A physiological nutrition method for secondary flowering of phalaenopsis combined with multi-spectrum diagnosis

Abstract

The present application provides a kind of secondary flowering physiological nutrition method of phalaenopsis combined with multispectral diagnosis, belongs to the technical field of multispectral diagnosis, the present application is collected to the leaf of phalaenopsis flower bud differentiation period by multispectral image, the concentration of anthocyanin and sugar nitrogen ratio spectrum characteristics are extracted using microstate entropy increase dynamics equation constraint, fusion feature vector, input multisource physiological nutrition diagnosis model, rely on dynamic hierarchical early exit mechanism and particle swarm optimization dynamic attenuation sliding window adaptive algorithm complete adaptive deep reasoning, combined with space-time high-order tensor conduction mechanism output nutrient regulation sequence, finally through comprehensive evaluation function to water and fertilizer execution mechanism executes hierarchical adaptive regulation, forms closed-loop nutrient regulation system, solves the technical problem that multisource heterogeneous characteristics cannot realize low-delay alignment fusion and adaptive deep reasoning in phalaenopsis secondary flowering nutrient diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of multispectral diagnostic technology, and more specifically, relates to a physiological nutrition method for Phalaenopsis orchids that combines multispectral diagnostics. Background Technology

[0002] Precise nutrient regulation for Phalaenopsis orchids to bloom twice is a core technological requirement in the field of high-end flower facility cultivation. Currently, it mainly relies on multi-source heterogeneous methods such as multispectral imaging, soil conductivity sensing, and leaf temperature sensing to collect information on the physiological state of plants, and uses deep learning models to diagnose and infer nutrient requirements. In existing technologies, multispectral data possesses high-dimensional broadband characteristics, while soil electrical conductivity and leaf temperature are low-dimensional scalars. These three types of data differ fundamentally in sampling frequency, data dimensionality, and physical dimensions. When these heterogeneous features are directly concatenated and fed into a unified inference model, the model cannot establish effective cross-modal alignment relationships in the unified latent space due to nonlinear coupling and dimensionality mismatch between the modal feature spaces. This leads to a significant decrease in the confidence of diagnostic inference as the degree of feature coupling intensifies. Simultaneously, existing deep learning inference architectures perform full-layer inference on all input samples. Under the dynamic physiological rhythm from Phalaenopsis orchid bud differentiation to secondary flowering, simple and complex samples are allocated the same computational depth, resulting in a severe waste of memory resources and neuron activation computation. Furthermore, the temporal window decay coefficient is usually manually fixed and cannot be dynamically adjusted according to the rhythm of different physiological stages, further reducing the utilization efficiency of temporal features. In other words, existing technologies suffer from the technical problem of failing to achieve low-latency alignment fusion and adaptive deep inference for multi-source heterogeneous features in the nutrient diagnosis of Phalaenopsis orchid secondary flowering. Summary of the Invention

[0003] In view of this, the present invention provides a physiological nutrition method for Phalaenopsis orchids in secondary flowering that combines multispectral diagnosis, which can solve the technical problem in the prior art that multi-source heterogeneous features cannot achieve low-latency alignment fusion and adaptive deep inference in the nutrient diagnosis of Phalaenopsis orchids in secondary flowering.

[0004] This invention is achieved as follows: This invention provides a physiological and nutritional method for the secondary flowering of Phalaenopsis orchids that combines multispectral diagnostics, comprising the following steps:

[0005] Multispectral images of Phalaenopsis flower bud differentiation period leaves were acquired in real time. Low-rank tensor spatial projection was performed on the original high-dimensional multispectral data to reduce the dimensionality of the data. The reduced band data was then stored in a streaming double buffer structure to complete band alignment and streaming caching.

[0006] The multispectral data in the convection buffer is constrained by the micro-state entropy increase dynamic equation, the energy boundary is suppressed for spectral noise, the leaf anthocyanin reflectance feature value and sugar-nitrogen ratio spectral feature vector are extracted, and the cosine similarity is compared with the standard feature vector library to output the anthocyanin concentration diagnostic value and sugar-nitrogen ratio diagnostic value.

[0007] Soil electrical conductivity and leaf temperature environmental parameters are collected. Multispectral image features, soil electrical conductivity and leaf temperature are fed into a shared latent space orthogonal decoupling operator to perform heterogeneous feature alignment and fusion, and output a unified low-dimensional orthogonal fused feature vector.

[0008] The unified low-dimensional orthogonal fusion feature vector is input into the multi-source physiological nutrition diagnosis model. Inference based on the dynamic hierarchical early exit mechanism is executed to output the nutrient demand diagnosis vector of Phalaenopsis orchid in the second flowering stage. At the same time, the spatiotemporal high-order tensor transmission mechanism is executed to model the spatial diffusion and temporal accumulation trajectory of nutrients and output the nutrient regulation sequence.

[0009] Calculate the comprehensive nutrient status assessment function value, and based on the interval to which the comprehensive assessment function value belongs, perform graded adaptive adjustment of the fertilizer concentration, fertilizer frequency and liquid volume parameters of the water and fertilizer irrigation actuator to complete closed-loop nutrient regulation;

[0010] Optionally, it also includes: integrating a multispectral image acquisition device, a soil conductivity sensor, a leaf temperature sensor, a multi-source physiological nutrition diagnostic model, and a water and fertilizer actuator into a Phalaenopsis orchid secondary flowering physiological nutrition regulation system, wherein the system is equipped with a readable storage medium, which stores program instructions, and the program instructions are used to execute the above methods when running in a computer.

[0011] Specifically, the low-rank tensor spatial projection involves constructing the original multispectral data into a third- or fourth-order tensor, performing a low-rank approximation on each mode of the tensor using Tucker decomposition, retaining principal components whose energy percentage exceeds the energy percentage threshold, discarding redundant bands, and reducing the data dimension from the original number of bands to the number of low-rank bands after dimensionality reduction.

[0012] Specifically, the streaming dual-buffer structure consists of two physically independent memory buffers. The acquisition thread continuously writes the latest dimensionality-reduced band data to one of the buffers, while the processing thread synchronously reads data from the other buffer to perform subsequent operations. The two buffers switch roles through a lock-free pointer atomic swap mechanism, and the switching frequency is dynamically matched with the acquisition frequency.

[0013] Specifically, the microscopic state entropy increase dynamic equation constraint abstracts the metabolic state of Phalaenopsis leaf cells into a thermodynamic open system, calculates the Shannon entropy value of the probability distribution corresponding to the multi-band reflectance sequence of the current acquisition frame as the microscopic metabolic state entropy of the current frame, establishes upper and lower bound constraints on the entropy increase rate between adjacent frames, and determines that when the change in reflectance of a certain band causes the entropy increase rate to exceed the upper bound constraint, it is an environmental noise interference, and weighted smoothing filtering is performed on the corresponding band.

[0014] Specifically, the cosine similarity comparison involves performing cosine similarity calculations on the extracted anthocyanin reflectance feature value and sugar-nitrogen ratio spectral feature vector with pre-stored feature vectors in a standard feature vector library. When the similarity value is lower than the lower similarity threshold, an abnormal state marker is triggered.

[0015] Specifically, the shared latent space orthogonal decoupling operator constructs independent linear projection matrices for the multispectral image feature vector, soil electrical conductivity scalar, and leaf surface temperature scalar, applies orthogonality constraints to each projection matrix, and makes the three features mutually orthogonal in the low-dimensional latent space after projection. Then, the three orthogonal projection results are spliced ​​together into a unified low-dimensional orthogonal fusion feature vector.

[0016] The structure of the multi-source physiological nutrition diagnostic model is as follows: the input end receives a unified low-dimensional orthogonal fusion feature vector, which is mapped to a low-dimensional shared latent space through an orthogonal projection matrix. The feature flow then enters the low-level inference module, the middle-level inference module, and the high-level inference module in sequence. Each inference module has an ultra-lightweight confidence classifier based on the state entropy determination method embedded at its end. When the confidence meets the conditions, it triggers early exit.

[0017] Specifically, the state entropy determination method calculates the Shannon information entropy value of the probability distribution of the current layer's output feature vector after Softmax normalization, compares the entropy value with a confidence threshold, and triggers early exit when the entropy value is lower than the confidence threshold.

[0018] The high-level inference module consists of deep convolutional layers containing a nonlinear gated jump network. The nonlinear gated jump network independently calculates the gate weights for each dimension of the input feature vector. The gate weights determine whether the feature energy of that dimension flows to the current layer or jumps to the residual path.

