Training method of neural network model, washing machine and control method

Abstract

The application discloses a kind of neural network model training method, washing machine and control method, neural network model training method: current in data set is converted into q-axis current and d-axis current;Respectively carry out FFT transformation, obtain current spectrum;Determine motor speed, calculate motor mechanical rotation frequency, extract the harmonic amplitude of 1 frequency and 2 frequency in current spectrum;With the extracted harmonic amplitude and motor speed as the input of neural network model, the output of neural network model includes weight, eccentricity and diagonal eccentricity;With the eccentricity, diagonal eccentricity and weight output by neural network model and corresponding actual value, the model is trained.The model obtained by training the model with the harmonic amplitude of 1 frequency and 2 frequency in current spectrum can predict eccentricity and diagonal eccentricity.At the same time, the extracted harmonic amplitude is more obvious than the change characteristics of current, and easier to train the model, improving the accuracy of model output.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically, it relates to a method for training a neural network model, a washing machine, and a control method. Background Technology

[0002] Currently, the detection of washing machine load weight, eccentricity, and diagonal eccentricity is mainly achieved by setting up sensors.

[0003] With the development of artificial intelligence technology in recent years, neural network models have been used to predict the weight of loads. The training of neural network models is mainly carried out using parameters such as motor speed and current.

[0004] Using the current in the motor driver directly as input to the neural network model requires the model to identify the characteristics of current changes. Therefore, the neural network model is complex, has a long training time, poor generalization ability, and yields large errors, making it difficult to achieve the expected results. Most importantly, existing neural network models cannot simultaneously predict both eccentricity and diagonal eccentricity.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background technology of this application, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

[0006] This invention proposes a neural network model training method, a washing machine, and a control method to solve the technical problems of existing washing machines that use parameters such as motor speed and current as inputs to the neural network model, resulting in large prediction errors and the inability to simultaneously predict eccentricity and diagonal eccentricity.

[0007] To achieve the above-mentioned invention / design objectives, the present invention adopts the following technical solution:

[0008] A method for training a neural network model for a washing machine, the method comprising:

[0009] Obtain the dataset; the dataset includes the detected motor speed and motor UV phase line current when several different loads, eccentric loads, and diagonally eccentric loads are placed inside the washing machine drum. , ;

[0010] The current of the UV phase line of the motor , The current is converted into q-axis and d-axis currents; the q-axis and d-axis currents are then segmented according to a set sampling frequency and a set sliding window segmentation time to obtain several frames of q-axis current signals. and d-axis current signal Perform FFT transformations on each to obtain the q-axis current spectrum. and d-axis current spectrum Where x represents the time of the x-th sliding window segmentation, and x is a natural number;

[0011] Determine the motor speed corresponding to the x-th sliding window segmentation time. Calculate the mechanical rotation frequency of the motor Extract the q-axis current spectrum. and d-axis current spectrum middle harmonic amplitude and The harmonic amplitude; based on the extracted harmonic amplitude and motor speed As input to the neural network model, the output of the neural network model includes weight, eccentricity, and diagonal eccentricity;

[0012] The neural network model is trained using the eccentricity, diagonal eccentricity, and weight output by the neural network model and their corresponding actual values ​​to obtain a trained neural network model.

[0013] The neural network model training method for washing machines described above involves dividing the extracted harmonic amplitude by the fundamental amplitude to obtain the harmonic energy. , l=1, 2; based on the harmonic energy and motor speed As input to the neural network model.

[0014] The neural network model training method for the washing machine described above, wherein the formula for the j-th neuron in the i-th hidden layer of the neural network model is as follows:

[0015] ;in, It is the output of the j-th neuron node in the previous layer. for The weight of the j-th neuron in the layer. for The bias of the j-th neuron in layer i, where i is the layer containing the neuron and j is the number of neurons in each layer. for The total number of neurons in the layer; The neuron's output is obtained after an activation function. Activation function For ReLU;

[0016] Mean Absolute Error Loss Function: ,in, This is the actual value. The neural network model outputs a value, M is the number of samples, and m is the mth sample. The actual value and the neural network model output value are substituted into the loss function to calculate the loss value of the current neural network model.

