Intelligent control method for cyclic utilization of washing water

Abstract

The invention provides an intelligent control method for cyclic utilization of washing water, which comprises the following steps: collecting multi-mode water quality data such as ultraviolet absorption spectrum, optical scattering intensity and electrochemical signals, combining context information such as clothes material, washing temperature and laundry detergent type, performing normalization, de-noising and temperature compensation pre-processing, and calculating the water quality of the washing water; inputting into a multi-scale convolution auto-encoder to realize decoupling separation of pollution features and background interference features; a gating fusion mechanism is utilized, key features are dynamically screened, a reusability discrimination vector is generated, a water quality score and a waterway control signal are output through a lightweight student model, intelligent switching of a water circulation path is realized, the water quality evaluation precision is improved, the method has interpretability and visual analysis capability on a discrimination process, and the method is suitable for popularization and application. The method is suitable for efficient utilization of water resources under diversified washing situations.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water quality evaluation and recycling control of intelligent washing equipment, and particularly relates to a laundry water recycling intelligent control method. BACKGROUND

[0002] At present, with the popularization of energy saving and intelligent household appliances, higher requirements are put forward for the recycling of laundry water in the field of household washing machines. In order to improve the utilization rate of water resources and reduce the overall water consumption, modern intelligent washing equipment generally integrates a water quality analysis module to realize the control decision of "recycling" and "discharge" by judging the quality of the backflow water.

[0003] However, under diversified washing needs and complex home environments, the above-mentioned prior art exposes the following outstanding problems: (1) The correlation processing of multiple index data is insufficient, and there is a typical nonlinear coupling relationship between water quality parameters such as detergent residue, turbidity and conductivity, the importance of which dynamically changes due to factors such as clothing material, washing temperature and washing agent type. Traditional methods mostly use static weighting or simple multivariate linear regression, which is difficult to establish an accurate fusion model for the coupling relationship in the actual washing scene, resulting in a large error in water quality evaluation results and insufficient discrimination accuracy; (2) The existing attention fusion mechanism is limited, and some machine learning schemes use global attention mechanisms to weight and fuse multi-dimensional water quality features. Although the dynamic feature allocation capability is improved, such mechanisms are sensitive to input noise, especially under the environmental interference brought by low-cost household sensors, the model fusion is easily misled by irrelevant features, and the calculation resource consumption is large, which is not suitable for resource-constrained embedded terminal deployment; (3) The generalization ability and scene adaptability are poor, and the existing methods mostly statically set water quality thresholds, weights or model structures, lacking sufficient use of environmental context information such as "clothing category", "washing agent type" and "water temperature change", and cannot adapt to diversified actual washing situations, resulting in large fluctuations in judgment accuracy in new environments, unfamiliar clothes or new cleaning agent placement, causing difficulty in achieving both water saving effect and washing quality. SUMMARY

[0004] The present application provides a laundry water recycling intelligent control method to solve the above-mentioned technical problems.

[0005] The technical solution of the present application is as follows: a laundry water recycling intelligent control method, comprising: S1: Collecting multi-source water quality sensing data in the laundry water backflow pipeline, the data including ultraviolet absorption spectrum, optical scattering intensity, electrochemical response signal, and simultaneously recording clothing material category, washing temperature and washing agent type code as context information; S2: Normalize and denoise the collected raw sensor data to eliminate the interference of sensor noise and environmental temperature drift on water quality feature expression, and generate a standardized water quality feature vector; S3: Concatenate the standardized water quality feature vector with the context information embedding vector, and input them into a multi-scale convolutional autoencoder network, extract local and global features of multi-modal data through grouped convolution structure, and introduce orthogonal constraint terms in the hidden space to realize decoupling and separation of pollution features and background interference features; S4: Input the decoupled key pollution feature vector and the context embedding vector into the context-aware gated fusion unit, dynamically select the most discriminative water quality feature combination through the gated recurrent mechanism, and generate a reusable degree discriminative feature vector; S5: Input the discriminative feature vector into the lightweight student model, output the washing water reusable degree score, and generate a waterway control signal combined with a preset threshold, which is used to trigger water circulation path switching or drainage operation; S6: Based on the soft label and intermediate layer activation map of the teacher model on the large-scale synthetic washing data set, perform knowledge distillation training on the student model to reduce the model parameter quantity and maintain the discriminative accuracy; S7: Use the gradient weighted class activation mapping technology to visualize the analysis of each reusable degree score result, and output the contribution weight distribution diagram of each water quality dimension to the current judgment result; S8: When detecting the change of the clothes material category or the washing agent type, dynamically adjust the input content of the context embedding vector, and automatically adapt the new feature combination strategy through the gated fusion mechanism, so as to improve the generalization ability of the model in diversified washing scenes.

[0006] The washing water recycling intelligent control method provided by the application has the following beneficial effects: (1) The application avoids the problems of static weight focusing and attention generalization deficiency through a novel multi-scale feature decoupling and gated dynamic fusion mechanism, significantly reduces the error range of water quality reusable degree judgment compared with the traditional weighted method, significantly improves the accuracy, and greatly improves the scientificity and effectiveness of water saving control; (2) The application uses a context-aware gating mechanism to realize dynamic feature selection, which can adaptively adjust the feature fusion strategy when facing different clothes, detergents and working condition changes, and is significantly better than the existing models relying on fixed rules or single attention mechanism, covering common household, commercial and multi-type washing scenes, and having strong compatibility and expansibility; (3) The application dynamically outputs waterway switching or drainage control instructions based on the reusable degree score, constructs a washing water automatic circulation control closed loop, greatly improves the automation and unattended level, and expands the intelligent upgrading space of the washing machine. Attached Figure Description Figure 1 A flowchart of an intelligent control method for recycling laundry water according to the present invention; Figure 2 This is a sub-flowchart of an intelligent control method for recycling laundry water according to the present invention; Figure 3 This is another sub-flowchart of an intelligent control method for recycling laundry water according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, the present invention provides an intelligent control method for recycling washing water, specifically including: S1: Collect multi-source water quality sensor data in the washing water return pipe. The data includes ultraviolet absorption spectrum, optical scattering intensity, and electrochemical response signal. Simultaneously record the clothing material type, washing temperature, and detergent type code as context information. S2: Normalize and denoise the collected raw sensor data to eliminate the interference of sensor noise and environmental temperature drift on the expression of water quality characteristics and generate a standardized water quality feature vector. S3: After concatenating the standardized water quality feature vector with the context information embedding vector, the vector is input into a multi-scale convolutional autoencoder network. Local and global features of multimodal data are extracted through grouped convolutional structures, and orthogonal constraint terms are introduced in the latent space to achieve decoupling and separation of pollution features and background interference features. S4: Input the decoupled key pollution feature vector and the context embedding vector into the context-aware gating fusion unit, and dynamically select the most discriminative water quality feature combination through the gating loop mechanism to generate a reusability discriminative feature vector. S5: Input the discriminant feature vector into the lightweight student model, output the reusability score of laundry water, and generate a water path control signal in combination with a preset threshold to trigger water circulation path switching or drainage operation. S6: Based on the soft labels and intermediate layer activation maps of the teacher model on a large-scale synthetic laundry dataset, knowledge distillation training is performed on the student model to reduce the number of model parameters and maintain discrimination accuracy; S7: Use gradient weighted class activation mapping technology to visualize and analyze the results of each reusability score, and output a weight distribution map of the contribution of each water quality dimension to the current judgment result; S8: When a change in clothing material type or laundry detergent type is detected, the input content of the context embedding vector is dynamically adjusted, and the new feature combination strategy is automatically adapted through a gating fusion mechanism to improve the model's generalization ability in diverse washing scenarios.

