A clothes water content detection method based on UWB spectrum

Abstract

The application discloses a clothes water content state detection method based on UWB spectrum, and particularly relates to the technical field of household appliance control state detection, wherein clothes echo is collected through an ultra-wideband signal, and a multipath scattering time-frequency response spectrum is extracted; a multi-source feature fusion normalization matrix is constructed by fusing spectrum features and drying process parameters; a clothes water content rate estimation value is obtained through a nonlinear regression model; a signed trajectory deviation is calculated in combination with an ideal drying attenuation curve; a background clutter deduction threshold and a filter coefficient update amount are generated by using frequency point sensitivity feedback; UWB spectrum parameter adaptive adjustment and stable convergence detection are realized; clothes echo features are obtained through an ultra-wideband signal; a multi-source feature fusion normalization matrix is constructed in combination with drying process parameters; clothes water content rate estimation is realized; and through frequency point sensitivity feedback and parameter adaptive adjustment, the accuracy, stability and continuity of clothes water content state detection in a dynamic drying process are improved.

Description

Technical Field

[0001] This invention relates to the field of home appliance control status detection technology, and more specifically, to a method for detecting the moisture content of clothing based on the UWB spectrum. Background Technology

[0002] With the development of clothing drying equipment, the drying process is gradually evolving towards dynamic adjustment based on the condition of the clothing. During the drying process, the internal moisture content of clothing continuously changes due to heating, tumbling, and moisture migration, resulting in significant dynamic changes in the drying state of clothing at different stages. Accurately determining the current moisture content of clothing is crucial for identifying the state of the drying process. However, clothing does not maintain a fixed position or shape while moving within the drying drum; instead, it continuously tumbles, unfolds, stacks, and changes position under the drum's rotation. Because of differences in clothing material, thickness, structure, and initial moisture distribution, the moisture change process varies in different areas of the clothing, causing the overall moisture content of the clothing to continuously change throughout the drying process.

[0003] In actual drying processes, it is necessary to acquire information reflecting changes in the dryness of clothing and determine the current drying stage based on this information. Since clothing is in continuous motion, the state information obtained during detection needs to maintain continuity and stability to more accurately reflect the overall moisture content trend of the clothing. The moisture content of clothing is influenced by multiple operational factors, and single state information is insufficient to comprehensively describe the moisture content changes during drying. Therefore, it is necessary to combine multi-dimensional state information related to the drying process to comprehensively judge the moisture content of the clothing, thereby improving the correspondence between the state detection results and the actual drying process. For drying equipment, accurately acquiring the moisture content information of clothing provides a data foundation for subsequent drying process adjustments, enabling the drying equipment to adjust its operation based on the actual drying state of the clothing. Therefore, this invention proposes a clothing moisture content detection method based on the UWB spectrum to solve the above problems. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting the moisture content of clothing based on UWB spectrum includes the following steps: S1, inside the drying equipment drum, an ultra-wideband signal transceiver module transmits ultra-wideband pulses and receives echo signals reflected by the clothes and the inner wall of the drum; multiple candidate frequency points and corresponding non-overlapping candidate frequency bands are preset; in the first detection cycle, the initial background clutter subtraction threshold corresponding to each candidate frequency point and the initial preprocessing filter coefficients corresponding to each candidate frequency band are used; in subsequent detection cycles, the updated corresponding parameters are used for preprocessing to extract the frequency domain attenuation, peak attenuation characteristics and root mean square time delay characteristics of each candidate frequency point to form a multipath scattering time-frequency response spectrum; S2, obtain drying process parameters including cylinder temperature, exhaust humidity and running time, and align them with the multipath scattering time-frequency response spectrum according to the sampling time and normalize them in a dimensionless manner to construct a multi-source feature fusion normalization matrix; S3, input the multi-source feature fusion normalization matrix into the preset nonlinear regression model to obtain the estimated value of clothing moisture content; S4, calculate the signed trajectory deviation between the actual change trajectory of the estimated moisture content of clothing within a continuous time window and the reference ideal drying decay curve; S5, determine the local moisture sensitivity of each candidate frequency point to the estimated moisture content of clothing, form a frequency point prediction contribution sequence based on the local moisture sensitivity and the corresponding frequency domain attenuation change, and generate the background clutter subtraction threshold update amount and the preprocessing filter coefficient update amount corresponding to each candidate frequency point based on the consistency of the frequency point prediction contribution sequence with the change direction of the reference ideal drying attenuation curve and the signed trajectory deviation. S6, feed the update back to S1; if the amplitude of the signed trajectory deviation decreases in the next detection period, retain the current update; otherwise, restore the parameters before the update and reduce the update step size in the next detection period; when the amplitude of the signed trajectory deviation is less than the preset convergence threshold for a consecutive preset number of detection periods, and each update value is less than the corresponding preset stability threshold for a consecutive preset number of detection periods, lock the current background clutter subtraction threshold and preprocessing filter coefficients.

[0005] In a preferred embodiment, in S1, the echo signal is segmented into continuous and non-overlapping time windows. For each time window, the frequency domain attenuation and the frequency domain peak attenuation amplitude of each candidate frequency point are extracted. The root mean square time delay feature is calculated based on the time delay and corresponding amplitude of each multipath component in the echo signal, and the frequency domain peak attenuation amplitude is used as the peak attenuation feature. The frequency domain attenuation, peak attenuation feature and root mean square time delay feature corresponding to multiple consecutive time windows are respectively subjected to median filtering to filter out instantaneous outliers caused by clothing rolling, falling and collision. The correspondence between the frequency domain attenuation, peak attenuation feature and root mean square time delay feature is maintained according to the sampling time of each time window.

[0006] In a preferred embodiment, baseline drift correction is performed before median filtering: the frequency domain peak attenuation amplitudes of multiple consecutive time windows are formed into a frequency domain peak attenuation amplitude sequence along the time axis and empirical mode decomposition is performed to obtain multiple intrinsic mode function components; the intrinsic mode function component with the lowest changing frequency is selected, and the correlation between the intrinsic mode function component and the temperature change sequence inside the cylinder within the same time interval is calculated. When the correlation reaches a preset thermal drift correlation threshold, the intrinsic mode function component is determined as the thermal drift baseline component and is subtracted point by point from the frequency domain peak attenuation amplitude sequence. The subtracted frequency domain peak attenuation amplitude sequence is used as the input for median filtering.

[0007] In a preferred embodiment, the dimensionless normalization in S2 includes: for each feature quantity, subtracting the minimum value of the feature quantity within a preset historical period from the current value, and then dividing by the difference between the maximum and minimum values ​​within the preset historical period, so that the corresponding feature quantity is mapped to the interval between 0 and 1; when the difference between the maximum and minimum values ​​is zero, the normalization result of the corresponding feature quantity is set to a preset neutral value in the first detection cycle, and the normalization result of the corresponding feature quantity in the previous detection cycle is used in subsequent detection cycles; the frequency domain attenuation of multiple candidate frequency points forms a frequency domain attenuation feature vector according to the candidate frequency point order, and together with the peak attenuation feature, the root mean square delay feature, and the drying process parameters, constitutes a multi-source feature fusion normalization matrix.

[0008] In a preferred embodiment, the preset nonlinear regression model is established by the drying equipment before leaving the factory through drying tests on standard test loads of various known materials and multiple initial moisture content levels. The frequency domain attenuation, peak attenuation characteristics, root mean square time delay characteristics, and drying process parameters of each candidate frequency point are collected simultaneously. The training multi-source feature fusion normalization matrix is ​​constructed according to the same time alignment and dimensionless normalization method as S2. The real moisture content obtained by the synchronous weighing method is used as the training label. The multi-layer feedforward neural network is trained through the backpropagation algorithm to obtain the preset nonlinear regression model that maps the training multi-source feature fusion normalization matrix to the estimated moisture content of the clothing.

[0009] In a preferred embodiment, an ideal drying attenuation curve library is pre-established. The ideal drying attenuation curve library includes multiple ideal drying attenuation curves obtained for different standard test loads and different initial moisture content ranges. Each ideal drying attenuation curve is generated by aligning the real moisture content trajectory obtained by synchronous weighing in multiple repeated drying tests under the corresponding test conditions with the time axis and taking the median value at the corresponding time. For each repeated drying test, the average value of the multi-source feature fusion normalization matrix within the initial preset time period is calculated according to each feature dimension to form the initial feature vector of the test. The average value of multiple initial feature vectors under the same test conditions is calculated according to the corresponding feature dimensions to form the initial feature template vector. After the current drying task is started, the current initial feature vector is formed in the same way. The distance between the current initial feature vector and each initial feature template vector is calculated, and the ideal drying decay curve associated with the initial feature template vector with the smallest distance value is determined as the reference ideal drying decay curve.

[0010] In a preferred embodiment, when calculating the signed trajectory deviation, short-term, medium-term, and long-term windows are established, and the difference between the rate of change of the estimated moisture content of clothing and the rate of change of the corresponding segment of the reference ideal drying attenuation curve within each of the three time windows is calculated. The reliability weight is determined according to the fluctuation variance of the frequency domain attenuation characteristic vector within each time window, wherein the larger the fluctuation variance, the smaller the corresponding reliability weight. The differences corresponding to the three time windows are weighted and summed according to the reliability weights, and the deviation direction of the signed trajectory deviation is determined according to the sign of the weighted summation result, and the deviation amplitude of the signed trajectory deviation is determined by its absolute value.