[0017] Using the chain rule, the loss function is calculated from the output layer to each weight and bias. For the mth layer, the calculation is as follows:

[0018] ;

[0019] wherein, is the weight vector of the mth layer, is the bias vector of the mth layer;

[0020] According to the calculated gradient, the weights and biases of the neural network model are updated:

[0021] ;

[0022] wherein, is the learning rate;

[0023] Iterative training is performed until the loss function converges or a predetermined number of training rounds is reached, resulting in a trained neural network model.

[0024] The neural network model training method for the washing machine as described above, the motor speed in the data set includes at least two different motor speeds and intermediate speeds between the two different motor speeds.

[0025] A control method for a washing machine, the method comprising:

[0026] Starting the dehydration;

[0027] Controlling the motor speed to reach a first speed;

[0028] Collecting the motor speed, the current of the UV phase line of the motor , Converting the current of the UV phase line of the motor , to q-axis current and d-axis current, respectively performing FFT transform on the q-axis current and d-axis current to obtain q-axis current spectrum and d-axis current spectrum ;

[0029] Determining the corresponding motor speed , calculating the motor mechanical rotation frequency ;

[0030] ;

[0031] extracting the q-axis current spectrum and the d-axis current spectrum of the q-axis current spectrum and the d-axis current spectrum of the q-axis current spectrum and the d-axis current spectrum; taking the extracted harmonic amplitudes and the motor speed as inputs of a trained neural network model, the trained neural network model outputting the weight, the eccentricity and the diagonal eccentricity; controlling the washing machine according to the weight, the eccentricity and the diagonal eccentricity;

[0032] The trained neural network model is a neural network model trained by the training method.

[0033] The control method of the washing machine as described above, judging whether the eccentricity and the diagonal eccentricity output by the trained neural network model are qualified, if yes, controlling the motor to speed up to the second speed for dehydration, if not, controlling the motor to run at the first speed.

[0034] The control method of the washing machine as described above, monitoring the eccentricity when the motor speed is at the second speed for dehydration, continuing dehydration until a dehydration end condition is reached when the eccentricity is qualified, controlling the motor to run at the first speed when the eccentricity is not qualified.

[0035] The control method of the washing machine as described above, the method of monitoring the eccentricity is:

[0036] Monitoring by the eccentricity output by the trained neural network model.

[0037] The control method of the washing machine as described above, the method of monitoring the eccentricity is:

[0038] Monitoring by the eccentricity output by a second neural network model;

[0039] The second neural network model is trained by the following method:

[0040] Obtaining a data set, the data set being the motor speed, the total vector current Is and the total vector voltage Vs corresponding to the detection when a plurality of different eccentricity loads are placed in the drum of the drum washing machine;

[0041] Taking the motor speed, the total vector current Is and the total vector voltage Vs as inputs of the second neural network model, the output of the second neural network model including the eccentricity;

[0042] The motor speed in the data set of the second neural network model is greater than the motor speed in the data set of the neural network model.

[0043] The control method of the washing machine as described above, the formula of the jth neuron of the ith hidden layer of the second neural network model is: ;

[0044] wherein, is the output of the jth neuron node of the previous layer, is the weight of the jth neuron of the ith layer, is the bias of the jth neuron of the ith layer, i is the layer where the neuron is located, and j is the number of neurons in each layer, is the total number of neurons in the ith layer; is the total number of neurons in the ith layer; is the output of the neuron after passing through the activation function , the activation function is ReLU; the mean absolute error loss function is: wherein, is the actual value, is the output value of the neural network model, M is the number of samples, and m is the mth sample; the actual value and the output value of the neural network model are substituted into the loss function to calculate the loss value of the current neural network model; The gradient of the loss function with respect to each weight and bias is calculated from the output layer using the chain rule, and for the ith layer, the calculation is as follows:

[0045] wherein, is the weight vector of the ith layer, is the bias vector of the ith layer;

[0046] ;

[0047] wherein, is the weight vector of the ith layer, is the bias vector of the ith layer; According to the calculated gradient, the weights and biases of the second neural network model are updated:

[0048] wherein, is the learning rate;

[0049] ;

[0050] wherein, is the learning rate;

[0051] Iterative training is performed until the loss function converges or a predetermined number of training rounds is reached, and a trained second neural network model is obtained.