[0010] Step S1: Collect multi-source water quality sensor data from the washing water return pipe. This data includes ultraviolet absorption spectrum, optical scattering intensity, and electrochemical response signal, and simultaneously records the clothing material type, washing temperature, and detergent type code as contextual information. Specifically, this includes: S1.1: Absorbance data of laundry water samples in the 200-300nm wavelength range were collected based on an ultraviolet absorption spectroscopy sensor to obtain preliminary characterization information on surfactant residual concentration; At the collection node of the washing water return pipe, the light signal response data in the 200-300nm wavelength band is acquired by the ultraviolet absorption spectroscopy sensor and used as the original absorbance time sequence input. A spectral radiative conversion algorithm (parameters: calibration coefficient matrix, incident light intensity reference value) is used to convert the current water sample's transmitted light flux into wavelength-resolved absorbance values ​​and form a full-band absorbance sequence. Furthermore, the absorbance curve drift is corrected by using a baseline correction algorithm (parameters: baseline fitting window length, spline smoothing order), eliminating the influence of dark current in the optical system and baseline drift, and obtaining the baseline-normalized ultraviolet absorbance spectrum. Furthermore, by using a characteristic wavelength selection method (parameter: predefined surfactant characteristic peak range within the target band of 200-300nm), the characteristic range of the baseline normalized spectrum is extracted, and a spectral subset containing characteristic peak data is generated. Furthermore, a multi-peak fitting algorithm (parameters: number of Gaussian functions, initial peak position, half-width constraint) is used to perform spectral deconvolution processing on the absorbance curve within the characteristic interval, decompose the single-peak parameters of each potential surfactant component, and calculate the peak area value as a concentration characterization factor. The residual concentration of surfactant is calculated using the Beer-Lambert law, as shown in the following formula: in, This refers to the surfactant concentration (mg / L). The absorbance value at the characteristic wavelength. molar absorption coefficient (L·mol) -1 ·cm -1 ), Optical path length (cm); The initial concentration value obtained by this formula serves as a quantitative characteristic of surfactant residue in laundry water, providing input for subsequent water quality reusability analysis. The above algorithm chain transforms the raw ultraviolet spectral signal into concentration characterization data with physicochemical significance, enabling a preliminary quantitative assessment of detergent residue. For example, in a household smart washing machine equipped with a UV absorption spectral sensor with an optical path length of 1.0 cm, absorbance curves in the 200-300 nm wavelength band were collected. After conversion using a calibration coefficient matrix and incident light intensity reference values, 101 sampling points of absorbance data were obtained at 1 nm intervals. A baseline correction algorithm with a spline smoothing order of 3 and a fitting window length of 10 nm was used to eliminate the 0.02 absorbance unit offset caused by the device's dark current. The defined surfactant characteristic peak range (225-235 nm) was extracted, and a double Gaussian peak fitting algorithm (initial peak positions were 228 nm and 232 nm, with a half-width constraint of 2 nm) was applied to decompose the peaks into peak areas of 0.15 and 0.12 absorbance·nm, respectively. The absorbance at the maximum peak position of 228 nm was selected as A=0.30, and the molar absorptivity ε=50000 L·mol⁻¹. -1 ·cm -1 The optical path length l = 1.0 cm, substituting into the formula... The residual surfactant concentration is calculated to be 6.0 × 10⁻⁶. -6 The calculated mass concentration is approximately 2.1 mg / L, which is lower than the preset residual limit of 3 mg / L for reusability. Therefore, the water sample is deemed to have the potential for reuse, and this concentration characteristic is sent to the water quality analysis module for further fusion processing. S1.2: Use an optical scattering sensor to detect the turbidity of the laundry water sample and collect the scattered light intensity signal to obtain quantitative data on the content and distribution of suspended particulate matter; The input is an optical scattering sensor installed in the return pipe of the washing machine. Its output is a time-continuous scattered light intensity electrical signal. The initial conditions include sensor calibration coefficients, signal-to-noise ratio estimates, and stability indicators of light source emission power. A multi-angle scattering light detection method (parameters: forward scattering angle 0°–10°, side scattering angle 80°–100°, back scattering angle 170°–180°) was adopted to achieve parallel acquisition of the scattering characteristics of suspended particles in different directions in laundry water samples; Furthermore, by using a light intensity correction algorithm based on Lambert-Beer's law (parameters: optical path length L=10mm, incident light wavelength λ=860nm), the optical scattering signal is normalized according to the incident light intensity, and the relative scattering intensity data matrix per unit length of optical path is obtained. Furthermore, a frequency domain Butterworth low-pass filtering algorithm (parameters: cutoff frequency f_c=5Hz, filter order n=4) is used to suppress high-frequency noise in the scattered light intensity signal and generate a smooth scattering intensity baseline curve. Furthermore, based on the Mie scattering theory inversion algorithm (parameters: refractive index m=1.33, particle radius estimation range r=0.1μm–50μm), the particle size distribution is calculated from the scattering intensity at multiple angles, and the distribution spectrum data of suspended particle volume fraction and particle size range are generated. By using the optical scattering particle distribution calculation formula, the results of the previous step are converted into quantified turbidity values, thereby achieving high-precision characterization of the suspended particle content and dispersion state of water samples. in, This is the integral value of total turbidity. Scattering angle The light intensity at that location, The incident light intensity is [value].

[0011] For example, in a 12kg capacity drum-type household washing machine, the return pipe is equipped with three sets of photodetectors: forward, lateral, and reverse. The sampling frequency is 100Hz, the light source is an 860nm infrared LED, and the optical path length is 10mm. After correction according to Lambert-Beer's law, the normalized average scattering intensity is 0.35 for the forward channel, 0.42 for the lateral channel, and 0.27 for the reverse channel. A fourth-order Butterworth low-pass filter with a cutoff frequency of 5Hz is applied to process the signal, improving the signal-to-noise ratio to 45dB. Mie scattering inversion calculations show that the volume fraction of suspended particles in the 0.5μm–5μm particle size range is 0.12%, and the volume fraction in the 5μm–50μm range is 0.08%. Substituting these values ​​into the turbidity integral formula, the total turbidity value T is calculated to be 0.021 (unit: m³). -1 Compared with the standard reusable water quality threshold of 0.025, the water quality is determined to meet the reuse conditions and the turbidity quantitative data is output to the multimodal feature splicing stage, which improves the accuracy of reusability judgment. S1.3: The pH value and conductivity signals of the washing water are collected through an integrated electrochemical sensor module to reflect the changes in ion concentration and acid-base balance in the water; S1.4: The clothing fiber type signal is collected by the clothing material recognition sensor and converted into clothing material category code by the encoder to construct part of the context feature input vector; S1.5: Collect the current washing water temperature data through the temperature sensor, and combine it with the preset detergent type recognition module to output the detergent type code to form a complete contextual information embedding vector.

