Blower bearing temperature prediction method based on frequency spectrum reconstruction neural network

Abstract

The invention provides an air blower bearing temperature prediction method based on a frequency spectrum reconstruction neural network. Accurate prediction of the future state of the air blower bearing temperature is achieved. According to the method, outlet air pressure, high-speed shaft driving end bearing temperature, a high-speed shaft vibration value, inlet air temperature, main motor current and historical blower main motor bearing temperature are collected through a sensor and serve as model input variables, and a prediction model of a frequency spectrum reconstruction neural network structure is established; a spectrum reconstruction layer is constructed to perform frequency domain feature extraction and smoothing processing on input variables, a spectrum projection attention mechanism is designed, a query matrix and a key matrix in an attention matrix are decomposed by using approximate orthogonal projection, different frequency feature weights are adaptively adjusted, and the problem of concept drift is effectively relieved; and model parameters are dynamically updated in combination with operation data of the air blower, a prediction model with the concept drift adaptive capacity is formed, and high-precision prediction of the future state of the temperature of the bearing of the air blower is achieved.

Description

Technical Field

[0001] The present invention designs a blower bearing temperature prediction method based on a spectrum reconstruction neural network, which realizes the accurate prediction of the future state of the blower motor bearing temperature. The blower bearing temperature is an important parameter reflecting the operating status of the equipment. It can directly reflect the stress condition, lubrication condition and health of the bearing, and is the key basis for judging whether the equipment is operating abnormally. As a typical large-scale rotating machinery, blowers are widely used in sewage treatment, petrochemical, electric power and other industries. Therefore, predicting the blower bearing temperature is of great significance for equipment status assessment, fault warning and maintenance strategy formulation, and belongs to the field of industrial equipment status monitoring and intelligent operation and maintenance technology. Background Art

[0002] Blowers, as a typical large-scale rotating mechanical equipment, are widely used in industries such as sewage treatment, electricity, petrochemicals, and metallurgy. They perform important functions such as gas transportation and ventilation in industrial production processes. During blower operation, the bearing components are in a state of high-speed rotation and heavy load for a long time. The temperature changes can reflect the stress, wear, and lubrication conditions of the bearings, and are one of the key parameters that characterize the operating status of the equipment. If abnormal bearing temperature is not discovered in time, it may lead to equipment failure, increased energy consumption, and even affect the continuity and safety of the overall operation of the system. Therefore, predicting the blower bearing temperature is an important means to ensure stable operation of the equipment and achieve status maintenance.

[0003] Existing methods for monitoring and predicting blower bearing temperature primarily rely on static modeling, empirical threshold determination, or single-variable analysis, making them difficult to adapt to the time-varying distribution of data during actual operation. Under the influence of factors such as multi-operating mode switching, equipment aging, or external environmental fluctuations, blower operating data is prone to concept drift, where the statistical characteristics of the input variables change over time, resulting in degraded model performance and unstable prediction results. Traditional methods generally lack the ability to identify and model concept drift, making it difficult to meet the prediction accuracy and reliability requirements under long-term operation and complex operating conditions.

[0004] The present invention designs a blower bearing temperature prediction method based on a spectrum reconstruction neural network. This method uses sensors to collect key variables in the blower operation process, establishes a blower bearing temperature prediction model based on a spectrum reconstruction neural network, and uses the operation process data to dynamically update the model parameters. This constructs a prediction model that can adapt to the concept drift of input variables and realizes accurate prediction of the future state of the bearing temperature. Summary of the Invention

[0005] The present invention obtains a blower bearing temperature prediction method based on spectrum reconstruction neural network. The method collects outlet air pressure, high-speed shaft drive end bearing temperature, high-speed shaft vibration value, inlet air temperature, main motor current, and historical blower main motor bearing temperature through sensors as input variables of the model, and establishes a blower bearing temperature prediction model based on spectrum reconstruction neural network. By smoothing the frequency projection vector and introducing the orthogonal spectrum projection attention mechanism, the concept drift problem of the input variables during operation is alleviated. The model parameters are dynamically updated in combination with the blower operation data, forming a prediction model with concept drift self-adaptation capability, thereby realizing accurate prediction of the future state of the bearing temperature.

[0006] The present invention adopts the following technical solutions and implementation steps:

[0007] A blower bearing temperature prediction method based on spectrum reconstruction neural network is characterized by: collecting blower data, establishing a blower bearing temperature prediction model, training blower bearing temperature prediction model parameters, and predicting the blower bearing temperature, including the following steps:

[0008] 1. Collect blower data:

