An artificial intelligence-based multi-parameter environment sensor chip data fusion method

Abstract

This invention provides a data fusion method for multi-parameter environmental sensor chips based on artificial intelligence, belonging to the field of agricultural product growth environment monitoring technology. This invention constructs a continuous data sequence using a multi-parameter environmental sensor chip, evaluates and preprocesses the data quality using anomaly jump matrices and interference anomaly matrices, constructs a multi-dimensional fusion algorithm to calculate convergence, stability, and disorder indices to quantify the data fusion state, and employs an adaptive calibration algorithm combined with jump compensation and anomaly compensation matrices to dynamically compensate for sensor baseline drift and sensitivity attenuation in real time. The number of attention approximation features is dynamically adjusted through a feature adjustment function, ultimately achieving classification prediction and quality assessment of agricultural product growth environment. This invention solves the technical problem that baseline drift and sensitivity attenuation in multi-parameter environmental sensors significantly reduce data fusion accuracy over time during long-term operation.

Description

Technical Field

[0001] This invention belongs to the field of agricultural product growth environment monitoring technology, and specifically relates to a data fusion method for multi-parameter environmental sensor chips based on artificial intelligence. Background Technology

[0002] In the field of agricultural environmental monitoring, multi-parameter environmental sensor chips, by integrating multiple sensor units such as temperature, pH, dissolved oxygen, and conductivity, can collect key indicator data of the agricultural product growth environment in real time, providing data support for precision agriculture. Traditional sensor calibration techniques mainly employ periodic manual calibration, fixed-cycle calibration, or static calibration methods based on simple linear compensation algorithms. These methods are widely deployed in long-term continuous monitoring applications such as greenhouses, farmland irrigation systems, and soil monitoring networks, correcting sensor outputs through preset compensation parameters. However, traditional calibration methods suffer from drawbacks such as the inability to track sensor performance degradation in real time and the inability to dynamically update fixed compensation parameters. This often leads to a gradual decrease in calibration accuracy and a continuous decline in data reliability during long-term operation. In existing technologies, due to the lack of effective real-time calibration parameter update mechanisms and adaptive compensation capabilities for sensor performance degradation, multi-parameter sensor systems inevitably experience baseline drift and sensitivity attenuation during long-term operation. This is especially true under harsh environmental conditions and high-frequency usage, where fixed calibration parameter compensation methods cannot adjust in a timely manner according to changes in actual sensor performance. In other words, existing technologies have the technical problem that baseline drift and sensitivity decay of multi-parameter environmental sensors during long-term operation lead to a significant decrease in data fusion accuracy over time. Summary of the Invention

[0003] In view of this, the present invention provides a data fusion method for multi-parameter environmental sensor chips based on artificial intelligence, which can solve the technical problem in the prior art where baseline drift and sensitivity decay of multi-parameter environmental sensors cause a significant decrease in data fusion accuracy over time.

[0004] This invention is implemented as follows: It provides an artificial intelligence-based multi-parameter environmental sensor chip data fusion method. This method collects temperature, pH, dissolved oxygen, and conductivity sensor data from a multi-parameter environmental sensor chip, establishes a sensor data acquisition matrix, and constructs a continuous data sequence using a time window sliding mechanism. The continuous data sequence is preprocessed, and anomaly detection matrices are used to detect abrupt changes. An interference anomaly matrix is ​​used to identify electromagnetic interference signals and cross-sensitivity signals, establishing a data quality assessment system. A multi-dimensional fusion algorithm is constructed to calculate the convergence index, stability index, and disorder index of the continuous data sequence. A deep fusion model is established to extract features and perform correlation analysis on the preprocessed continuous data sequence. The number of attention approximate features is dynamically adjusted based on the convergence index, stability index, and disorder index using a feature adjustment function. An adaptive calibration algorithm is used to compensate for sensor baseline drift and sensitivity attenuation. An environmental state recognition model is established, and a pattern matching algorithm is used to classify and predict the growth environment of agricultural products, outputting environmental quality assessment results.

[0005] Specifically, the time window sliding mechanism involves performing time-series processing on the sensor data acquisition matrix, converting temperature sensor data, pH sensor data, dissolved oxygen sensor data, and conductivity sensor data into a data sequence with temporal continuity.

[0006] The abnormal jump matrix refers to a two-dimensional array structure used to detect abnormal sudden changes in sensor data. The matrix elements represent the jump detection results of each parameter at different time points, which are used to identify equipment failures or external interference.

[0007] The interference anomaly matrix refers to a data structure used to identify electromagnetic interference and cross-sensitivity between multi-parameter sensors. The matrix rows and columns correspond to different parameters, and the element values ​​represent the interference intensity and type between parameters.

[0008] The data quality assessment system is established based on the detection results of the abnormal jump matrix and the interference anomaly matrix. It is used to evaluate the reliability and validity of sensor data and provide quality assurance for subsequent data processing steps.

[0009] The calculation steps of the convergence index are specifically a quantitative indicator of the consistency of multi-parameter sensor data during the fusion process. It is obtained by calculating the reciprocal of the standard deviation of the rate of change of each parameter within a continuous time window. The larger the value, the better the convergence of data fusion.

[0010] Specifically, the stability index is a quantitative indicator of the degree of fluctuation of sensor measurements over a time period. It is obtained by calculating the ratio of the mean to the standard deviation of the parameters and is used to evaluate the reliability and reproducibility of the sensor output.

[0011] Specifically, the calculation steps of the disorder index are a quantitative indicator of the degree of mutual influence and interference between multiple parameters. It is obtained by calculating the condition number of the correlation coefficient matrix between parameters. The smaller the value, the better the coordination between parameters.

[0012] The deep fusion model is structured as a multi-layer attention network based on the Performer architecture, comprising an input embedding layer, a multi-head attention layer, a feedforward neural network layer, and an output projection layer. The attention approximation mechanism employs a random feature mapping method to reduce computational complexity.

[0013] The feature adjustment function is used to adjust the number of attention approximation features in the deep fusion model. The adjustment factor is calculated based on the convergence index, stability index, and disorder index. The balance between computational efficiency and fusion accuracy is optimized by dynamically adjusting the number of attention approximation features.