[0011] In a preferred embodiment, in S5, while keeping the other inputs in the multi-source feature fusion normalization matrix unchanged, the normalized frequency domain attenuation of the current candidate frequency point is increased and the preset perturbation amount is decreased respectively. The ratio of the difference between the estimated moisture content of clothing obtained after inputting the preset nonlinear regression model twice to twice the preset perturbation amount is used as the local moisture content sensitivity of the current candidate frequency point. Within a continuous time window, the normalized frequency domain attenuation change value of the current candidate frequency point between each adjacent sampling time is multiplied by the corresponding local water content sensitivity to obtain the frequency point prediction contribution corresponding to each sampling interval, and a frequency point prediction contribution sequence is formed in chronological order. The proportion of the number of sampling intervals in which the frequency point prediction contribution sequence and the reference ideal dry attenuation curve have the same direction of change within the corresponding sampling interval to the total number of sampling intervals is determined as the degree of consistency of change direction. Based on the local water content sensitivity, the degree of consistency of change direction, and the signed trajectory deviation, the background clutter subtraction threshold update amount of the current candidate frequency point and the preprocessing filter coefficient update amount of the corresponding candidate frequency band are determined.

[0012] In a preferred embodiment, in S6, the background clutter subtraction threshold update amount for each candidate frequency point corresponds to a first stable threshold, and the preprocessing filter coefficient update amount for each candidate frequency band corresponds to a second stable threshold. The first and second stable thresholds are determined by pre-calibration experiments and are used to characterize the maximum allowable single detection cycle update amount when the corresponding background clutter subtraction threshold and preprocessing filter coefficient enter a stable state. The preset convergence threshold is determined by a standard load drying experiment and is used to characterize the upper limit of the amplitude of the signed trajectory deviation allowed by the actual change trajectory of the clothing moisture content estimate relative to the reference ideal drying attenuation curve.

[0013] In a preferred embodiment, the method further includes a drying control step: determining the target value of heating power and the target value of drum speed based on the preset moisture content range where the estimated moisture content of the clothes obtained in S3 falls; determining the allowable single-detection cycle variation of the target value of heating power and the target value of drum speed based on the amplitude of the signed trajectory deviation obtained in S4, wherein when the amplitude of the signed trajectory deviation exceeds a preset deviation threshold, the allowable single-detection cycle variation is reduced, and the drying equipment is controlled according to the defined target value of heating power and the target value of drum speed.

[0014] The technical effects and advantages of this invention are as follows: This invention acquires echo signals reflected from clothing and the inner wall of the drying drum using an ultra-wideband signal transceiver module. It then extracts frequency domain attenuation, peak attenuation characteristics, and root-mean-square time delay characteristics from multiple candidate frequency points to form a multipath scattering time-frequency response spectrum. This method allows for continuous acquisition of echo response changes caused by tumbling, posture changes, and moisture content variations during the drying process, thereby obtaining characteristic information reflecting the moisture content changes of the clothing and improving the continuity and stability of moisture content detection in dynamic drying environments.

[0015] This invention aligns the frequency attenuation, peak attenuation characteristics, and root mean square time delay characteristics in the multipath scattering time-frequency response spectrum with drying process parameters including drum temperature, exhaust humidity, and running time, according to the sampling time, and performs dimensionless normalization processing to construct a multi-source feature fusion normalization matrix. Then, it obtains the estimated moisture content of clothing through a preset nonlinear regression model. By fusing feature information from different sources, the determination of clothing moisture content no longer relies on a single feature change, but rather comprehensively utilizes the ultra-wideband spectral response and drying process parameters, improving the matching degree between the estimated moisture content of clothing and the actual drying state.

[0016] This invention calculates the signed trajectory deviation between the estimated moisture content of clothing and the reference ideal drying attenuation curve. It then combines this deviation with the local moisture content sensitivity of each candidate frequency point, the predicted contribution sequence of the frequency point, and the consistency of the change direction to generate background clutter subtraction threshold updates and preprocessing filter coefficient updates, which are fed back to subsequent detection cycles. Simultaneously, by judging the deviation changes in the update results, the update parameters are retained when the detection results tend to stabilize; if the expected effect is not achieved, the parameters are restored and the update process is adjusted; and the current parameters are locked after the convergence condition is met. This ensures stable operation of the ultra-wideband spectrum detection process and improves the reliability of clothing moisture content detection during continuous drying. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1This is a schematic diagram of a method for detecting the moisture content of clothing based on the UWB spectrum, as described in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 The following examples were obtained: Example 1: A method for detecting the moisture content of clothing based on UWB spectrum, comprising the following steps: S1, Inside the drying drum, an ultra-wideband (UWB) signal transceiver module transmits UWB pulses and receives echo signals reflected from the clothing and the inner wall of the drum. The UWB pulses, after reflection from the clothing and the inner wall of the drum, form multipath components with different propagation delays and amplitudes. The UWB signal transceiver module samples the echo signals according to a preset sampling frequency, forming an echo sampling sequence arranged according to the sampling time, and retaining the time position corresponding to each echo sample value. Multiple candidate frequency points and corresponding, non-overlapping candidate frequency bands are preset. Candidate frequency points are frequency positions within the UWB operating frequency range used to extract frequency domain attenuation, and candidate frequency bands are frequency ranges containing the corresponding candidate frequency points and used for preprocessing filtering. Each candidate frequency band corresponds to a set of preprocessing filter coefficients, and each candidate frequency point corresponds to a background clutter subtraction threshold. The preprocessing filter coefficients are applied to the corresponding candidate frequency band, and the background clutter subtraction threshold is applied to the corresponding candidate frequency point.

[0020] The first detection cycle uses the initial background clutter subtraction threshold for each candidate frequency point and the initial preprocessing filter coefficients for each candidate frequency band. Subsequent detection cycles use updated parameters for preprocessing. The initial background clutter subtraction threshold and initial preprocessing filter coefficients are preset and stored before the start of the first detection cycle. Subsequent detection cycles read the background clutter subtraction threshold and preprocessing filter coefficients retained from the feedback of the previous detection cycle. The background clutter subtraction threshold is used to subtract background clutter components from the corresponding candidate frequency point, and the preprocessing filter coefficients are used to filter the echo signal within the corresponding candidate frequency band. The echo signals after background clutter subtraction and preprocessing filtering are arranged according to the original sampling time and serve as a common data source for frequency domain attenuation, peak attenuation characteristics, and root mean square delay characteristics.

[0021] In S1, echo signals are segmented into continuous and non-overlapping time windows. These time windows are arranged consecutively according to the sampling time of the echo signals, with no overlapping sampling intervals between adjacent time windows. The time window length is a preset length, and all time windows in the same drying task use the same time window length, ensuring that the time-domain and frequency-domain characteristics obtained from adjacent time windows have the same time scale. For each time window, the frequency-domain attenuation and the attenuation amplitude of the main peak at each candidate frequency point are extracted. The echo signal within the current time window is converted from the time domain to the frequency domain to obtain the echo frequency-domain amplitude corresponding to each frequency position, and the echo frequency-domain amplitude corresponding to each candidate frequency point is read. The echo frequency-domain amplitude of each candidate frequency point is compared with the corresponding transmit reference frequency-domain amplitude, and the resulting amplitude attenuation is used as the frequency-domain attenuation of the corresponding candidate frequency point.

[0022] In the frequency domain response corresponding to the current time window, the dominant frequency peak is determined, and the echo frequency domain amplitude corresponding to the dominant frequency peak is compared with the corresponding transmit reference frequency domain amplitude. The resulting amplitude attenuation is taken as the attenuation amplitude of the dominant frequency peak, and this attenuation amplitude is used as the peak attenuation characteristic. The frequency domain attenuation and peak attenuation characteristic of each candidate frequency point are established with a sampling time correspondence with the time window in which the corresponding data is generated. Based on the time delay and corresponding amplitude of each multipath component in the echo signal, the root mean square time delay characteristic is calculated. The time delay of each multipath component is the propagation time delay corresponding to the echo component formed by different propagation paths, and the corresponding amplitude of each multipath component is the echo amplitude formed by the corresponding propagation path. The corresponding amplitude of each multipath component within the same time window is squared to obtain the power characterization value of each multipath component. The power characterization value of each multipath component is then divided by the sum of the power characterization values ​​of all multipath components to obtain the time delay weight corresponding to each multipath component.

[0023] The time delay of each multipath component is multiplied by its corresponding time delay weight and summed to obtain the weighted average time delay. Then, the difference between the time delay of each multipath component and the weighted average time delay is calculated. The square of this difference is multiplied by its corresponding time delay weight and summed. The square root of the sum is used as the root mean square (RMS) time delay feature. The obtained RMS time delay feature maintains a corresponding sampling time relationship with the frequency domain attenuation and peak attenuation features within the same time window. Median filtering is applied to the frequency domain attenuation, peak attenuation feature, and RMS time delay features corresponding to multiple consecutive time windows to filter out instantaneous outliers caused by clothing tumbling, falling, and collisions. The frequency domain attenuation values ​​obtained for the same candidate frequency point within consecutive time windows are arranged according to sampling time to form a frequency domain attenuation sequence. Similarly, the peak attenuation feature and RMS time delay feature obtained within consecutive time windows are arranged according to sampling time to form a peak attenuation feature sequence and an RMS time delay feature sequence, respectively.

[0024] Median filtering windows are set on the frequency domain attenuation sequence, peak attenuation feature sequence, and root mean square delay feature sequence, respectively. The continuous feature values ​​covered by the median filtering window are arranged in order of magnitude, and the median is obtained. The median is used as the filtering result for the corresponding sampling position of the median filtering window. The median filtering window moves sequentially according to the sampling time, ensuring that the various features corresponding to the continuous time window undergo median filtering while maintaining the arrangement relationship between the filtering results and the original sampling time. The correspondence between frequency domain attenuation, peak attenuation feature, and root mean square delay feature is maintained according to the sampling time of each time window. The frequency domain attenuation of each candidate frequency point corresponding to the same sampling time is arranged in candidate frequency order and combined with the peak attenuation feature and root mean square delay feature corresponding to the same sampling time to form time-frequency feature data. Then, the time-frequency feature data of continuous sampling times are arranged in chronological order to form a multipath scattering time-frequency response spectrum.