[0052] Compared with the prior art, the advantages and positive effects of the present application are: the neural network model training method of the washing machine: obtaining a data set, the data set including the corresponding detected motor speed, motor UV phase current when a washing machine drum is placed with several different loads, eccentric loads, and diagonal eccentric loads 、 ; to change the current of the motor's UV phase line , The current is converted to q-axis and d-axis currents. FFT transformations are performed on both q-axis and d-axis currents to obtain the current spectrum. The motor speed is determined, and the motor's mechanical rotational frequency is calculated. Harmonic amplitudes at 1st and 2nd harmonic frequencies are extracted from the current spectrum. The extracted harmonic amplitudes and motor speed are used as inputs to a neural network model. The output of the neural network model includes weight, eccentricity, and diagonal eccentricity. The neural network model is trained using the eccentricity, diagonal eccentricity, and weight outputs along with their corresponding actual values, resulting in a trained neural network model. The eccentricity spectrum shows a significant increase in the 1st frequency component and a slight increase in the 2nd frequency component. The diagonal spectrum also shows a significant increase in the 2nd frequency component. Therefore, by performing FFT transformations on the q-axis and d-axis currents to obtain the current spectrum and extracting the harmonic amplitudes at 1st and 2nd harmonic frequencies, the trained neural network model can predict weight, eccentricity, and diagonal eccentricity. Furthermore, using the extracted harmonic amplitudes is more effective than directly using the current's variation characteristics, making it easier to train the neural network model and improving the accuracy of the model's output.

[0053] The control method for washing machines uses a trained neural network model that can predict the weight, eccentricity, and diagonal eccentricity of the load. Furthermore, the extracted harmonic amplitude is more significant than the direct use of current variation characteristics, making it easier to train the neural network model and improving the accuracy of the model output, thus enabling more precise control of the washing machine.

[0054] Other features and advantages of the present invention will become clearer after reading the detailed embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of a specific embodiment of the present invention, showing the load evenly arranged inside the cylinder.

[0057] Figure 2 yes Figure 1 Top view of area 1.

[0058] Figure 3 This is a schematic diagram of a specific embodiment of the present invention with the load arranged in an eccentric state.

[0059] Figure 4 yesFigure 3 Top view of region one.

[0060] Figure 5 is a schematic diagram of a specific embodiment of the present application load arranged in a diagonal eccentric state.

[0061] Figure 6 is Figure 5 Top view of region one.

[0062] Figure 7 is Figure 5 Top view of region two.

[0063] Figure 8 is a motor speed change diagram.

[0064] Figure 9 is a model diagram of a first neural network model of a specific embodiment of the present application.

[0065] Figure 10 is a flow chart of a specific embodiment of the present application washing machine control method.

[0066] Figure 11 is a model diagram of a second neural network model of a specific embodiment of the present application.

[0067] Figure 12 is a flow chart of a control method of a specific embodiment of the present application washing machine. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0069] In the description of the present application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore, it cannot be understood as a limitation of the present application.

[0070] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. In the description of embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0071] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0072] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0073] Weighing is essential for washing machines to determine the weight of clothes. This allows for the calculation of appropriate water levels, washing time, and detergent dosage based on the weight and type of clothing, ensuring thorough cleaning while conserving water and detergent. Therefore, weighing in washing machines provides more efficient and reliable washing. Off-center and diagonal off-center detection identifies the distribution of clothes within the machine, preventing vibrations and noise caused by uneven distribution, reducing damage to internal components, and extending the machine's lifespan. Off-center detection during high-speed spin-drying helps prevent drum collisions caused by sudden imbalances in the drum when spinning out clothes with poor water permeability.