[0019] Among them, the multi-source physiological nutrition diagnosis model embeds a particle swarm optimization dynamic decay sliding window adaptive algorithm. Specifically, when initializing the particle swarm, each particle carries two positional dimensions: decay coefficient and window width. The weighted value of forward diagnosis accuracy and inference calculation time is used as the fitness function. The output optimal decay coefficient is applied to the temporal attention weight matrix to perform adaptive exponential decay on historical multispectral frame information.

[0020] Specifically, the spatiotemporal high-order tensor transmission mechanism involves constructing a high-order tensor along the time axis and the plant spatial location axis from the nutrient diagnosis vector within the historical time window, performing Tucker decomposition on the tensor to extract the temporal modality factor matrix and the spatial modality factor matrix, reconstructing the nonlinear spatiotemporal evolution trajectory through tensor product, and outputting a sequence of nutrient demand prediction vectors for each spatial region within the future control step.

[0021] Among them, the comprehensive nutrient status assessment function value Normalized results of anthocyanin concentration diagnostic values Normalized results of glucose-nitrogen ratio diagnostic values Normalized deviation from nutrient regulation sequence The weighted calculation is obtained, and the formula is: ,in , , These are weighted coefficients, and the sum of the three is 1.

[0022] Specifically, the hierarchical adaptive adjustment is when Less than the lower threshold If nutrient deficiency is detected, adjust the fertilizer concentration to a high level and increase the fertilization frequency to the highest level; when exist and If the nutrient level is slightly low, adjust the fertilizer concentration to a medium level; when exist and When the nutrient levels are deemed suitable, only minor adjustments are made to the application rate; when... Not less than If nutrient excess is detected, the fertilizer concentration should be reduced and the fertilization interval extended.

[0023] Among them, the training dataset of the multi-source physiological nutrition diagnostic model was established in a controlled cultivation experimental greenhouse. Multispectral images, soil electrical conductivity and leaf temperature data were collected synchronously, and chemical detection results of leaf anthocyanin content and sugar-nitrogen ratio were collected synchronously as annotation labels to construct a multimodal time-series labeled dataset.

[0024] The energy percentage threshold ranges from 90% to 99%; the low-dimensional band dimension ranges from 5 to 20; the similarity lower limit threshold ranges from 0.80 to 0.92; the low-dimensional spatial dimension of the unified low-dimensional orthogonal fusion feature vector ranges from 8 to 32; the confidence threshold ranges from 0.10 to 0.40; the attenuation coefficient search ranges from 0.85 to 0.99; the window width search ranges from 4 to 32 frames; and the tensor order of the spatiotemporal higher-order tensor ranges from 3 to 5. The range is 0.20 to 0.35. The range is 0.45 to 0.60. The range is 0.70 to 0.85. The range is 0.3 to 0.5. The range is 0.2 to 0.4. The range is 0.1 to 0.3.

[0025] This invention constructs a shared latent space orthogonal decoupling operator to map three heterogeneous features—multispectral image features, soil conductivity, and leaf temperature—to a unified low-dimensional latent space via independent orthogonal projection. This eliminates linear coupling between heterogeneous features, ensuring that the three modes are on a linearly separable orthogonal basis before fusion, fundamentally eliminating feature alignment errors caused by dimensional mismatch and dimensional differences. Furthermore, a dynamic hierarchical early exit mechanism is implemented, using the state entropy judgment method to evaluate the feature distribution confidence at the end of each inference module in real time. This triggers early exit for simple samples at low or middle levels, concentrating deep memory allocation and neuron activation calculations on complex, pathological samples that truly require deep extraction, avoiding redundant occupation of computational resources by full-layer inference. Simultaneously, a particle swarm optimization dynamic decay sliding window adaptive algorithm uses swarm intelligence to search for the optimal temporal decay coefficient and window width in the two-dimensional parameter space, dynamically adjusting the temporal attention weight according to the rhythm of different physiological stages of the Phalaenopsis orchid, avoiding overfitting or underutilization of historical information caused by fixed window parameters. In summary, this invention solves the technical problem mentioned in the background art that multi-source heterogeneous features cannot achieve low-latency alignment fusion and adaptive deep inference in the nutrient diagnosis of Phalaenopsis orchids during secondary flowering. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a time-series curve showing the diagnostic values ​​of anthocyanin concentration and sugar-nitrogen ratio.

[0028] Figure 3 The structure diagram of the cross-correlation matrix of the latent space after orthogonal projection of the three heterogeneous features is shown.

[0029] Figure 4 The activation frequency distribution of each inference layer during the inference process of 120 plants is shown in the figure.

[0030] Figure 5 This is a comparison of the optimal decay coefficient and window width for the particle swarm optimization algorithm to converge in the early stage of flower bud differentiation and near the flowering stage.

[0031] Figure 6 This is a flowchart illustrating the overall architecture of the Phalaenopsis orchid secondary flowering physiological nutrition regulation system of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0033] like Figure 1 The diagram shown is a flowchart of a physiological and nutritional method for secondary flowering of Phalaenopsis orchids combined with multispectral diagnostics provided by this invention. This method includes the following steps:

[0034] S01. Real-time acquisition of multispectral images of Phalaenopsis orchid leaves during the bud differentiation period, with the acquisition band range covering 400-1000nm and the acquisition frequency being 1-10Hz. Low-rank tensor spatial projection dimensionality reduction is performed on the original multispectral high-dimensional data, and the dimensionality-reduced band data is stored in a streaming double buffer structure to complete band alignment and streaming buffering.

[0035] S02. The multispectral data in the convection buffer is constrained by the micro-state entropy increase dynamic equation, the energy boundary is suppressed for spectral noise, the leaf anthocyanin reflectance feature value and sugar-nitrogen ratio spectral feature vector are extracted, and the cosine similarity is compared with the standard feature vector library to output the anthocyanin concentration diagnostic value and sugar-nitrogen ratio diagnostic value.

[0036] S03. Collect soil electrical conductivity and leaf temperature environmental parameters, send multispectral image features, soil electrical conductivity and leaf temperature into the shared latent space orthogonal decoupling operator, perform heterogeneous feature alignment and fusion, and output a unified low-dimensional orthogonal fusion feature vector.

[0037] S04. Input the unified low-dimensional orthogonal fusion feature vector into the multi-source physiological nutrition diagnosis model, execute inference based on the dynamic hierarchical early exit mechanism, output the nutrient demand diagnosis vector of the second flowering stage of Phalaenopsis orchid, and simultaneously execute the spatiotemporal high-order tensor transmission mechanism to model the spatial diffusion and temporal accumulation trajectory of nutrients, and output the nutrient regulation sequence.

[0038] S05. Calculate the comprehensive nutrient status assessment function value, and according to the interval to which the comprehensive assessment function value belongs, perform graded adaptive adjustment of the fertilizer concentration, fertilizer frequency and liquid volume parameters of the water and fertilizer irrigation actuator to complete the closed-loop nutrient regulation.

[0039] Optionally, it also includes: S06, integrating a multispectral image acquisition device, a soil conductivity sensor, a leaf surface temperature sensor, the multi-source physiological nutrition diagnostic model, and a water and fertilizer actuator into a Phalaenopsis orchid secondary flowering physiological nutrition regulation system, wherein the system is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above method when running in a computer.

[0040] The low-rank tensor spatial projection refers to constructing the original multispectral data into a third- or fourth-order tensor, performing low-rank approximation on each mode of the tensor using Tucker decomposition, retaining principal component components with energy proportions exceeding a set threshold, and discarding redundant bands. This achieves lossless compression and dimensionality reduction of high-dimensional band data on the front-end hardware side. After dimensionality reduction, the data dimension is reduced from the original number of bands to 5-20 dimensions. The energy proportion threshold is determined by singular value decomposition analysis experiments on multispectral data of different Phalaenopsis orchid sample batches, and the suggested range is 90%-99%.

[0041] The streaming dual-buffer structure refers to setting up two physically independent memory buffers. The acquisition thread continuously writes the latest dimensionality-reduced band data to one buffer, while the processing thread synchronously reads data from the other buffer to perform subsequent operations. The two buffers switch roles through a lock-free pointer atomic swap mechanism, eliminating mutual exclusion and waiting between the acquisition thread and the processing thread during the transmission of multispectral data streams. This achieves decoupling of the three-stage pipeline of spectral acquisition, image preprocessing, and model inference, reducing throughput latency from milliseconds to microseconds in the traditional synchronous lock architecture. The specific switching frequency is dynamically matched with the acquisition frequency.

