A banana sprout recognition method and system based on multi-dimensional electrochemical feature fusion

Abstract

The application discloses a banana sprout growth point recognition method and system based on multi-dimensional electrochemical feature fusion. The method comprises the following steps: performing wide-frequency impedance scanning on banana bulbs; selecting a single / double CPE equivalent circuit according to Nyquist pattern morphology, and accurately extracting electrochemical features by using a frequency domain decoupling and step-by-step cross-constraint iteration strategy; after the features are logarithmized, the features are input into a heterogeneous fusion architecture composed of linear discrimination and nonlinear tree models to suppress moisture drift and mine nonlinear coupling relationships; and finally, the growth point target is determined based on a soft voting mechanism weighted fusion. The application overcomes the low-frequency physical mismatch limitation, decouples moisture and lignification camouflage interference, and realizes high-robustness and accurate recognition of the growth point.

Description

Technical Field

[0001] This invention belongs to the technical field of smart agriculture sensing, and specifically relates to a precise identification method for plant tissues based on electrochemical impedance spectroscopy (EIS), specifically a banana sucker identification method and system based on multidimensional electrochemical feature fusion. Background Technology

[0002] Precise sucker removal is a crucial agronomic step in ensuring the yield and healthy growth of the mother plant. Accurate identification of the sucker's growth point is the core prerequisite for successful sucker removal. In current agricultural practices, the identification methods for banana sucker growth points mainly fall into two categories: manual cutting and surface machine vision recognition. Both have significant technical drawbacks: manual cutting is labor-intensive and destructive, easily damaging the mother plant's root system and affecting its normal growth; while machine vision technology is non-contact, it can only capture surface visual features and cannot penetrate the soil and bulb epidermis to detect internal physiological conditions, leading to identification errors and a dual risk of either damaging the mother plant or incomplete sucker removal.

[0003] Electrical impedance spectroscopy (EIS), as a non-destructive testing method capable of probing the microstructure within biological tissues, has shown great potential in plant physiological monitoring due to its non-invasive and highly sensitive characteristics, providing a new technical approach for the accurate identification of banana sucker growth points. However, when applying this technology to the identification of complex banana bulb tissues, existing techniques still face several unresolved technical problems, limiting its practical application. These problems are specifically reflected in the following three aspects:

[0004] First, existing physical models suffer from severe mismatch issues in equivalent circuit modeling. Traditional Randle's equivalent circuit model typically assumes ideal capacitance at the electrode interface and semi-infinite diffusion, resulting in a 45° diagonal line characteristic in the Nyquist plot. However, actual banana bulb tissue, especially the growing point tissue, often exhibits a finite-length diffusion characteristic with a "vertical rise" in its impedance spectrum at low frequencies, and the cell membrane interface shows a significant non-ideal double-layer effect. Directly using traditional models for fitting fails to accurately characterize these non-ideal properties, leading to significant errors in the extracted electrical parameters, distorted physical meaning, and an inability to provide accurate feature bases for subsequent identification.

[0005] Secondly, existing technologies have weak anti-interference capabilities and are unable to cope with individual differences in living biological tissues. Significant differences in water content among different plants cause large fluctuations in single electrical characteristics such as ohmic resistance, resulting in baseline drift. Furthermore, the complex internal structure of banana bulbs, with its highly lignified central tube tissue exhibiting high impedance characteristics, can easily be confused with the growth point tissue, forming a "structural camouflage." Relying solely on single-dimensional electrical parameters is insufficient for effective differentiation.

[0006] Finally, existing recognition algorithms have limitations, making it difficult to balance generalization ability and nonlinear fitting ability. Single linear models (such as Linear Discriminant Analysis (LDA)) struggle to handle the complex nonlinear boundaries of tissue transition zones, while single nonlinear models (such as XGBoost) are prone to overfitting under the small sample conditions common in agricultural data.

[0007] In summary, developing a banana sucker identification method that is adapted to the complex biological tissue characteristics of banana bulbs and possesses high robustness and high accuracy has become a pressing technical problem to be solved in the field of smart agriculture. Summary of the Invention

[0008] The main objective of this invention is to overcome the shortcomings and deficiencies of existing technologies and provide a banana sucker identification method and system based on the fusion of multi-dimensional electrochemical features. This invention solves the technical problems that traditional equivalent circuit models cannot fit the low-frequency non-ideal electrochemical features of banana bulb tissue, the identification method has weak resistance to moisture interference, and it is difficult to distinguish between the lignified tube and the growth point. It achieves non-destructive, accurate, and highly robust identification of banana sucker growth points, providing reliable technical support for automated sucker removal operations.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a banana bud identification method based on multidimensional electrochemical feature fusion, comprising the following steps:

[0011] S1. Perform a wideband frequency scan on the banana bulb tissue to be tested to obtain complex impedance data;

[0012] S2. Based on the geometric topology of the measured Nyquist plot, an equivalent circuit model is selected to fit the complex impedance data. A local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling is adopted to extract multidimensional electrochemical characteristic parameters.

[0013] S3. Perform a base-10 logarithmic spatial reconstruction preprocessing on the ohmic and polarization resistance features and relaxation time constant features with physical dimensions among the multidimensional electrochemical characteristic parameters, so as to reduce the nonlinear multiplicative error caused by the drastic fluctuation of the water content of individual plants into a linear additive bias, thereby obtaining a standard logarithmic feature vector that eliminates the order-of-magnitude barrier.