[0012] Step S2: Normalize and denoise the collected raw sensor data to eliminate the interference of sensor noise and environmental temperature drift on the expression of water quality characteristics, and generate a standardized water quality feature vector. Specifically, this includes: S2.1: Perform normalization processing based on the minimum-maximum scaling algorithm on the ultraviolet absorption spectrum, optical scattering intensity and electrochemical response signal collected from the laundry water return pipe to eliminate the difference in dimensions of different sensors and obtain a normalized water quality sensing data vector under a unified dimension. S2.2: The normalized ultraviolet absorption spectrum data is denoised based on the wavelet threshold denoising algorithm to remove high-frequency sensor noise and obtain the denoised ultraviolet absorption spectrum feature sequence. S2.3: Perform mean filtering based on a sliding window on the denoised optical scattering intensity signal to suppress instantaneous optical path interference noise and obtain smoothed optical scattering intensity time series data; S2.4: Based on the temperature compensation model, the pH value and conductivity data in the electrochemical response signal are subjected to environmental temperature drift correction processing to eliminate the influence of water temperature fluctuation on the electrochemical measurement results and obtain temperature-corrected electrochemical sensing data. S2.5: The normalized and denoised ultraviolet absorption spectral feature sequence, the smoothed optical scattering intensity time series data, and the temperature-corrected electrochemical sensing data are spliced ​​together in a multimodal manner to form a standardized water quality feature vector, which is used as the input of a multi-scale convolutional autoencoder network. For the normalized and denoised ultraviolet absorption spectral feature sequence, the smoothed optical scattering intensity time series data, and the temperature-corrected electrochemical sensing data, a multimodal feature stitching algorithm (parameter: stitching order is ultraviolet features, optical features, and electrochemical features) is used to achieve dimensional unification and structural alignment of cross-sensor data. This algorithm performs matrix concatenation operations along the feature dimension to ensure that the time-series sampling points from different data sources correspond one-to-one, thereby forming a joint matrix of time-feature two-dimensional structure. Furthermore, a high-dimensional vectorized representation of the two-dimensional joint matrix is ​​achieved through a vectorization transformation algorithm (parameter: flattened by column, each column represents the combined features of each sensor at the same sampling point), and a multimodal feature vector of fixed length is obtained to be compatible with the input requirements of subsequent convolutional structures. Furthermore, by using a normalization and recalibration algorithm (parameters: Z-score normalization, mean based on the global mean of the training dataset, standard deviation based on the global sample standard deviation), the numerical scale of each modality data after concatenation is made consistent, and a fusion feature representation with good statistical stability is generated. Furthermore, by using the principal component preservation mapping algorithm (parameter: retain the first k principal components, the value of k is determined based on the cumulative contribution rate ≥ 95%), redundant feature dimensions are removed from the concatenated feature vectors while maintaining the original main information, thus reducing the computational load of the subsequent network. Through the above feature splicing and compression processing, the spectral features, scattering intensity features and electrochemical features of the previous step are transformed into a standardized water quality feature vector of a unified dimension, so as to realize the efficient fusion and semantic alignment of multimodal water quality information at the input end. For example, in a laundry water reflux detection task, the UV absorption spectral feature sequence, after noise reduction, is a 200-sampling-point × 5-dimensional band absorbance; the optical scattering intensity time series data, after smoothing, is a 200-sampling-point × 3-dimensional scattering intensity index; and the electrochemical sensing data, after temperature correction, is a 200-sampling-point × 2-dimensional (pH value and conductivity) index. In this task, the three types of data are sequentially concatenated according to their feature dimensions to form a 200 × 10-dimensional joint matrix, and a feature vector of length 2000 is obtained through vectorization. This feature vector is then Z-score normalized, and the mean is taken as... Standard deviation is taken After normalization, the values ​​of each feature dimension all fall within the interval [-2, 2]. Then, principal component preserving mapping is used to perform eigenvalue decomposition on the feature matrix, selecting the first... Each principal component has a cumulative contribution rate. %, to obtain a 50-dimensional compressed vector as a standardized water quality feature vector input to a multi-scale convolutional autoencoder network.