[0042] The microscopic entropy increase kinetic equation constraint refers to abstracting the metabolic state of Phalaenopsis leaf cells into a thermodynamic open system, defining a probability distribution corresponding to the multi-band reflectance sequence of the current spectral acquisition frame, calculating the Shannon entropy value of the probability distribution as the microscopic metabolic state entropy of the current frame, establishing upper and lower bound constraints on the entropy increase rate between adjacent frames, and determining environmental noise interference when the entropy increase rate exceeds the upper bound constraint due to a change in reflectance of a certain band. Weighted smoothing filtering is then performed on the corresponding band of the frame to enhance the expression ability of weak reflectance features of anthocyanins and sugar-nitrogen ratio under low signal-to-noise ratio conditions. The upper and lower bound constraint parameters are determined by joint analysis of multiple batches of Phalaenopsis orchid bud differentiation period leaf spectral acquisition experiments combined with cell metabolic physiological data, and the initial range is suggested to be 0.01 to 0.15 entropy change rate per frame.

[0043] The standard feature vector library refers to the set of anthocyanin and sugar-nitrogen ratio spectral feature vectors extracted from Phalaenopsis orchid samples with different nutrient levels after pre-collecting multispectral images and processing them for dimensionality reduction and noise constraint. The similarity value output by the cosine similarity comparison ranges from -1 to 1. When the similarity value is lower than the set lower threshold, an abnormal state marker is triggered. The lower threshold is determined by controlled cultivation experiments in the laboratory, and the suggested range is 0.80 to 0.92.

[0044] The shared latent space orthogonal decoupling operator refers to constructing independent linear projection matrices for the multispectral image feature vector, soil electrical conductivity scalar, and leaf surface temperature scalar, respectively, applying orthogonality constraints to each projection matrix to make the three features mutually orthogonal in the low-dimensional latent space after projection, eliminating the linear coupling between heterogeneous features, and then splicing the three orthogonal projection results into a unified low-dimensional orthogonal fusion feature vector. The low-dimensional space dimension is recommended to be 8 to 32 dimensions, and the specific dimension is determined by the cross-modal reconstruction error minimization experiment.

[0045] The multi-source physiological nutrition diagnostic model is a deep learning-based fusion inference model that integrates a dynamic decay sliding window adaptive algorithm based on particle swarm optimization. The specific structure of the multi-source physiological nutrition diagnostic model is as follows: the model input receives a unified low-dimensional orthogonal fusion feature vector. First, the multi-source heterogeneous features are forcibly mapped to an extremely low-dimensional shared latent space using an orthogonal projection matrix (8-32 dimensions recommended), achieving hard alignment of cross-modal features. Then, the feature flow sequentially enters a three-level structure: a low-level inference module, a mid-level inference module, and a high-level inference module. The low-level inference module consists of 2-3 fully connected layers, responsible for extracting simple linear spectral features. The mid-level inference module consists of 2-3 nonlinear layers with residual jump connections. The high-level inference module consists of 3-5 deep convolutional layers with nonlinear gated jump networks. The control jump network performs element-wise gating modulation on the feature stream, controlling the distribution of feature energy to different deep channels. Each inference module embeds an ultra-lightweight confidence classifier based on state entropy determination. This classifier calculates the distribution entropy value of the current layer's feature vector and compares it with a preset confidence threshold. When the feature distribution entropy is lower than the threshold, it is determined that the current feature meets the confidence requirement for secondary flowering physiological diagnosis, triggering a jump instruction to allow the signal to bypass subsequent higher-level inference modules and exit early, directly outputting the nutrient demand diagnosis vector. If the feature distribution entropy exceeds the threshold, it continues to propagate along the feature stream to higher layers until full-layer inference is completed or early exit is triggered again at a higher layer. The hierarchical early exit mechanism allows regular samples to complete inference only through low or middle layers, and deep memory allocation is activated only when pathological or complex spectral features appear, achieving adaptive adjustment of memory allocation and activation weight allocation between neurons. The model embeds a particle swarm optimization dynamic decay sliding window adaptive algorithm to control the decay window of the model's temporal attention mechanism. When initializing the particle swarm, each particle carries two positional dimensions: decay coefficient and window width. The fitness function is a weighted value of forward diagnostic accuracy and inference computation time. Particles update their speed and position based on their individual historical best position and the group's global best position. The suggested search range for the decay coefficient is 0. The recommended search range for the window width is 4 to 32 frames, with a value of 85 to 0.99. The final output optimal attenuation coefficient is applied to the temporal attention weight matrix, and adaptive exponential attenuation is performed on the historical multispectral frame information. The weights decrease between adjacent temporal frames according to the optimal attenuation coefficient. The attention weight of the historical frame to the current diagnosis decreases exponentially with the increase of time distance. The spatiotemporal high-order tensor transmission mechanism is executed in parallel at the output of the high-level inference module. The historical nutrient diagnosis vector sequence is constructed into a spatiotemporal high-order tensor. The nonlinear evolution trajectory of nutrients in the spatial diffusion dimension and the temporal accumulation dimension is explicitly modeled through tensor convolution, and the nutrient regulation sequence is output. The tensor order is recommended to be 3 to 5.The steps for establishing the training dataset for the multi-source physiological and nutritional diagnostic model specifically include: cultivating Phalaenopsis orchid samples with different nutrient levels in a controlled cultivation experimental greenhouse, covering all physiological stages from flower bud differentiation to secondary flowering; collecting multispectral images, soil conductivity, and synchronous data on leaf surface temperature; simultaneously collecting chemical detection results of leaf anthocyanin content and sugar-nitrogen ratio as annotation labels; and recording corresponding temperature, humidity, and light parameters to construct a multimodal time-series labeled dataset. The dataset size is recommended to be no less than 5000 time-series sample groups. The training steps for the multi-source physiological and nutritional diagnostic model specifically include: jointly supervising the confidence classifiers of each layer and the final nutrient diagnostic output using a multivariate cross-entropy loss function; employing gradient pruning to prevent gradient explosion in deep convolutional layers, with a recommended gradient pruning threshold range of 0.5–5.0; and using a cosine annealing learning rate scheduling strategy, with an initial learning rate range of [insert range here]. ~ Early termination should be based on the combined indicators of diagnostic accuracy and computation time on the validation set. The number of training rounds should not exceed 500. The confidence threshold for early termination of each layer should be determined by multiple rounds of ablation experiments on the validation set, with an initial range of 0.80 to 0.95.

[0046] The multi-source physiological nutrition diagnostic model brings the following technical benefits to the entire scheme: Orthogonal hard alignment of the shared latent space eliminates alignment errors caused by differences in sampling frequency and dimension of multi-source heterogeneous features, making the information of each modality in a unified linearly separable space before feature fusion; The dynamic hierarchical early exit mechanism enables the network to autonomously determine the computation depth according to the complexity of feature distribution, thereby prioritizing the allocation of limited memory resources and neuron activation computation to pathological samples that truly require deep extraction, significantly reducing the inter-layer data transmission and activation computation overhead of conventional samples; The particle swarm optimization dynamic decay sliding window adaptive algorithm searches for the optimal temporal decay parameters through swarm intelligence, enabling the model to adaptively adjust the temporal dependence length of different physiological stages of Phalaenopsis orchids, avoiding overfitting or underutilization of historical information due to fixed window parameters; The spatiotemporal high-order tensor transmission mechanism replaces implicit circular memory with explicit tensor modeling, making the nonlinear evolution trajectory of nutrient spatial diffusion and temporal accumulation have an interpretable tensor component correspondence within the model, improving the physiological interpretability and temporal robustness of nutrient regulation sequence prediction.

[0047] The principle of the particle swarm optimization dynamic decay sliding window adaptive algorithm is as follows: A two-dimensional continuous search space is constructed using the temporal attention decay coefficient and the window width. The position of each particle in the particle swarm corresponds to a set of parameter configurations. Particles update their motion speed based on their historical best position and global best position according to inertia weight, individual learning coefficient, and social learning coefficient. The suggested range for inertia weight is 0.4–0.9, and the suggested ranges for individual and social learning coefficients are both 1.5–2.5. The suggested number of particles is 20–80. These coefficients are determined through multiple rounds of iterative experiments on the validation set accuracy-time joint fitness. The algorithm is executed online at a fixed frequency after training, with diagnostic inference performed at regular intervals. The batch re-optimization and update of the attention weight matrix allows the temporal decay parameters to be dynamically adjusted according to the growth rhythm of the Phalaenopsis orchid. The recommended re-optimization frequency is once every 100-500 inference batches. The technical effects of the algorithm are as follows: it avoids the overfitting of the model to specific physiological stages caused by manually fixing the temporal decay coefficient; it dynamically finds the optimal balance between accuracy and computation time through group collaborative search; it enables the model to maintain efficient and accurate nutrient status diagnosis under different physiological rhythms from the Phalaenopsis orchid flower bud differentiation period to the secondary flowering period; and it avoids the problem of single gradient optimization methods easily getting trapped in local optima on non-convex fitness surfaces, thus enhancing the global search capability of temporal parameter optimization.