[0025] In the multipath scattering time-frequency response spectrum, the time dimension corresponds to the sampling time of a continuous time window, the candidate frequency point dimension corresponds to multiple candidate frequency points, the frequency domain attenuation represents the frequency domain response change of each candidate frequency point over time, the peak attenuation feature and the root mean square delay feature are kept in correspondence with the frequency domain attenuation at the same sampling time, and the multipath scattering time-frequency response spectrum is composed of time domain information and frequency domain information obtained from multiple consecutive time windows.

[0026] S1 also includes baseline drift correction before median filtering: the frequency domain peak attenuation amplitudes of multiple consecutive time windows are formed into a frequency domain peak attenuation amplitude sequence along the time axis and subjected to empirical mode decomposition to obtain multiple intrinsic mode function (IMF) components. Each IMF component is arranged according to the same sampling time as the frequency domain peak attenuation amplitude sequence, enabling each IMF component to establish a sampling-time correspondence with the cylinder temperature change sequence within the same time interval. The IMF component with the lowest frequency of change is selected, and the correlation between the IMF component and the cylinder temperature change sequence within the same time interval is calculated. The IMF component with the lowest frequency of change is used as the low-frequency baseline change component to be determined, and is not directly used as the thermal drift baseline component.

[0027] When calculating the correlation, the intrinsic mode function component with the lowest frequency of change and the temperature change sequence inside the cylinder are sampled using the same amount of data at the same sampling time. The average values ​​of the two sequences are calculated separately. Each sample value in the two sequences is subtracted from the corresponding average value. The two differences obtained at the same sampling time are multiplied and summed. The result is normalized by the square root of the product of the sum of the squares of the differences between the two sequences. The absolute value of the resulting correlation coefficient is used as the correlation degree. If either sequence does not show any effective change within the corresponding time interval, making normalization calculation impossible, the intrinsic mode function component with the lowest frequency of change is not determined as the thermal drift baseline component.

[0028] A preset thermal drift correlation threshold is used to determine whether the intrinsic mode function component with the lowest changing frequency has a correlation with the temperature change sequence inside the cylinder that satisfies the thermal drift determination condition. When the correlation reaches the preset thermal drift correlation threshold, the intrinsic mode function component with the lowest changing frequency is determined as the thermal drift baseline component; when the correlation does not reach the preset thermal drift correlation threshold, the original frequency domain main peak attenuation amplitude sequence is retained, and thermal drift baseline component subtraction is not performed. The thermal drift baseline component is subtracted point by point from the frequency domain main peak attenuation amplitude sequence, and the subtracted frequency domain main peak attenuation amplitude sequence is used as the input for median filtering. During point-by-point subtraction, the frequency domain main peak attenuation amplitude at each sampling time is subtracted from the thermal drift baseline component value at the same sampling time. The result forms the baseline drift corrected frequency domain main peak attenuation amplitude sequence according to the original sampling time order, and the baseline drift corrected frequency domain main peak attenuation amplitude continues to be used as the peak attenuation feature of the corresponding sampling time. After baseline drift correction, the sampling time and number of data for the peak attenuation feature remain unchanged, and the time correspondence between the frequency domain attenuation and root mean square delay features of each candidate frequency point in the same time window is maintained. Then, median filtering is performed together to form a multipath scattering time-frequency response spectrum.

[0029] S2. Acquire drying process parameters including drum temperature, exhaust humidity, and runtime. Align these parameters with the multipath scattering time-frequency response spectrum at the sampling time and perform dimensionless normalization to construct a multi-source feature fusion normalization matrix. Drum temperature is the temperature inside the drum obtained at the corresponding sampling time; exhaust humidity is the exhaust humidity obtained at the corresponding sampling time; and runtime is the accumulated runtime relative to the start time of the current drying task. All three drying process parameters retain their corresponding sampling times. The frequency domain attenuation, peak attenuation characteristic, and root mean square delay characteristic of each candidate frequency point in the multipath scattering time-frequency response spectrum also retain their corresponding sampling times. Using the sampling times in the multipath scattering time-frequency response spectrum as a unified time identifier, establish a correspondence between drum temperature, exhaust humidity, and runtime at the same sampling time and their corresponding frequency domain attenuation, peak attenuation characteristic, and root mean square delay characteristic. After time alignment, all feature quantities corresponding to the same sampling time form a set of feature data to be normalized, and different sampling times are arranged in chronological order.

[0030] The dimensionless normalization in S2 includes: for each feature quantity, subtracting the minimum value of the corresponding feature quantity within a preset historical period from the current value, and then dividing by the difference between the maximum and minimum values ​​of the corresponding feature quantity within the preset historical period, thus mapping the corresponding feature quantity to the interval between 0 and 1. The preset historical period is a continuous time range up to and including the current sampling time. The values ​​of the same feature quantity obtained within the preset historical period are compared, with the largest value being taken as the maximum and the smallest value as the minimum. The frequency domain attenuation is statistically analyzed for maximum and minimum values ​​for each candidate frequency point, and each candidate frequency point is compared only using its own frequency domain attenuation within the preset historical period. Peak attenuation features, root mean square delay features, cylinder temperature, exhaust humidity, and running time are also statistically analyzed for maximum and minimum values ​​using feature quantities with the same name. Feature quantities with different names do not share maximum and minimum values, thus enabling different physical quantities to be dimensionlessly normalized according to their own range of variation.

[0031] For any feature quantity, the current value at the current sampling moment is subtracted from the corresponding minimum value; the difference represents the position of the current value relative to the lower limit of the corresponding range of change. The corresponding maximum value is subtracted from the corresponding minimum value; the difference represents the range of change of the corresponding feature quantity within a preset historical period. The normalized result is obtained by dividing the previous difference by the next difference. The preset historical period includes the current sampling moment. The current value is not greater than the corresponding maximum value and not less than the corresponding minimum value. When the corresponding maximum value and the corresponding minimum value are not equal, the normalized result is located in the interval between 0 and 1. When the current value is equal to the corresponding minimum value, the normalized result is located in the lower limit of the interval between 0 and 1; when the current value is equal to the corresponding maximum value, the normalized result is located in the upper limit of the interval between 0 and 1; when the current value is between the corresponding minimum value and the corresponding maximum value, the normalized result is located within the interval between 0 and 1. Frequency domain attenuation, peak attenuation characteristics, root mean square time delay characteristics, cylinder temperature, exhaust humidity, and running time have a unified numerical scale after being normalized using the same dimensionless structure. When the difference between the maximum and minimum values ​​is zero, the normalized result of the corresponding feature quantity is set to a preset neutral value in the first detection period, and the normalized result of the corresponding feature quantity in the previous detection period is used in subsequent detection periods. When the difference between the maximum and minimum values ​​is zero, the maximum and minimum values ​​of the corresponding feature quantity within a preset historical period are equal, and no further division operation is performed; the corresponding zero-difference processing is directly initiated. The preset neutral value is a fixed normalized value that is pre-set and located in the interval between 0 and 1. It is only used for the corresponding feature quantity whose maximum and minimum values ​​are zero in the first detection period. When the difference between the maximum and minimum values ​​is zero in subsequent detection periods, the normalized result with the same name and feature position in the previous detection period is read as the current normalized result. When using the normalized result of the corresponding feature quantity in the previous detection period, the feature position of the corresponding feature quantity in the subsequent multi-source feature fusion normalization matrix is ​​not changed.

[0032] The frequency domain attenuation values ​​of multiple candidate frequency points are arranged in order to form a frequency domain attenuation feature vector, which, together with the peak attenuation feature, root mean square delay feature, and drying process parameters, constitutes a multi-source feature fusion normalization matrix. When forming the frequency domain attenuation feature vector, the frequency domain attenuation value of each candidate frequency point at the same sampling time is first normalized dimensionlessly. Then, the normalized frequency domain attenuation values ​​are arranged sequentially according to a preset candidate frequency point order, ensuring that each candidate frequency point always corresponds to a fixed position in the frequency domain attenuation feature vector. The same candidate frequency point arrangement order is used at different sampling times, and the same position in the frequency domain attenuation feature vector always corresponds to the same candidate frequency point. The number of candidate frequency points remains constant in the same drying task, ensuring that the frequency domain attenuation feature vectors formed at different sampling times have the same data length. The frequency domain attenuation feature vector, normalized peak attenuation feature, normalized root mean square delay feature, normalized drum temperature, normalized exhaust humidity, and normalized running time at the same sampling time are arranged sequentially according to a fixed feature order to form a set of multi-source fusion feature data corresponding to the sampling time. The drying process parameters maintain a fixed arrangement of cylinder temperature, exhaust humidity, and running time, and the same sequence of all features is used at different sampling times.

[0033] The multi-source fusion feature data corresponding to consecutive sampling times are arranged row by row according to the chronological order of sampling times to form a multi-source feature fusion normalization matrix. Each row in the multi-source feature fusion normalization matrix corresponds to a sampling time, and each column corresponds to a fixed feature position. The normalized frequency domain attenuation of each candidate frequency point occupies an independent column, while the peak attenuation feature, root mean square delay feature, cylinder temperature, exhaust humidity, and running time occupy corresponding fixed columns, enabling subsequent processing to read feature quantities with consistent names according to fixed column positions. When new sampling times are added consecutively, a set of multi-source fusion feature data corresponding to the new sampling time is added to the multi-source feature fusion normalization matrix according to the same feature arrangement order, while maintaining the temporal relationship between the already formed data rows. This ensures that the multi-source feature fusion normalization matrix simultaneously maintains the fixed arrangement relationship between candidate frequencies, the fixed arrangement relationship between different types of features, and the temporal relationship between consecutive sampling times.