[0013] Step S3: The standardized water quality feature vector and the context information embedding vector are concatenated and then input into a multi-scale convolutional autoencoder network. Local and global features of the multimodal data are extracted through a grouped convolutional structure, and orthogonal constraint terms are introduced into the latent space to decouple and separate pollution features from background interference features. Figure 2 As shown, it specifically includes: S3.1: Perform a concatenation operation on the standardized water quality feature vector and the context information embedding vector to form a unified input feature tensor, which serves as the input to the multi-scale convolutional autoencoder network, ensuring that the model can simultaneously perceive the intrinsic water quality features and external context conditions. The standardized water quality feature vector generated by S2 processing and the context information embedding vector composed of clothing material type, washing temperature and detergent type are used to achieve strict alignment of the two types of inputs in the dimensional space by using a feature dimension unified mapping method (parameters: alignment dimension = 128, zero padding strategy = tail padding fixed constant 0). Furthermore, by using the feature tensor concatenation algorithm (parameter: concatenation axis = feature dimension axis), the dimension-aligned standardized water quality feature vector and the context information embedding vector are concatenated in a positional order to generate a composite feature matrix, so as to ensure that environmental conditions and water quality signals are expressed together in the same input frame; Furthermore, a positional encoding fusion method (parameters: encoding period = 64, encoding function = sine and cosine mixture) is adopted to add positional information encoding to the concatenated composite feature matrix, thereby explicitly preserving the relative positional information of the original features in the time series, so as to enhance the temporal sensitivity of the subsequent convolutional structure; Furthermore, channel-level normalization is performed on the composite feature matrix with positional encoding by using a feature normalization processing algorithm (method: LayerNorm, parameter: normalization dimension = full feature dimension) to obtain a unified input tensor with zero mean and one variance, thereby eliminating the impact of differences in feature value ranges on the stability of network training. By using tensor encapsulation and batch construction, the normalized unified input tensor is transformed into batch input data that can be directly loaded by a multi-scale convolutional autoencoder network, achieving the technical effect of the model simultaneously perceiving the intrinsic characteristics of water quality and external context conditions. For example, in a specific washing cycle, the standardized water quality feature vector has a dimension of 128, and its numerical range has been normalized to [ ] after S2 processing. [1,1], the context information embedding vector has a dimension of 32. A dimension-unified mapping method is used to extend the context vector to 128 dimensions, and padding with constants of 0 at the end is used to fill in the non-original dimension positions. The composite feature matrix combines the two types of vectors on the concatenation axis to form a feature column vector of size 256×1. The position encoding is calculated using sine and cosine functions with a period of 64. The position encoding formula for the i-th dimension is: and ,in For time step position index, As the encoding dimension, the function assigns values ​​alternately based on the parity of index i; the matrix after position encoding is normalized by LayerNorm, and the mean reaches [value missing]. The variance reached Finally, input tensors with a batch size of 32 are generated and fed into the MS-CAE model. In this scenario test, the convolutional autoencoder can improve the discrimination accuracy of the contaminated feature separation task while preserving the correlation between the original features and the context conditions, verifying the technical effect of this sub-step processing chain in improving the effectiveness of feature fusion and the adaptability of network input; S3.2: Based on the grouped convolutional structure in the multi-scale convolutional autoencoder network, the local feature extraction and global feature modeling of the input feature tensor are performed on the multi-modal data to obtain the multi-scale local-global feature expression vector, so as to enhance the model's ability to perceive complex water quality conditions. S3.3: Introduce orthogonal constraint terms in the latent space of the multi-scale convolutional autoencoder network to decouple and separate contamination features and background interference features on the extracted multi-scale local-global feature representation vectors, generating contamination feature vectors and background interference feature vectors with clear physical meaning, so as to reduce feature redundancy and improve the model's discriminative ability. The multi-scale local-global feature representation vector obtained by S3.2 processing is used as the input object. A constrained optimization method based on latent space feature orthogonalization (parameters: orthogonality weight λ=0.15, iteration batch size B=64) is adopted to achieve linear independent separation of contamination feature vector and background interference feature vector in the feature space dimension. Furthermore, by constructing a dual-channel feature mapping operator (parameters: mapping dimension d=128, activation function type ReLU), the multi-scale feature representation vectors are projected onto the contamination feature subspace and the background interference feature subspace respectively, to obtain the preliminary separated feature matrix; Furthermore, Gram matrix operations and the Frobenius norm constraint method are employed (parameter: norm threshold ε=1e). 3) To achieve a numerical evaluation of the correlation between the two types of feature matrices and obtain the inner product results between each channel, the process formula is as follows: in The pollution feature matrix, The background interference feature matrix, Let be the Gram matrix of both. This is the matrix transpose. Furthermore, by adding an orthogonality loss term, the loss term is calculated using the following formula. : in is the Frobenius norm of the Gram matrix; Furthermore, the loss items By incorporating a multi-scale convolutional autoencoder into the overall loss function and employing stochastic gradient descent to update network parameters, a clear physical separation between contamination features and background interference features is achieved. By using the above orthogonal constraint and subspace mapping processing method, the multi-scale local-global feature vectors of the previous step are transformed into two sets of linearly independent and low-redundancy feature vectors, which effectively separates pollution information from background noise, thereby improving the feature discriminativeness and result interpretability of the model in water quality reusability discrimination. For example, in a single laundry cycle detection process, the standardized water quality feature vector has a dimension of 512. After grouped convolution processing, a 256-dimensional multi-scale local-global feature representation vector is obtained. A dual-channel feature mapping operator is used to map this vector into two 128-dimensional subspace representations: the contamination feature matrix P and the background interference feature matrix B. The Gram matrix is ​​calculated as follows: The Frobenius norm of the matrix multiplication result is 0.002, satisfying the orthogonality threshold ε=1e 3. Based on the formula: The orthogonality loss value is 4e. 6. During training iterations, this loss value is fused with the main discriminant loss (weight coefficient 0.15). After 15 iterations, the correlation coefficient between the contamination features and the background interference features is reduced to below 0.01, resulting in the final output contamination feature vector. S3.4: Based on the contamination feature vector and the background interference feature vector, calculate their orthogonality loss value in the latent space to optimize the parameter configuration of the multi-scale convolutional autoencoder network, thereby enhancing the decoupling degree between contamination features and background interference features and improving the robustness and interpretability of the model. S3.5: The decoupled and separated pollution feature vector is used as the output result and then used as the input of the subsequent context-aware gating fusion unit to support the dynamic generation of the feature vector for the reusability of laundry water, thereby achieving high-precision assessment and feature fusion of water quality status. The contaminated feature vectors obtained by decoupling and separating the multi-scale convolutional autoencoder network in the latent space are used as the main processing object and input into the feature cache management unit for high-precision storage configuration to ensure the consistency and traceability of numerical representation in subsequent calculations. A feature normalization recalibration algorithm (parameter range: normalization target mean is set to 0, variance is set to 1) is adopted to achieve consistency of pollution feature vectors of different batches on the statistical scale and generate a numerical sequence with uniform distribution characteristics, so as to avoid the discrimination shift of the subsequent gating fusion mechanism due to the difference in feature distribution. Furthermore, by using a feature importance weighted mapping algorithm (weight initialization based on a multi-scale feature contribution rate matrix), the differentiated emphasis of contamination features in each dimension is achieved, resulting in the optimal weighted contamination feature vector in the latent space discrimination dimension, providing a highly discriminative input signal for context-aware fusion. Furthermore, a feature sparsity constraint optimization method is adopted (the L1 norm constraint coefficient is set to 0.001). Under the premise of ensuring the integrity of the discrimination information, the feature dimensions with high redundancy are set to zero or compressed to reduce the computational redundancy of input features and improve the real-time computation performance of the gated loop unit. By rearranging the feature vector dimension for encoding, the contaminated feature vectors, which have been weighted and sparsified, are transformed into structured feature tensors, ensuring their alignment with the context embedding vectors in the input channel structure and achieving seamless splicing in the input stage. Through the above algorithm chain, the decoupled output of the multi-scale convolutional autoencoder is transformed into a formatted, weighted, sparsified, and channel-aligned pollution feature vector, achieving a water quality assessment feature input effect with high robustness and high discriminative power. For example, in a household smart washing machine application scenario, the ultraviolet absorption spectrum collected by the reflux pipe is decoupled to obtain a 256-dimensional contamination feature vector, which is then stored in 64-bit floating-point format by the feature cache management unit. The normalized recalibration parameter is set to the mean. and variance By performing a linear transformation on each dimension, the mean is obtained to be close to And the variance is The feature distribution within the range. The importance weighting matrix is ​​derived from the sensitivity analysis of features to reusability scores, with weight coefficients ranging from... to The 256-dimensional features are weighted and adjusted using matrix multiplication. In the sparsity constraint optimization, L1 norm constraints are employed. The contribution rate is lower than The features are set to zero, reducing the effective dimension to 180. The dimension rearrangement encoding strategy maps the 180-dimensional contamination features to the input channel consistent with the context vector according to the isomorphic structure of the context embedding vector, so that they can be directly called by the subsequent gating fusion unit.