[0048] The comprehensive nutrient status assessment function value is calculated by weighting the anthocyanin concentration diagnostic value, the sugar-nitrogen ratio diagnostic value, and the nutrient regulation sequence deviation value, and is denoted as... , Calculated by the following formula: ;in The results are normalized for the diagnostic values ​​of anthocyanin concentration. The result is a normalized result of the glucose-nitrogen ratio diagnostic value. This represents the normalized value of the nutrient regulation sequence bias. , , These are weighting coefficients, with the sum of the three being 1. The specific weights are determined by analyzing the contribution of each indicator to the success rate of secondary flowering in multiple cultivation experiments, with suggested initial ranges of 0.3–0.5, 0.2–0.4, and 0.1–0.3 respectively. If nutrient deficiency is detected, the fertilizer concentration will be increased to a high level, and the frequency of fertilization will be increased to the highest level. The lower limit threshold; when If the nutrient level is slightly low, adjust the fertilizer concentration to the medium range and maintain the standard fertilization frequency; When the nutrient levels are determined to be within the appropriate range, the current fertilizer concentration and frequency are maintained, with only minor adjustments made to the liquid application rate; when If nutrient excess is detected, the fertilizer concentration is reduced to a low level and the fertilization interval is extended; , , The thresholds were determined by statistical analysis of flowering rate and physiological indicators under various nutrient levels in multiple batches of Phalaenopsis orchid secondary flowering experiments. The suggested initial ranges are 0.20–0.35, 0.45–0.60, and 0.70–0.85, respectively.

[0049] The aforementioned high-order spatiotemporal tensor transmission mechanism refers to constructing a high-order tensor along the time axis and the plant spatial location axis from the nutrient diagnosis vector within the historical time window, performing Tucker decomposition on the tensor to extract the temporal modality factor matrix and the spatial modality factor matrix, reconstructing the nonlinear spatiotemporal evolution trajectory through tensor product, and outputting the nutrient demand prediction vector sequence for each spatial region within the future control step.

[0050] The nonlinear gated jump network refers to embedding element-wise Sigmoid gate units between deep convolutional layers. The gate unit independently calculates the gate weight for each dimension of the input feature vector. The gate weight determines whether the feature energy of that dimension flows to the current layer or jumps to the residual path, realizing the dynamic distribution of feature flow between deep channels and avoiding unnecessary inter-layer data transfer overhead caused by all features passing through all deep convolutions.

[0051] The state entropy determination method refers to calculating the Shannon information entropy value of the probability distribution of the current layer output feature vector after Softmax normalization, comparing the entropy value with a preset confidence threshold. The lower the entropy value, the more concentrated the feature distribution, and the higher the corresponding diagnostic confidence. When the threshold is reached, early exit is triggered. The threshold is determined by the validation set ablation experiment, and the suggested range is 0.10 to 0.40 (in units of natural logarithmic entropy).

[0052] Optionally, this solution also provides a computer-based physiological and nutritional regulation system for Phalaenopsis orchids' secondary flowering. The system integrates a multispectral image acquisition device, a soil conductivity sensor, a leaf surface temperature sensor, and a water and fertilizer actuator. The system is equipped with a readable storage medium that stores program instructions. When the program instructions are run on the computer, they are used to drive the above-mentioned devices to collaboratively execute all the steps of this method.

[0053] The specific implementation of step S01 is as follows. During the differentiation period of Phalaenopsis flower buds, the leaves are scanned in real time using a multispectral image acquisition device. The acquisition band range covers 400–1000 nm, and the acquisition frequency is set to 1–10 Hz. The raw multispectral data acquired in each frame is a high-dimensional broadband matrix, and direct transmission and processing would bring huge computational and storage overhead. Therefore, each frame of raw data is constructed as a third- or fourth-order tensor. Tucker decomposition is used to perform low-rank approximation on each mode of the tensor. The principle is to decompose the high-order tensor into the product of the core tensor and each mode factor matrix. By retaining the principal component components whose energy proportion exceeds the energy proportion threshold (recommended range 90%–99%) and discarding redundant bands, lossless compression and dimensionality reduction of high-dimensional band data is achieved on the front-end hardware side. After dimensionality reduction, the data dimension is reduced from the original number of bands to 5–20 dimensions. The dimensionality-reduced band data is written into a streaming double-buffered structure. This structure sets up two physically independent memory buffers. The acquisition thread continuously writes data to one buffer through a lock-free pointer atomic swap mechanism, while the processing thread synchronously reads data from the other buffer. The switching frequency of the two buffers is dynamically matched with the acquisition frequency, eliminating mutual exclusion waiting between the acquisition thread and the processing thread. This decouples the three stages of spectral acquisition, image preprocessing, and model inference into a pipeline structure, reducing the throughput latency from milliseconds to microseconds in the traditional synchronous lock architecture.

[0054] The specific implementation of step S02 is as follows: The multi-band reflectance sequence of the current frame after dimensionality reduction is read from the streaming buffer and treated as a discrete probability distribution. The Shannon entropy value of this distribution is calculated as the microscopic metabolic state entropy of the current frame, reflecting the physical modeling idea of ​​abstracting the metabolic state of leaf cells into a thermodynamic open system. Upper and lower bound constraints are established for the entropy growth rate between adjacent frames. The upper and lower bound parameters are jointly determined by multiple batches of leaf spectral experiments during flower bud differentiation and cell metabolic physiological data. It is suggested that the initial range be 0.01–0.15 entropy change rate per frame. When the reflectance change of a certain band causes the entropy growth rate to exceed the upper bound constraint, it is determined that the corresponding band of that frame is interfered with by environmental noise. Weighted smoothing filtering is performed on that band, thereby enhancing the expression ability of weak reflectance features of anthocyanins and sugar-nitrogen ratio under low signal-to-noise ratio conditions. After noise suppression is completed, anthocyanin reflectance feature value and sugar-nitrogen ratio spectral feature vector are extracted from the current frame and compared with the feature vector in the pre-built standard feature vector library using cosine similarity. The similarity value ranges from -1 to 1. When the similarity is lower than the lower limit threshold (suggested range 0.80 to 0.92), an abnormal state marker is triggered. Under normal circumstances, the anthocyanin concentration diagnostic value and sugar-nitrogen ratio diagnostic value are output.

[0055] The specific implementation of step S03 is as follows. Soil conductivity scalars and leaf temperature scalars are simultaneously acquired using a soil conductivity sensor and a leaf temperature sensor. These, along with the multispectral image feature vector output from step S02, are fed into a shared latent space orthogonal decoupling operator. The operator constructs independent linear projection matrices for each of the three inputs and applies orthogonality constraints to each projection matrix during training. Mathematically, this means ensuring that the inner product between the projected components of the three features in the low-dimensional latent space is zero, thus eliminating linear coupling between heterogeneous features. The recommended dimension of the low-dimensional latent space is 8–32 dimensions, with the specific dimension determined through experiments minimizing cross-modal reconstruction errors. The three orthogonal projection results are then concatenated to form a unified low-dimensional orthogonal fusion feature vector, providing linearly separable cross-modal input for the subsequent inference model.