[0014] S4. Construct a heterogeneous fusion architecture based on heterogeneous integration. This architecture cascades a linear discriminant path for capturing the global linear chassis and a nonlinear fine-grained path for uncovering local nonlinear details. The linear discriminant path uses a linear discriminant analysis algorithm to establish a stable macroscopic judgment baseline, while the nonlinear fine-grained path uses a gradient boosting tree algorithm to keenly capture the complex nonlinear coupling relationship between the parameters of open-circuit finite-length diffusion elements and constant phase angle elements, thereby decoupling and identifying the structural camouflage between highly lignified cylinders and dense growth points. The standard logarithmic feature vector is input into this heterogeneous fusion architecture, and the linear prediction confidence and nonlinear prediction confidence are calculated and output independently through the two paths.

[0015] S5. Based on the adaptive soft voting mechanism, perform high-order weighted fusion of the linear prediction confidence and the nonlinear prediction confidence, and accurately determine whether the tissue to be tested is the target banana sucker growth point based on the final output joint decision probability.

[0016] As a preferred technical solution, in step S1, a three-electrode probe is used to perform wideband frequency scanning on the banana bulb tissue to be tested.

[0017] The three-electrode probe includes a working electrode, a reference electrode, and a counter electrode. The working electrode is inserted into the geometric center of the banana bulb tissue to be tested. The reference electrode and the counter electrode work together to form a detection circuit. A portable electrochemical workstation is used to perform frequency scanning. The wide frequency range is 1 Hz to 1 MHz.

[0018] As a preferred technical solution, the equivalent circuit model is a Cole single CPE equivalent circuit model or a Cole double CPE equivalent circuit model, and the topology is selected based on the frequency domain response characteristics of the Nyquist plot and the Bode plot.

[0019] Among them, for the tissue under test exhibiting a single relaxation characteristic, the selected Cole single CPE equivalent circuit model is composed of three components connected in series in terms of topology: the solution resistance, the parallel branch consisting of the charge transfer resistance and the constant phase angle element, and the open-circuit finite-length diffusion element.

[0020] For the test tissue exhibiting double relaxation superposition characteristics, the selected Cole double CPE equivalent circuit model strictly cascades the solution resistance, the first-stage charge transfer resistor and the constant phase angle element in parallel branch, the second-stage charge transfer resistor and the constant phase angle element in parallel branch, and the open-circuit finite-length diffusion element connected in series at the end; the constant phase angle element is used to accurately characterize the frequency diffusion effect caused by biomembrane roughness, and the open-circuit finite-length diffusion element is used to characterize the boundary blocking effect induced by low-frequency polarized ions in the extremely dense cellular confined space.

[0021] As a preferred technical solution, in step S2, the local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling specifically refers to:

[0022] Select the corresponding equivalent circuit model based on the single capacitive arc or double capacitive arc characteristics presented by the impedance spectrum;

[0023] Local prefitting is used to independently fit the mid-to-high frequency polarization semicircle and the low frequency diffusion oblique line, extract the initial parameter values ​​and load them into the equivalent circuit model, and set all component parameters in the equivalent circuit model to a locked state;

[0024] The parameters of the low-frequency open-circuit finite-length diffusion element are released first for fitting, and the parameters of the open-circuit finite-length diffusion element are locked when the low-frequency fitting residual meets the preset convergence condition.

[0025] Subsequently, the parameters of the mid-to-high frequency branches were released and fitted with the topological geometry of the Nyquist plot and the Bode plot;

[0026] Based on the dynamic feedback of local residuals, the constraint locking and releasing adjustment steps of parameters in each frequency band are executed alternately until the global weighted chi-square error converges to... Less than orders of magnitude.

[0027] As a preferred technical solution, the multidimensional electrochemical characteristic parameters include ohmic internal resistance, charge transfer resistance, double-layer capacitance parameters, and diffusion impedance parameters.

[0028] As a preferred technical solution, the calculation process of the linear discrimination path in step S4 is as follows:

[0029] Construct inter-class scatter matrix and intra-class scatter matrix based on the training set;

[0030] The optimal projection vector is calculated based on the Fisher criterion to maximize the ratio of between-class divergence to within-class divergence.

[0031] The input logarithmic feature vector is projected onto the linear space generated by the optimal projection vector to obtain the projection value;

[0032] The projected values ​​are mapped to linear discriminant probabilities using the Sigmoid function to characterize the confidence that a sample belongs to a growth point on the global impedance scale.

[0033] As a preferred technical solution, the calculation process of the nonlinear fine-grained pathway in step S4 is as follows:

[0034] To explore the nonlinear coupling relationship between diffusion impedance parameters and double-layer capacitance parameters;

[0035] A target function containing a loss function and a regularization term is constructed, and the target function is approximated by a second-order Taylor expansion to isolate and extract the complex nonlinear coupling characteristics between the parameters of the open-circuit finite-length diffusion element and the constant phase angle parameter.

[0036] A Bayesian optimization module is introduced, which uses a Gaussian process as a surrogate model to perform global iterative optimization of the hyperparameter combination of learning rate, maximum tree depth and regularization coefficient of the XGBoost algorithm, and outputs the optimal nonlinear discrimination probability.

[0037] As a preferred technical solution, in step S5, the linear and nonlinear discrimination probabilities are weighted and fused based on an adaptive soft voting mechanism, specifically as follows:

[0038] ;

[0039] In the formula, The weighting coefficients of the linear path are represented. Represents the weighting coefficients of the nonlinear path; and These represent the linear and nonlinear discriminant probabilities, respectively; if the final confidence level... If the value exceeds the preset threshold, the current detection location is determined to be the banana sucker growth point.

[0040] Secondly, the present invention provides a banana bud identification system based on multidimensional electrochemical feature fusion, including a data acquisition unit, a data processing unit and an output interaction unit, which are connected in sequence; the data acquisition unit is equipped with a portable electrochemical workstation and a customized three-electrode probe assembly, which is configured to penetrate the test tissue site of the banana bulb to perform a broadband impedance scanning task.