[0034] S3: Input the multi-source feature fusion normalization matrix into the preset nonlinear regression model to obtain the estimated moisture content of the clothing. The multi-source feature fusion normalization matrix serves as the input data set for the preset nonlinear regression model, and is read row by row according to the sampling time. Each row corresponds to a set of normalized feature data formed at a sampling time. The preset nonlinear regression model performs a feedforward calculation on the row of normalized feature data corresponding to the current sampling time, and outputs the estimated moisture content of the clothing corresponding to the current sampling time. The input position in each row of normalized feature data is consistent with the fixed column position of the multi-source feature fusion normalization matrix. The normalized frequency domain attenuation of each candidate frequency point occupies the corresponding input position, and the normalized peak attenuation feature, normalized root mean square time delay feature, normalized cylinder temperature, normalized exhaust humidity, and normalized running time occupy the corresponding input positions. Different sampling times are all entered into the preset nonlinear regression model according to the same input position relationship, so that features with the same name are always input into the same position in the preset nonlinear regression model.

[0035] The pre-defined nonlinear regression model is established by conducting drying tests on various known materials under standard test loads and multiple initial moisture content levels before the drying equipment leaves the factory. For each known material, standard test loads are dried under multiple initial moisture content levels, and feature data for model training and the actual moisture content corresponding to the feature data sampling time are continuously obtained during the drying process. This ensures that the training data includes corresponding data for different materials, different initial moisture contents, and different drying times. The frequency domain attenuation, peak attenuation characteristics, root mean square delay characteristics, and drying process parameters of each candidate frequency point are collected synchronously. At each sampling time, the frequency domain attenuation, peak attenuation characteristics, root mean square delay characteristics, drum temperature, exhaust humidity, and running time of each candidate frequency point are obtained, and a correspondence between the same group of data is established according to the sampling time. A multi-source feature fusion normalization matrix for training is constructed using the same time alignment and dimensionless normalization method as S2. The feature names, candidate frequency order, time alignment, dimensionless normalization processing, and feature column order used during training are all consistent with those used when forming the multi-source feature fusion normalization matrix in the actual detection process. The normalized feature data corresponding to the same sampling time are arranged as a row in the multi-source feature fusion normalization matrix used for training.

[0036] In the training multi-source feature fusion normalization matrix, each row corresponds to a set of model training inputs, and each column corresponds to a fixed input feature position. During training, the training multi-source feature fusion normalization matrix is ​​read in row order, and the normalized feature values ​​in the current row are input into the multi-layer feedforward neural network in fixed column order, ensuring that the training and actual detection phases use the same data structure and input position relationship. The true moisture content obtained by synchronous weighing is used as the training label. The true moisture content obtained by synchronous weighing is associated with the corresponding row in the training multi-source feature fusion normalization matrix according to the sampling time corresponding to the weighing result. The normalized feature data of one row at the same sampling time forms a training sample only with the true moisture content at the same sampling time.

[0037] The true moisture content serves as the target output value corresponding to the current training input of the multilayer feedforward neural network (MLN), and is used to evaluate the deviation between the current output of the MLN and the true moisture content. Training labels at different sampling times are paired with corresponding rows in the multi-source feature fusion normalization matrix used for training, according to temporal correspondence, without overlapping the true moisture content from different sampling times. The MLN is trained using a backpropagation algorithm. The MLN includes an input part that receives training feature data, an intermediate part that performs layer-by-layer feedforward operations, and an output part that outputs the moisture content value of the clothing. Each input position corresponds to a fixed feature column in the multi-source feature fusion normalization matrix used for training, and the output part outputs a moisture content value corresponding to the current training sample. The number of hidden layers, the number of nodes in each hidden layer, and the number of training iterations of the MLN are configured based on the standard test load and feature dimensions.

[0038] During the feedforward operation, the normalized feature values ​​from the current training sample are fed into their corresponding input positions. The intermediate part performs a weighted combination of the outputs from the previous layer according to the internal connection parameters of the multi-layer feedforward neural network, and then performs a nonlinear transformation on the weighted combination result. The result is then passed to the next layer until the output part forms the moisture content output value corresponding to the current training sample. The data transmission direction is from the input part to the intermediate part and then to the output part, and the output of the subsequent layer is not reversed as the current feedforward input during the feedforward operation.

[0039] The moisture content output value obtained from the multi-layer feedforward neural network is compared with the actual moisture content corresponding to the current training sample to obtain the output deviation of the moisture content output value relative to the actual moisture content. The backpropagation algorithm propagates the deviation information from the output part to the previous layer sequentially based on the output deviation, and adjusts the corresponding internal connection parameters according to the deviation information of each layer. This adjusted internal connection parameter changes the moisture content output value when subsequent training samples perform feedforward operations. During training, the processes of training sample input, layer-by-layer feedforward operation, obtaining moisture content output value, comparing with the actual moisture content, backpropagating the output deviation, and adjusting internal connection parameters are repeatedly executed, gradually forming a non-linear mapping relationship between the internal connection parameters and each normalized input feature and the actual moisture content. Standard test loads of different known materials and training samples corresponding to multiple initial moisture content levels are all processed according to the same relationship in adjusting the internal connection parameters. A preset nonlinear regression model is obtained by mapping the multi-source feature fusion normalization matrix used for training to the estimated moisture content of clothing. After training, the network connection relationship of the multi-layer feedforward neural network, the internal connection parameters formed during training, and the fixed input feature position relationship are retained. The trained multi-layer feedforward neural network is used as the preset nonlinear regression model, and the preset nonlinear regression model is no longer retrained using the current detection results during the normal clothing moisture content detection process.

[0040] In actual testing, the multi-source feature fusion normalization matrix formed by S2 is input into a preset nonlinear regression model row by row according to the sampling time, and the normalized feature data of each row in the current row is sent to the same fixed input position as in the training phase. The preset nonlinear regression model performs feedforward operation using the internal connection parameters retained after training, outputs the estimated moisture content of clothing at the current sampling time, and arranges the estimated moisture content of clothing obtained from consecutive sampling times according to the sampling time. The estimated moisture content of clothing corresponds to the current sampling time input into the preset nonlinear regression model, so that subsequent consecutive time windows can read multiple consecutive estimated moisture content of clothing according to the sampling time and form the actual change trajectory of the estimated moisture content of clothing.

[0041] S4 calculates the signed trajectory deviation between the actual change trajectory of the estimated moisture content of clothing within a continuous time window and the reference ideal drying decay curve. The estimated moisture content of clothing obtained from continuous sampling times is arranged in chronological order to form the actual change trajectory of the estimated moisture content, and the sampling time corresponding to each estimated moisture content is retained. An ideal drying decay curve library is pre-established, containing multiple ideal drying decay curves obtained for different standard test loads and different initial moisture content ranges. Different standard test loads and different initial moisture content ranges are combined to form corresponding test conditions. Multiple repeated drying tests are performed under the same corresponding test conditions, and the corresponding ideal drying decay curve is formed using the actual moisture content trajectory obtained from the multiple repeated drying tests. Each ideal drying decay curve is generated by aligning the actual moisture content trajectory obtained from synchronous weighing in the multiple repeated drying tests under the corresponding test conditions along the time axis and taking the median value at the corresponding time. Each repeated drying test uses the start time of the drying task as the starting point. The true moisture content obtained by synchronous weighing is arranged according to the time position relative to the start time of the drying task, so that the true moisture content trajectories formed by different repeated drying tests under the same corresponding test conditions have corresponding time positions.

[0042] After aligning the timeline, the true moisture content obtained from multiple repeated drying tests is read at each corresponding time point. The true moisture content at the same corresponding time point is arranged according to its numerical value, and the median value is obtained. Then, the median values ​​of consecutive corresponding time points are arranged in chronological order to form an ideal drying decay curve. When the number of true moisture contents involved in the median value calculation is odd, the true moisture content in the middle position after sorting is taken. When the number of true moisture contents involved in the median value calculation is even, the average of the two true moisture contents in the middle position after sorting is taken. Each ideal drying decay curve saves the corresponding standard test load, initial moisture content range, and time position corresponding to each median value. Multiple ideal drying decay curves formed under different corresponding test conditions are jointly stored to form an ideal drying decay curve library. The reference ideal drying decay curve is an ideal drying decay curve determined from the ideal drying decay curve library. For each repeated drying test, the mean of the multi-source feature fusion normalization matrix within the initial preset time period is calculated according to each feature dimension to form the initial feature vector of the test. For each feature dimension in the multi-source feature fusion normalization matrix, all normalized values ​​of the corresponding feature dimension within the initial preset time period are read, and the sum of all normalized values ​​is divided by the number of sampling times involved in the calculation to obtain the mean of the corresponding feature dimension.

[0043] The means of each feature dimension are arranged sequentially according to the fixed feature order in the multi-source feature fusion normalization matrix to form the initial feature vector of the experiment. Each vector position in the initial feature vector always corresponds to a normalized feature quantity with the same name. For multiple initial feature vectors under the same test condition, the means are calculated according to their corresponding feature dimensions to form the initial feature template vector. For the same vector position, the corresponding values ​​from multiple initial feature vectors under the same test condition are read, summed, and then divided by the number of initial feature vectors involved in the calculation to obtain the template mean for the corresponding feature dimension.

[0044] The template mean values ​​of each feature dimension are arranged according to the fixed feature order of the initial feature vector to form an initial feature template vector. This initial feature template vector is then correlated with the ideal drying decay curve formed under the same test conditions. After the current drying task starts, the current initial feature vector is formed in the same way. The multi-source feature fusion normalization matrix within the initial preset time period of the current drying task is read, and the mean value for each feature dimension is calculated separately. These mean values ​​are then arranged according to the same fixed feature order as the initial feature vector to form the current initial feature vector.

[0045] Calculate the distance between the current initial feature vector and each initial feature template vector, and determine the ideal drying attenuation curve associated with the initial feature template vector with the smallest distance value as the reference ideal drying attenuation curve. The current initial feature vector and each initial feature template vector use the same vector length and the same feature position relationship. When calculating the distance value, the corresponding values ​​are compared according to the same feature position, and the current initial feature vector and all initial feature template vectors use the same distance value calculation rules. After obtaining the distance values ​​between the current initial feature vector and each initial feature template vector, compare all distance values, read the ideal drying attenuation curve associated with the initial feature template vector with the smallest distance value, and determine the read ideal drying attenuation curve as the reference ideal drying attenuation curve.