[0014] The adaptive calibration algorithm compensates for sensor baseline drift and sensitivity decay, and updates calibration parameters in real time using jump compensation matrix and anomaly compensation matrix.

[0015] The jump compensation matrix refers to a set of compensation parameters used to correct sudden changes in sensor data. It is established through statistical analysis of historical data and includes the jump compensation coefficients of each parameter under different environmental conditions.

[0016] The anomaly compensation matrix refers to the parameter adjustment matrix used to correct abnormal sensor outputs. It is established based on reference data under normal operating conditions and includes compensation algorithm parameters corresponding to various abnormal situations.

[0017] The pattern matching algorithm is used to process the environmental state recognition model. By comparing the degree of matching between real-time monitoring data and preset patterns, it realizes the automatic classification of agricultural product growth environment and prediction of future state.

[0018] Specifically, the training dataset establishment step of the deep fusion model involves collecting historical data from multi-parameter sensors under different agricultural environments. The data covers the four seasons of spring, summer, autumn, and winter, as well as different crop growth cycles. Each sample contains continuous monitoring data from temperature sensor, pH sensor, dissolved oxygen sensor, and conductivity sensor.

[0019] The process includes establishing a feedback optimization mechanism after the environmental condition identification model is established. This mechanism adjusts the data fusion parameters based on the environmental quality assessment results and improves the accuracy of multi-parameter sensor data processing and system stability through iterative optimization.

[0020] This invention addresses the technical problem of traditional fixed calibration methods' inability to dynamically track sensor performance degradation by constructing an adaptive calibration algorithm and combining jump compensation and anomaly compensation matrices to update calibration parameters in real time. The invention employs a compensation parameter set established based on historical data statistical analysis, which can automatically identify baseline drift and sensitivity attenuation phenomena according to the actual operating state of the sensor. By dynamically adjusting the compensation algorithm parameters, it achieves real-time compensation for sensor performance degradation, effectively maintaining stable data fusion accuracy during long-term operation and overcoming the insufficient compensation capability caused by fixed calibration parameters in existing technologies. In summary, this invention, by establishing a real-time dynamic calibration mechanism, solves the technical problem that baseline drift and sensitivity attenuation in multi-parameter environmental sensors significantly reduce data fusion accuracy over time during long-term operation. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a neural network structure diagram of the deep fusion model involved in the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] like Figure 1 The diagram shown is a flowchart of a multi-parameter environmental sensor chip data fusion method based on artificial intelligence provided by the present invention. This method includes the following steps:

[0025] S01. Collect temperature sensor data, pH sensor data, dissolved oxygen sensor data, and conductivity sensor data from a multi-parameter environmental sensor chip. Combine the temperature sensor data, pH sensor data, dissolved oxygen sensor data, and conductivity sensor data to establish a sensor data acquisition matrix. Process the sensor data acquisition matrix through a time window sliding mechanism to construct a continuous data sequence.

[0026] S02. Preprocess the continuous data sequence, use an anomaly transition matrix to detect abrupt changes in the continuous data sequence, use an interference anomaly matrix to identify electromagnetic interference signals and cross-sensitive signals in the continuous data sequence, and establish a data quality assessment system based on the detection results of the anomaly transition matrix and the interference anomaly matrix.

[0027] S03. Construct a multi-dimensional fusion algorithm, evaluate the stability of the data fusion process by calculating the convergence index of the continuous data sequence, quantify the consistency of sensor measurement results by using the stability index of the continuous data sequence, and measure the coordination between the temperature sensor data, pH sensor data, dissolved oxygen sensor data, and conductivity sensor data by using the disorder index of the continuous data sequence.

[0028] S04. Establish a deep fusion model to extract features and perform correlation analysis on the preprocessed continuous data sequence. The number of attention approximation features in the deep fusion model is dynamically adjusted according to the convergence index, the stability index, and the disorder index through a feature adjustment function.

[0029] S05. An adaptive calibration algorithm is used to compensate for the baseline drift and sensitivity attenuation of the sensor, and the calibration parameters of the adaptive calibration algorithm are updated in real time using the jump compensation matrix and the anomaly compensation matrix.

[0030] S06. An environmental state identification model is established through the fused continuous data sequence. A pattern matching algorithm is used to process the environmental state identification model to classify and predict the growth environment of agricultural products and output the environmental quality assessment results.

[0031] S07. Optionally, it also includes establishing a feedback optimization mechanism to adjust the data fusion parameters based on the environmental quality assessment results, and improve the accuracy of multi-parameter sensor data processing and system stability through iterative optimization.

[0032] The convergence index is a quantitative indicator of the consistency of multi-parameter sensor data during fusion. It is obtained by calculating the reciprocal of the standard deviation of the rate of change of each parameter within a continuous time window; a larger value indicates better data fusion convergence. The stability index is a quantitative indicator of the fluctuation of sensor measurements over a time period. It is obtained by calculating the ratio of the parameter mean to the standard deviation and is used to evaluate the reliability and reproducibility of sensor output. The disorder index is a quantitative indicator of the degree of mutual influence and interference between multiple parameters. It is obtained by calculating the condition number of the correlation coefficient matrix between parameters; a smaller value indicates better coordination between parameters. The abnormal jump matrix is ​​a two-dimensional array structure used to detect abnormal abrupt changes in sensor data. Matrix elements represent the jump detection results of each parameter at different time points, used to identify equipment failures or external interference. The interference anomaly matrix is ​​a data structure used to identify electromagnetic interference and cross-sensitivity between multi-parameter sensors. The matrix rows and columns correspond to different parameters, and the element values ​​represent the interference intensity and type between parameters. The jump compensation matrix is ​​a set of compensation parameters used to correct abrupt changes in sensor data. It is established through historical data statistical analysis and includes the jump compensation coefficients of each parameter under different environmental conditions. The anomaly compensation matrix is ​​a parameter adjustment matrix used to correct abnormal sensor outputs. It is established based on reference data under normal operating conditions and includes compensation algorithm parameters corresponding to various abnormal situations.