[0014] Step S4: The decoupled key pollution feature vector and the context embedding vector are input into the context-aware gating fusion unit. The most discriminative water quality feature combination is dynamically selected through a gating loop mechanism to generate a reusability discriminative feature vector. For example... Figure 3 As shown, it specifically includes: S4.1: The decoupled key pollution feature vector is concatenated with the context embedding vector composed of clothing material category, washing temperature and detergent type encoding to form a joint feature input tensor that integrates context information; S4.2: Based on the gated recurrent unit (GRU), sequence modeling is performed on the joint feature input tensor to extract the temporal dependencies of each water quality dimension under different washing scenarios, so as to generate a context-aware feature weight allocation vector; S4.3: The feature weight allocation vector is transformed nonlinearly using a gated activation function to select the subset of water quality features with the greatest discriminative power under the current context, so as to construct a dynamic feature selection mask matrix; For the context-aware feature weight assignment vector output by the gated recurrent unit, a sigmoid-based gated activation function is used (parameter: gate threshold). saturation range This enables nonlinear compression and suppression of different water quality characteristic dimensions within the weighted domain; By performing activation transformations on each element in the weight allocation vector, response values ​​between 0 and 1 are obtained, in order to preserve high response features under the current context and weaken the influence of low response features. Furthermore, a threshold filtering algorithm (parameter: (Selecting a threshold for the feature response) enables hard determination of the feature response value after nonlinear transformation, constructing a preliminary feature selection Boolean vector, and setting thresholds for response values ​​greater than a certain threshold. Feature locations are marked as valid, and the rest are marked as invalid. The hard decision process can use the following formula to calculate the filter mask: in, Assign values ​​to the weights of the i-th feature dimension. Let it be the response value after passing through the gated activation function. Let i be the i-th element of the mask matrix; Furthermore, a response reweighting algorithm based on the mutual information criterion (parameter: (As the minimum mutual information threshold), perform feature mutual information analysis on the initial mask vector, eliminate pre-selected features that are not sufficiently related to the context conditions, and normalize the response values ​​of features with high correlation so that they are proportionally weighted in the output mask matrix; The final dynamic feature selection mask matrix is ​​obtained by response reweighting and mutual information filtering. The high-weight positions correspond to the water quality feature dimensions with the strongest discriminative ability in the current context. They can be directly multiplied element-wise with the original key pollution feature vector to achieve context-adaptive feature filtering. By using the above-mentioned nonlinear compression, hard decision screening and mutual information reweighting combination based on gated activation function, the weight allocation result of the previous step is transformed into a sparse dynamic feature selection mask matrix with context discrimination, so as to achieve accurate screening and redundancy suppression of key water quality features in the process of calculating the reusability of laundry water. For example, in the intelligent water recirculation system of a household drum washing machine, the decoupled key contamination feature vector has a length of 64 dimensions, and the weight allocation vector output by the GRU ranges from [-0.8, 0.9]. A Sigmoid gated activation function is used, with a saturation range of... Setting the weights to [-4, 4] maps the weight assignment values ​​to the response range [0.02, 0.98] and sets the feature response selection threshold. The value is 0.55. After performing hard decision, approximately 27 features are retained. The mutual information between the retained features and the contextual conditions (clothing material, detergent type) is calculated, and features below this value are discarded. The 8-dimensional feature vector with a weight of 0.15 is used, and the remaining 19-dimensional feature vectors are normalized to generate the final mask matrix, where the weight distribution has a mean of 0.68 and a variance of 0.12. This mask matrix is ​​then element-wise multiplied with the original contaminated feature vector to obtain a weighted feature vector. S4.4: Perform element-wise multiplication of the dynamic feature selection mask matrix with the original key contamination feature vector to generate a context-adaptive weighted feature representation vector; The input consists of a dynamic feature selection mask matrix and the original key contamination feature vector, along with the nonlinear screening results from the previous sub-step and the unweighted contamination features. An element-wise multiplication method is used (parameter: the dimension of the mask matrix is ​​exactly the same as the dimension of the contaminated feature vector) to achieve the local weighting effect of the mask matrix on each component of the contaminated feature vector; Furthermore, by employing a sparsity filtering strategy (parameter: the low-weight threshold in the mask matrix is ​​set to 0.1), the low-discrimination feature components are suppressed, and a preliminary sparsity feature weighted vector is obtained. Furthermore, by using a normalization coefficient calculation method (parameter: the coefficient is the inverse of the L2 norm of the weighted vector), the magnitude of the weighted vector under different cyclic scenarios is standardized, and normalized weighted vector data is generated. Furthermore, a principal component preserving projection algorithm is adopted (parameter: the projection matrix is ​​calculated based on the covariance matrix of the original pollution features) to reduce the dimensionality of the weighted feature vector while preserving the main discriminant components, and obtain a dimensionality-reduced weighted representation vector; Through the above element-wise multiplication and subsequent normalization, sparsification and dimensionality reduction processing, the result of the mask matrix in the previous step is transformed into a context-adaptive weighted feature representation, thereby achieving dynamic adaptation of feature representation and enhanced discrimination capability. For example, in a drum-type smart washing machine with a rated capacity of 8kg, the original key pollution feature vector includes four types of features: ultraviolet absorption intensity, conductivity, pH value, and optical scattering intensity. Each type has 256-dimensional time-series data, forming a vector with a total dimension of 1024. The dynamic feature selection mask matrix is ​​output by GRU, and after being filtered by a gated activation function, weight coefficients with values ​​ranging from [0,1] are generated on the same dimension of 1024. In the implementation process, the element-wise multiplication calculation formula can be expressed as: in, For weighted feature representation vectors, This is the original key pollution feature vector. Select a mask matrix for dynamic features, symbol This represents the Hadamard element-wise product. After the operation, components with weights less than 0.1 are set to zero to achieve sparsity. The formula for calculating the normalization coefficient is: in, The features are sparsified, and after normalization, a feature vector with uniform amplitude is obtained. The projection matrix is ​​selected to retain the first 200 principal components, achieving a cumulative variance contribution rate of over 95%, and ultimately outputting a 200-dimensional context-adaptive weighted feature representation vector. S4.5: Perform fully connected mapping and normalization on the weighted feature representation vector to output a reusability discriminant feature vector for subsequent reusability score calculation in the lightweight student model; The input is the context-adaptive weighted feature representation vector generated in step S4.4. The data structure is a floating-point high-dimensional vector containing a set of contaminated feature parameters after dynamic feature selection masking and element-wise weighting. A multi-layer fully connected mapping algorithm (parameters: the number of hidden layer nodes is 128 and 64 respectively, and the activation function is ReLU) is used to achieve non-linear feature combination and dimensionality compression of the input weighted feature representation vector; Furthermore, by using a batch normalization method (parameters: mean calculation batch size is 32, variance stabilization factor is 1e-5), the scale adjustment and mean centering of the feature distribution are achieved, suppressing the internal covariate shift. Furthermore, the L2 norm normalization algorithm (parameter: norm order p=2) is used to achieve the normalization transformation of the eigenvectors in Euclidean space, and the normalized vectors are obtained. ; Furthermore, a linear output layer mapping algorithm (parameter: output dimension is M, M is the length of the reusability discriminant feature vector) is adopted to realize the linear projection from the feature space to the target discriminant space and generate the reusability discriminant feature vector; By using the output layer Softmax normalization process, the linearly mapped M-dimensional vector is converted into a normalized probability distribution of each water quality discrimination dimension in the current context, so as to realize the reusability discrimination feature input that can be directly received by the subsequent evaluation module. Through the above fully connected mapping and normalization processing, the weighted feature representation vector is transformed into a numerically stable and scale-consistent reusability discriminant feature vector, realizing a standardized expression of the state of laundry water and supporting high-precision scoring calculation of the downstream lightweight student model. For example, in a real-world application scenario of a smart home washing machine, the input weighted feature representation vector has a dimension of 256, and the feature components are derived from the fusion features of ultraviolet absorption spectrum, optical scattering intensity, and electrochemical response signal. The fully connected mapping uses a first-layer node count of 128 and a second-layer node count of 64, employing the ReLU function for nonlinear combination. Batch normalization uses a batch size of 32, a variance stabilization factor of 1e-5, and calculates the global mean and variance from the most recent 32 samples. L2 normalization normalizes the batch-normalized output vector. The L2 norm is adjusted to 1, and the normalization calculation formula is: ,in These are the feature vectors after batch normalization. Given its L2 norm length, a unit vector of length 64 is obtained. This vector is then linearly mapped by the output layer to a discrimination space of length 16. Finally, after Softmax normalization, the values ​​of each output component are between 0.05 and 0.12, with a sum of 1, representing the relative contribution weight of each water quality discrimination dimension in the current washing context.