[0046] When calculating signed trajectory deviation, short-term, medium-term, and long-term windows are established. Each window ends at the current sampling time and covers consecutive historical sampling times preceding it. The short-term window is shorter than the medium-term window, and vice versa. Fixed window lengths are used for the same drying task. The difference between the estimated rate of change of clothing moisture content within each of the three time windows and the corresponding segment of the ideal drying decay curve is calculated. For any time window, the estimated clothing moisture content at the start and end sampling times is read. The estimated moisture content at the start sampling time is subtracted from the estimated moisture content at the end sampling time, and then divided by the time interval between the end and start sampling times to obtain the estimated rate of change of clothing moisture content within the corresponding time window.

[0047] Two moisture content values ​​corresponding to the start and end sampling times of the same time window are read from the reference ideal drying decay curve. The moisture content value corresponding to the start time position is subtracted from the moisture content value corresponding to the end time position, and then divided by the same time interval to obtain the change rate of the corresponding segment of the reference ideal drying decay curve. The change rate of the estimated moisture content of clothing is calculated using the same time direction and the same change rate calculation method as the change rate of the corresponding segment of the reference ideal drying decay curve. The change rate difference of the short-time window is obtained by subtracting the change rate of the short-time segment corresponding to the reference ideal drying decay curve from the change rate of the estimated moisture content of clothing within the short-time window; the change rate difference of the medium-time window is obtained by subtracting the change rate of the medium-time segment corresponding to the reference ideal drying decay curve from the change rate of the estimated moisture content of clothing within the medium-time window; and the change rate difference of the long-time window is obtained by subtracting the change rate of the long-time segment corresponding to the reference ideal drying decay curve from the change rate of the estimated moisture content of clothing within the long-time window.

[0048] Reliability weights are determined based on the variance of the frequency domain attenuation feature vector within each time window, with larger variance corresponding to smaller reliability weights. For short, medium, and long time windows, all frequency domain attenuation feature vectors arranged according to sampling time are read within the corresponding time window, maintaining the same candidate frequency order for frequency domain attenuation feature vectors at different sampling times. For any time window, the normalized frequency domain attenuation is read for each candidate frequency point at consecutive sampling times within the corresponding time window. The normalized frequency domain attenuation of the same candidate frequency point is then summed and divided by the number of sampling times used in the calculation to obtain the average normalized frequency domain attenuation of the corresponding candidate frequency point within the corresponding time window. The average normalized frequency domain attenuation of all candidate frequency points is arranged in candidate frequency order to form the average frequency domain attenuation feature vector for the corresponding time window.

[0049] For each sampling moment within the corresponding time window, the frequency domain attenuation feature vector is calculated by subtracting the average normalized frequency domain attenuation corresponding to the same candidate frequency point from each normalized frequency domain attenuation in the frequency domain attenuation feature vector. The squared differences are then summed to obtain the discrete value of the frequency domain attenuation feature vector at the corresponding sampling moment. The discrete values ​​of the frequency domain attenuation feature vectors at all sampling moments within the corresponding time window are summed and divided by the number of frequency domain attenuation feature vectors involved in the calculation. The result is used as the variance of the frequency domain attenuation feature vector within the corresponding time window.

[0050] The fluctuation variances of the frequency domain attenuation feature vectors corresponding to the short-term, medium-term, and long-term windows are obtained using the same calculation method described above. These three fluctuation variances are calculated using the same candidate frequency points, the same feature arrangement relationship, and the same degree of dispersion. A larger fluctuation variance in the frequency domain attenuation feature vector indicates a greater overall dispersion of the normalized frequency domain attenuation of each candidate frequency point within the corresponding time window relative to the corresponding average frequency domain attenuation feature vector. The fluctuation variances of the frequency domain attenuation feature vectors corresponding to the short-term, medium-term, and long-term windows are then input into the same preset monotonically decreasing reliability weight determination relationship to obtain the initial reliability weights for the short-term, medium-term, and long-term windows, respectively. The preset monotonically decreasing reliability weight determination relationship satisfies the following condition: when the fluctuation variance of the input frequency domain attenuation feature vector increases, the output initial reliability weight decreases; conversely, when the fluctuation variance of the input frequency domain attenuation feature vector decreases, the output initial reliability weight increases.

[0051] The initial reliability weights for the short-term window, medium-term window, and long-term window are summed to obtain the total initial reliability weight. Then, each of the three initial reliability weights is divided by the total initial reliability weight to obtain the short-term window reliability weight, medium-term window reliability weight, and long-term window reliability weight. After normalization, the sum of the short-term window reliability weight, medium-term window reliability weight, and long-term window reliability weight is 1, maintaining the relationship that the larger the variance of the frequency domain attenuation feature vector, the smaller the corresponding reliability weight.

[0052] The differences corresponding to the three time windows are weighted and summed according to reliability weights. The difference in the rate of change of the short-term window is multiplied by the short-term window reliability weight, the difference in the rate of change of the medium-term window is multiplied by the medium-term window reliability weight, and the difference in the rate of change of the long-term window is multiplied by the long-term window reliability weight. The three products are then added together to obtain the weighted sum. The sign of the signed trajectory deviation is determined according to the sign of the weighted sum, and its absolute value is used to determine the magnitude of the signed trajectory deviation. When the weighted sum is greater than 0, the signed trajectory deviation has one direction; when the weighted sum is less than 0, the signed trajectory deviation has the opposite direction; when the weighted sum is equal to 0, the magnitude of the signed trajectory deviation is 0. The absolute value of the weighted sum is used as the magnitude of the signed trajectory deviation, and the sign of the weighted sum is retained as the direction of the signed trajectory deviation. This ensures that the signed trajectory deviation simultaneously carries the degree and direction of deviation of the actual drying rate of change relative to the reference ideal drying decay curve.

[0053] S5. Determine the local moisture content sensitivity of each candidate frequency point to the estimated moisture content of clothing. Based on the local moisture content sensitivity and the corresponding frequency domain attenuation change, form a frequency point prediction contribution sequence. Then, based on the consistency of the frequency point prediction contribution sequence with the reference ideal drying attenuation curve and the signed trajectory deviation, generate the background clutter subtraction threshold update amount for each candidate frequency point and the preprocessing filter coefficient update amount for each candidate frequency band. When calculating the local moisture content sensitivity, the normalized frequency domain attenuation of the current candidate frequency point is used as the single perturbation object, while keeping the normalized frequency domain attenuation, peak attenuation characteristics, root mean square time delay characteristics, and drying process parameters in the multi-source feature fusion normalization matrix unchanged, except for the current candidate frequency point. Increase and decrease the preset perturbation amount of the normalized frequency domain attenuation of the current candidate frequency point respectively, forming two sets of multi-source feature fusion normalization matrices with different perturbation inputs only for the current candidate frequency point. The two sets of perturbation-induced multi-source feature fusion normalization matrices are input into a preset nonlinear regression model to obtain the estimated moisture content of clothing under the increased perturbation state and the estimated moisture content of clothing under the decreased perturbation state. The difference between the two is divided by twice the preset perturbation amount to obtain the local moisture content sensitivity of the current candidate frequency point.

[0054] Within a continuous time window, the normalized frequency domain attenuation of the current candidate frequency point at adjacent sampling times is read. The normalized frequency domain attenuation of the previous sampling time is subtracted from the normalized frequency domain attenuation of the next sampling time to obtain the normalized frequency domain attenuation change of the current candidate frequency point within the corresponding sampling interval. The normalized frequency domain attenuation change within each sampling interval is multiplied by the local moisture content sensitivity corresponding to the current candidate frequency point to obtain the frequency point prediction contribution for each sampling interval. These contributions are then arranged in the order of sampling times to form the frequency point prediction contribution sequence for the current candidate frequency point. The moisture content change corresponding to each sampling interval is read from the reference ideal drying attenuation curve. The direction of change of moisture content for the corresponding sampling interval of the reference ideal drying attenuation curve is determined based on the direction of moisture content change at adjacent time positions.

[0055] The predicted contribution direction of each sampling interval in the frequency prediction contribution sequence is compared with the change direction of the corresponding sampling interval of the reference ideal dry attenuation curve. When the two directions are consistent, the corresponding sampling interval is included in the consistent direction count. The consistent direction count is then divided by the total number of sampling intervals within the continuous time window to obtain the consistency degree of change direction. Based on the local water content sensitivity, consistency degree of change direction, and signed trajectory deviation of the current candidate frequency, the background clutter subtraction threshold update amount corresponding to the current candidate frequency is generated according to the preset update relationship. The preset update relationship is used to establish the correspondence between local water content sensitivity, consistency degree of change direction, signed trajectory deviation, and background clutter subtraction threshold update amount. The preset update relationship includes input quantity determination, update direction determination, and update amplitude determination.

[0056] When the input is determined, the local moisture content sensitivity of the current candidate frequency point is used as the frequency point response capability parameter, the consistency of the change direction is used as the frequency point prediction reliability parameter, and the signed trajectory deviation is used as the deviation parameter of the current drying process. When the update direction is determined, the background clutter subtraction threshold adjustment direction is determined based on the sign of the signed trajectory deviation, and the correspondence between the frequency domain attenuation change corresponding to the current candidate frequency point and the estimated change in clothing moisture content is determined based on the sign of the local moisture content sensitivity. When the consistency of the change direction between the frequency point prediction contribution sequence and the reference ideal drying attenuation curve reaches a preset consistency condition, the influence of the update result corresponding to the current candidate frequency point on the next detection cycle is increased; when the consistency of the change direction is lower than the preset consistency condition, the influence of the update result corresponding to the current candidate frequency point on the next detection cycle is decreased. When determining the update amplitude, the update amount is determined jointly based on the absolute value of the local moisture content sensitivity, the consistency of the change direction, and the amplitude of the signed trajectory deviation. The larger the absolute value of the local moisture content sensitivity, the more significant the impact of the current candidate frequency point on the estimated moisture content of clothing; the higher the consistency of the change direction, the more consistent the predicted contribution of the current candidate frequency point with the trend of the reference ideal drying attenuation curve; and the larger the amplitude of the signed trajectory deviation, the more significant the deviation of the current detection result from the reference ideal drying attenuation curve.