[0056] The specific implementation of step S04 is as follows. A unified low-dimensional orthogonal fusion feature vector input multi-source physiological nutrition diagnostic model is established. The model structure includes a three-level structure: a low-level inference module, a mid-level inference module, and a high-level inference module. The low-level inference module consists of 2-3 fully connected layers, the mid-level inference module consists of 2-3 nonlinear layers with residual jump connections, and the high-level inference module consists of 3-5 deep convolutional layers with nonlinear gated jump networks. An ultra-lightweight confidence classifier is embedded at the end of each inference module to calculate the Shannon information entropy of the current layer's output feature vector after Softmax normalization. When the entropy value is lower than the confidence threshold (suggested range 0.10-0.40), early exit is triggered, so that simple samples can complete inference only through the low-level or mid-level layers, and deep calculations are only activated for pathological samples. The model incorporates a particle swarm optimization algorithm with a dynamic decay sliding window. The particle swarm searches for optimal parameters in a two-dimensional space defined by the decay coefficient (suggested search range 0.85–0.99) and the window width (suggested search range 4–32 frames). The fitness function is a weighted average of forward diagnostic accuracy and inference computation time. Optimization is performed every 100–500 inference batches. The optimal decay coefficient is applied to the temporal attention weight matrix, causing the weight of historical frames to decrease exponentially with increasing time distance. A high-order spatiotemporal tensor propagation mechanism is executed in parallel at the output of the high-level inference module. It constructs the historical nutrient diagnostic vector sequence into 3rd–5th order spatiotemporal high-order tensors, extracts the temporal and spatial modality factor matrices through Tucker decomposition, and reconstructs the nonlinear spatiotemporal evolution trajectory using tensor product, outputting the nutrient regulation sequence.

[0057] The specific implementation method of step S05 is as follows. Nutrient status comprehensive evaluation function value. Normalized results of anthocyanin concentration diagnostic values Normalized results of glucose-nitrogen ratio diagnostic values Normalized deviation from nutrient regulation sequence By weighting factor , , (The sum of the three is 1; initial ranges are suggested as 0.3–0.5, 0.2–0.4, and 0.1–0.3 respectively.) The weighted calculation is as follows: .when Less than When the recommended range is 0.20–0.35, it indicates severe nutrient deficiency. Increase the fertilizer concentration to the higher range and raise the fertilization frequency to the highest level. fall into ( When the recommended range is 0.45–0.60, the nutrient level is considered slightly low; adjust the fertilizer concentration to a medium level and maintain the standard application frequency. fall into ( When the recommended range is 0.70–0.85, the nutrient content is considered suitable, and only minor adjustments are needed to the application rate; when… Not less than If nutrient excess is detected, the fertilizer concentration is reduced to a low level and the fertilization interval is extended to form a closed-loop nutrient regulation.

[0058] It should be noted that the key technologies of this invention include: the shared latent space orthogonal decoupling operator, by applying orthogonal projection constraints to the three heterogeneous features, eliminates linear coupling in the latent space, enabling cross-modal alignment to be completed at the input end with hard constraints, fundamentally avoiding interference caused by dimensional mismatch and dimensional differences to subsequent inference layers; the dynamic hierarchical early exit mechanism uses the state entropy judgment method to evaluate the concentration of feature distribution in each layer in real time, dynamically allocating inference depth according to the intrinsic complexity of the samples, so that limited memory resources and neuron activation calculations are concentrated on pathological samples that truly require deep feature extraction; and the particle swarm optimization dynamic decay sliding window adaptive algorithm uses swarm intelligence to globally search for the optimal temporal decay parameters in the two-dimensional parameter space, so that the temporal attention weights are dynamically adjusted with the physiological rhythm of the orchid, avoiding overfitting or underutilization of historical information caused by fixed window parameters. The synergistic effect of the three is reflected in the following aspects: orthogonal alignment ensures the quality of fused features, the early exit mechanism realizes the dynamic allocation of inference efficiency on this basis, and particle swarm optimization further ensures the model's adaptability to different physiological stages in the temporal dimension. Together, the three constitute a closed loop of end-to-end adaptive optimization from feature input, inference depth to temporal parameters, enabling the overall system to achieve a dynamic optimal balance between nutrient diagnosis accuracy and computational efficiency.

[0059] It should be noted that under the dynamic physiological rhythm from Phalaenopsis flower bud differentiation to secondary flowering, the spectral reflectance characteristics of different physiological stages exhibit significant non-stationarity in the temporal dimension. When the system needs to model continuous time frames, the contribution of nutrient diagnosis information in historical frames to the current frame decreases rapidly with increasing time distance. However, the rate of decay itself differs fundamentally between the early stage of flower bud differentiation, the rapid development stage, and the stage near flowering. The spectral characteristics near flowering are highly dependent on recent historical frames but contribute almost nothing to distant historical frames, while the opposite is true in the early stage of flower bud differentiation. The reason for the above technical problem is that the cell metabolic rate of Phalaenopsis varies at different physiological stages, causing a phased jump in the temporal autocorrelation length of the leaf multispectral reflectance characteristics. The optimal decay coefficient of the temporal attention weight is not a fixed constant but dynamically drifts within a continuous interval with the physiological rhythm. A fixed decay coefficient cannot track this drift process, resulting in a systematic bias in the model's weighting of historical information at certain physiological stages. The common solution to the aforementioned technical problems is to manually fine-tune the decay coefficient on the validation set after model training using grid search or random search, and then deploy the model after determining a fixed value that performs optimally on average across multiple physiological stages. However, this method essentially involves a static trade-off between accuracy rates across multiple physiological stages. The diagnostic accuracy at a specific stage will inevitably be lower than the accuracy corresponding to the optimal parameters at that stage. Furthermore, due to individual differences in the physiological rhythms of different batches and varieties of Phalaenopsis orchids, the generalization ability of the manually tuned fixed value for new batches is limited. In addition, grid search has a computational cost that increases exponentially with dimensionality when the parameter space is large, making online dynamic updates impossible. This invention constructs a two-dimensional continuous search space using the temporal attention decay coefficient and the window width. It employs a particle swarm optimization dynamic decay sliding window adaptive algorithm, driven by inertia weights, individual learning coefficients, and social learning coefficients to collaboratively optimize within the search space. The fitness function simultaneously considers forward diagnostic accuracy and inference computation time. This avoids the problem of single gradient optimization methods easily getting trapped in local optima on non-convex fitness surfaces. Furthermore, through an online update mechanism that re-optimizes every 100-500 inference batches, the temporal decay parameter dynamically adjusts to follow the actual growth rhythm of the Phalaenopsis orchid. In the early stages of flower bud differentiation, it automatically searches for a larger window width and a slower decay rate to fully utilize historical information from the past. Near flowering, it automatically converges to a smaller window width and a faster decay rate to focus on recent spectral features. Thus, it maintains diagnostic accuracy close to the optimal parameters for each physiological stage, fundamentally solving the problem of systematic diagnostic bias caused by the inability of a fixed decay coefficient to track physiological rhythm drift.

[0060] Specifically, the principle of this invention is as follows: The core logic that enables this invention to solve the above-mentioned technical problems lies in the linkage of the following internal mechanisms. First, the multispectral image feature vector, soil electrical conductivity scalar, and leaf surface temperature scalar are each used to construct independent linear projection matrices, and orthogonality constraints are applied to each projection matrix to make the three features mutually orthogonal in the low-dimensional latent space. This operation mathematically guarantees that the inner product between the projection components of the three modes is zero, that is, it eliminates the linear coupling components, so that the subsequent fused feature vector has clear modal discriminability in the latent space. The model does not need to consume additional layer capacity to separate the aliased heterogeneous information, thus aligning the cross-modal features at the input end to complete the hard constraint, providing a reliable unified input for the subsequent inference layer. Secondly, the dynamic hierarchical early exit mechanism is based on the state entropy determination method. At the end of each inference module, the normalized Shannon information entropy of the current layer's output feature vector is calculated. A low entropy value means that the feature distribution is concentrated and the diagnostic confidence is high. At this time, a jump instruction is triggered so that the signal bypasses the deeper inference module and directly outputs the nutrient diagnosis vector. A high entropy value means that the features have not been fully separated and need to continue to be transmitted to deeper layers. The rationale for this mechanism is that the multispectral reflectance features of Phalaenopsis orchids in the normal nutrient range are often concentrated, and low-level linear extraction can reach the confidence threshold. However, the spectral feature distribution under pathological or extreme nutrient levels is diffuse, and deep nonlinear gated convolution is required to complete fine separation. Therefore, allocating the calculation depth as needed conforms to the distribution law of the sample's inherent complexity. Furthermore, the particle swarm optimization algorithm drives collaborative particle search in a two-dimensional continuous search space through inertia weights, individual learning coefficients, and social learning coefficients. It uses a weighted average of forward diagnostic accuracy and inference computation time as the fitness function, avoiding the problem of single gradient optimization methods easily getting trapped in local optima on non-convex fitness surfaces. The optimal decay coefficient found is applied to the temporal attention weight matrix, causing the attention weight of historical frames to decrease exponentially with increasing time distance, thus preserving recent effective temporal information while suppressing interference from irrelevant historical frames from the distant past. Finally, the spatiotemporal high-order tensor transmission mechanism constructs historical nutrient diagnostic vectors as high-order tensors. Through Tucker decomposition, it extracts temporal and spatial modality factor matrices, and explicitly models the nonlinear evolution trajectory of nutrients in the spatial diffusion and temporal accumulation dimensions using tensor convolution. This gives the nutrient regulation sequence prediction an interpretable tensor component correspondence, improving the robustness of temporal prediction. The four mechanisms mentioned above form a logically connected closed loop: orthogonal feature alignment provides high-quality input to the inference layer, the early exit mechanism allocates inference depth as needed, particle swarm optimization ensures the global optimality of temporal parameters, and the tensor transmission mechanism outputs a nutrient regulation sequence with spatiotemporal interpretability. The four mechanisms work together to ensure the synchronous realization of nutrient diagnosis for Phalaenopsis orchids' secondary flowering in three dimensions: low latency, high accuracy, and adaptive computation depth. From a principle perspective, this closes the two major gaps in existing technologies: heterogeneous feature alignment and adaptive inference depth.