[0033] The deep fusion model is structured as a multi-layer attention network based on the Performer architecture, comprising an input embedding layer, a multi-head attention layer, a feedforward neural network layer, and an output projection layer. The attention approximation mechanism employs a stochastic feature mapping method to reduce computational complexity. The model has a total of 2.85 million parameters and supports a maximum sequence length of 2048 time steps. The training dataset for the deep fusion model was established by collecting historical multi-parameter sensor data under different agricultural environments. The data covers all four seasons and different crop growth cycles, totaling 780,000 valid samples. Each sample contains 48 hours of continuous monitoring data from the temperature sensor, pH sensor, dissolved oxygen sensor, and conductivity sensor, along with corresponding environmental quality labels and abnormal event markers. The data was divided into training, validation, and test sets in a 7:2:1 ratio. The training steps of the deep fusion model include first normalizing and serializing the original data, generating training sample pairs using the sliding window method, then updating the model parameters using the Adam optimizer with a learning rate of 0.0001, a batch size of 64, 200 training epochs, and a cosine annealing learning rate scheduling strategy. The model performance is monitored on the validation set, and an early stopping mechanism is triggered when the validation loss does not improve for 15 consecutive epochs. Finally, the model achieves a fusion accuracy of 94.7% on the test set.

[0034] The feature adjustment function is used to adjust the number of attention approximation features of the deep fusion model. The adjustment factor α is calculated based on the convergence index, the stability index, and the disorder index. When α∈[0.156, 0.432), a low feature number mode is adopted to set the number of attention approximation features to 128. When α∈[0.432, 0.778), a medium feature number mode is adopted to set the number of attention approximation features to 256. When α∈[0.778, 1.000], a high feature number mode is adopted to set the number of attention approximation features to 512. By dynamically adjusting the number of attention approximation features, a balance between computational efficiency and fusion accuracy is achieved.

[0035] The specific implementation methods of the above steps are described in detail below.

[0036] The specific implementation of step S01 involves first synchronously acquiring raw measurement data from temperature, pH, dissolved oxygen, and conductivity sensors via a data acquisition interface module. The sampling frequency for each sensor is set to 1Hz to ensure data timestamp consistency. Then, the data from the four types of sensors is organized into a four-dimensional data matrix according to time series, where rows represent time nodes and columns represent different sensor types. Next, a time window sliding mechanism is applied to segment the data matrix. The time window length is set to 60 minutes, and the sliding step size is set to 5 minutes. Overlapping window design ensures data continuity, ultimately constructing a continuous multi-dimensional time series dataset. The purpose of this step is to convert discrete multi-source sensor data into a structured continuous data sequence, laying the foundation for subsequent data fusion processing.

[0037] The specific implementation of step S02 involves constructing an anomaly jump matrix using a threshold-based anomaly jump identification algorithm. This is achieved by calculating the rate of change of data between adjacent time points in a continuous data sequence. When the rate of change exceeds a set threshold, a jump point is marked. The jump threshold is set to 5℃ / min for the temperature sensor, 0.5pH / min for the pH sensor, 2mg / L / min for the dissolved oxygen sensor, and 100μS / cm / min for the conductivity sensor. Simultaneously, a frequency domain analysis method is used to construct an interference anomaly matrix. The time domain signal is converted to a frequency domain signal using a Fast Fourier Transform (FFT) to identify 50Hz power frequency interference and other electromagnetic interference signals. When the frequency domain amplitude exceeds 150% of the normal signal amplitude, it is determined to be an interference signal. Based on the detection results of the anomaly jump matrix and the interference anomaly matrix, a comprehensive data quality assessment system is established, and a data reliability index is calculated. Data with a reliability index below 0.7 is marked as low-quality data. This step improves data quality through a multi-layered anomaly detection mechanism, providing reliable input data for accurate data fusion.

[0038] The specific implementation of step S03 involves constructing a multi-dimensional fusion algorithm evaluation system based on statistical principles. The convergence index is calculated by selecting the rate of change of each sensor parameter within a continuous time window, calculating the standard deviation of the rate of change sequence, and then taking the reciprocal of the standard deviation as the convergence index. The convergence index typically ranges from 0.1 to 10; a larger value indicates better convergence of the data fusion. The stability index is obtained by calculating the ratio of the mean to the standard deviation of the measured values ​​of each sensor parameter within a time period. A ratio greater than 3 indicates stable and reliable sensor output. The disorder index is based on the principle of multi-parameter correlation analysis. By constructing a Pearson correlation coefficient matrix between parameters, the condition number of this matrix is ​​calculated as a disorder index. A condition number less than 10 indicates good coordination between parameters, while a condition number greater than 100 indicates severe mutual interference between parameters. This step evaluates the stability and reliability of the data fusion process through a quantitative index system, providing a scientific basis for parameter adjustment of the fusion algorithm.

[0039] The specific implementation of step S04 involves constructing a deep fusion model based on the Performer architecture for feature extraction and correlation analysis. The model first converts multi-dimensional sensor data into high-dimensional feature vectors through an input embedding layer, with the embedding dimension set to 512. Then, a multi-head attention mechanism is used to calculate the correlation weights between different sensor parameters, with 8 attention heads. A random feature mapping method is used to achieve approximate linear complexity for attention computation. The feedforward neural network layer uses a two-layer fully connected structure with a hidden layer dimension of 2048, and the GELU activation function is used. The output projection layer maps the fused features to the target dimension output. The feature adjustment function calculates an adjustment factor α based on a weighted combination of convergence, stability, and disorder indices, with weights set to 0.4, 0.35, and 0.25, respectively. The number of attention approximation features is dynamically adjusted based on different intervals of α values ​​to achieve an adaptive balance between computational efficiency and accuracy. This step uses deep learning technology to uncover complex nonlinear correlations between multi-parameter sensor data, improving the accuracy and robustness of data fusion.

[0040] Step S05 specifically involves applying an adaptive calibration algorithm based on Kalman filtering to compensate for sensor baseline drift and sensitivity decay. The algorithm establishes a sensor state-space model, treating baseline drift and sensitivity decay as system state variables, and estimates the state transition matrix and observation matrix parameters using historical data. Baseline drift compensation calculates the drift rate using a linear trend estimation method; the baseline drift compensation threshold for temperature sensors is set to 0.1℃ / day, and for pH sensors, it is set to 0.02pH / day. Sensitivity decay compensation employs an exponential decay model, with the decay time constant set according to the sensor type, typically between 30 and 180 days. A jump compensation matrix stores compensation coefficients for various abnormal jumps, while an anomaly compensation matrix contains correction parameters for different anomaly modes. Both matrices dynamically update the calibration parameters based on real-time detection results. This step eliminates system errors during long-term sensor use through an adaptive compensation mechanism, ensuring the long-term stability and accuracy of the data fusion results.