[0009] Taking the blower as the research object, the outlet air pressure, high-speed shaft drive end bearing temperature, high-speed shaft vibration value, inlet air temperature, main motor current, and historical blower main motor bearing temperature are selected as the input variables of the blower bearing temperature prediction model, and the blower main motor bearing temperature is used as the output variable of the blower bearing temperature prediction model; among them, the main motor refers to the three-phase asynchronous motor that provides direct driving power for the blower, and its output shaft is connected to the high-speed shaft to drive the entire blower system; the high-speed shaft refers to the high-speed rotating shaft section connecting the main motor output shaft and the impeller transmission system, which is mainly used to transmit the rotational power of the main motor; the outlet air pressure data at time t is collected by using a pressure transmitter installed in the outlet pipe as a vector x 1 (t) = [x 1 (t-L+1),…,x 1 (t-1),x 1 (t)], the resistance temperature sensor arranged on the bearing seat is used to collect the bearing temperature data of the high-speed shaft drive end at time t as vector x 2 (t) = [x 2 (t-L+1),…,x 2 (t-1),x 2 (t)], the high-speed shaft vibration value data at time t is collected using a three-axis acceleration sensor as vector x 3 (t) = [x 3 (t-L+1),…,x 3 (t-1),x 3(t)], the temperature transmitter is used to collect the inlet air temperature data at time t as vector x 4 (t) = [x 4 (t-L+1),…,x 4 (t-1),x 4 (t)], the AC current transformer is used to collect the main motor current data at time t as vector x 5 (t) = [x 5 (t-L+1),…,x 5 (t-1),x 5 (t)], the resistance temperature sensor is used to collect the historical blower main motor bearing temperature data at time t as vector x 6 (t) = [x 6 (t-L+1),…,x 6 (t-1),x 6 (t)], construct the input sample matrix as X(t)=[x 1 (t) T ,…,x n (t) T ,…,x 6 (t) T ] T , x n (t) = [x n (t-L+1),…,x n (t-1),x n (t)], n=1,2,…,6, t=L+1,L+2,…,N+L+1, N is the number of sample sequences, L is the number of samples of the input variable, and T represents the transpose of the matrix;

[0010] 2. Establish a blower bearing temperature prediction model:

[0011] A blower bearing temperature prediction model based on spectrum reconstruction neural network is constructed. The spectrum reconstruction neural network consists of an input layer, a spectrum reconstruction layer, a spectrum projection attention layer, and an output layer.

[0012] Input layer: Input the input sample matrix into the model row by row. The output of the input layer at time t is X(t) = [x 1 (t) T ,x 2 (t) T ,…,x 6 (t) T ] T ;

[0013] Spectrum reconstruction layer: Perform real cosine transform on the input sample matrix. Each row of input samples is mapped to a real vector consisting of M frequency projection values. The calculation formula is as follows:

[0014]

[0015] in, represents the mth frequency projection value of the nth row input sample at time t, cos(·) represents the cosine function, F n (t) represents the nth frequency projection vector of the input sample matrix at time t, m = 1, 2, ..., M; the frequency projection vector is smoothed, and the calculation formula of the nth smoothed frequency projection vector is as follows:

[0016] a n (t)=0.9a n (t-1)+0.1||F n (t)-F n (t-1)|| (2)

[0017] b n (t) = 0.9b n (t-1)+0.1(1-||F n (t)-F n (t-1)||) (3)

[0018]

[0019] Among them, a n (t) represents the weighted cumulative value of the change amplitude of the nth frequency projection vector at time t. The change amplitude refers to the Euclidean distance between the frequency projection vector at the current moment and the previous moment. n (t-1) represents the weighted cumulative value of the change amplitude of the nth frequency projection vector at the moment before time t, F n (t-1) represents the nth frequency projection vector of the input sample matrix at the moment before time t, represents the mth frequency projection value of the nth row of the input sample matrix at the moment before time t, ||·|| represents the Euclidean norm operation of the vector, b n (t) represents the weighted cumulative value of the change stability of the nth frequency projection vector at time t. The change stability refers to the difference between the Euclidean distance between the current frequency projection vector and the previous moment and 1. b n (t-1) represents the weighted cumulative value of the stability of the n-th frequency projection vector change at the moment before time t, represents the smoothed nth frequency projection vector, represents the mth frequency projection value of the nth row input sample at time t after smoothing, α n (t) represents the smoothing coefficient corresponding to the nth frequency projection vector, α(t) = [α 1 (t),α 2 (t),…,α6 (t)] T , α(t) represents the smoothing coefficient vector at time t; the smoothed frequency projection vectors are combined into a frequency embedding matrix, and a linear transformation operation is performed on the frequency embedding matrix to obtain the word unit expression matrix. The calculation formula is as follows:

[0020] Z1(t)=E(t)W1(t)+b1(t) (6)

[0021] Among them, E(t) represents the frequency embedding matrix at time t, Z1(t) represents the word expression matrix at time t, represents the mth row vector of the word expression matrix, represents the element in the mth row and ith column of the word expression matrix, W1(t) represents the frequency embedding weight matrix at time t, represents the nth row vector of the frequency embedding weight matrix, represents the element in the nth row and ith column of the frequency embedding weight matrix, b1(t) represents the frequency embedding bias vector at time t, represents the i-th frequency embedding bias term, D is the embedding dimension of the word unit expression, i = 1, 2, ..., D;