[0041] The data processing unit is configured to implement the banana sucker recognition method based on multidimensional electrochemical feature fusion;

[0042] The output interaction unit is used to visually display the output results of the data processing unit.

[0043] As a preferred technical solution, the output interaction unit is configured with a communication interface, which is used to output a growth point position signal to guide automated agricultural equipment to perform bud removal.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] (1) This invention proposes an improved equivalent circuit model (Cole single CPE equivalent circuit model or Cole double CPE equivalent circuit model) that introduces a finite-length diffusion element to adapt to the finite-length diffusion characteristics of the low-frequency "vertical rise" of banana bulb tissue. At the same time, through a distribution cross-constraint fitting strategy based on frequency domain decoupling, it solves the problem of traditional model parameter divergence and inability to fit non-ideal electrochemical characteristics, and the fitting residual ( Reduced to The extracted 11-dimensional electrochemical characteristic parameters, on an order of magnitude, can truly reflect the microstructural characteristics of banana bulb tissue.

[0046] (2) By performing logarithmic preprocessing on the electrochemical characteristic parameters, the multiplicative error caused by moisture fluctuations is transformed into an additive error. Combined with the projection operation of the LDA linear discrimination pathway, this additive error can be effectively offset, decoupled from the interference of "false high impedance caused by drying", and realize the universal identification of banana plants with different water contents, which greatly improves the robustness of the method of the present invention.

[0047] (3) The present invention constructs a heterogeneous fusion architecture based on heterogeneous integration, which combines the dual advantages of linear discriminant model with strong anti-interference and generalization ability and nonlinear model with fine-grained identification and excellent fitting ability. It can effectively distinguish between highly lignified central tube tissue and growth point tissue. After testing and verification, the overall identification accuracy of the test set of this method reaches 93.0%, and the false alarm rate of growth point identification is extremely low.

[0048] (4) The identification method of the present invention, combined with SHAP analysis, proves that the identification basis is highly consistent with the biophysical mechanism of "small growth point cells, high density, and no gaps". It is not a simple algorithm fitting. The identification strategy is not just a black box fitting at the level of data statistical correlation, but objectively maps the real dielectric response law in the process of plant living tissue evolving from primary meristem to secondary lignification. At the methodological level, it realizes a perfect scientific closed loop between data intelligent decision-making and plant electrophysiological mechanism. The model has high scientificity and interpretability.

[0049] (5) The identification system of the present invention uses a portable electrochemical workstation with a customized three-electrode probe assembly. The overall structure is lightweight and can meet the mobile detection needs in banana planting fields. At the same time, the output interaction unit of the system can be linked with automated agricultural equipment to realize the integrated operation of identification and bud removal, thereby improving the level of intelligence in banana planting. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is the overall flowchart of the banana sucker recognition method based on multidimensional electrochemical feature fusion of the present invention;

[0052] Figure 2 This is a topology diagram of the Cole single CPE equivalent circuit model adapted to plant bulb tissue for this invention;

[0053] Figure 3 This is a topology diagram of the Cole double CPE equivalent circuit model adapted to plant bulb tissue for this invention;

[0054] Figure 4 The figure shows the impedance spectrum fitting effect of the present invention on a single time constant sample based on the Cole single CPE equivalent circuit model.

[0055] Figure 5 The figure shows the impedance spectrum fitting effect of the dual time constant sample based on the Cole double CPE equivalent circuit model of the present invention. Part (a) is the global fitting effect and part (b) is the local fitting effect.

[0056] Figure 6 This is a ranking chart of the contribution of electrochemical features analyzed by the SHAP method in this invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0058] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0059] like Figure 1As shown in the figure, this embodiment provides a banana sucker identification method based on multidimensional electrochemical feature fusion, which includes the following steps:

[0060] S1. Perform a wideband frequency scan on the banana bulb tissue to be tested to obtain complex impedance data;

[0061] Furthermore, in step S1, a three-electrode probe is used to perform a wide-frequency domain scan on the banana bulb tissue to be tested; the three-electrode probe includes a working electrode, a reference electrode, and a counter electrode. The working electrode is inserted into the geometric center of the banana bulb tissue to be tested, and the reference electrode and the counter electrode cooperate to form a detection circuit. A portable electrochemical workstation is used to perform the frequency scanning operation; the frequency range of the wide-frequency domain is 1Hz to 1MHz.

[0062] S2. Based on the measured geometric topology of the Nyquist plot (i.e., the degree of flattening distortion of the mid-to-high frequency polarization semicircle and the vertical blocking characteristics of the low-frequency diffusion tail), the Cole single-CPE or double-CPE equivalent circuit model is selected to fit the complex impedance data. A local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling is adopted to extract multi-dimensional electrochemical characteristic parameters. Specifically, the local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling is as follows:

[0063] S21. For samples exhibiting a single capacitive arc characteristic in their impedance spectrum, this invention establishes a stepwise cross-constraint fitting method based on frequency domain decoupling, specifically including:

[0064] S211. Use the local prefitting tool to independently fit the mid-to-high frequency polarization semicircle and the low frequency diffusion oblique line, extract the corresponding prior initial values ​​and load them into the global model, and force the locking of all component parameters in the initial state.

[0065] S212, Prioritize the release of open-circuit finite-length diffusion elements ( The low-frequency region is approximated using the parameters of the theoretical curve. During this process, targeted dynamic fine-tuning is performed based on the deviation characteristics between the measured trajectory and the theoretical curve: if the slope of the low-frequency tail deviates, the diffusion index (WP) is adjusted to match the blocking characteristics of the physical boundary; if the inflection point frequency of the diffusion trajectory shifts from tilted to vertical, the diffusion time constant (WT) is corrected; if the extension span of the low-frequency region on the real axis is inconsistent, the diffusion drag (WR) is fine-tuned. When the low-frequency fitted curve highly coincides with the measured trajectory and the relative error of the parameters converges to a minimum, the diffusion parameter set is locked.