[0041] The specific implementation of step S06 involves constructing an environmental state identification model using the fused continuous data sequence and employing a support vector machine (SVM) classification algorithm to achieve intelligent identification of the agricultural product growth environment. The environmental state classification includes five levels: excellent, good, average, poor, and severe. The classification feature vector consists of fused temperature, pH, dissolved oxygen, and conductivity data, along with their statistical characteristics, with a feature dimension of 32. The SVM uses a radial basis function kernel with a kernel parameter γ set to 0.01 and a regularization parameter C set to 100. A pattern matching algorithm predicts environmental quality by calculating the similarity between the current environmental state and historical standard patterns. The similarity calculation uses the cosine similarity method, with a similarity threshold set to 0.85. The environmental quality assessment results include the current environmental quality level, quality score, key influencing factors, and improvement suggestions. This step transforms complex multi-parameter sensor data into intuitive environmental quality assessment results, providing scientific decision support for agricultural production management.

[0042] The specific implementation of step S07 involves establishing a feedback optimization mechanism based on reinforcement learning principles to adjust data fusion parameters according to the accuracy of the environmental quality assessment results. The optimization mechanism uses a Q-learning algorithm to construct a parameter adjustment strategy. The state space includes the current fusion parameter configuration, data quality indicators, and assessment accuracy, while the action space includes increasing, decreasing, or maintaining various fusion parameters. The reward function is designed based on the matching degree between the assessment results and the actual environmental state; a matching degree higher than 90% is given a positive reward, and a matching degree lower than 70% is given a negative reward. The learning rate is set to 0.01, the discount factor is set to 0.95, and the exploration rate adopts an ε-greedy strategy, with an initial exploration rate of 0.3, gradually decreasing to 0.05 as training progresses. The iterative optimization process performs parameter updates every 100 assessment cycles, continuously improving the fusion algorithm performance through accumulated experience. This step achieves self-learning and continuous optimization of the data fusion system through a closed-loop feedback mechanism, enhancing the system's adaptability and long-term stability.

[0043] like Figure 2 As shown, the deep fusion model adopts a Performer variant structure based on the Transformer architecture, specifically including an input embedding layer, a position encoding layer, a multi-layer encoder module, and an output projection layer. The input embedding layer maps four-dimensional sensor data into a 512-dimensional feature vector through a linear transformation and adds learnable position encoding information to identify the positional relationships of the time series. The encoder module contains six identical encoding layers, each containing a multi-head self-attention sub-layer and a feedforward neural network sub-layer, using residual connections and layer normalization techniques to stabilize the training process. The multi-head self-attention mechanism uses eight attention heads, each with a 64-dimensional dimension. A random feature mapping method reduces the computational complexity of attention from quadratic to linear, and the random feature dimensions are dynamically set according to the feature adjustment function. The feedforward neural network uses a two-layer fully connected structure, with the middle layer having a 2048-dimensional dimension. The activation function is GELU, and the dropout ratio is set to 0.1 to prevent overfitting. The output projection layer maps the encoded feature vectors to environmental state classification results and uses a softmax activation function to output the probability distribution of each category.

[0044] The training dataset was established by first selecting representative agricultural production areas across the country, including typical farmland environments in different climate zones such as the North China Plain, the Yangtze River Basin, and South China. Data collection covered the complete growth cycle of major crops such as wheat, rice, corn, and vegetables, from pre-planting soil preparation to post-harvest environmental monitoring. A grid-based sensor deployment scheme was adopted, with each monitoring point equipped with a complete four-parameter sensor combination. Data was collected every 5 minutes, continuously monitoring for 48 hours to form a complete sample. Data preprocessing included outlier removal, missing value imputation, and data normalization to ensure data quality and consistency. Sample labeling was conducted by agricultural environmental experts based on a five-level quality assessment considering factors such as crop growth status, soil health indicators, and environmental suitability. The dataset was stratified according to chronological order and geographical distribution, with a training set, validation set, and test set ratio of 7:2:1 to ensure balanced data distribution across subsets. The training process employed a phased strategy: first, pre-training was performed to learn general sensor data representations; then, fine-tuning was conducted to optimize the environmental state classification task; and finally, adversarial training was used to improve the model's robustness and generalization ability.

[0045] The key technical ideas of this invention are mainly reflected in the following aspects. The first key technical idea is the construction of a multi-dimensional data quality assessment system. Through a dual detection mechanism of anomaly jump matrix and interference anomaly matrix, it can accurately identify various anomalies in sensor data. Compared with traditional single-threshold detection methods, this system can effectively distinguish different anomaly sources such as equipment failure, environmental mutations, and electromagnetic interference, significantly improving the accuracy and reliability of data quality assessment and providing high-quality input assurance for subsequent data fusion. The second key technical idea is a dynamic feature adjustment mechanism based on convergence index, stability index, and disorder index. By quantitatively evaluating multiple dimensions of the data fusion process, it achieves adaptive adjustment of the number of attention approximate features. Compared with traditional methods with a fixed number of features, this mechanism can automatically balance computational efficiency and fusion accuracy according to data characteristics, significantly reducing computational complexity while ensuring fusion accuracy, making it particularly suitable for resource-constrained embedded environments. The third key technical approach is the collaborative design of adaptive calibration algorithms and feedback optimization mechanisms. The adaptive calibration algorithm is based on the Kalman filtering principle to compensate for the baseline drift and sensitivity attenuation of the sensor, while the feedback optimization mechanism uses reinforcement learning to adjust the fusion parameters according to the evaluation results. Compared with traditional static calibration methods, this collaborative mechanism can enable the sensor system to learn itself and continuously optimize, effectively addressing the performance degradation problem during long-term use.