[0061] 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.

[0062] The specific implementation method of step S01 is as follows.

[0063] Multispectral images of Phalaenopsis flower buds during differentiation were acquired in real time, covering a wavelength range of 400–1000 nm and a frequency range of 1–10 Hz. The raw multispectral data were then constructed into a third-order tensor. ,in The height is the number of pixels in the image. The image width in pixels. This represents the original number of bands. Perform Tucker decomposition on each mode of the tensor:

[0064] ;

[0065] In the formula, For the kernel tensor, , , These are the low-rank factor matrices for the three modes, , , Principal component dimensions retained for each mode Indicates the tensor along the first Modal product operation, The principal component dimension of each mode is determined by the energy percentage threshold of the corresponding mode singular value decomposition. It is determined that the empirical value is 90%–99%, meaning that the proportion of the cumulative sum of squares of singular values ​​retained exceeds [a certain percentage]. Principal components. Band dimensions after dimensionality reduction. The data is compressed to 5-20 dimensions. After dimensionality reduction, the data is stored in a streaming double-buffered structure, with the buffer... With buffer Physically independent, the acquisition thread continuously writes to one buffer while the processing thread reads from the other. The two buffers switch roles through a lock-free pointer atomic swap mechanism to complete band alignment and streaming caching.

[0066] The specific implementation method of step S02 is as follows.

[0067] Apply microscopic state entropy increase dynamic constraints to the multispectral data in the streaming buffer. Define the first... The probability distribution is obtained after normalizing the multi-band reflectance sequence of frames. ,in For the first Normalized probability value of band reflectivity and Calculate the Shannon entropy for the current frame:

[0068] ;

[0069] In the formula, For the first The entropy of the microscopic metabolic state of a frame, measured in nanots (nat). The rate of increase of entropy between adjacent frames is defined as... The constraint interval is The initial empirical range is 0.01–0.15 NAT per frame, where This is the lower bound of the entropy growth rate. The upper bound of the entropy growth rate was determined by a combination of spectral data collection from multiple batches of Phalaenopsis orchid bud differentiation period leaf samples and cellular metabolic physiological data. When the corresponding band is determined to be affected by environmental noise, a weighted smoothing filter is applied to the corresponding band of that frame:

[0070] ;

[0071] In the formula, For the first Band 1 Reflectance after frame filtering For the first Band 1 Frame original reflectivity, For symmetric weighting coefficients to satisfy , The filter radius is then used. Subsequently, the anthocyanin reflectance feature vector is extracted. Spectral eigenvectors of sugar-nitrogen ratio , and the corresponding reference vector in the standard feature vector library and Perform cosine similarity comparison:

[0072] ;

[0073] ;

[0074] In the formula, For anthocyanin cosine similarity, Cosine similarity of sugar-nitrogen ratio Represents the dot product of two vectors. The Euclidean norm of the vectors is used to determine the similarity values, which range from 1 to 2. Both equations have dimensionless ratios on both sides of the equal sign. or Below the lower threshold An abnormal state marker is triggered when the value reaches an empirical range of 0.80–0.92. A standard feature vector library is pre-established through controlled cultivation experiments, covering Phalaenopsis orchid samples with different nutrient levels. After dimensionality reduction and noise constraint, the feature vector is extracted, and anthocyanin concentration diagnostic values ​​are output. Diagnostic value of glucose-nitrogen ratio Both are normalized dimensionless quantities obtained after corresponding cosine similarity mapping, and their values ​​range from 1 to 2. .

[0075] The specific implementation method of step S03 is as follows.

[0076] Soil electrical conductivity Leaf surface temperature , fuse multispectral feature vectors , and Construct independent linear projection matrices respectively Projection vector , Apply orthogonality constraints to each projection matrix. The three features are in The three orthogonal projections in the latent space are mutually orthogonal, eliminating linear coupling between heterogeneous features. The three orthogonal projection results are then concatenated into a unified low-dimensional orthogonal fusion feature vector.

[0077] $\boldsymbol{z}=[W_{ms}\boldsymbol{f}_{ms}\;\|\;\boldsymbol{w}_{ec}c_{ec}\;\|\;\boldsymbol{w}_T c_T]\in\mathbb{R}^{3d}$;

[0078] In the formula, To unify low-dimensional orthogonal fusion feature vectors, for An identity matrix of order 1. This represents a vector concatenation operation. The empirical range is 8–32, with specific values ​​determined by experiments minimizing cross-modal reconstruction error. Because , and Each segment maps the original quantity into dimensionless latent space features through its own dimensional projection matrix or projection vector. After splicing the three segments... Each component is a dimensionless characteristic value with uniform dimensions.

[0079] The specific implementation method of step S04 is as follows.

[0080] Will The input multi-source physiological and nutritional diagnostic model sequentially passes through a low-level inference module, a mid-level inference module, and a high-level inference module. A non-linear gated jump network is embedded in the high-level inference module to process the feature vector input to the current layer. Perform element-wise gated modulation:

[0081] ;

[0082] In the formula, The output feature vector after gating For learnable gated parameter vectors, For element-wise Sigmoid function, Output the feature vector for the current convolutional layer. The residual path feature vector, For element-wise multiplication, For the first Layer feature dimension This represents the hierarchical number of the reasoning module. Both sides of the equals sign are... A dimensionless eigenvector with consistent dimensions.

[0083] Each inference module calculates its exit confidence score at the end based on the state entropy determination method. Let the current... The probability distribution is obtained after the layer output features are normalized by Softmax. Then the exit judgment entropy is:

[0084] ;

[0085] In the formula, For the first Layer exit entropy, measured in NAT. For the first After the layer output is normalized by Softmax, the th The probability value of the class. The total number of diagnostic categories, For category numbering. When Early exit triggered at any time For the first The layer confidence threshold, empirically ranging from 0.10 to 0.40 nat, was determined by validation set ablation experiments.

[0086] The model incorporates a particle swarm optimization dynamic decay sliding window adaptive algorithm, where the particle position vector is... ,in For the first The temporal attention decay coefficient carried by each particle has an empirical range of 0.85 to 0.99. For the first The width of the temporal sliding window carried by each particle, empirically ranging from 4 to 32 frames. Number the particles. The formulas for updating particle velocity and position are:

[0087] ;

[0088] ;

[0089] In the formula, For the first Particle number The velocity vector of the next iteration The inertial weight has an empirical range of 0.4 to 0.9. For individual learning coefficients, For social learning coefficient, and The empirical range is 1.5 to 2.5. , for Uniformly distributed random numbers within the range For the first The optimal position in particle history. The optimal position for the entire group. The iteration number is given, and the total number of particles empirically ranges from 20 to 80. The fitness function is:

[0090] ;

[0091] In the formula, For fitness value, To validate the diagnostic accuracy of the validation set, the value range is: , The time taken for inference calculations is expressed in seconds (s). , The weighted coefficients are and satisfy the following conditions: ,because and Different dimensions and Each carries a proportionality factor that unifies the dimensions of the two normalized quantities. This is a dimensionless fitness value. The optimal attenuation coefficient. Applying to the temporal attention weight matrix, the historical [number]th The attention weight of a frame to the current frame is: ,in The current frame number. For historical frame number, The weight is the frame interval, and it decreases exponentially as the frame interval increases. This is a dimensionless weight value.

[0092] The spatiotemporal higher-order tensor propagation mechanism constructs a higher-order tensor from the nutrient diagnosis vector sequence within the historical time series window. ,in The length of the historical time series window. This refers to the number of spatial locations of the plant. This is the dimension of the nutrient diagnosis vector. (For...) Perform Tucker decomposition to extract the time mode factor matrix With spatial mode factor matrix ,in The low-rank dimension preserved for time modes. To preserve the low-rank dimension of the spatial modes, the nutrient demand prediction vector sequence for each spatial region within the future control step size is reconstructed through tensor product, which is the nutrient control sequence. The tensor order is empirically 3 to 5.