[0015] Step S5: Input the discriminative feature vector into the lightweight student model, output a score for the reusability of laundry water, and generate a water path control signal based on a preset threshold to trigger water circulation path switching or drainage operation. Specifically, this includes: S5.1: The decoupled key pollution feature vector and context embedding vector are concatenated to construct a fusion input vector for student model reasoning, which is used to represent the water quality status and context conditions in the current laundry scenario. S5.2: Input the fused input vector into the lightweight student model optimized based on knowledge distillation technology, and perform nonlinear mapping through a multi-layer fully connected network to extract reusability discriminative features and obtain a reusability score vector; S5.3: Perform normalized activation function processing on the reusability score vector to generate a standardized laundry water reusability score, which is used to quantify whether the current laundry water meets the reuse standard; S5.4: Based on the preset reusability scoring threshold, the standardized score is compared and judged to determine whether the conditions for reuse of laundry water are met, and the corresponding judgment result signal is generated; The input object is the reusability score of laundry water after being processed by the normalized activation function, which is stored in the standardized scoring variable. The scoring range is [0,1], and it has the ability to quantify whether the current water quality meets the reuse standards. A threshold comparison algorithm (parameters: standardized score, preset reusability score threshold τ) is used to achieve accurate comparison between the score value and the threshold. Furthermore, by using a difference calculation method (parameters: standardized score, threshold τ), the degree to which the score deviates from the threshold is quantified, and a difference indicator is obtained. ,in For standardized scoring, The preset threshold; Furthermore, a sign function discrimination algorithm (parameter: difference indicator Δ) is used to logically determine the sign of the difference and generate a binary logic signal. This is used for subsequent result mapping; Furthermore, by using the conditional mapping method (parameter: logic signal J), the conversion from the logic signal to the judgment result signal is achieved. The time mapping is "reuse feasible", in The time mapping is "reuse is not feasible"; By using threshold comparison and logical mapping processing, the standardized score from the previous step is transformed into a judgment result signal that meets the water quality reusability discrimination standard, thereby realizing the logical trigger condition output of the waterway control strategy. For example, in a certain circulating water test, the standardized score output by the lightweight student model was: The preset reusability rating threshold τ is set to After calculating the difference, we obtain... The result of the sign function discrimination is The mapping is "reuse feasible," and the system generates a corresponding judgment result signal and transmits it to the control module, triggering the water path switching valve to execute the loop path holding action. In another scenario, the standardized score is... When, the difference is Sign function discrimination result The value is mapped to "reuse is not feasible". Based on this, a drainage command signal is generated to control the water circuit switching valve to perform drainage operation, effectively preventing the reuse of water that does not meet the standards. S5.5: Generate a water circuit control signal based on the judgment result signal. The water circuit control signal includes a water circulation path switching command or a drainage operation command, which is used to control the water circuit switching valve to perform corresponding actions and realize intelligent recycling control of washing water.

[0016] Step S6: Based on the soft labels and intermediate layer activation maps of the teacher model on a large-scale synthetic laundry dataset, knowledge distillation training is performed on the student model to reduce the number of model parameters while maintaining discrimination accuracy. Specifically, this includes: S6.1: Based on the soft label distribution output by the teacher model on a large-scale synthetic laundry dataset, perform KL divergence minimization optimization on the output probability distribution of the student model to improve the student model's ability to approximate the decision boundary of the teacher model. S6.2: Channel alignment and spatial interpolation are performed on the activation maps output by the teacher model and the student model in the intermediate layer to achieve consistency in feature mapping across network structures and provide a structured alignment basis for subsequent knowledge transfer. S6.3: Based on the aligned intermediate layer activation map, a channel attention mechanism is used to calculate the guiding weight of the teacher model's activation features on the corresponding features of the student model, and a weighted knowledge transfer loss term is generated to enhance the student model's learning ability of key features. S6.4: Linearly weighted fusion of KL divergence loss and weighted knowledge transfer loss is performed to construct a composite distillation loss function, and end-to-end backpropagation training is performed on the student model based on the function to simultaneously optimize the model's discriminative ability and generalization performance. The input conditions include: the soft label probability distribution output by the teacher model and the intermediate layer weighted activation map after channel alignment and spatial interpolation, which serve as dual supervision signals for training the student model; A linear weighting strategy with weight normalization is adopted (parameter: soft label loss weights). ∈[0,1], Feature distillation loss weights This enables the fusion construction of the KL divergence loss term and the weighted knowledge transfer loss term; Furthermore, a composite distillation loss function is constructed through weighted summation. The calculation formula is as follows: in, Let p be the probability distribution p of the teacher model and q be the probability distribution q of the student model. The weighted knowledge transfer loss value between the teacher feature map F and the student feature map G; The batch gradient descent optimization algorithm (parameters: learning rate η, batch size M) is used to implement end-to-end backpropagation training and update of the student model parameter set Θ using the composite loss function; Furthermore, gradient vectors are generated through an automatic differentiation mechanism. And use gradient clipping method (threshold γ) to control the risk of gradient explosion; Furthermore, the adaptive moment estimation of the Adam optimizer (parameter: first-order decay rate) is combined with... With second-order decay rate Simultaneously update the weights and bias parameters to ensure the synergistic optimization effect of the KL divergence term and the characteristic distillation term; By using a composite distillation loss function and an end-to-end training strategy, the results of soft label supervision and feature-level knowledge transfer are transformed into the weight distribution optimization of the student model, thereby achieving a simultaneous improvement in discriminative ability and generalization performance. For example, in a real-world application scenario, the soft label loss weight α = 0.6, the feature distillation loss weight β = 0.4, the learning rate η = 0.001, the batch size M = 64, the gradient clipping threshold γ = 5.0, and the first-order decay rate... Second-order decay rate During the training iteration process, for a certain batch of samples, the calculated KL divergence loss... Weighted knowledge transfer loss The combined distillation loss value is = After calculating the gradient using backpropagation and updating the student model parameters, the mean absolute error of the reuse score on the validation set decreased significantly compared to before training, and the fluctuation of the model's scores under different washing scenarios was significantly reduced, showing that the composite distillation training significantly improved the generalization ability while maintaining accuracy. S6.5: Dynamically adjust the temperature coefficient of the soft label output by the teacher model during training to control the smoothness of the predicted probability distribution during knowledge distillation, thereby improving the robustness and interpretability of the student model on low-confidence samples.