[0046] The synergistic effect of these key technological approaches forms a complete intelligent data fusion solution, exhibiting significant comprehensive advantages compared to existing technologies. Multi-dimensional data quality assessment provides a reliable evaluation basis for dynamic feature adjustment, while dynamic feature adjustment provides the optimal computational configuration for the deep fusion model. Adaptive calibration and feedback optimization ensure the long-term stable operation of the system, forming an organically unified technical system. This collaborative mechanism not only solves key problems in traditional multi-parameter sensor data fusion, such as low accuracy, high computational complexity, and poor long-term stability, but more importantly, it achieves a technological leap from passive data processing to proactive intelligent optimization. This provides crucial technical support for the development of agricultural IoT and precision agriculture, possessing broad application prospects and significant practical value.

[0047] It should be noted that this invention also addresses the following technical problems: In the process of multi-parameter environmental sensor data fusion, there is a common problem of fixed feature extraction quantities and an inability to adaptively adjust according to data quality and environmental changes. Traditional fusion methods often employ static fusion algorithms based on Kalman filtering, weighted averaging, or simple neural networks, which struggle to achieve a balance between computational efficiency and fusion accuracy. This leads to wasted computational resources or insufficient fusion accuracy when facing environments of varying complexity. This invention, by constructing a feature adjustment function based on convergence, stability, and disorder indices, realizes a dynamic adjustment mechanism for the number of attention-approximate features in a deep fusion model. It can automatically determine the most suitable feature extraction complexity based on key indicators such as data fusion convergence, sensor measurement consistency, and parameter coordination, effectively solving the technical problem that traditional fixed-feature-quantity fusion methods cannot adapt to changing environmental conditions. Furthermore, multi-parameter sensor systems are susceptible to electromagnetic interference and cross-sensitivity between parameters in complex environments. Traditional data processing methods lack effective interference identification and suppression mechanisms, resulting in low reliability of the fusion results. This invention establishes an interference anomaly matrix to accurately identify electromagnetic interference signals and cross-sensitive signals, and combines a data quality assessment system to quantify the degree of interference. This effectively suppresses the influence of various interference signals during data fusion, significantly improving the anti-interference capability and reliability of multi-parameter sensor data fusion.

[0048] Specifically, the principle of this invention is as follows: The fundamental principle behind its ability to solve the performance degradation compensation problem of multi-parameter environmental sensors during long-term operation lies in the construction of a complete adaptive calibration system and a real-time parameter update mechanism. First, this invention establishes a set of jump compensation coefficients for each parameter under different environmental conditions through a jump compensation matrix. This matrix, constructed based on historical data statistical analysis, accurately reflects the performance change patterns of the sensor under various operating states, providing a precise reference for baseline drift compensation. Second, the anomaly compensation matrix establishes a parameter library for compensation algorithms corresponding to various abnormal situations by recording reference data under normal operating conditions. This matrix can effectively identify the specific manifestations of sensor sensitivity attenuation, providing targeted correction schemes for sensitivity compensation. Third, the adaptive calibration algorithm, by monitoring the deviation between the sensor output characteristics and historical normal states in real time, can dynamically determine the current degree and type of sensor performance degradation and automatically select the corresponding compensation strategy for correction. This real-time calibration mechanism conforms to the physical laws of sensor performance degradation; that is, the baseline drift and sensitivity attenuation of the sensor usually exhibit a gradual and regular change process. By continuously tracking these changes and making timely compensation adjustments, the impact of performance degradation on measurement accuracy can be effectively offset. Meanwhile, the real-time update of the compensation parameters ensures that the calibration algorithm is always synchronized with the current performance status of the sensor, avoiding the problem of compensation effect decay caused by the inability of traditional fixed parameter calibration methods to adapt to performance changes, thus maintaining the stable accuracy of data fusion throughout the entire service life.

[0049] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0050] The specific implementation of step S01 is to establish a sensor data acquisition matrix, as shown below:

[0051]

[0052] In the formula, D matrix For sensor data acquisition matrix; T i P represents the temperature sensor data at time i; i The data is the pH sensor data at time i; O i C represents the dissolved oxygen sensor data at time i; i is the conductivity sensor data at time i; n is the total number of sampling points.

[0053] The construction of continuous data sequences is achieved through a sliding time window mechanism, as shown below:

[0054] S j ={D matrix [j:j+w-1,:]};

[0055] In the formula, S j Let w be the data sequence for the j-th time window; w is the length of the time window, with a value of 60; and j is the starting position of the window, with a step size of 5.

[0056] The specific implementation of step S02 involves constructing an abnormal transition matrix, and the transition detection function is expressed as follows:

[0057]

[0058] In the formula, J ij For anomalous jump matrix elements; θ j Let θ1 = 5℃ / min, θ2 = 0.5pH / min, θ3 = 2mg / L / min, and θ4 = 100μS / cm / min be the threshold values ​​for the j-th type of sensor.

[0059] The interference anomaly matrix is ​​constructed through frequency domain analysis and is represented as follows:

[0060]

[0061] In the formula, I kl The elements are the disturbance anomaly matrix elements; FFT is the Fast Fourier Transform function; f l For the l-th frequency component; μ kl For the k-th type sensor at frequency f l The average amplitude of the normal signal at that location.

[0062] The data quality assessment system is represented by the following function:

[0063]

[0064] In the formula, Q quality β1 = 0.6 is the weight for jump detection; β2 = 0.4 is the weight for interference detection; m sensor =4 represents the number of sensor types; f num This represents the total number of frequency components.

[0065] The specific implementation of step S03 involves calculating a multi-dimensional fusion evaluation index. The convergence index is calculated as follows:

[0066]

[0067] In the formula, α conv σ is the convergence exponent; rate The standard deviation of the rate of change of each parameter within a continuous time window; ∈ conv To prevent small constants from being divided by zero, the value is set to 0.001.

[0068] The rate of change of the parameter is calculated as follows:

[0069]

[0070] In the formula, r k Δt represents the rate of change of the k-th parameter; Δt is the sampling time interval, which is 1 minute.

[0071] The standard deviation of the rate of change is calculated as follows:

[0072]

[0073] In the formula, The mean rate of change; m param The number of parameters is 4.

[0074] The mean rate of change is calculated as follows:

[0075]

[0076] The stability index is calculated as follows:

[0077]

[0078] In the formula, α stab For stability index; μ param σ is the average value of the sensor measurements over the time period. param The standard deviation of the sensor measurements; ∈ stab To prevent small constants from being divided by zero, the value is set to 0.001.