[0066] S213. Release the mid-to-high frequency branch parameters to fit the polarization semicircle. The system performs physical parameter mapping based on the topological geometry of the Nyquist plot: fine-tuning the solution internal resistance by anchoring the real intercept in the high-frequency region. Accurately calibrate the charge transfer resistor based on the transverse span diameter of the capacitive arc. To address the longitudinal collapse and flattening of the polarized semicircle, the dispersion shape factor (CPE-P) is modified to restore the non-ideal polarization behavior caused by interface roughness; and the double-layer polarization constant (CPE-T) is dynamically calibrated by combining the phase angle peak characteristics of the Bode plot.

[0067] S214. Based on local residual dynamic feedback, the constraints and releases of the above parameters are executed alternately. This step-by-step iterative mechanism avoids the correlation divergence trap in high-dimensional parameter space from the methodological level, enabling the global theoretical fitting curve to achieve an extremely accurate fit with the measured data, and the global weighted chi-square error ( Converging to This order of magnitude ensures the physical authenticity of the extracted core electrophysiological features.

[0068] S22. For complex samples (such as highly lignified cylinders) exhibiting overlapping or separated characteristics of double capacitive arcs in their impedance spectra, this invention utilizes the Cole double CPE equivalent circuit model and continues the step-by-step cross-constraint fitting logic based on frequency domain decoupling, specifically including:

[0069] S221. Using local prefitting tools, independent feature extraction is performed on the high-frequency polarization semicircle, the mid-frequency additional polarization semicircle, and the low-frequency diffusion oblique line. The prior initial values ​​of each physical relaxation process are anchored and loaded into the global model, and fully locked in the initial state.

[0070] S222, Prioritize the release of the end-open-circuit finite-length diffusion element ( The parameter set is dynamically fine-tuned based on the low-frequency measured trajectory deviation: the diffusion index (WP) is adjusted to match the tilt evolution of the low-frequency tail; the diffusion time constant (WT) is corrected to accurately anchor the characteristic inflection point of the diffusion trajectory pulling towards the vertical capacitive barrier; the diffusion drag (WR) is fine-tuned to approximate its effective extension span on the real axis; after the low-frequency physical barrier boundary fitting converges, the parameter set is locked.

[0071] S223. Synchronously release the parameters of the dual mid-to-high frequency parallel branches to analyze the overlapping polarization semicircles. The system performs high-dimensional physical mapping based on the Nyquist topological geometry and Bode frequency response characteristics: the solution internal resistance is calibrated by the real-axis intercept in the high-frequency region. The charge transfer resistance of protoplast membranes was precisely quantified based on the lateral span of the first high-frequency capacitive arc. Furthermore, the first-order dispersion shape factor (CPE1-P) was corrected based on the degree of longitudinal flattening distortion; and the polarization resistance of the lignified additional interface was rigorously calibrated according to the geometric extension and collapse characteristics of the secondary mid-frequency capacitive arc. The second-order dispersion shape factor (CPE2-P) is used. Simultaneously, the peak frequencies of the two phase angles presented in the Bode plot are used to dynamically constrain and calibrate the polarization constants of the two-level double layers, CPE1-T and CPE2-T, respectively.

[0072] S224. Finally, relying on the dynamic feedback of local residuals, the constraints and releases of the above-mentioned multiple relaxation parameters are executed alternately. This deeply customized step-by-step iterative mechanism avoids the cross-correlation divergence trap that is very easy to occur in the 11-dimensional high-order parameter space from the methodological level, so that the global theoretical fitting curve and the double circular arc measured data achieve a tight physical fit, and the global weighted chi-square error ( Converging to This ensures high-fidelity decoupling and extraction of dielectric features at multiple interfaces in complex structures, even at orders of magnitude below the limit.

[0073] like Figure 2 As shown, the Cole single CPE equivalent circuit model for analyzing the single relaxation process is illustrated, and the functions and roles of each component are strictly defined as follows:

[0074] Solution resistance ( ): Used to characterize the macroscopic ohmic resistance of the ion-conducting medium composed of intracellular and extracellular fluids, vascular bundle sap, and external test buffer.

[0075] Charge transfer resistor-constant phase angle element parallel branch (Rct, CPE): used to analyze the charge transfer polarization process of electrochemical excitation signal across the protoplast cell membrane, the dominant biological interface. Rct precisely quantifies the physical energy barrier of ion transmembrane crossing, while CPE accurately characterizes the non-ideal capacitance frequency diffusion effect caused by the physical roughness and microscopic heterogeneity of the cell membrane surface.

[0076] Open-circuit finite-length diffusion element connected in series at the end ( ): Specifically designed for rigorous characterization of anomalous vertically confined diffusion behavior of polarized ions in extremely dense, gapless cellular tissues such as growth points, caused by their contact with rigid physical barriers formed by the cell wall.

[0077] like Figure 3 As shown, the topology of the Cole double CPE equivalent circuit model used to analyze complex double relaxation processes is illustrated, and the functions and roles of each component are strictly defined as follows:

[0078] Solution resistance ( ): Used to characterize the macroscopic ohmic internal resistance of the ion-conducting medium composed of intracellular and extracellular fluids, vascular bundle sap, and external test buffer;

[0079] First charge transfer resistor-constant phase angle element parallel branch ( , ): Used to analyze the charge transfer polarization process across the protoplast cell membrane, a core biological interface, using electrochemical excitation signals. The energy barrier for ion transmembrane crossing was quantified, while CPE1 characterizes the non-ideal capacitance frequency diffusion effect caused by the physical roughness of the cell membrane surface.