[0017] Step S7: Utilizing gradient-weighted class activation mapping (JEM) technology, visualize and analyze each reusability score result, outputting a weighted distribution map of the contribution of each water quality dimension to the current judgment result. Specifically, this includes: S7.1: Perform gradient backpropagation calculation on the feature map of the last convolutional layer of the lightweight student model during the reusability scoring process to obtain the sensitivity gradient information of each water quality dimension to the final scoring result; S7.2: Based on the acquired gradient information, perform a global average pooling operation on the feature map to generate weight coefficients for each channel, representing the contribution intensity of different water quality feature channels to the current scoring result; S7.3: Perform a weighted summation operation on the weight coefficients of each channel and the corresponding original feature map to generate a class activation heatmap, highlighting the water quality feature areas that play a dominant role in the reusability assessment; S7.4: Perform normalization and color mapping processing on the generated class activation heatmap to generate a visual image that intuitively presents the distribution of the influence of water quality dimensions such as ultraviolet absorption intensity, optical scattering signal, and conductivity on the judgment result; For two-dimensional matrix data generated by class activation heatmap, the minimum-maximum normalization algorithm (parameters: input pixel intensity range [0,255], normalization target range [0,1]) is used to realize the linear mapping of response values ​​of different water quality characteristic channels, so that the contribution weight of each channel can be compared under a unified dimension. Furthermore, by using the histogram equalization method (parameter: number of bins 256), the local contrast of the normalized weight map is enhanced, highlighting the subtle differences in the low-response region, and the enhanced weight distribution matrix is ​​obtained. Furthermore, a pseudo-color mapping algorithm (parameter: Colormap type is Jet, mapping interval is [0,1]) is adopted to convert the weight distribution matrix into an RGB color space representation and generate a corresponding color encoding matrix to intuitively distinguish the contribution strength of different water quality feature dimensions; Furthermore, by color inversion and brightness adjustment processing (parameters: brightness scaling factor 0.9, inversion threshold 0.5), the overall visual contrast is optimized, and high-recognition color image data for user visualization interfaces is generated. The optimized color image is resampled using a bilinear interpolation algorithm (parameter: target resolution 1920×1080) to ensure detail rendering on high-resolution display devices and output the final visualization image file. For example, in a single washing cycle, the original matrix size of the class activation heatmap is 128×128, with a minimum response value of 20 and a maximum response value of 240. When performing min-max normalization, the pixel intensity... Mapped to The normalized matrix elements are distributed within the range [0,1]. After histogram equalization, the distribution density of the low-response region (original value 0.1-0.3) is increased by about 35%, making the weak-contribution features clearly visible in the image. Pseudo-color mapping marks the highest response value region as red (RGB≈[255,0,0]) and the lowest response value region as dark blue (RGB≈[0,0,128]), with intermediate response values ​​transitioning with a continuous color gradient. Color inversion mapping the original red high-response region to cyan improves the color distinction between different features, and brightness scaling controls the overall light output within a 0.9x range to avoid glare in bright areas. Bilinear interpolation maintains neighborhood smoothness during resolution conversion, improving the naturalness of the transition at the edges of the enlarged image. The resulting high-resolution visualization image on the OLED display shows a cyan main patch in the main contribution area of ​​ultraviolet absorption intensity, a green transition layer in the secondary contribution area of ​​optical scattering signal, and a blue background with low conductivity contribution. Users can intuitively see the spatial distribution of the impact of each water quality dimension on the current reusability score, achieving a dual optimization of interpretability and visual clarity. S7.5: Output the visualized image along with the current reusability score to the user interaction module to improve system transparency and user understanding and trust in the water quality judgment logic.

[0018] Step S8: When a change in clothing material type or laundry detergent type is detected, the input content of the context embedding vector is dynamically adjusted, and a new feature combination strategy is automatically adapted through a gating fusion mechanism to improve the model's generalization ability in diverse washing scenarios. Specifically, this includes: S8.1: Real-time monitoring of clothing material category code and laundry detergent type code; based on classification and recognition algorithm, determine whether the current washing context has changed, and generate context change detection signal; S8.2: Based on the context change detection signal, perform an incremental update operation on the original context embedding vector, and use the embedding space mapping algorithm to generate a new context embedding vector that adapts to the current clothing material and laundry detergent type, so as to realize the dynamic expression of context information; S8.3: Input the updated context embedding vector into the context-aware gated fusion unit, and calculate the dynamic weight allocation coefficient between the current context information and key pollution features based on the gated loop mechanism to generate a context-aware feature weighted vector; S8.4: Based on the fusion calculation of the key pollution feature vectors output by the feature weighted vector and the multi-scale convolutional autoencoder, a nonlinear activation function is used to generate a discriminative feature vector that is adapted to the current washing scenario, so as to improve the model's recognition accuracy of the current water quality status. S8.5: Input the discriminative feature vector adapted to the current washing scenario into the lightweight student model, perform reusability score calculation, and combine it with dynamically adjusted thresholds to generate water path control signals adapted to the current context, so as to realize intelligent switching control of the washing water circulation path.