[0057] The preset update relationship can be implemented using a proportional mapping-based update method, combining the aforementioned input quantities according to a preset ratio to form the update quantity; alternatively, a piecewise mapping-based update method can be used, determining the corresponding update amplitude based on the local water content sensitivity, the consistency of change direction, and the signed trajectory deviation within their respective intervals; or a nonlinear mapping-based update method can be used, causing the update quantity to change nonlinearly with the input quantity. In specific implementation, when the consistency of change direction is high and the signed trajectory deviation amplitude is large, the background clutter subtraction threshold adjustment range is increased, enhancing the retention of effective water content change information at the current candidate frequency point in the next detection cycle; when the consistency of change direction is low, the background clutter subtraction threshold adjustment range is decreased, keeping the adjustment process corresponding to the current candidate frequency point stable.

[0058] S6 feeds back the update obtained in S5 to S1. Specifically, the background clutter subtraction threshold update for each candidate frequency point is applied to the current background clutter subtraction threshold for that candidate frequency point, and the preprocessing filter coefficient update for each candidate frequency band is applied to the current preprocessing filter coefficient for that candidate frequency band. This yields the updated background clutter subtraction threshold and updated preprocessing filter coefficients for the next detection cycle. In the next detection cycle, S1 to S4 are re-executed using the updated background clutter subtraction threshold and preprocessing filter coefficients, and the signed trajectory deviation amplitude for the next detection cycle is obtained. The signed trajectory deviation amplitude for the next detection cycle is compared with the signed trajectory deviation amplitude for the detection cycle before this update to determine whether the update has caused the actual change trajectory of the clothing moisture content estimate to converge towards the reference ideal drying attenuation curve.

[0059] If the amplitude of the signed trajectory deviation in the next detection cycle decreases, the current update is retained; otherwise, the parameters before the update are restored and the update step size for the next detection cycle is reduced. Specifically, if the amplitude of the signed trajectory deviation in the next detection cycle is less than the amplitude of the signed trajectory deviation in the previous detection cycle, the updated background clutter subtraction threshold and preprocessing filter coefficients are retained. If the amplitude of the signed trajectory deviation in the next detection cycle is greater than or equal to the amplitude of the signed trajectory deviation in the previous detection cycle, the background clutter subtraction threshold and the preprocessing filter coefficients are restored to their previous values. After restoring the parameters before the update, the update amounts of the background clutter subtraction threshold and the preprocessing filter coefficients generated by S5 in the next detection cycle are reduced by a smaller adjustment margin, ensuring that the parameter changes in the next detection cycle are less than those in the previous update process. The update step size is used to limit the actual adjustment ratio of the background clutter subtraction threshold update and the preprocessing filter coefficient update within a single detection period. When the update step size decreases, the changes in the background clutter subtraction threshold and the preprocessing filter coefficient within a single detection period decrease synchronously.

[0060] When the amplitude of the signed trajectory deviation is less than a preset convergence threshold for a consecutive preset number of detection cycles, and each update value is less than its corresponding preset stability threshold for a consecutive preset number of detection cycles, the current background clutter subtraction threshold and preprocessing filter coefficients are locked. Here, "consecutive preset number of detection cycles" means that the judgment condition is met for multiple consecutive detection cycles. If the judgment condition is not met in any intermediate detection cycle, the number of consecutive detection cycles that meet the condition is re-accumulated.

[0061] The background clutter subtraction threshold update amounts for each candidate frequency point correspond to the first stable threshold, and the preprocessing filter coefficient update amounts for each candidate frequency band correspond to the second stable threshold. To determine whether the background clutter subtraction threshold has reached a stable state, the absolute value of the background clutter subtraction threshold update amount obtained in the current detection cycle for each candidate frequency point is compared with the corresponding first stable threshold. To determine whether the preprocessing filter coefficient has reached a stable state, the absolute value of the preprocessing filter coefficient update amount obtained in the current detection cycle for each candidate frequency band is compared with the corresponding second stable threshold. The first and second stable thresholds are determined through pre-calibration experiments and are used to characterize the maximum allowable update amount of the corresponding parameter within a single detection cycle. A smaller first stable threshold indicates that the background clutter subtraction threshold needs to reach a smaller change state before locking; a smaller second stable threshold indicates that the preprocessing filter coefficient needs to reach a smaller change state before locking. The preset convergence threshold is determined through a standard load drying experiment and is used to characterize the upper limit of the allowable signed trajectory deviation amplitude of the actual change trajectory of the clothing moisture content estimate relative to the reference ideal drying attenuation curve. When determining the preset convergence threshold, the amplitude of multiple signed trajectory deviations between the actual change trajectory of the estimated moisture content of clothing during the standard load drying test and the reference ideal drying decay curve is obtained, and the upper limit of the allowable signed trajectory deviation amplitude is determined according to the deviation range corresponding to the predetermined stable drying state.

[0062] When determining the locking condition, within a preset number of consecutive detection cycles, each detection cycle must simultaneously meet the following conditions: the amplitude of the signed trajectory deviation is less than a preset convergence threshold; the absolute value of the background clutter subtraction threshold update for all candidate frequency points is less than the corresponding first stability threshold; and the absolute value of the preprocessing filter coefficient update for all candidate frequency bands is less than the corresponding second stability threshold. Once these conditions are met, the current background clutter subtraction threshold and preprocessing filter coefficients are locked, ensuring that subsequent detection cycles continue to use the locked parameters for echo signal processing, and reducing fluctuations in detection results caused by continuous parameter adjustments.

[0063] If the amplitude of the signed trajectory deviation does not reach the preset convergence threshold within a continuous detection cycle, or if the update amounts of the background clutter subtraction threshold and the preprocessing filter coefficients do not meet the corresponding stability conditions, the current parameter locking operation is not performed, and the parameter adaptive adjustment state continues. The number of consecutive non-converged detection cycles is used to record the duration for which the locking conditions are not met for multiple consecutive detection cycles. When any detection cycle meets other state changes besides the locking conditions, the number of consecutive non-converged detection cycles is recorded again. When the number of consecutive non-converged detection cycles reaches the preset readjustment condition, it indicates that the current background clutter subtraction threshold and preprocessing filter coefficients have not formed a stable combination matching the current clothing state. In this case, the update step size in the next detection cycle is reduced, and the corresponding update amounts are regenerated based on the local moisture sensitivity, the consistency of change direction, and the signed trajectory deviation obtained in S5. When reducing the update step size, the adjustment range of the background clutter subtraction threshold update amount and the preprocessing filter coefficient update amount within a single detection cycle is also reduced, so that the range of subsequent parameter changes gradually decreases, avoiding a continuous increase in the signed trajectory deviation due to continuous large-amplitude adjustments.

[0064] During the readjustment process, the background clutter subtraction threshold and preprocessing filter coefficients obtained in the current detection cycle are retained as the basic parameters for the next detection cycle, and adjustments are made only within a limited range based on the new update. If the preset convergence threshold is not met after multiple readjustments, the background clutter subtraction threshold and preprocessing filter coefficients corresponding to the current detection cycle are kept unchanged, and the estimated moisture content of the clothing and the signed trajectory deviation obtained based on the current parameters continue to be output, so that the drying process can continue to run without detection interruption due to parameter adjustment failure. When the closed-loop adjustment process can reach the convergence condition, the corresponding parameters are locked; when the closed-loop adjustment process cannot quickly reach the convergence condition due to changes in clothing status, environmental changes, or signal disturbances, the update amplitude is limited and stable parameters are maintained to avoid continuous parameter oscillation, thus keeping the entire feedback process stable.

[0065] It should be noted that the process also includes a drying control step: based on the preset moisture content range where the estimated moisture content of the clothing obtained in S3 falls, target values ​​for heating power and drum speed are determined. The preset moisture content range is established in advance through standard load drying tests. During this process, standard load drying tests are conducted under different initial moisture content conditions, and the actual moisture content, heating power, and drum speed operating status at each drying stage are recorded. Control parameters corresponding to different moisture content ranges are determined based on drying efficiency and drying process stability. The continuously changing moisture content range of the clothing is divided into multiple preset moisture content ranges, and a corresponding correlation between the target heating power value and the target drum speed value is established for each preset moisture content range. This correlation uses the moisture content range as input and the target heating power value and the target drum speed value as output.

[0066] During the current drying task, the estimated moisture content of the clothing obtained in step S3 is acquired, and the preset moisture content range to which the estimated moisture content belongs is determined. When the estimated moisture content falls within a preset moisture content range, the target heating power and target drum speed corresponding to that range are read and used as the initial control target values ​​for the current detection cycle. As the estimated moisture content decreases, the corresponding preset moisture content range changes, and the target heating power and target drum speed are adjusted synchronously according to their correlation, ensuring that the control state of the drying equipment corresponds to the actual moisture content of the clothing. The allowable single-detection-cycle variation of the target heating power and target drum speed is determined based on the amplitude of the signed trajectory deviation obtained in step S4. This allowable single-detection-cycle variation limits the maximum range of variation of the target heating power and target drum speed between adjacent detection cycles, preventing excessive adjustments to the equipment control parameters due to short-term fluctuations in the estimated moisture content of the clothing.