[0080] Second charge transfer resistor-constant phase angle element parallel branch ( , ): This is used for deep decoupling of the additional interfacial polarization effect of the secondary high-impedance physical interface constructed by the high lignification of the cell wall or the formation of thick-walled tissue. This branch is specifically designed to strip away the additional relaxation process caused by the structural insulation barrier.

[0081] Open-circuit finite-length diffusion element connected in series at the end ( ): Specifically designed for rigorous characterization of the anomalous vertically confined diffusion behavior of polarized ions in extremely dense, gapless cellular tissues such as growth points, caused by their contact with physical barriers formed by cell walls.

[0082] S3. Perform a base-10 logarithmic space reconstruction preprocessing on the multidimensional electrochemical characteristic parameters to obtain a standardized logarithmic feature vector.

[0083] Specifically, in step S3, addressing the significant physical barriers spanning five orders of magnitude in the original electrochemical characteristic parameters, this embodiment performs [further analysis] on all resistance and time constant characteristics with physical dimensions. Spatial reconstruction. This step transforms the nonlinear multiplicative error caused by drastic fluctuations in the water content of individual plants into a linear additive bias from the algorithm's underlying layer. This eliminates the magnitude barrier and greatly improves the robustness of subsequent models against baseline drift in physiological states.

[0084] A deep analysis of the quantitative data in Table 1 reveals that the extracted electrochemical parameters and microscopic mechanism deductions achieve precise numerical verification. Firstly, regarding fluid conductivity, the highly lignified central cylinder exhibits an extremely significant high ohmic resistance (mean value reaching...). ), while the meristematic growth point, rich in free water, has extremely low internal resistance (only ), This verified the insulating and blocking effect of lignin deposition on ion migration. Secondly, regarding interfacial polarization dispersion, the shape factor of the growth point tissue... The mean value (0.67) is significantly lower than that of mature cortical tissue (0.78), and this drop in value clearly reflects the extremely complex boundary membrane pore distribution and strong physical roughness of primary meristem tissue. More importantly, at the level of confined diffusion kinetics, the diffusion time constant of growth point tissue is... The value of 4.18s is extremely small, much lower than that of loose cortical tissue (97.24s). This confirms the previous inference from an absolute quantitative mathematical perspective: the dense growth point cells constitute an extremely narrow microscopic confined space, forcing free ions to reach the rigid boundary and trigger a capacitive barrier effect within a very short migration time.

[0085] Table 1. Descriptive statistics of core electrochemical characteristic parameters of three typical tissue types (mean ± standard deviation)

[0086]

[0087] However, global statistical analysis also reveals serious data engineering vulnerabilities. Features from different physical dimensions exhibit significant scale differences in absolute numerical space. For example, dissipation resistance, which reflects the difficulty of ion diffusion at extremely low frequencies... Its average value can soar to The enormous order of magnitude; and the internal resistance of the solution With capacitance parameters Just hovering and The extremely small intervals. This physical scale barrier, spanning up to ten orders of magnitude, and the enormous variance and standard deviation caused by individual fluctuations in sample water content, profoundly reveal that directly inputting the original electrochemical parameters into machine learning algorithms will inevitably lead to gradient weight imbalance and multiplicative drift risks. Therefore, this mathematical analysis verifies the absolute necessity and urgency of subsequently performing logarithmic spatial preprocessing.

[0088] S4. Construct a heterogeneous fusion algorithm architecture. This architecture addresses two major challenges in biological tissue impedance characteristics: "global water interference" and "local structural camouflage." It designs parallel processing pathways, and the specific data processing and calculation process is as follows:

[0089] S41, Data Definition;

[0090] Let the 11-dimensional eigenvectors after logarithmic preprocessing be... ,in Corresponding ohmic internal resistance and charge transfer resistance The logarithm of Corresponding to constant phase angle element (CPE) and finite length diffusion element ( The logarithm of the parameter.

[0091] S42. First Pathway: Global Linear Discriminant Pathway (LDA) – Establishment of the Macro-linear Decision-Making Framework;

[0092] This approach utilizes the Linear Discriminant Analysis (LDA) algorithm to establish an extremely robust macroscopic decision baseline in the logarithmic feature space. First, the system constructs an inter-class scatter matrix based on the training set. With the intra-class scatter matrix Calculate the optimal projection vector based on Fisher's criterion. This maximizes the objective function in the following equation, in order to find the linear boundary with the highest global discriminative power:

[0093] ;

[0094] In the formula, Represents the projection vector. Indicates its transpose; This represents the inter-class scatter matrix, used to characterize the distance between the centers of growth points and non-growth point samples; This represents the intra-class scatter matrix, used to characterize the degree of dispersion within samples of the same class.

[0095] Next, the input feature vector Projecting onto this linear space yields the projected values. :

[0096] ;

[0097] In the formula, This represents the optimal projection vector that maximizes the above objective function; This represents the logarithmized feature vector of the input; This represents the bias term. Here, due to the input features... The logarithmic transformation has been performed, and the impedance multiplicative error caused by the fluctuation of water content in biological tissues is mathematically converted into an additive bias, which is effectively canceled out by the above linear projection operation.

[0098] Finally, the projected values ​​are mapped to probabilities using the Sigmoid function. This characterizes the confidence level that a sample belongs to a "growth point" on the global impedance scale.