[0019] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0020] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart control method for recycling washing water, characterized in that, Includes the following steps: S1: Collect multi-source water quality sensor data in the washing water return pipe. The data includes ultraviolet absorption spectrum, optical scattering intensity and electrochemical response signal, and simultaneously record clothing material type, washing temperature and detergent type code as context information. S2: Normalize and denoise the multi-source water quality sensing data to generate a standardized water quality feature vector; S3: The standardized water quality feature vector and the context information embedding vector are concatenated and then input into a multi-scale convolutional autoencoder network. Local and global features of multimodal data are extracted through grouped convolutional structures, and orthogonal constraint terms are introduced in the latent space to decouple and separate pollution features from background interference features. S4: The decoupled key pollution feature vector and the context information are embedded into the vector input context-aware gating fusion unit. The most discriminative water quality feature combination is dynamically selected through the gating loop mechanism to generate a reusability discriminative feature vector. S5: Input the reusability discrimination feature vector into the lightweight student model, output the reusability score of laundry water, and generate a water circuit control signal in combination with a preset threshold. S6: Based on the soft labels and intermediate layer activation maps of the teacher model on a large-scale synthetic laundry dataset, perform knowledge distillation training on the lightweight student model; S7: Use gradient-weighted class activation mapping technology to visualize and analyze the results of each reusability score, and output a weight distribution map of the contribution of each water quality dimension to the current judgment result.

2. The intelligent control method for recycling washing water according to claim 1, characterized in that, Following step S7, the following is also included: S8: When a change in clothing material category or laundry detergent type is detected, the input content of the context embedding vector is dynamically adjusted, and the new feature combination strategy is automatically adapted through a gating fusion mechanism.

3. The intelligent control method for recycling washing water according to claim 1, characterized in that, Step S1 specifically includes: The absorbance data of laundry water samples in the 200-300nm wavelength range were collected using an ultraviolet absorption spectroscopy sensor to obtain preliminary characterization information on surfactant residual concentration. Turbidity of laundry water samples is detected by using an optical scattering sensor, and scattered light intensity signals are collected to obtain quantitative data on the content and distribution of suspended particulate matter. The pH value and conductivity signals of the washing water are collected through an integrated electrochemical sensor module. The sensor for identifying clothing material collects signals of clothing fiber type and converts them into clothing material category codes through an encoder, which are then used to construct part of the context feature input vector. The current washing water temperature data is collected by a temperature sensor, and the detergent type code is output by a preset detergent type identification module to form a complete contextual information embedding vector.

4. The intelligent control method for recycling washing water according to claim 3, characterized in that, The method of using an optical scattering sensor to detect turbidity in laundry water samples includes collecting multi-angle scattered light signals, including forward, lateral, and reverse scattering, correcting with Lambert-Beer's law, Butterworth low-pass filtering, and Mie scattering theory inversion, quantifying the turbidity distribution of suspended particulate matter in the water sample, and determining reusability based on the total turbidity integral value.

5. The intelligent control method for recycling washing water according to claim 1, characterized in that, Step S2 specifically includes: The ultraviolet absorption spectrum, optical scattering intensity and electrochemical response signal collected from the laundry water return pipe were normalized based on the minimum-maximum scaling algorithm to obtain a normalized water quality sensing data vector under a unified dimension. The normalized ultraviolet absorption spectrum data is denoised using a wavelet threshold denoising algorithm to obtain the denoised ultraviolet absorption spectrum feature sequence. The normalized optical scattering intensity signal is subjected to mean filtering based on a sliding window to obtain smoothed optical scattering intensity time series data; Based on the temperature compensation model, the pH value and conductivity data in the electrochemical response signal are subjected to environmental temperature drift correction to obtain temperature-corrected electrochemical sensing data. The noise-reduced ultraviolet absorption spectral feature sequence, the smoothed optical scattering intensity time series data, and the temperature-corrected electrochemical sensing data are spliced ​​together in a multimodal manner to form a standardized water quality feature vector.

6. The intelligent control method for recycling washing water according to claim 1, characterized in that, Step S3 specifically includes: A concatenation operation is performed on the standardized water quality feature vector and the context information embedding vector to form a unified input feature tensor; Based on the grouped convolutional structure in the multi-scale convolutional autoencoder network, the unified input feature tensor is subjected to local feature extraction and global feature modeling of multimodal data to obtain a multi-scale local-global feature representation vector. An orthogonal constraint term is introduced into the latent space of the multi-scale convolutional autoencoder network to perform decoupling and separation processing of contamination features and background interference features on the multi-scale local-global feature representation vector, generating contamination feature vector and background interference feature vector. Based on the pollution feature vector and the background interference feature vector, calculate their orthogonality loss value in the latent space.

7. The intelligent control method for recycling washing water according to claim 6, characterized in that, The context information embedding vector consists of a group of codes for clothing material type, washing temperature and detergent type. It is fused with feature dimension unified mapping and position encoding to concatenate standardized water quality features and context embedding vector on the dimension axis. After being normalized by LayerNorm, it is used as input to the convolutional autoencoder.

8. The intelligent control method for recycling washing water according to claim 1, characterized in that, Step S4 specifically includes: The decoupled key pollution feature vectors are concatenated with the context embedding vectors formed by encoding clothing material type, washing temperature, and detergent type to form a joint feature input tensor that integrates context information. Based on the gated recurrent unit, sequence modeling is performed on the joint feature input tensor to extract the temporal dependencies of each water quality dimension under different washing scenarios and generate a context-aware feature weight allocation vector. By using a gated activation function to perform a nonlinear transformation on the feature weight allocation vector, a subset of water quality features with the greatest discriminative power under the current context is selected, and a dynamic feature selection mask matrix is ​​constructed. The dynamic feature selection mask matrix is ​​multiplied element-wise with the original key pollution feature vector to generate a context-adaptive weighted feature representation vector. Perform fully connected mapping and normalization on the weighted feature representation vector to output a reusability discrimination feature vector.

9. The intelligent control method for recycling washing water according to claim 8, characterized in that, In step S4, the gated recurrent unit performs sequence modeling on the concatenated feature tensor, outputs a context-aware feature weight allocation vector, forms a feature response mask through a nonlinear transformation of a Sigmoid-type gated activation function, reweights and removes low-correlation features based on the mutual information criterion, and outputs a dynamic feature selection mask matrix.

10. The intelligent control method for recycling washing water according to claim 1, characterized in that, Step S5 specifically includes: The decoupled key pollution feature vector and the context embedding vector are concatenated to construct a fusion input vector for student model inference; The fused input vector is input into the lightweight student model optimized based on knowledge distillation technology. Nonlinear mapping is performed through a multi-layer fully connected network to extract reusability discrimination features and obtain a reusability score vector. The reusability score vector is processed by a normalized activation function to generate a standardized laundry water reusability score. The standardized reusability score of laundry water is compared and judged based on a preset reusability score threshold, and a corresponding judgment result signal is generated. A water circuit control signal is generated based on the judgment result signal. The water circuit control signal includes a water circulation path switching command or a drainage operation command, which is used to control the water circuit switching valve to perform corresponding actions.

Images