[0074] A method for training a neural network model for a washing machine, comprising:

[0075] Obtain the dataset.

[0076] The dataset includes the detected motor speed and motor UV phase line current when several different loads, eccentric loads, and diagonally eccentric loads are placed inside the washing machine drum. , .

[0077] exist Figures 1-7 In this example, the cylinder is divided into three zones at 120-degree angles, and the counterweights are placed according to these zones.

[0078] In some embodiments, the load is increased by 2x500g per area each time, and the counterweights are placed from left to right according to the working condition requirements. That is, 1000g of all areas are placed with 500g counterweights in positions 1 and 2 at the same time; 2000g are placed with 500g counterweights in positions 1, 2, 3, and 4 at the same time; when the load exceeds 4000g, new counterweights are stacked on the counterweights in positions 1 and 2, and so on.

[0079] The eccentric block (red block in the figure) is fixedly placed in position 4 of each area, and is increased by 200g.

[0080] When measuring the eccentric load condition, first configure the counterweights, and measure the corresponding eccentric counterweight for each area respectively. Place the eccentric block on the load block in position 4, and increase by 200g.

[0081] When testing the diagonal eccentricity, for testing area one, place the diagonal eccentricity in position 2 and align the outer cylinder wall, and place the same weight eccentric block in position 7 of area two and align the inner cylinder wall. Other areas are done in the same way.

[0082] Convert the current of the UV phase line of the motor 、 to the q-axis current and the d-axis current.

[0083] Collect the current of the UV phase line of the motor through a sampling resistor 、 , and convert it to -axis coordinate system 、 .

[0084] .

[0085] Convert -axis coordinate system 、 to , dq-axis coordinate system.

[0086] ;

[0087] wherein, is the electrical angle of the motor rotor.

[0088] The motor speed in the data set includes at least two different motor speeds and intermediate speeds between the two different motor speeds.

[0089] In Figure 8In the example of the motor speed control, the data is collected after the motor speed is stabilized at 90 rpm, during the acceleration process to 120 rpm, after the motor speed is stabilized at 120 rpm, during the deceleration process to 90 rpm, and after the motor speed is stabilized at 90 rpm.

[0090] Divide the sample data set: divide the collected data into a training set (80%), a validation set (10%), and a test set (10%). The training set is used for model training, the validation set is used for adjusting model hyperparameters, and the test set is used for evaluating the final performance of the model.

[0091] Preprocess the data set, normalize and standardize the data, etc., to make the data have similar scales and distributions, to speed up model convergence.

[0092] Data feature extraction: filter, sliding window segmentation, and FFT extraction of q-axis and d-axis current harmonic spectrum features.

[0093] (1) Perform first-order low-pass filtering on the collected data to remove noise data.

[0094] (2) The sampling frequency is set to the set sampling frequency, and the sliding window segmentation time is set to the set sliding window segmentation time.

[0095] (3) Divide the q-axis current and d-axis current according to the set sampling frequency and set sliding window segmentation time to obtain a number of q-axis current signal frames and d-axis current signal frames , respectively, and perform FFT transformation to obtain q-axis current spectrum and d-axis current spectrum ; where x represents the xth sliding window segmentation time, and x is a natural number.

[0096] ;

[0097] .

[0098] Fundamental frequency identification: determine the motor speed corresponding to the xth sliding window segmentation time , and calculate the motor mechanical rotation frequency .

[0099] Harmonic positioning: extract the harmonic amplitude of and the harmonic amplitude of from the q-axis current spectrum and the d-axis current spectrum .

[0100] : the main characteristic frequency of eccentricity (OOB).

[0101] Characteristic frequency of off-diagonal eccentricity (DOOB).

[0102] Spectrum performance of off-diagonal eccentricity (DOOB): Component significantly increases, Component slightly increases.

[0103] Spectrum performance of off-diagonal eccentricity (DOOB): Component significantly increases.

[0104] The extracted harmonic amplitudes and the motor speed are taken as inputs of the neural network model, and the output of the neural network model includes the weight, the eccentricity and the off-diagonal eccentricity.