[0099] S43, Second Pathway: Local Nonlinear Fine-Grained Pathway (BO-XGBoost) – Microstructure Identification;

[0100] This pathway employs the Extreme Gradient Boosting Tree (XGBoost) algorithm, specifically designed to mine diffusion impedance (…). ) and double-layer capacitance ( The nonlinear coupling relationship between the lignified cylinder and the growth point is used to distinguish them. The objective function of the model is... By loss function and regularization term constitute:

[0101] ;

[0102] To accurately capture the nonlinear manifold with microscopic features, the objective function is approximated by a second-order Taylor expansion:

[0103] ;

[0104] in and These are the first and second derivatives of the loss function, respectively. This represents the prediction score of the t-th tree for the i-th sample; This represents the regularization term, used to control the number and weight of leaf nodes in the tree, preventing the model from overfitting on small samples.

[0105] Furthermore, in the model building phase, a Bayesian optimization module is introduced, using a Gaussian process as a surrogate model to optimize the hyperparameter combination of XGBoost (including the learning rate). Maximum tree depth (max_depth), regularization coefficient Perform global iterative optimization to output the optimal nonlinear discrimination probability. .

[0106] S5. Based on the adaptive soft voting mechanism, perform high-order weighted fusion of the linear prediction confidence and the nonlinear prediction confidence, and accurately determine whether the tissue to be tested is the target banana sucker growth point based on the final output joint decision probability.

[0107] In step S5, the formula for weighted fusion calculation is as follows:

[0108] ;

[0109] In the formula, The weighting coefficients of the linear path are represented. The weighting coefficients representing the nonlinear path (in this embodiment) ); and Let represent the linear and nonlinear discriminant probabilities, respectively. If the final confidence level... If the value is greater than a preset threshold (e.g., 0.5), the current detection location is determined to be a growth point tissue.

[0110] In another embodiment of this application, a banana bud recognition system based on multidimensional electrochemical feature fusion is provided. The system includes a data acquisition unit, a data processing unit, and an output interaction unit, which are connected sequentially:

[0111] The data acquisition unit is equipped with a portable electrochemical workstation and a customized three-electrode probe assembly. The three-electrode probe is configured to penetrate the test tissue of the banana bulb to perform a broadband impedance scanning task.

[0112] The data processing unit integrates modules for equivalent circuit fitting, feature logarithmic preprocessing, and fusion classification and recognition algorithms; the data processing unit is configured to implement the following method:

[0113] Wideband frequency scanning was performed on the banana bulb tissue to be tested to obtain complex impedance data;

[0114] Based on the geometric topology of the measured Nyquist plot (i.e. the degree of flat distortion of the mid-to-high frequency polarization semicircle and the vertical blocking characteristics of the low frequency diffusion tail), the Cole single CPE or double CPE equivalent circuit model is selected to fit the complex impedance data. The multidimensional electrochemical characteristic parameters are extracted by adopting the local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling.

[0115] The ohmic and polarization resistance features with physical dimensions and the relaxation time constant features in the multidimensional electrochemical characteristic parameters are subjected to a base-10 logarithmic spatial reconstruction preprocessing to reduce the nonlinear multiplicative error caused by the drastic fluctuation of the water content of individual plants into a linear additive bias, thereby obtaining a standard logarithmic feature vector that eliminates the order-of-magnitude barrier.

[0116] A heterogeneous fusion architecture based on heterogeneous integration is constructed, which cascades a linear discrimination path for capturing the global linear chassis and a nonlinear fine-grained path for mining local nonlinear details. The linear discrimination path uses a linear discriminant analysis algorithm to establish a stable macroscopic judgment baseline, while the nonlinear fine-grained path uses a gradient boosting tree algorithm to keenly capture the complex nonlinear coupling relationship between the parameters of open-circuit finite-length diffusion elements and constant-phase-angle elements, thereby decoupling and identifying the structural camouflage between highly lignified cylinders and dense growth points. The standard logarithmic feature vector is input into this heterogeneous fusion recognition architecture, and the linear prediction confidence and nonlinear prediction confidence are calculated and output independently through the two paths.

[0117] Based on an adaptive soft voting mechanism, a high-order weighted fusion is performed on the linear prediction confidence and the nonlinear prediction confidence, and the joint decision probability of the final output is used to accurately determine whether the tissue under test is the target banana sucker growth point.

[0118] The output interaction unit is used to visualize the tissue category results and is equipped with a communication interface to output growth point location signals to guide automated agricultural equipment to perform bud removal.

[0119] In a more specific example, the process for accurately identifying the growth point of a banana sucker is as follows:

[0120] (1) System setup;

[0121] A detection system was constructed, comprising an electrochemical workstation (such as CHI660e) and a three-electrode probe. The working electrode (WE) was inserted into the geometric center of the banana bulb tissue to be tested, and the reference electrode (RE) and counter electrode (CE) were placed in appropriate positions to form a circuit.

[0122] (2) Data acquisition and fitting;

[0123] AC impedance scanning of the tissue was performed in the frequency range of 1 Hz to 1 MHz. The acquired complex impedance data was imported into the built-in algorithm module. Targeting the vertical rise (finite-length diffusion) characteristic of the impedance spectrum of growth point tissue in the low-frequency band, the program automatically invoked the algorithm containing... The Cole equivalent circuit model of a single CPE (open-circuit finite-length diffusion) element is presented. During parameter identification, a high-precision 11-dimensional electrochemical characteristic parameter is extracted using a local prefitting strategy based on frequency domain decoupling and a step-by-step cross-constraint iterative optimization strategy.

[0124] The following is a brief explanation of the topology selection and fitting results in this step, in conjunction with the attached diagram:

[0125] like Figure 2 As shown, for tissues exhibiting a single relaxation characteristic (such as typical growth points), the Cole single CPE equivalent circuit model is selected for fitting. Figure 4 The impedance spectrum fitting results of the single time constant sample shown clearly demonstrate that the theoretical fitting curve and the measured discrete points exhibit extremely high topological overlap across the entire frequency band, and the open-circuit finite-length diffusion element at the low frequency end accurately quantifies the physical boundary blocking behavior of ions.