[0079] The disorder index is calculated as follows:

[0080]

[0081] In the formula, α turb The disorder index; cond(R) corr ) represents the condition number of the correlation coefficient matrix between parameters; ∈ turb To prevent small constants from being divided by zero, the value is set to 0.001.

[0082] The correlation coefficient matrix is ​​represented as follows:

[0083]

[0084] In the formula, ρ ij Let be the Pearson correlation coefficient between the i-th parameter and the j-th parameter.

[0085] The Pearson correlation coefficient is calculated as follows:

[0086]

[0087] In the formula, X itLet be the measured value of the i-th parameter at time t; Let be the mean of the i-th parameter; N is the length of the time series.

[0088] The condition number is calculated as follows:

[0089]

[0090] In the formula, λ max (R corr ) represents the largest eigenvalue of the correlation coefficient matrix; λ min (R corr ) is the smallest eigenvalue of the correlation coefficient matrix.

[0091] The specific implementation of step S04 involves constructing a feature adjustment function, and the adjustment factor is calculated as follows:

[0092] α adj =w conv ·α conv +w stab ·α stab +w turb ·α turb ;

[0093] In the formula, α adj For adjustment factor; w conv =0.4 is the convergence exponent weight; w stab =0.35 is the stability index weight; w trub =0.25 is the weight of the disorder index.

[0094] The attention approximation feature quantity adjustment function is expressed as follows:

[0095]

[0096] In the formula, N feat This represents the number of attention-approximate features.

[0097] The specific implementation of step S05 is to establish an adaptive calibration algorithm. The baseline drift compensation function is expressed as follows:

[0098] D calib [i, j] = D matrix [i, j]-δ drift [j]·t i ;

[0099] In the formula, D calib [i, j] represents the calibrated sensor data; δ drift [j] represents the baseline drift rate of the j-th type of sensor; t i Let be the relative time at time i.

[0100] The sensitivity attenuation compensation function is expressed as follows:

[0101]

[0102] In the formula, D sens [i, j] represents the sensor data after sensitivity compensation; γ decay [j] represents the initial sensitivity coefficient of the j-th type of sensor; τ j Let be the decay time constant of the j-th type of sensor.

[0103] The jump compensation matrix update function is expressed as follows:

[0104] K jump [i, j] = K jump [i, j]+η jump ·(J ij -K jump [i, j]);

[0105] In the formula, K jump [i, j] are elements of the jump compensation matrix; η jump The learning rate is set to 0.1 to compensate for jumps.

[0106] The anomaly compensation matrix update function is expressed as follows:

[0107] K anom [u, v] = K anom [u, v] + η anom ·(A uv -K abom [u, v]);

[0108] In the formula, K anom [u, v] are elements of the anomaly compensation matrix; A uv The detected abnormal intensity; η anom The abnormality compensation learning rate is set to 0.05.

[0109] Among them, A uv The data was obtained using statistical analysis, including: Step 1: calculating the magnitude of the deviation of the sensor output relative to the normal baseline; Step 2: statistically analyzing the intensity distribution of the abnormal deviation using a sliding window; Step 3: normalizing the abnormal intensity to a range of 0 to 1 as A. uv The value, u, represents the sensor type index, and v represents the anomaly type index.

[0110] The final calibration data is calculated as follows:

[0111] D final [i, j] = D sens [i, j]·(1+K) jump [i, j])·(1+K anom[i, j]);

[0112] In the formula, D final [i, j] represents the final calibrated sensor data.

[0113] The specific implementation method of step S06 is the same as described above, and will not be repeated in detail here.

[0114] The specific implementation of step S07 is to establish a feedback optimization mechanism. The Q-learning reward function is expressed as follows:

[0115]

[0116] In the formula, R reward For reward value; acc eval To ensure the accuracy of environmental quality assessments.

[0117] The Q-value update function is expressed as follows:

[0118] Q(s t a t )=Q(s t a t )+η learn ·[R reward +γ disc ·max a′ Q(s t+1 ,a′)-Q(s t a t )];

[0119] In the formula, Q(s) t a t ) represents state s t Take action a t Q value; η learn γ is the learning rate, with a value of 0.01; disc s is the discount factor, with a value of 0.95; t+1 a represents the next state; a′ represents the possible actions for the next state.

[0120] It should be explained that the convergence index formula is based on statistical variance theory. It quantifies the convergence of the data fusion process by calculating the reciprocal of the standard deviation of the rate of change of multiple parameters. The formula for calculating the rate of change is as follows:

[0121] The smaller the standard deviation, the more consistent the trend of each parameter. The larger the convergence index, the better the fusion effect. Compared with the traditional fixed-weight fusion method, this index can dynamically reflect the real-time status of data fusion, provide a quantitative basis for convergence evaluation, thereby achieving adaptive adjustment of fusion parameters and significantly improving the stability and accuracy of data fusion.

[0122] The stability index formula is based on the coefficient of variation principle. It assesses the stability of data by calculating the ratio of the mean to the standard deviation of sensor measurements. The formula is as follows: This index can effectively identify the fluctuation characteristics of sensor output. Compared with the traditional simple threshold judgment method, this index provides a more accurate and continuous quantitative assessment of stability, providing a scientific basis for data quality evaluation and sensor status monitoring, and effectively avoiding fusion errors caused by data instability.

[0123] The disorder index formula is based on the linear algebraic condition number theory. It measures the compatibility among multiple parameters by calculating the reciprocal of the condition number of the correlation coefficient matrix between parameters. The formula for calculating the condition number is as follows: The condition number reflects the ill-conditioned nature of the matrix. The larger the condition number, the more serious the mutual interference between parameters. The smaller the disorder index, the worse the coordination. Compared with the traditional simple correlation analysis, this formula can comprehensively evaluate the overall correlation characteristics between multiple parameters, provide a global coordination assessment for multi-sensor data fusion, and effectively reduce the negative impact of parameter interference on the fusion results.