[0105] The eccentricity, the off-diagonal eccentricity and the weight output by the neural network model are trained with the corresponding actual values to obtain a trained neural network model.

[0106] In some embodiments, the extracted harmonic amplitudes are divided by the fundamental amplitude to obtain harmonic energy , , l=1, 2; the harmonic energy and the motor speed are taken as inputs of the neural network model.

[0107] Dividing each harmonic amplitude by the fundamental amplitude to obtain the relative harmonic energy (in percentage form) can eliminate the influence of motor power difference.

[0108] Using the harmonic amplitudes of the extracted data as inputs of the neural network is more obvious than directly using the current variation characteristics, and is easier for the neural network model to fit.

[0109] In the example of Figure 9 , the input layer of the first neural network model consists of 3 neurons, respectively corresponding to the speed, the q-axis current harmonic amplitude and the d-axis current harmonic amplitude of the collected sample data; the output layer consists of 3 neurons, respectively corresponding to the eccentricity (OOB), the off-diagonal eccentricity (DOOB) and the weight of the sample calibration data.

[0110] Formula of the jth neuron of the ith hidden layer of the neural network model: .

[0111] Wherein, is the output of the jth neuron node of the previous layer, is the weight of the jth neuron of the layer, is the bias of the jth neuron of the layer, i is the layer where the neuron is located, j is the number of neurons in each layer, is the total number of neurons in the layer. is the total number of neurons in the layer. is the total number of neurons in the layer. is the total number of neurons in the layer.

[0112] The neuron's output is obtained after an activation function. Activation function It is ReLU.

[0113] Mean Absolute Error Loss Function: ,in, This is the actual value. The output value of the neural network model is M, where M is the number of samples and m is the m-th sample.

[0114] Substituting the actual values ​​and the output values ​​of the neural network model into the loss function, the loss value of the current neural network model is calculated. This value reflects the degree of difference between the model's prediction and the actual situation.

[0115] Starting from the output layer, using the chain rule, calculate the gradient of the loss function with respect to each weight and bias. For the ... The layer is calculated as follows:

[0116] ;

[0117] in, For the first Layer weight vector, For the first Layer bias vector;

[0118] Based on the calculated gradient, the weights and biases of the neural network model are updated using gradient descent as follows:

[0119] ;

[0120] in, It is the learning rate.

[0121] Iterative training continues until the loss function converges or the predetermined number of training epochs is reached. The model's performance is evaluated on the validation set, and hyperparameters are tuned to prevent overfitting. Finally, the model's generalization ability is evaluated on the test set.

[0122] The trained neural network model is obtained.

[0123] A method for controlling a washing machine:

[0124] Dehydration begins.

[0125] Control the motor speed to reach the first speed.

[0126] In some embodiments, the first rotational speed is a distributed rotational speed.

[0127] Collect motor speed and motor UV phase line current. , The current in the UV phase line of the motor 、 Converting into q-axis current and d-axis current, performing FFT transform on q-axis current and d-axis current respectively to obtain q-axis current spectrum and d-axis current spectrum .

[0128] Determining corresponding motor speed , calculating motor mechanical rotation frequency .

[0129] ;

[0130] Extracting harmonic amplitude in q-axis current spectrum and d-axis current spectrum . Harmonic amplitude.

[0131] Taking extracted harmonic amplitude and motor speed as input of trained neural network model, trained neural network model outputs weight, eccentricity and diagonal eccentricity.

[0132] According to weight, eccentricity and diagonal eccentricity, control the washing machine.

[0133] Wherein, the trained neural network model is a first neural network model trained by the above training method.

[0134] According to the eccentricity and the diagonal eccentricity output by the trained first neural network model, it is judged whether it is qualified or not, if yes, the motor speed is controlled to the second speed for dehydration, if not, the motor is controlled to run at the first speed.

[0135] In some embodiments, the second speed is a high-speed dehydration speed.

[0136] When the motor speed is at the second speed for dehydration, the eccentricity is monitored, and when the eccentricity is qualified, the dehydration continues until the dehydration end condition is reached, and when the eccentricity is unqualified, the motor is controlled to run at the first speed.