[0067] When determining the allowable change per single detection cycle, the amplitude of the signed trajectory deviation is used as the input for control stability judgment, and a preset correspondence is established between the amplitude of the signed trajectory deviation and the allowable change per single detection cycle. When the amplitude of the signed trajectory deviation is small, it indicates that the actual change trajectory of the estimated moisture content of the clothing deviates little from the reference ideal drying attenuation curve. Therefore, a larger allowable change per single detection cycle is maintained, allowing the target values ​​of heating power and drum speed to respond quickly to changes in the condition of the clothing. When the amplitude of the signed trajectory deviation exceeds the preset deviation threshold, it indicates that the actual change trajectory of the estimated moisture content of the clothing deviates significantly from the reference ideal drying attenuation curve. Therefore, the allowable change per single detection cycle is reduced, slowing down the adjustment process of the target values ​​of heating power and drum speed. The preset deviation threshold is determined through a standard load drying test and is used to characterize the critical value of the signed trajectory deviation amplitude that needs to limit the rate of change of the equipment control parameters. The allowable change per single detection cycle can be determined by proportional limitation, segmented limitation, or nonlinear limitation.

[0068] When using a proportional limiting method, the allowable change per detection cycle is determined based on the proportional relationship between the amplitude of the signed trajectory deviation and a preset deviation threshold. This reduces the allowable change as the deviation amplitude increases. When using a segmented limiting method, the amplitude of the signed trajectory deviation is divided into multiple deviation intervals, with different intervals corresponding to different allowable changes per detection cycle. When the amplitude of the signed trajectory deviation enters a larger deviation interval, a smaller allowable change per detection cycle is used. When using a nonlinear limiting method, the allowable change per detection cycle changes nonlinearly with the amplitude of the signed trajectory deviation, reducing the adjustment speed of control parameters under large deviation conditions. The drying equipment is controlled according to the defined target values ​​for heating power and drum speed. Specifically, the target value for heating power, after being limited by the allowable change per detection cycle, is sent to the heating control module, and the target value for drum speed, after being limited by the allowable change per detection cycle, is sent to the drive control module, causing the drying equipment to operate according to the control parameters corresponding to the current detection cycle.

[0069] In subsequent testing cycles, the target values ​​for basic heating power and drum speed are re-determined based on the estimated moisture content of the clothing obtained in S3. The allowable variation in a single testing cycle is adjusted based on the signed trajectory deviation obtained in S4, so that the control process of the drying equipment keeps dynamically matched with the actual drying state of the clothing. This drying control process only changes the execution parameters of the drying equipment, without changing the background noise subtraction threshold and preprocessing filter coefficients used for ultra-wideband spectrum feedback adjustment in S5 and S6, so that the equipment control closed loop and the detection parameter optimization closed loop remain independent of each other.

[0070] Implementation Method 1: A drying control implementation method based on estimated moisture content of clothing. In one implementation method, a standard cotton load is used to test the drying equipment. It should be noted that the following parameter values ​​are used to illustrate one implementation process of the technical solution and do not constitute a limitation on the scope of protection. In actual operation, adjustments can be made according to the structure of the drying equipment, the material of the clothing, and the load.

[0071] The actual moisture content variation range corresponding to different drying stages was obtained through standard load drying tests. Based on the moisture content of the clothing, a correspondence was established between preset moisture content ranges and target values ​​for heating power and drum speed. For example, the moisture content of the clothing was divided into: High moisture content range: 60%–100%; medium moisture content range: 30%–60%; low moisture content range: 0–30%. The corresponding relationships are established as shown in Table 1. Table 1. Correlation between preset moisture content range, target heating power, and target drum speed.

[0072] After the current drying task starts, S3 obtains the estimated moisture content of the clothes, for example: the estimated moisture content is 72%. Since 72% is in the high moisture content range of 60% to 100%, the target values ​​for heating power and drum speed are determined as follows: 1800W; 55r / min. Then, S4 obtains the current signed trajectory deviation amplitude to determine whether the current drying trajectory deviates from the reference ideal drying decay curve. For example: current signed trajectory deviation amplitude: 0.08; preset deviation threshold: 0.05. Because 0.08 > 0.05, it indicates that the current actual drying trajectory deviates beyond the allowable range. Therefore, the allowable single-detection cycle variation of the target heating power and drum speed is reduced, so that the equipment control parameters are adjusted according to a smaller variation range. For example: original allowable single-detection cycle variation: heating power: 200W; drum speed: 10r / min; after limitation: heating power: 50W; drum speed: 3r / min.

[0073] Implementation Method 2: Parameter Update Implementation Method Based on Frequency Sensitivity Feedback. In one implementation method, the ultra-wideband signal transceiver module sets five candidate frequency points: 3.5GHz, 4.0GHz, 4.5GHz, 5.0GHz, and 5.5GHz. The following parameter values ​​are used to illustrate a specific implementation process and are not intended to limit fixed parameters in actual applications. In actual operation, they can be adjusted according to the structure of the drying equipment, the material of the clothing, and the signal environment. During the current detection cycle, the clothing echo signal is acquired through the ultra-wideband signal transceiver module, and the multipath scattering time-frequency response spectrum is obtained through S1, and then a multi-source feature fusion normalization matrix is ​​formed through S2. The multi-source feature fusion normalization matrix is ​​input into a preset nonlinear regression model to obtain the current estimated value of the clothing moisture content. For example, if the estimated value of the clothing moisture content obtained in the current detection cycle is 42%, according to the preset moisture content range, the estimated value of the clothing moisture content is in the medium moisture content range of 30% to 60%. Therefore, the current drying control parameters are determined to be a target heating power of 1400W and a target drum speed of 45r / min.

[0074] The signed trajectory deviation amplitude for the current detection cycle is calculated using S4, yielding a value of 0.06. The preset deviation threshold is 0.05. Since 0.06 is greater than 0.05, it indicates a significant deviation between the actual change trajectory of the current estimated moisture content of the clothing and the reference ideal drying attenuation curve. Therefore, the allowable single-detection-cycle variation of the target heating power and drum speed is limited, reducing the adjustment speed of the equipment control parameters. In the S5 frequency sensitivity feedback process, a candidate frequency of 4.5GHz is used as an example. The normalized frequency domain attenuation corresponding to the current candidate frequency is 0.60, and the preset perturbation is set to 0.01. While keeping other inputs in the multi-source feature fusion normalization matrix unchanged, the normalized frequency domain attenuation of the current candidate frequency is adjusted to 0.61 and 0.59 respectively, and then input into the preset nonlinear regression model. When the normalized frequency domain attenuation is 0.61, the estimated moisture content of the clothing is 41.8%; when the normalized frequency domain attenuation is 0.59, the estimated moisture content of the clothing is 42.2%.

[0075] Calculate the local water content sensitivity of the current candidate frequency point based on the output results of the two models above: (41.8-42.2)÷(2×0.01)=-20; The above calculation result of local moisture content sensitivity is used to represent the local influence of the change in normalized frequency domain attenuation of the current candidate frequency point on the estimated moisture content of clothing. This value is only used to illustrate the calculation process. In actual operation, the local moisture content sensitivity changes according to different candidate frequencies and the current state of clothing. Subsequently, the change value of normalized frequency domain attenuation between adjacent sampling times of the current candidate frequency point is obtained within a continuous time window, and this change value is multiplied by the local moisture content sensitivity to form the frequency point prediction contribution sequence corresponding to the current candidate frequency point. Based on the consistency of the frequency point prediction contribution sequence with the change direction of the reference ideal drying attenuation curve within the corresponding time interval, the consistency of the change direction of the current candidate frequency point is determined.

[0076] Based on the local moisture content sensitivity, consistency of change direction, and signed trajectory deviation of the current candidate frequency, a background clutter subtraction threshold update is generated for the current candidate frequency. For example, if the background clutter subtraction threshold for the current candidate frequency is 0.35, and the background clutter subtraction threshold update obtained after S5 calculation is -0.02, then the background clutter subtraction threshold used in the next detection cycle is adjusted to 0.33. Simultaneously, according to the preprocessing filter coefficient update relationship of the corresponding candidate frequency band, if the current preprocessing filter coefficient is 0.80, and the preprocessing filter coefficient update is 0.03, then the preprocessing filter coefficient used in the next detection cycle is adjusted to 0.83. Through the above process, different candidate frequencies are differentiated and adjusted according to their own influence on the estimated moisture content of clothing, thereby enhancing the moisture-sensitive frequency band and reducing the impact of background interference frequency bands on the detection results.

[0077] Implementation Method 3: Closed-Loop Convergence Locking Implementation Method. In one implementation method, after the drying equipment completes one detection cycle, the detection parameters for the next detection cycle are adjusted based on the background noise subtraction threshold update and preprocessing filter coefficient update obtained in S5. The following parameter values ​​are used to illustrate the closed-loop adjustment and locking process and are not intended to limit the fixed parameters in actual operation. Initially, the current background noise subtraction threshold is set to 0.35, and the current preprocessing filter coefficient is set to 0.80. After the first detection cycle runs, the signed trajectory deviation amplitude corresponding to the estimated moisture content of the clothing is obtained through S4 as 0.12. According to S5, the background noise subtraction threshold update is -0.02, and the preprocessing filter coefficient update is 0.03. Therefore, the background noise subtraction threshold used in the next detection cycle is adjusted to 0.33, and the preprocessing filter coefficient is adjusted to 0.83.

[0078] The second detection cycle re-executes the detection process using the updated parameters and obtains the signed trajectory deviation amplitude again. For example, the signed trajectory deviation amplitude obtained in the second detection cycle is 0.07, which still has not reached the preset convergence state, so parameter adjustment continues. After calculation by S5, the background clutter subtraction threshold is adjusted to 0.32, and the preprocessing filter coefficient is adjusted to 0.84. The third detection cycle continues to run using the adjusted parameters and obtains a signed trajectory deviation amplitude of 0.03. If the preset convergence threshold is set to 0.05, then 0.03 is less than 0.05, indicating that the actual change trajectory of the current clothing moisture content estimate has entered the allowable deviation range. Meanwhile, if the absolute value of the background clutter subtraction threshold update obtained in the current detection cycle is 0.005 and the first stable threshold is set to 0.01, then 0.005 is less than 0.01, satisfying the background clutter subtraction threshold stability condition; if the absolute value of the current preprocessing filter coefficient update is 0.006 and the second stable threshold is set to 0.01, then 0.006 is less than 0.01, satisfying the preprocessing filter coefficient stability condition.