[0124] The feature adjustment function dynamically controls the number of attention features by weighting and combining three evaluation indices. The calculation formula is α. adj =w conv ·α conv +w stab ·α stab +w turb ·α turb Different weighting coefficients reflect the importance of each index in feature adjustment. Compared with the traditional method with a fixed number of features, this function can automatically balance computational complexity and fusion accuracy according to data characteristics. When the data quality is good, the number of features is reduced to improve computational efficiency, and when the data quality is poor, the number of features is increased to ensure fusion accuracy, thus realizing intelligent optimization of computing resources.

[0125] The baseline drift compensation function is based on the principle of linear regression. It eliminates the system offset during long-term sensor use through a time-dependent linear compensation term. The compensation formula is D. calib [i, j] = D matrix [i, j]-δ drift [j]·t i Compared to traditional periodic calibration methods, this function can achieve continuous real-time drift compensation, effectively improving the accuracy and consistency of long-term sensor measurements and reducing data fusion errors caused by baseline drift.

[0126] The sensitivity attenuation compensation function is based on an exponential attenuation model, which describes the attenuation of sensor sensitivity over time using an exponential function. The compensation formula is as follows: Compared to traditional segmented calibration methods, this function can continuously track changes in sensor performance, achieve precise dynamic sensitivity compensation, effectively extend the sensor's lifespan and maintain measurement accuracy, and provide a reliable guarantee for long-term stable data fusion.

[0127] The data quality assessment system formula is based on the multivariate anomaly detection theory. It establishes quality assessment indicators by comprehensively considering the results of jump detection and interference detection. The calculation formula is as follows: Compared to traditional single detection methods, this system can comprehensively evaluate multiple dimensions of data quality, providing reliable quality assurance for subsequent data fusion and significantly improving the robustness and accuracy of the system.

[0128] The Q-learning reward function is based on reinforcement learning theory. It guides parameter optimization by setting a clear reward mechanism, and the update formula is Q(s). t a t )=Q(s t a t )+η learn ·[R reward +γ disc ·max a′ Q(s t+1 ,a′)-Q(s t a t This function, compared to traditional static parameter settings, can automatically adjust fusion parameters based on system performance feedback, enabling the system to learn and continuously optimize itself. This significantly improves the adaptability and long-term stability of the data fusion system, providing an intelligent solution for multi-parameter sensor applications in complex environments.

[0129] To better understand and implement this invention, a specific application scenario, Example 2, is provided below: The technical team first deployed multi-parameter environmental sensor chip nodes in each greenhouse. Each node integrates a temperature sensor, a pH sensor, a dissolved oxygen sensor, and a conductivity sensor. The temperature sensor has a measurement range of -20 to 60°C, an accuracy of ±0.1°C, and a response time of 2 seconds. The pH sensor has a measurement range of pH 3.0 to 11.0, an accuracy of ±0.02pH, and a calibration cycle of 30 days. The dissolved oxygen sensor has a measurement range of 0 to 20 mg / L, an accuracy of ±0.1 mg / L, and a stabilization time of 30 seconds. The conductivity sensor has a measurement range of 0 to 20000 μS / cm, an accuracy of ±1%, and a temperature compensation range of 0 to 50°C.

[0130] The sensor data acquisition frequency was set to once every 30 seconds, with 6 sensor nodes set up in each greenhouse, for a total of 216 nodes operating simultaneously. A sensor data acquisition matrix was established, with dimensions of 216×4×2880, corresponding to the number of nodes, the number of sensor types, and the number of daily acquisitions. A sliding time window mechanism was set with a window length of 120 time steps and a sliding step size of 30 time steps to construct a continuous data sequence for subsequent analysis.

[0131] During the data preprocessing stage, the technical team established an anomaly transition matrix to detect abrupt changes. The threshold parameters for the anomaly transition matrix are shown in Table 1.

[0132] Table 1. Threshold Parameters for Abnormal Jump Detection

[0133] Sensor type Single-step jump threshold Continuous jump threshold Detection window length Temperature sensor 2.5℃ 5.0℃ 6 time steps pH sensor 0.3 pH 0.8 pH 8 time steps Dissolved oxygen sensor 0.8 mg / L 1.5 mg / L 10 time steps conductivity sensor 150μS / cm 300μS / cm 12 time steps

[0134] The interference anomaly matrix is ​​used to identify electromagnetic interference signals and cross-sensitivity signals. The matrix structure is a 4×4 symmetric matrix, with diagonal elements representing the interference intensity of each sensor itself and off-diagonal elements representing the cross-interference intensity between sensors. The baseline parameters of the interference anomaly matrix were obtained through seven consecutive days of benchmark testing, as shown in Table 2.

[0135] Table 2 Inter-sensor interference intensity benchmark matrix

[0136] Sensor type Temperature sensor pH sensor Dissolved oxygen sensor conductivity sensor Temperature sensor 0.023 0.045 0.012 0.038 pH sensor 0.045 0.031 0.067 0.089 Dissolved oxygen sensor 0.012 0.067 0.019 0.025 conductivity sensor 0.038 0.089 0.025 0.042

[0137] Based on the detection results of the abnormal jump matrix and the interference anomaly matrix, the system establishes a data quality assessment system. The assessment indicators include three dimensions: data integrity, data accuracy, and data consistency, with weights of 0.35, 0.40, and 0.25 for each dimension, respectively.

[0138] The technical team constructed a multi-dimensional fusion algorithm to calculate key indices for continuous data sequences. The convergence index was obtained by calculating the reciprocal of the standard deviation of the rate of change of each parameter within a continuous time window. After 15 days of testing, the average convergence index was 0.847. The stability index was obtained by calculating the ratio of the parameter mean to the standard deviation. The stability indices for the four sensors were: temperature sensor 12.3, pH sensor 8.7, dissolved oxygen sensor 6.9, and conductivity sensor 11.2. The disorder index was obtained by calculating the condition number of the correlation coefficient matrix between parameters. The average disorder index during system operation was 2.34.

[0139] The deep fusion model employs a multi-layer attention network based on the Performer architecture. The model contains eight attention layers with a hidden dimension of 512 and a feedforward network dimension of 2048. The input embedding layer maps the raw sensor data into a 512-dimensional vector representation. The multi-head attention layer uses eight attention heads, each with a dimension of 64. The feedforward neural network layer uses the GELU activation function with a dropout rate of 0.1. The output projection layer maps the feature vectors to the predicted environmental state.