[0137] The method for monitoring the eccentricity is:

[0138] Using the eccentricity output by the trained first neural network model for monitoring.

[0139] In the example of Figure 10 , the control method of the washing machine comprises:

[0140] S1, start dehydration.

[0141] S2, control the motor speed to reach the first speed.

[0142] ​S3, collect data, extract current harmonic amplitude, and take the extracted current harmonic amplitude and motor speed as the input of the first neural network model.

[0143] S4, the first neural network model outputs weight, eccentricity and diagonal eccentricity.

[0144] S5, judge whether the eccentricity and diagonal eccentricity are qualified, if yes, go to step S6, otherwise, go to step S2.

[0145] S6, control the motor to the second speed for dewatering.

[0146] S7, monitor the eccentricity by the first neural network model.

[0147] S8, judge whether the eccentricity is qualified, if yes, go to step S9, otherwise, go to step S2.

[0148] S9, reach the dewatering end condition, if yes, go to step S10, otherwise, go to step S7.

[0149] S10, dewatering is finished.

[0150] In some embodiments, the method for monitoring the eccentricity is:

[0151] The eccentricity output by the second neural network model is used for monitoring.

[0152] The second neural network model is obtained by the following method:

[0153] Obtain a data set, the data set is the motor speed, total vector current Is and total vector voltage Vs corresponding to the detection when a plurality of different eccentricity loads are placed in the drum of the drum washing machine;

[0154] Take the motor speed, total vector current Is and total vector voltage Vs as the input of the second neural network model, and the output of the second neural network model includes eccentricity.

[0155] The motor speed in the second neural network data set is greater than the motor speed in the neural network data set.

[0156] In some embodiments, the motor speed in the second neural network data set is 1200 rpm.

[0157] Divide the sampling data set: divide the collected data into training set (80%), validation set (10%) and test set (10%), the training set is used for model training, the validation set is used for adjusting model hyperparameters, and the test set is used for evaluating the final performance of the model.

[0158] Data preprocessing: normalize, standardize and other operations are performed on the data to make the data have similar scale and distribution, so as to accelerate the convergence of the model.

[0159] exist Figure 11 In the example, the input layer of the second neural network model consists of three neurons, corresponding to the rotation speed of the collected sample data, The total vector current Is and the total vector voltage Vs are in the coordinate system; the output layer consists of one neuron, corresponding to the sample calibration data eccentricity (OOB).

[0160] The formula for the j-th neuron in the i-th hidden layer of the second neural network model is as follows:

[0161] .

[0162] in, It is the output of the j-th neuron node in the previous layer. for The weight of the j-th neuron in the layer. for The bias of the j-th neuron in layer i, where i is the layer containing the neuron and j is the number of neurons in each layer. This represents the total number of neurons in the layer.

[0163] The neuron's output is obtained after an activation function. Activation function It is ReLU.

[0164] Mean absolute error loss function: ,in, This is the actual value. The output value of the neural network model is M, where M is the number of samples and m is the m-th sample.

[0165] Substituting the actual values ​​and the output values ​​of the neural network model into the loss function, the loss value of the current neural network model is calculated. This value reflects the degree of difference between the model's prediction and the actual situation.

[0166] Starting from the output layer, using the chain rule, calculate the gradient of the loss function with respect to each weight and bias. For the ... The layer is calculated as follows:

[0167] ;

[0168] in, For the first Layer weight vector, For the first Layer bias vector.

[0169] Based on the calculated gradient, update the weights and biases of the second neural network model.

[0170] ;

[0171] wherein, is the learning rate.

[0172] The training is iterated until the loss function converges or a predetermined number of training rounds is reached, the performance of the model is evaluated on the validation set, and the hyperparameters are adjusted to prevent overfitting. The generalization ability of the model is finally evaluated on the test set.

[0173] A second neural network model is obtained after the training is completed.