[0093] The specific implementation method of step S05 is as follows.

[0094] Nutrient status comprehensive assessment function value The calculation formula is:

[0095] ;

[0096] In the formula, This is the value of the comprehensive nutrient status assessment function. The results are normalized for the diagnostic values ​​of anthocyanin concentration. The result is a normalized result of the glucose-nitrogen ratio diagnostic value. The normalized value of the nutrient regulation sequence bias is denoted as , and all three are . Dimensionless quantity , , As weighting coefficients, the three satisfy... The initial empirical ranges were 0.3–0.5, 0.2–0.4, and 0.1–0.3, respectively, determined by analyzing the contribution of each indicator to the success rate of secondary flowering in multiple batches of cultivation experiments. Both sides of the equation are dimensionless quantities with consistent dimensions. The interval to which it belongs performs hierarchical adaptive adjustment: when When nutrient deficiency is determined to be severe, the fertilizer concentration is increased to a high level, and the fertilization frequency is increased to the highest frequency level; when If the nutrient level is slightly low, adjust the fertilizer concentration to the medium range and maintain the standard fertilization frequency; When the nutrient levels are determined to be within the appropriate range, the current fertilizer concentration and frequency are maintained, with only minor adjustments made to the liquid application rate; when If nutrient excess is detected, the fertilizer concentration is reduced to a low level and the fertilization interval is extended. Threshold , , The initial empirical ranges were determined by statistical analysis of flowering rate and physiological indicators under various nutrient levels in multiple batches of Phalaenopsis orchid secondary flowering experiments, which were 0.20–0.35, 0.45–0.60, and 0.70–0.85, respectively.

[0097] The specific implementation method of step S06 is as follows.

[0098] A multispectral image acquisition device, a soil conductivity sensor, a leaf temperature sensor, a multi-source physiological nutrition diagnostic model, and a water and fertilizer actuator are integrated into a physiological nutrition regulation system for Phalaenopsis orchids during their second flowering. The system is equipped with a readable storage medium containing program instructions. When the program instructions are run on a computer, they drive the aforementioned devices to collaboratively execute all the steps of the above method, thereby achieving closed-loop regulation.

[0099] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0100] This embodiment was conducted in a controlled cultivation experimental greenhouse, with the period from Phalaenopsis orchid bud differentiation to the pre-second flowering stage as the monitoring window. The plant variety was Phalaenopsis orchid of the white-flowered large-flowered series, and a total of 120 plants were selected as test subjects. Each plant was under normal cultivation and management conditions, and the water and fertilizer application was managed by the control system of this method throughout the process.

[0101] During system initialization, the multispectral image acquisition device scans the blades in real time, covering a wavelength range of 400–1000 nm, with an acquisition frequency set to 5 Hz. The original multispectral data is constructed as a third-order tensor, and the energy percentage threshold of Tucker decomposition is set to 95%. After dimensionality reduction, the data dimension is compressed from the original 76 dimensions to 12 dimensions. The dimensionality-reduced band data is written into a streaming double-buffered structure, and the lock-free pointer atom exchange frequency dynamically matches the 5 Hz acquisition frequency, ensuring no mutual exclusion or waiting between the acquisition thread and the processing thread.

[0102] After dimensionality reduction and caching, the system calculates the Shannon entropy value for each frame's 12-dimensional band reflectance sequence. The upper bound constraint parameter for the entropy growth rate between adjacent frames is set to 0.08 per frame. On the third day of continuous monitoring, the system detected an entropy growth rate of 0.11 in the 550nm band, exceeding the upper bound constraint, which was determined to be ambient light interference. Weighted smoothing filtering was performed on this band, with the filter weights allocated according to the inverse of the reflectance variance. After filtering, anthocyanin reflectance feature values ​​and sugar-nitrogen ratio spectral feature vectors were extracted and compared with a standard feature vector library using cosine similarity. The lower limit threshold for cosine similarity was set to 0.86. The average similarity of this batch of samples was 0.91, with no abnormality markers triggered. Anthocyanin concentration diagnostic values ​​and sugar-nitrogen ratio diagnostic values ​​were output. The temporal trend of the diagnostic values ​​for each frame during the diagnostic process is shown below. Figure 2 As shown.

[0103] The soil conductivity sensor collected an average soil conductivity value of 1.8 for the current batch of soil. The leaf surface temperature sensor collected an average leaf surface temperature of 24.3℃, which, along with the multispectral image feature vector, was fed into a shared latent space orthogonal decoupling operator. The operator projects the three feature vectors into a 12-dimensional low-dimensional latent space. Orthogonality constraints are achieved by performing QR decomposition on the projection matrix. The three orthogonal projection results are then concatenated to form a 36-dimensional unified low-dimensional orthogonal fusion feature vector. Figure 3 As shown, after orthogonal projection, the cross-correlation matrix of the three features in the latent space exhibits an approximately identity matrix structure, indicating that linear coupling has been effectively eliminated and the modal information is on a linearly separable orthogonal basis.

[0104] A unified low-dimensional orthogonal fusion feature vector input multi-source physiological nutrition diagnostic model is proposed. The model's low-level inference module consists of three fully connected layers, the mid-level inference module consists of three nonlinear layers with residual jump connections, and the high-level inference module consists of four deep convolutional layers with nonlinear gated jump networks. The confidence threshold of the confidence classifier at the end of each inference module is set to 0.25. Figure 4 As shown, in the continuous inference of 120 plants, the activation frequency distribution of each inference module indicates that about 73% of the sample frames are triggered to exit prematurely in the low-level inference module, about 19% of the sample frames are triggered to exit in the middle level, and only about 8% of the complex pathological sample frames need to enter the high-level inference module to complete the full-level inference. This allows the memory allocation and neuron activation calculation to be concentrated on the samples that really need deep feature extraction, avoiding the redundant occupation of computing resources by full-level inference.

[0105] The particle swarm optimization algorithm with dynamic decaying sliding window is run online after model deployment. The number of particles is set to 40, the initial inertia weight is 0.72, and both the individual learning coefficient and the social learning coefficient are set to 2.0. The fitness function is a weighted average of forward diagnostic accuracy and inference computation time, with a weight ratio of 7:3. Optimization is performed again every 200 inference batches. Figure 5 As shown, in the early stage of flower bud differentiation, the optimal decay coefficient for algorithm convergence is 0.97, and the window width is 24 frames. However, as the flowering period approaches, the optimal decay coefficient shrinks to 0.88, and the window width is shortened to 8 frames. As the decay coefficient decreases from 0.97 to 0.88, the effective memory length of the temporal attention weights gradually shortens, and the weight of historical long-term frames quickly approaches zero. The model's attention to recent spectral frames then becomes more concentrated, making the diagnostic output curve more sensitive to the current physiological state. This aligns with the physiological law that the cell metabolic rate accelerates and short-term spectral changes are more informative as the flowering period approaches.

[0106] The spatiotemporal high-order tensor transmission mechanism constructs the nutrient diagnosis vectors of the past 16 frames into a fourth-order spatiotemporal high-order tensor. The Tucker decomposition extracts the temporal modality factor matrix (dimension 16×4) and the spatial modality factor matrix (dimension 12×4). The nonlinear spatiotemporal evolution trajectory is reconstructed through tensor product, and the nutrient demand prediction vector sequence of each spatial region within the next three control steps is output, providing a feedforward control basis for the water and fertilizer execution agency.

[0107] In this embodiment, the weighting coefficients of the nutrient status comprehensive evaluation function are set as follows: , , The grading threshold is set to , , The execution status of nutrient regulation parameters at each time point during the test in this embodiment is shown in Table 1:

[0108] Table 1. Nutrient Regulation Parameter Execution Record Table

[0109]

[0110] As shown in Table 1, the system detected on day 1. The value is 0.24, which is lower than If nutrient deficiency is detected, immediately trigger high-concentration and highest-frequency fertilization; as nutrients are gradually replenished, on the 5th day... The value rose to 0.47, entering the slightly low range, and the system automatically adjusted the fertilizer concentration to the medium level; on the 10th day... The value rose to 0.61, the system determined that the nutrients were suitable, and switched to only making minor adjustments to the application rate; Day 15 The value reached 0.79, exceeding... The system determines that there is an excess of nutrients and automatically reduces the fertilizer concentration and extends the fertilization interval; on the 20th day... The value dropped back to 0.58, and the system returned to a state of slight adjustment within the appropriate range. This dynamic adjustment process indicates that, as... The transition of values ​​between different intervals allows the system to adaptively switch control commands for the water and fertilizer actuators, demonstrating the real-time correction capability of the closed-loop control mechanism for nutrient status deviations. This verifies in principle that the comprehensive evaluation function, through multi-index weighted fusion, can compress multi-dimensional physiological diagnostic information into a single comparable value, eliminating the risk of misjudgment caused by the coupling between indicators in traditional single-index threshold determination methods.