[0079] After the convergence and stabilization conditions are met for a preset number of detection cycles, the current background clutter subtraction threshold and preprocessing filter coefficients are locked. For example, if the following conditions are met for three consecutive detection cycles: the amplitude of the signed trajectory deviation is less than 0.05, the absolute value of the background clutter subtraction threshold update is less than 0.01, and the absolute value of the preprocessing filter coefficient update is less than 0.01, then the current parameter state is locked. At this time, the locked background clutter subtraction threshold is 0.32, and the preprocessing filter coefficient is 0.84. Subsequent detection cycles continue to use the above parameters to perform ultra-wideband echo signal processing. If the locking condition is not met within consecutive detection cycles, the update step size is reduced according to the non-convergence processing method in S6, and the parameters are adjusted according to the new detection results, so that the parameter adjustment process gradually reduces the change amplitude and avoids continuous oscillation of the background clutter subtraction threshold and preprocessing filter coefficients. Through the above implementation method, dynamic optimization and stable locking in the adaptive adjustment process of ultra-wideband spectrum parameters can be realized, so that the detection parameters can be adapted to different clothing states and different drying stages, while ensuring that the subsequent clothing moisture content detection process remains stable.

[0080] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0081] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0083] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting the moisture content of clothing based on UWB spectrum, characterized in that, Includes the following steps: S1, emits an ultra-wideband pulse inside the drum of the drying equipment and receives the echo signal reflected by the clothes and the inner wall of the drum; preset multiple candidate frequency points and one-to-one corresponding and non-overlapping candidate frequency bands, and only the first detection cycle uses the initial background clutter subtraction threshold and initial preprocessing filter coefficient to extract the frequency domain attenuation, peak attenuation characteristics and root mean square time delay characteristics to form a multipath scattering time-frequency response spectrum. S2, obtain the temperature inside the cylinder, exhaust humidity and running time, and align them with the time-frequency response spectrum of multipath scattering to construct a multi-source feature fusion normalization matrix; S3. Input the results of S2 into the preset nonlinear regression model to obtain the estimated moisture content of the clothing. S4, calculate the signed trajectory deviation between the actual change trajectory of the estimated moisture content of clothing and the reference ideal drying decay curve; S5, determine the local water content sensitivity of each candidate frequency point, and form a frequency point prediction contribution sequence by combining the corresponding frequency domain attenuation changes. Generate background clutter deduction threshold update amount and preprocessing filter coefficient update amount according to the consistency of change direction and signed trajectory deviation. S6 will send the update data back to S1; The update step size is dynamically determined based on the amplitude of the signed trajectory deviation in the next detection cycle; when the amplitude of the signed trajectory deviation and each update quantity are less than the preset convergence threshold and the corresponding preset stability threshold for a consecutive preset number of detection cycles, the current background clutter subtraction threshold and preprocessing filter coefficients are locked.

2. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, In S1, the echo signal is segmented into segments by continuous and non-overlapping time windows. For each time window, the frequency domain attenuation and the frequency domain main peak attenuation amplitude of each candidate frequency point are extracted. The root mean square time delay characteristics are calculated based on the time delay and corresponding amplitude of each multipath component in the echo signal. The frequency domain main peak attenuation amplitude is used as the peak attenuation characteristic. The frequency domain attenuation, peak attenuation characteristics, and root mean square delay characteristics corresponding to multiple consecutive time windows are respectively subjected to median filtering to filter out instantaneous outliers caused by clothing rolling, falling, and collisions, and the correspondence between frequency domain attenuation, peak attenuation characteristics, and root mean square delay characteristics is maintained according to the sampling time of each time window.

3. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 2, characterized in that, Baseline drift correction is performed before median filtering: the frequency domain peak attenuation amplitudes of multiple consecutive time windows are formed into a frequency domain peak attenuation amplitude sequence along the time axis and empirical mode decomposition is performed to obtain multiple intrinsic mode function components; the intrinsic mode function component with the lowest changing frequency is selected, and the correlation between the intrinsic mode function component and the temperature change sequence inside the cylinder in the same time interval is calculated. When the correlation reaches the preset thermal drift correlation threshold, the intrinsic mode function component is determined as the thermal drift baseline component and is subtracted point by point from the frequency domain peak attenuation amplitude sequence. The subtracted frequency domain peak attenuation amplitude sequence is used as the input for median filtering.

4. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, The dimensionless normalization in S2 includes: for each feature quantity, subtracting the minimum value of the feature quantity within a preset historical period from the current value, and then dividing by the difference between the maximum and minimum values ​​within the preset historical period, so that the corresponding feature quantity is mapped to the interval between 0 and 1; when the difference between the maximum and minimum values ​​is zero, the normalization result of the corresponding feature quantity is set to a preset neutral value in the first detection cycle, and the normalization result of the corresponding feature quantity in the previous detection cycle is used in subsequent detection cycles; the frequency domain attenuation of multiple candidate frequency points forms a frequency domain attenuation feature vector according to the order of candidate frequency points, and together with the peak attenuation feature, root mean square delay feature and drying process parameters, constitutes a multi-source feature fusion normalization matrix.

5. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, The preset nonlinear regression model is established by drying tests conducted on various known materials and multiple initial moisture content levels before the drying equipment leaves the factory. The frequency domain attenuation, peak attenuation characteristics, root mean square time delay characteristics, and drying process parameters of each candidate frequency point are collected simultaneously. The training multi-source feature fusion normalization matrix is ​​constructed according to the same time alignment and dimensionless normalization method as S2. The real moisture content obtained by the synchronous weighing method is used as the training label. The multi-layer feedforward neural network is trained through the backpropagation algorithm to obtain the preset nonlinear regression model that maps the training multi-source feature fusion normalization matrix to the estimated moisture content of the clothing.

6. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, An ideal drying attenuation curve library is pre-established, which includes multiple ideal drying attenuation curves obtained for different standard test loads and different initial moisture content ranges. Each ideal drying attenuation curve is generated by aligning the real moisture content trajectory obtained by synchronous weighing in multiple repeated drying tests under the corresponding test conditions with the time axis and taking the median value at the corresponding time. For each repeated drying test, the average value of the multi-source feature fusion normalization matrix within the initial preset time period is calculated according to each feature dimension to form the initial feature vector of the test. The average value of multiple initial feature vectors under the same test conditions is calculated according to the corresponding feature dimensions to form the initial feature template vector. After the current drying task is started, the current initial feature vector is formed in the same way. The distance between the current initial feature vector and each initial feature template vector is calculated, and the ideal drying decay curve associated with the initial feature template vector with the smallest distance value is determined as the reference ideal drying decay curve.

7. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 6, characterized in that, When calculating the signed trajectory deviation, short-term, medium-term, and long-term windows are established. The difference between the rate of change of the estimated moisture content of clothing and the rate of change of the corresponding segment of the reference ideal drying attenuation curve within each of the three time windows is calculated. The reliability weight is determined according to the fluctuation variance of the frequency domain attenuation characteristic vector within each time window, where the larger the fluctuation variance, the smaller the corresponding reliability weight. The differences corresponding to the three time windows are weighted and summed according to the reliability weights. The direction of the signed trajectory deviation is determined according to the sign of the weighted summation result, and the magnitude of the signed trajectory deviation is determined by its absolute value.

8. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, In S5, while keeping the other inputs in the multi-source feature fusion normalization matrix unchanged, the normalized frequency domain attenuation of the current candidate frequency point is increased and decreased by the preset perturbation amount respectively. The ratio of the difference between the estimated moisture content of clothing obtained after inputting the preset nonlinear regression model twice to twice the preset perturbation amount is used as the local moisture content sensitivity of the current candidate frequency point. Within a continuous time window, the normalized frequency domain attenuation change value of the current candidate frequency point between each adjacent sampling time is multiplied by the corresponding local water content sensitivity to obtain the frequency point prediction contribution corresponding to each sampling interval, and a frequency point prediction contribution sequence is formed in chronological order. The proportion of the number of sampling intervals in which the frequency point prediction contribution sequence and the reference ideal dry attenuation curve have the same direction of change within the corresponding sampling interval to the total number of sampling intervals is determined as the degree of consistency of change direction. Based on the local water content sensitivity, the degree of consistency of change direction, and the signed trajectory deviation, the background clutter subtraction threshold update amount of the current candidate frequency point and the preprocessing filter coefficient update amount of the corresponding candidate frequency band are determined.

9. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, In S6, the background clutter subtraction threshold update amount of each candidate frequency point corresponds to the first stable threshold, and the preprocessing filter coefficient update amount of each candidate frequency band corresponds to the second stable threshold. The first stable threshold and the second stable threshold are determined by pre-calibration experiments and are used to characterize the maximum single detection cycle update amount allowed when the corresponding background clutter subtraction threshold and preprocessing filter coefficient enter a stable state. The preset convergence threshold is determined through a standard load drying test and is used to characterize the upper limit of the amplitude of the signed trajectory deviation allowed by the actual change trajectory of the estimated moisture content of clothing relative to the reference ideal drying decay curve.

10. The method for detecting the moisture content of clothing based on UWB spectrum according to claim 1, characterized in that, It also includes a drying control step: based on the preset moisture content range where the estimated moisture content of the clothes obtained in S3 is located, the target values ​​of heating power and drum speed are determined respectively; based on the amplitude of the signed trajectory deviation obtained in S4, the allowable single detection cycle variation of the target values ​​of heating power and drum speed is determined, wherein when the amplitude of the signed trajectory deviation exceeds the preset deviation threshold, the allowable single detection cycle variation is reduced, and the drying equipment is controlled according to the limited target values ​​of heating power and drum speed.

Images