[0174] The weight, eccentricity and diagonal eccentricity are predicted by the first neural network model during the process of low-speed rotation of the drum before dehydration, at which time the amplitude-frequency characteristics of the d-axis and q-axis currents are more obvious, and the real-time requirement of detection is not high. Using the amplitude-frequency characteristics of the current as the input layer variable of the first neural network model is more suitable and can improve the prediction accuracy. In the high-speed stage of dehydration, the real-time requirement of calculating OOB is high, and the The total vector current Is and the total vector voltage Vs in the coordinate system are used as the input layer of the feedforward neural network model to meet the real-time requirement.

[0175] According to the different states of the drum at high and low speeds, two neural network models are used to detect to output higher detection accuracy under the premise of meeting the timeliness requirement.

[0176] In the example of Figure 12 , the control method of the washing machine is:

[0177] S1, start dehydration.

[0178] S2, control the motor speed to reach the first speed.

[0179] S3, collect data and extract current harmonic amplitude, and use the extracted current harmonic amplitude and motor speed as the input of the first neural network model.

[0180] S4, the first neural network model outputs the weight, eccentricity and diagonal eccentricity.

[0181] S5, judge whether the eccentricity and diagonal eccentricity are qualified, if yes, go to step S6, otherwise, go to step S2.

[0182] S6, control the motor to speed up to the second speed for dehydration.

[0183] S7, monitor the eccentricity by the second neural network model.

[0184] S8, judge whether the eccentricity is qualified, if yes, go to step S9, otherwise, go to step S2.

[0185] S9, reach the dehydration end condition, if yes, go to step S10, otherwise, go to step S7.

[0186] S10, the dehydration ends.

[0187] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, the technical solutions recorded in the foregoing examples can still be modified, or some of the technical features can be replaced by equivalents, by those of ordinary skill in the art; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions claimed by the present application.

Claims

1. A method of training a neural network model for a washing machine, the method comprising: The method comprises: Obtaining a data set; the data set comprising detected motor speed, motor UV phase current when placing several different loads, eccentric loads, diagonal eccentric loads in the drum of a washing machine , ; The current of the UV phase line of the motor , The current is converted into q-axis and d-axis currents; the q-axis and d-axis currents are then segmented according to a set sampling frequency and a set sliding window segmentation time to obtain several frames of q-axis current signals. and d-axis current signal Perform FFT transformations on each to obtain the q-axis current spectrum. and d-axis current spectrum Where x represents the time of the x-th sliding window segmentation, and x is a natural number; determining the motor speed corresponding to the xth sliding window partition time , calculating the motor mechanical rotation frequency ; Extract the q-axis current spectrum and d-axis current spectrum middle harmonic amplitude and The harmonic amplitude; among which, The characteristic frequency of the eccentricity The characteristic frequency is diagonally eccentric. extracted harmonic amplitudes and motor speed as input to a neural network model, the output of the neural network model including weight, eccentricity, and diagonal eccentricity; The eccentricity, the diagonal eccentricity and the weight output by the neural network model are used to train the neural network model with corresponding actual values, to obtain a trained neural network model. 2.The method of claim 1, wherein, Dividing the extracted harmonic amplitudes by the fundamental amplitude gives the harmonic energy , l = 1, 2; with the harmonic energy and motor speed as input to the neural network model. 3.The method of claim 1 or 2, wherein, The formula of the jth neuron of the ith hidden layer of the neural network model: ; wherein, is the output of the jth neuron node of the previous layer, is the output of the jth neuron node of the previous layer, is the weight of the jth neuron of the i layer, is the output of the jth neuron node of the previous layer, is the bias of the jth neuron of the i layer, i is the layer where the neuron is located, and j is the number of neurons in each layer, is the output of the jth neuron node of the previous layer, is the total number of neurons in the i layer. the output of the neuron through an activation function , the activation function is ReLU; The mean absolute error loss function is: wherein, is an actual value, is a neural network model output value, M is a sample number, and m is an mth sample. The actual values and the neural network model output values are substituted into a loss function, to calculate a loss value of the current neural network model; Using the chain rule, starting from the output layer, the gradients of the loss function with respect to each weight and bias are computed for the i-th layer as The gradients are computed as follows: ； ； wherein, is the first layer weight vector, is the first layer bias vector, is the second layer weight vector, is the second layer bias vector. According to the calculated gradient, the weight and the bias of the neural network model are updated: ； ; wherein, is the learning rate; Iterative training is performed until the loss function converges or a predetermined number of training rounds is reached, to obtain the trained neural network model. 4.The method of claim 1, wherein, The motor speeds in the data set include at least two different motor speeds and intermediate speeds between the two different motor speeds.