[0140] The attention approximation mechanism employs a stochastic feature mapping method, dynamically adjusting the number of approximate features through a feature adjustment function. The feature adjustment function is calculated as adjustment factor α = 0.4 × convergence exponent + 0.3 × stability exponent / 10 + 0.3 × (1 / disorder exponent). The number of attention approximation features is dynamically set according to the range of α values: 128 for α ∈ [0.156, 0.432), 256 for α ∈ [0.432, 0.778), and 512 for α ∈ [0.778, 1.000].

[0141] The model training dataset contains 1.57 million valid samples from 180 consecutive days of historical monitoring data, with each sample containing 48 hours of continuous monitoring data. The data is divided into training, validation, and test sets in a 7:2:1 ratio. The model is trained using the Adam optimizer with a learning rate of 0.0001 and a batch size of 64. After 150 training epochs, the validation loss converged.

[0142] An adaptive calibration algorithm is used to compensate for sensor baseline drift and sensitivity decay. The jump compensation matrix is ​​established based on 30-day sliding statistics, and the compensation coefficients are shown in Table 3.

[0143] Table 3 Sensor Jump Compensation Coefficient Matrix

[0144] Environmental conditions Temperature sensor pH sensor Dissolved oxygen sensor conductivity sensor High temperature and humidity 0.92 0.88 0.85 0.91 High temperature and low humidity 0.95 0.93 0.90 0.94 Low temperature and high humidity 0.89 0.86 0.82 0.87 Low temperature and low humidity 0.93 0.91 0.88 0.92

[0145] The anomaly compensation matrix is ​​established based on reference data under normal working conditions. It includes three types of parameters: sensor zero drift compensation, range drift compensation, and nonlinear error compensation. Each type of parameter has five levels of compensation coefficients set for different degrees of anomaly.

[0146] The environmental state identification model establishes 12 typical environmental state patterns, including environmental characteristic patterns for different growth stages such as suitable growth period, flowering period, fruiting period, and maturity period. The pattern matching algorithm calculates the similarity score between real-time monitoring data and preset patterns. A successful match is considered when the similarity exceeds 0.85, and the system automatically outputs corresponding environmental management suggestions. Traditional methods often use simple threshold judgments or weighted averages when processing multi-sensor data, which cannot effectively handle mutual interference between sensors and data quality issues. This invention, through a deep fusion model and adaptive calibration mechanism, achieves intelligent and precise multi-parameter environmental monitoring, significantly improving the overall performance of the agricultural environmental monitoring system.

[0147] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0148] Table 4. Variable Explanation Table (Part 1)

[0149]

[0150] Table 5. Variable Explanation Table (Part Two)

[0151]

[0152]

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based multi-parameter environmental sensor chip data fusion method, characterized by, The temperature sensor data, the pH sensor data, the dissolved oxygen sensor data and the conductivity sensor data of the multi-parameter environmental sensor chip are collected, a sensor data collection matrix is established, and a continuous data sequence is constructed through a time window sliding mechanism; the continuous data sequence is preprocessed, a sudden change matrix is used to detect the mutation points, an interference anomaly matrix is used to identify electromagnetic interference signals and cross-sensitivity signals, and a data quality evaluation system is established; a multi-dimensional fusion algorithm is constructed to calculate the convergence index, the stability index and the turbulence index of the continuous data sequence; A deep fusion model is established to extract features and analyze the correlation of the preprocessed continuous data sequence, and the number of attention approximation features is dynamically adjusted according to the convergence index, the stability index and the turbulence index through a feature adjustment function; an adaptive calibration algorithm is used to compensate for the baseline drift and sensitivity attenuation of the sensor; an environmental state recognition model is established, a pattern matching algorithm is used to realize the classification and prediction of the agricultural product growth environment, and the environmental quality evaluation results are output; The convergence index is a quantitative indicator of the consistency of the multi-parameter sensor data in the fusion process, which is obtained by calculating the reciprocal of the standard deviation of the parameter change rate in the continuous time window, and the larger the value, the better the data fusion convergence; The stability index is a quantitative indicator of the fluctuation degree of the sensor measurement value in a time period, which is obtained by calculating the ratio of the parameter mean value to the standard deviation, and is used to evaluate the reliability and reproducibility of the sensor output; The turbulence index is a quantitative indicator of the mutual influence and interference degree between the parameters, which is obtained by calculating the condition number of the correlation coefficient matrix between the parameters, and the smaller the value, the better the coordination between the parameters.

2. The multi-parameter environmental sensor chip data fusion method of claim 1, wherein, The time window sliding mechanism is a time sequence processing of the sensor data collection matrix, which converts the temperature sensor data, the pH sensor data, the dissolved oxygen sensor data and the conductivity sensor data into data sequences with time continuity.

3. The multi-parameter environmental sensor chip data fusion method of claim 2, wherein, The sudden change matrix is a two-dimensional array structure used to detect abnormal mutation phenomena in sensor data, and the matrix elements represent the jump detection results of each parameter at different time points, which are used to identify device failures or external disturbances.

4. The multi-parameter environmental sensor chip data fusion method of claim 3, wherein, The interference anomaly matrix is a data structure used to identify electromagnetic interference and cross-sensitivity between multi-parameter sensors, and the matrix rows and columns correspond to different parameters, and the element values represent the interference intensity and type between parameters.

5. The multi-parameter environmental sensor chip data fusion method of claim 4, wherein, The data quality evaluation system is established based on the detection results of the sudden change matrix and the interference anomaly matrix, and is used to evaluate the reliability and effectiveness of the sensor data to provide quality assurance for subsequent data processing steps.

6. The multi-parameter environmental sensor chip data fusion method of claim 5, wherein, The structure of the deep fusion model is a multi-layer attention network based on the Performer architecture, which includes an input embedding layer, a multi-head attention layer, a feedforward neural network layer and an output projection layer, and the attention approximation mechanism uses a random feature mapping method to reduce the computational complexity.

7. The multi-parameter environmental sensor chip data fusion method of claim 6, wherein, The feature adjustment function is used for adjusting the attention approximate feature quantity of the deep fusion model, an adjustment factor is calculated based on a convergence index, a stability index and a disorder index, and balance optimization of calculation efficiency and fusion precision is realized by dynamically adjusting the attention approximate feature quantity.

Images