change exceeds the preset rate of change threshold. Extract the characteristic moments of the same event in the response time series of different sensors. The least squares optimization algorithm is adopted to minimize the time deviation of the same event after time delay correction at different sensor response times, and the time delay correction parameters of each sensor are iteratively optimized. Subtract the inherent response delay and dynamic correction delay from the original timestamp to obtain the corrected synchronization timestamp.

3. The method for regulating the growth environment of edible fungi based on multi-source sensor data according to claim 1, characterized in that, The event-driven spatiotemporal alignment and 3D mapping also include: The time axis is divided into preset time periods. The first and second derivatives of the measured values ​​are calculated in each time period. When the environmental parameters change steadily, the linear interpolation method is used to interpolate the low-frequency sampling data. When the environmental parameters change drastically, the cubic spline interpolation method is used to interpolate all low-frequency sampled sensor data so that the interpolation target frequency reaches the highest sampling frequency among all sensors. The three-dimensional spatial coordinates of each sensor are read from the sensor network configuration database, and the three-dimensional space of the cultivation workshop is divided into a regular cubic grid. For point sensors, the spatial grid cell to which they belong is determined based on the installation coordinates. For area sensors, a perspective projection transformation model is used to convert the pixel coordinates of the image plane into spatial coordinates based on the camera calibration parameters. Add a spatial labeling field to each sensor data record to obtain a multi-source sensor dataset with complete spatiotemporal labeling.

4. The method for regulating the growth environment of edible fungi based on multi-source sensor data according to claim 1, characterized in that, The multidimensional reliability assessment includes: Extract historical measurement data of each sensor within a preset time window from the time series database, and calculate the mean, standard deviation, and coefficient of variation of the measured values; The trend curve is obtained by eliminating short-term fluctuations using moving average filtering, and the drift rate is obtained by calculating the deviation between the original measurement sequence and the trend curve. Detect abnormal jumps in historical measurement sequences and count the number of jumps. Use the Z-score normalization method to calculate the relative position of the sensor in the population. The historical stability reliability score was calculated using a weighted scoring method that combines the coefficient of variation, drift rate, jump frequency, and Z-score for population comparison. Retrieve other similar sensors within the spatial neighborhood of the target sensor from a fully spatiotemporally labeled multi-source sensor dataset, and calculate the median and interquartile range of the neighborhood measurements. A Gaussian function is used to map the bias to a confidence score, thus obtaining the spatial consistency confidence score.

5. The method for regulating the growth environment of edible fungi based on multi-source sensor data according to claim 1, characterized in that, The environmental parameter fusion dataset obtained using fuzzy logic reasoning includes: Read the device file information of each sensor, calculate the ratio of the current usage time of the sensor to its designed lifespan, and use a multi-factor weighted model to calculate the device health score; Historical stability confidence score, spatial consistency confidence score, and equipment health score are used as input variables for fuzzy logic inference. Fuzzy sets and membership functions are defined for each input variable. A fuzzy inference rule base is established. The activation intensity of each rule is calculated by traversing the inference rule base. The fuzzy sets of the consequents of all activated rules are weighted and aggregated. The output fuzzy set is defuzzified to obtain the comprehensive confidence weight value. A threshold screening method is applied to the comprehensive reliability weight. For reliable sensor data with a reliability weight higher than the threshold, a weighted average fusion algorithm is used to calculate the fused environmental parameter values.

6. The method for regulating the growth environment of edible fungi based on multi-source sensor data according to claim 1, characterized in that, The spatiotemporal kriging interpolation includes: The fusion values ​​of various environmental parameters at all times and in all spatial grid cells are extracted from the environmental parameter fusion dataset to form a discrete set of spatiotemporal sampling points; Calculate the spatiotemporal distance and the square of the difference in measurements between all sampling point pairs, and use anisotropy measures to calculate the spatiotemporal distance; Half of the squared difference in the measured values ​​is taken as the variability function value, and empirical variability functions are obtained by grouping and statistically analyzing the spatiotemporal distance. The theoretical variogram model is fitted to the empirical variogram, and the Kriging equations are constructed and solved for the target spatiotemporal location to be interpolated to obtain the weight coefficients of each sampling point.

7. A system for regulating the growth environment of edible fungi based on multi-source sensor data, characterized in that, A method for regulating the growth environment of edible fungi based on multi-source sensor data, as described in any one of claims 1-9, includes: The multi-source sensor network module is used to deploy multi-source sensor networks, unify time synchronization, perform event-driven spatiotemporal alignment and 3D mapping, and obtain a complete spatiotemporally labeled multi-source sensor dataset. The evaluation and data fusion module, based on a multi-source sensor dataset with complete spatiotemporal annotation, performs multi-dimensional reliability evaluation and environmental correction, and uses fuzzy logic reasoning to obtain an environmental parameter fusion dataset. The feature extraction module performs spatiotemporal kriging interpolation based on the environmental parameter fusion dataset and uses a multi-source data collaborative correction method to obtain a heterogeneous microenvironment feature database. The multi-parameter coupled prediction module performs cross-modal attention fusion based on a heterogeneous microenvironment feature database to obtain environmental change prediction results; The growth stage identification and evaluation module performs multi-scale convolution and temporal attention classification based on the environmental change prediction results to obtain the growth stage identification results; The regulation strategy generation module, based on the growth stage identification results, performs deep reinforcement learning and multi-objective optimization to obtain a set of regional collaborative regulation strategy instructions; The adaptive optimization module performs online incremental learning and optimization based on the partitioned collaborative control strategy instruction set to obtain an adaptively optimized partitioned collaborative control strategy.

Images