5. A control method of a laundry machine, characterized by, The method is: Start dehydration; Control the motor speed to reach a first speed; The motor speed and the current of the motor's UV phase lines are collected. , The current of the UV phase line of the motor , The current is converted into q-axis and d-axis currents, and then subjected to FFT transformation on the q-axis and d-axis currents respectively to obtain the q-axis current spectrum. and d-axis current spectrum ; determining a corresponding motor speed , calculating a motor mechanical rotation frequency ; ; extracting the q-axis current spectrum and the d-axis current spectrum of the harmonic amplitudes and harmonic amplitudes; wherein, is the characteristic frequency of the eccentricity, is the characteristic frequency of the diagonal eccentricity; extracted harmonic amplitudes and motor speed as input to a trained neural network model that outputs weight, eccentricity, and diagonal eccentricity; Control the washing machine according to the weight, the eccentricity and the diagonal eccentricity; The trained neural network model is a neural network model trained by the training method of any one of claims 1-4.

6. The control method of a washing machine according to claim 5, characterized in that, According to the eccentricity and the diagonal eccentricity output by the trained neural network model, it is judged whether it is qualified, if yes, the motor is controlled to speed up to a second speed for dehydration, if not, the motor is controlled to run to the first speed.

7. The control method of the washing machine according to claim 6, characterized in that, When the motor speed is at the second speed for dehydration, the eccentricity is monitored, and when the eccentricity is qualified, the dehydration continues until the dehydration end condition is reached, and when the eccentricity is unqualified, the motor is controlled to run to the first speed.

8. The control method of the washing machine according to claim 7, characterized in that, The method for monitoring the eccentricity is: The eccentricity output by the trained neural network model is used for monitoring.

9. The control method of the washing machine according to claim 7, characterized in that, The method for monitoring the eccentricity is: The eccentricity output by the second neural network model is used for monitoring; The second neural network model is trained by the following method: A data set is obtained, the data set is the corresponding detected motor speed, total vector current Is and total vector voltage Vs when a plurality of different eccentricity loads are placed in the drum of the drum washing machine; The motor speed, the total vector current Is and the total vector voltage Vs are used as the input of the second neural network model, and the output of the second neural network model includes the eccentricity; The motor speed in the data set of the second neural network model is greater than the motor speed in the data set of the neural network model.

10. The control method of the washing machine according to claim 9, characterized in that, a formula of a jth neuron of an ith hidden layer of the second neural network model: ; wherein, is the output of the jth neuron node of the previous layer, is the output of the jth neuron node of the previous layer, is the weight of the jth neuron of the i layer, is the output of the jth neuron node of the previous layer, is the bias of the jth neuron of the i layer, i is the layer where the neuron is located, and j is the number of neurons in each layer, is the output of the jth neuron node of the previous layer, is the total number of neurons in the i layer. the output of the neuron through an activation function , the activation function is ReLU; The mean absolute error loss function is: wherein, is an actual value, is a neural network model output value, M is a sample number, and m is an mth sample. The actual values and the neural network model output values are substituted into a loss function, to calculate a loss value of the current neural network model; Using the chain rule, starting from the output layer, the gradients of the loss function with respect to each weight and bias are computed for the i-th layer as follows: The gradients are computed as follows: ; ; wherein, is the first layer weight vector, is the first layer bias vector; According to the calculated gradient, the weight and the bias of the neural network model are updated: ; ; wherein, is the learning rate; Iterative training is performed until the loss function converges or a predetermined number of training rounds is reached, to obtain the trained neural network model.

Images