[0126] like Figure 3 As shown, for complex structures exhibiting double relaxation superposition characteristics (such as highly lignified cylindrical bodies), the Cole double CPE equivalent circuit model is selected for fitting. Figure 5 The fitting results of the dual-time-constant sample impedance spectrum shown further validate the effectiveness of this topology. Figure 5 As can be seen from the global fitting effect in part (a) and the local high-frequency amplification fitting effect in part (b), the model depicts two overlapping interfacial polarization processes caused by the lignified cell wall and the protoplast membrane, and also achieves the fit with the steep capacitive tail at the low frequency end.

[0127] (3) Pretreatment;

[0128] To eliminate numerical differences between different physical dimensions, the system automatically adjusts the resistance values ​​(units) in the above parameters. ) and time constant (unit) or )implement Logarithmic transformation is used to construct standardized feature vectors.

[0129] (4) Identification and judgment;

[0130] The processed feature vectors are input into the pre-trained VotingFusion model. The LDA branch inside the model outputs probabilities. BO-XGBoost branch output probability .calculate ;like The system outputs a "growth point detected" signal, triggering the bud removal mechanism; otherwise, it outputs "non-growth point".

[0131] (5) Verification results;

[0132] To verify the effectiveness of this embodiment, the model recognition results were compared with the results of paraffin section microscopic observation, which is considered the "gold standard." The confusion matrix of the test set samples was obtained as shown in Table 2. Calculations from the confusion matrix show that the method of this invention achieves an accuracy rate of 20 / (20+4)×100%=83.3% in identifying banana sucker growth points and 47 / (47+1)×100%=97.9% in identifying non-growth points. The overall recognition accuracy of the test set reaches (20+47) / 72×100%≈93.0%, and the false alarm rate for growth point identification is only 1 / (1+47)×100%≈2.1%. Experimental results indicate that the method has an extremely low false alarm rate for growth point identification, with an overall recognition accuracy of 93.0%, effectively overcoming the misjudgment problem of traditional methods under complex physiological conditions.

[0133] To further break the black-box nature of heterogeneous fusion recognition algorithms and verify their underlying physical science, this embodiment introduces the SHAP game theory framework for quantitative attribution analysis. Figure 6 This is a ranking of the contribution of electrochemical features analyzed using the SHAP method in this invention. The global feature importance ranking intuitively reveals that the core features that dominate the classification decision of heterogeneous tissues exhibit a significant gradient distribution. The top three core physical features are, in order, logarithmic ohmic resistance (Log_R0), first-order polarization dispersion index (CPE1-P), and second-order polarization dispersion index (CPE2-P), while logarithmic charge transfer resistance (Log_R1 / Rct), which characterizes the transmembrane polarization barrier, follows closely in fourth place. This graph objectively confirms that the fusion model constructed in this invention, by comprehensively utilizing the above-mentioned multidimensional electrical features, effectively eliminates the interference of water fluctuations and lignification structure camouflage in banana corm tissue, proving the scientific rationality of the features extracted and the identification strategy of this invention.

[0134] Table 2. Confusion matrix of the test set sample set

[0135]

[0136] The core innovations of this invention are reflected in two main aspects. First, it proposes a modeling and parameter extraction method using a Cole single CPE equivalent circuit model or a Cole double CPE equivalent circuit model adapted to the complex tissues of banana bulbs. By introducing a finite-length diffusion element to characterize the low-frequency boundary blocking effect, and adopting a fitting strategy of "locking the diffusion time constant," it solves the problem that traditional models cannot fit the "vertical tail" characteristics of biological tissues and the problem of parameter divergence. Second, it constructs an interference-resistant and high-precision heterogeneous multi-model fusion identification system for plant tissues. By reconstructing the feature space through logarithmic processing, it eliminates the order-of-magnitude barrier and multiplicative drift. At the same time, it establishes a heterogeneous model fusion strategy based on VotingFusion, which effectively decouples water content fluctuations from tissue structure characteristics, and achieves highly robust identification of banana sucker growth points under complex physiological conditions.

[0137] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0139] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for identifying banana suckers based on the fusion of multidimensional electrochemical features, characterized in that, Includes the following steps: S1. Perform a wideband frequency scan on the banana bulb tissue to be tested to obtain complex impedance data; S2. Based on the geometric topology of the measured Nyquist plot, an equivalent circuit model is selected to fit the complex impedance data. A local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling is adopted to extract multidimensional electrochemical characteristic parameters. S3. Perform a base-10 logarithmic spatial reconstruction preprocessing on the ohmic and polarization resistance features and relaxation time constant features with physical dimensions among the multidimensional electrochemical characteristic parameters, so as to reduce the nonlinear multiplicative error caused by the drastic fluctuation of the water content of individual plants into a linear additive bias, thereby obtaining a standard logarithmic feature vector that eliminates the order-of-magnitude barrier. S4. Construct a heterogeneous fusion architecture based on heterogeneous integration. This architecture cascades a linear discriminant path for capturing the global linear chassis and a nonlinear fine-grained path for uncovering local nonlinear details. The linear discriminant path uses a linear discriminant analysis algorithm to establish a stable macroscopic judgment baseline, while the nonlinear fine-grained path uses a gradient boosting tree algorithm to capture the complex nonlinear coupling relationship between the parameters of open-circuit finite-length diffusion elements and constant phase angle elements, thereby decoupling and identifying the structural camouflage between highly lignified cylinders and dense growth points. The standard logarithmic feature vector is input into this heterogeneous fusion architecture, and the linear prediction confidence and nonlinear prediction confidence are calculated and output independently through the two paths. S5. Based on the adaptive soft voting mechanism, perform high-order weighted fusion of the linear prediction confidence and the nonlinear prediction confidence, and determine whether the tissue to be tested is the target banana sucker growth point based on the final output joint decision probability.

2. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 1, characterized in that, In step S1, a three-electrode probe is used to perform a wideband frequency scan on the banana bulb tissue to be tested; The three-electrode probe includes a working electrode, a reference electrode, and a counter electrode. The working electrode is inserted into the geometric center of the banana bulb tissue to be tested. The reference electrode and the counter electrode work together to form a detection circuit. A portable electrochemical workstation is used to perform frequency scanning. The wide frequency range is 1 Hz to 1 MHz.

3. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 1, characterized in that, In step S2, the equivalent circuit model is either a Cole single CPE equivalent circuit model or a Cole double CPE equivalent circuit model, and the topology is selected based on the frequency domain response characteristics of the Nyquist plot and the Bode plot. Among them, for the tissue under test exhibiting a single relaxation characteristic, the selected Cole single CPE equivalent circuit model is composed of three components connected in series in terms of topology: the solution resistance, the parallel branch consisting of the charge transfer resistance and the constant phase angle element, and the open-circuit finite-length diffusion element. For the test tissue exhibiting double relaxation superposition characteristics, the selected Cole double CPE equivalent circuit model cascades the solution resistance, the first-stage charge transfer resistor and the constant phase angle element in parallel branch, the second-stage charge transfer resistor and the constant phase angle element in parallel branch, and the open-circuit finite-length diffusion element connected in series at the end; the constant phase angle element is used to characterize the frequency diffusion effect caused by biomembrane roughness, and the open-circuit finite-length diffusion element is used to characterize the boundary blocking effect induced by low-frequency polarized ions in the extremely dense cellular confined space.

4. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 1, characterized in that, In step S2, the local prefitting and step-by-step cross-constraint iterative optimization strategy based on frequency domain decoupling is specifically as follows: Select the corresponding equivalent circuit model based on the single capacitive arc or double capacitive arc characteristics presented by the impedance spectrum; Local prefitting is used to independently fit the mid-to-high frequency polarization semicircle and the low frequency diffusion oblique line, extract the initial parameter values ​​and load them into the equivalent circuit model, and set all component parameters in the equivalent circuit model to a locked state; The parameters of the low-frequency open-circuit finite-length diffusion element are released first for fitting, and the parameters of the open-circuit finite-length diffusion element are locked when the low-frequency fitting residual meets the preset convergence condition. Subsequently, the parameters of the mid-to-high frequency branches were released and fitted using the topological geometry of the Nyquist plot and Bode plot; Based on the dynamic feedback of local residuals, the constraint locking and releasing adjustment steps of parameters in each frequency band are executed alternately until the global weighted chi-square error converges to... Less than orders of magnitude.

5. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 1, characterized in that, The multidimensional electrochemical characteristic parameters include ohmic internal resistance, charge transfer resistance, double layer capacitance, and diffusion impedance.

6. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 1, characterized in that, In step S4, the calculation process of the linear discriminant path is as follows: Construct inter-class scatter matrix and intra-class scatter matrix based on the training set; The optimal projection vector is calculated based on the Fisher criterion to maximize the ratio of between-class divergence to within-class divergence. The input logarithmic feature vector is projected onto the linear space generated by the optimal projection vector to obtain the projection value; The projected values ​​are mapped to linear discriminant probabilities using the Sigmoid function to characterize the confidence that a sample belongs to a growth point on the global impedance scale.

7. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 6, characterized in that, In step S4, the calculation process of the nonlinear fine-grained pathway is as follows: To explore the nonlinear coupling relationship between diffusion impedance parameters and double-layer capacitance parameters; A target function containing a loss function and a regularization term is constructed, and the target function is approximated by a second-order Taylor expansion to isolate and extract the complex nonlinear coupling characteristics between the parameters of the open-circuit finite-length diffusion element and the constant phase angle parameter. A Bayesian optimization module is introduced, which uses a Gaussian process as a surrogate model to perform global iterative optimization of the hyperparameter combination of learning rate, maximum tree depth and regularization coefficient of the XGBoost algorithm, and outputs the optimal nonlinear discrimination probability.

8. The banana sucker identification method based on multidimensional electrochemical feature fusion according to claim 7, characterized in that, In step S5, the linear and nonlinear discriminant probabilities are weighted and fused based on an adaptive soft voting mechanism, specifically as follows: ； In the formula, The weighting coefficients of the linear path are represented. Represents the weighting coefficients of the nonlinear path; and These represent linear and nonlinear discrimination probabilities, respectively. If the final confidence level If the value exceeds the preset threshold, the current detection location is determined to be the banana sucker growth point.

9. A banana bud recognition system based on multidimensional electrochemical feature fusion, characterized in that, It includes a data acquisition unit, a data processing unit, and an output interaction unit, which are connected in sequence. The data acquisition unit is equipped with a portable electrochemical workstation and a customized three-electrode probe assembly. The three-electrode probe is configured to penetrate the test tissue site of the banana bulb to perform a broadband impedance scanning task. The data processing unit is configured to implement the banana sucker recognition method based on multidimensional electrochemical feature fusion as described in any one of claims 1-8; The output interaction unit is used to visually display the output results of the data processing unit.

10. A banana bud recognition system based on multidimensional electrochemical feature fusion according to claim 9, characterized in that, The output interaction unit is equipped with a communication interface, which is used to output a growth point position signal to guide automated agricultural equipment to perform bud removal.

Images