[0111] In terms of neural network structure, the data flow of the multi-source physiological nutrition diagnosis model in this embodiment is as follows: a unified low-dimensional orthogonal fusion feature vector (36 dimensions) is input into the orthogonal projection layer and compressed to 16 dimensions. It is then passed sequentially through a low-level inference module (3 fully connected layers with hidden layer dimensions of 64, 32, and 16), a mid-level inference module (3 nonlinear layers with residual jump connections, hidden layer dimensions of 32, 16, and 16), and a high-level inference module (4 deep convolutional layers with 32, 64, 64, and 32 channels). At the end of each layer, the confidence classifier performs Shannon entropy determination. If the threshold is met, the model exits early and outputs a 12-dimensional nutrient demand diagnosis vector. For alternative network structures, one approach is to use a gated recurrent unit network (GRN) to replace the high-level inference module. The temporal frame sequence is directly input into the GRN (hidden layer dimension 64, number of layers 2). The reset and update gates inside the GRN perform soft gating on the historical hidden states. The output sequence is projected onto a fully connected layer as a 12-dimensional nutrient demand diagnostic vector. Its feature extraction target is highly consistent with the nonlinear gated jump network of the high-level inference module of the main scheme, both targeting the non-stationary dynamic modeling of temporal spectral features. Those skilled in the art can directly replace the high-level convolutional layers with the GRN without redesigning the loss function. Another approach is to use a multi-head self-attention mechanism to replace the temporal attention weight matrix. The historical frame sequence is constructed as a sequence tensor and input into the multi-head self-attention mechanism (number of heads 4, key dimension 16). The attention weight matrix is ​​obtained by Softmax normalization of the inner product of the query matrix and the key matrix. The output weighted vector is mapped to a nutrient demand diagnostic vector through a fully connected layer. Its convergence target and particle swarm optimization decay attention mechanism both point to the weighted aggregation of historical frames, which is an equivalent replacement.

[0112] like Figure 6 The diagram shown is a system architecture flowchart of the overall solution of the present invention, which illustrates the complete data flow from multispectral image acquisition, dimensionality reduction caching, noise suppression, feature extraction, heterogeneous fusion, inference diagnosis to closed-loop control.

[0113] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0114] Table 2. Variable Explanation Table (Part 1)

[0115]

[0116] Table 3. Variable Explanation Table (Part Two)

[0117]

[0118] 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. A method for assessing the physiological and nutritional status of Phalaenopsis orchids during secondary flowering, combined with multispectral diagnostics, characterized in that... Includes the following steps: Multispectral images of Phalaenopsis flower bud differentiation period leaves were acquired in real time. Low-rank tensor spatial projection was performed on the original high-dimensional multispectral data to reduce the dimensionality of the data. The reduced band data was then stored in a streaming double buffer structure to complete band alignment and streaming caching. The multispectral data in the convection buffer is constrained by the micro-state entropy increase dynamic equation, the energy boundary is suppressed for spectral noise, the leaf anthocyanin reflectance feature value and sugar-nitrogen ratio spectral feature vector are extracted, and the cosine similarity is compared with the standard feature vector library to output the anthocyanin concentration diagnostic value and sugar-nitrogen ratio diagnostic value. Soil electrical conductivity and leaf temperature environmental parameters are collected. Multispectral image features, soil electrical conductivity and leaf temperature are fed into a shared latent space orthogonal decoupling operator to perform heterogeneous feature alignment and fusion, and output a unified low-dimensional orthogonal fused feature vector. The unified low-dimensional orthogonal fusion feature vector is input into the multi-source physiological nutrition diagnosis model. Inference based on the dynamic hierarchical early exit mechanism is executed to output the nutrient demand diagnosis vector of Phalaenopsis orchid in the second flowering stage. At the same time, the spatiotemporal high-order tensor transmission mechanism is executed to model the spatial diffusion and temporal accumulation trajectory of nutrients and output the nutrient regulation sequence. Calculate the comprehensive nutrient status assessment function value, and based on the interval to which the comprehensive assessment function value belongs, perform graded adaptive adjustment of the fertilizer concentration, fertilizer frequency and liquid volume parameters of the water and fertilizer irrigation actuator to complete closed-loop nutrient regulation.

2. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 1, characterized in that, The low-rank tensor spatial projection specifically involves constructing the original multispectral data into a third- or fourth-order tensor, performing a low-rank approximation on each mode of the tensor using Tucker decomposition, retaining principal component components whose energy percentage exceeds the energy percentage threshold, discarding redundant bands, and reducing the data dimension from the original number of bands to the number of low-dimensional bands.

3. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 2, characterized in that, The streaming dual-buffer structure specifically sets up two physically independent memory buffers. The acquisition thread continuously writes the latest dimensionality reduction band data to one of the buffers, while the processing thread synchronously reads data from the other buffer to perform subsequent operations. The two buffers switch roles through a lock-free pointer atomic swap mechanism, and the switching frequency is dynamically matched with the acquisition frequency.

4. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 3, characterized in that, The microscopic state entropy increase dynamic equation constraint specifically involves abstracting the metabolic state of Phalaenopsis leaf cells into a thermodynamic open system, calculating the Shannon entropy value of the probability distribution corresponding to the multi-band reflectance sequence of the current acquisition frame as the microscopic metabolic state entropy of the current frame, establishing upper and lower bound constraints on the entropy increase rate between adjacent frames, and determining environmental noise interference when the entropy increase rate caused by the reflectance change of a certain band exceeds the upper bound constraint, and performing weighted smoothing filtering on the corresponding band.

5. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 4, characterized in that, The cosine similarity comparison specifically involves performing cosine similarity calculations on the extracted anthocyanin reflectance feature value and sugar-nitrogen ratio spectral feature vector with pre-stored feature vectors in the standard feature vector library. When the similarity value is lower than the lower similarity threshold, an abnormal state marker is triggered.

6. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 5, characterized in that, The shared latent space orthogonal decoupling operator specifically constructs independent linear projection matrices for the multispectral image feature vector, soil electrical conductivity scalar, and leaf surface temperature scalar, applies orthogonality constraints to each projection matrix, making the three features mutually orthogonal in the low-dimensional latent space after projection, and then splices the three orthogonal projection results into a unified low-dimensional orthogonal fused feature vector.

7. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 6, characterized in that, The structure of the multi-source physiological nutrition diagnostic model is as follows: the input end receives a unified low-dimensional orthogonal fusion feature vector, which is mapped to a low-dimensional shared latent space through an orthogonal projection matrix. The feature flow then enters the low-level inference module, the middle-level inference module, and the high-level inference module in sequence. Each inference module has an ultra-lightweight confidence classifier based on the state entropy determination method embedded at its end. When the confidence meets the conditions, it is triggered to exit early.

8. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 7, characterized in that, The state entropy determination method specifically calculates the Shannon information entropy value of the probability distribution of the current layer's output feature vector after Softmax normalization, compares the entropy value with a confidence threshold, and triggers early exit when the entropy value is lower than the confidence threshold.

9. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 8, characterized in that, The high-level inference module consists of deep convolutional layers containing a nonlinear gated jump network. The nonlinear gated jump network independently calculates the gate weight for each dimension of the input feature vector. The gate weight determines whether the feature energy of that dimension flows to the current layer or jumps to the residual path.

10. The method for physiological nutrition of Phalaenopsis orchids undergoing secondary flowering combined with multispectral diagnosis according to claim 9, characterized in that, The multi-source physiological nutrition diagnosis model embeds a particle swarm optimization dynamic decay sliding window adaptive algorithm. Specifically, when initializing the particle swarm, each particle carries two positional dimensions: decay coefficient and window width. The weighted value of forward diagnostic accuracy and inference computation time is used as the fitness function. The output optimal decay coefficient is applied to the temporal attention weight matrix to perform adaptive exponential decay on historical multispectral frame information.

Images