[0022] Encoding layer: It consists of three stacked encoding modules, where s = 1, 2, 3 represents the number of the sth encoding module. Each encoding module performs frequency projection attention operation, residual connection operation, layer normalization operation, feedforward operation, residual connection operation, and layer normalization operation on the input word expression matrix in sequence. The frequency projection attention operation constructs the frequency projection matrix to achieve the division of attention representation in the new and old mode subspaces and enhance the model's adaptability to concept drift. The calculation formula of the frequency projection matrix is as follows:

[0023] P(t)=E(t)α(t)diag(Z1(t) T Z1(t) T (7)

[0024] Wherein, P(t) represents the frequency projection matrix at time t, P(t)=[P 1 (t) T ,...,P m (t) T ,...,P M (t) T ] T , P m (t) represents the mth row vector of the first frequency projection matrix at time t, P m (t)=[P m,1 (t),...,P m,j (t),...,P m,M(t)],P m,j (t) represents the element in the mth row and jth column of the frequency projection matrix at time t, diag(·) represents extracting the diagonal elements of the matrix and outputting them as column vectors, j = 1, 2, …, M. The calculation formulas for the query matrix, key matrix, and value matrix of the sth encoding module are:

[0025] Q s (t) = Z s (t)W Q,s (t) (8)

[0026] K s (t) = Z s (t)W K,s (t) (9)

[0027] V s (t) = Z s (t)W V,s (t) (10)

[0028] Among them, Q s (t) represents the query matrix of the s-th encoding module at time t, represents the mth row vector of the query matrix of the sth encoding module, represents the element in the mth row and ith column of the query matrix of the sth encoding module, W Q,s (t) represents the query weight matrix of the s-th encoding module at time t, represents the i-th row vector of the query weight matrix of the s-th encoding module, K represents the element in row i and column l of the query weight matrix of the s-th encoding module, s (t) represents the key matrix of the s-th encoding module at time t, represents the mth row vector of the sth encoding module key matrix, represents the element in the mth row and ith column of the sth encoding module key matrix, W K,s (t) represents the key weight matrix of the s-th encoding module, represents the i-th row vector of the key weight matrix of the s-th encoding module, Represents the element in row i and column l of the key weight matrix of the sth encoding module, V s (t) represents the value matrix of the s-th coding module at time t, represents the mth row vector of the sth encoding module value matrix, Represents the element in the mth row and ith column of the sth encoding module value matrix, W V,s (t) represents the value weight matrix of the s-th coding module at time t, represents the i-th row vector of the weight matrix of the s-th encoding module value, Represents the element in the i-th row and l-th column of the weight matrix of the s-th encoding module value, l = 1, 2, .., D; the calculation formula for the frequency projection attention output is as follows:

[0029] A 1,s (t) = softmax(P(t)Q s (t)K s (t) T P(t) T ) (11)

[0030] A 2,s (t) = softmax((IP(t))Q s (t)K s (t) T (IP(t)) T ) (12)

[0031] O s (t)=(A 1,s (t)+A 2,s (t))V s (t) (13)

[0032] Among them, A 1,s (t) represents the first attention matrix of the s-th encoding module at time t, represents the mth row vector of the first attention matrix of the sth encoding module at time t, represents the element of the mth row and jth column of the first attention matrix of the sth encoding module at time t, I represents the identity matrix, A 2,s (t) represents the second attention matrix of the s-th encoding module at time t, represents the mth row vector of the second attention matrix of the sth encoding module at time t, represents the element of the mth row and jth column of the second attention matrix of the sth encoding module at time t. Softmax(·) represents the normalization of the input matrix in the row dimension so that the sum of the elements in each row is 1. s (t) represents the attention output matrix of the s-th encoding module at time t, represents the mth row vector of the attention output matrix of the sth encoding module, Represents the element of the mth row and ith column of the attention output matrix of the sth encoding module; the attention output matrix O of the sth encoding module s (t) Perform residual connection operation and layer normalization operation to obtain the residual update output matrix R of the sth encoding module s (t), represents the mth row vector of the residual update output matrix of the sth encoding module, Represents the element in the mth row and ith column of the residual update output matrix of the sth encoding module; the matrix after the residual update is input into the feedforward neural network. The feedforward neural network structure consists of a first linear mapping layer with an input dimension of D and an output dimension of D, and a second linear mapping layer with an input dimension of D and an output dimension of D. A nonlinear activation layer is connected between the two. The nonlinear activation function is a linear rectification function, which is used to reconstruct the dimension and map the word unit expression vector, where the dimension D is equal to the embedding dimension of the word unit expression; the calculation formula of the feedforward neural network output of the sth encoding module is as follows:

[0033]

[0034] in, represents the feedforward network output matrix of the s-th encoding module, represents the mth row vector of the feedforward network output matrix of the sth encoding module, represents the element of the mth row and ith column of the feedforward network output matrix of the sth encoding module, ReLU(·) represents the linear rectification function, W 2,s (t) represents the first linear mapping weight matrix of the s-th encoding module, represents the i-th row vector of the first linear mapping weight matrix of the s-th encoding module, represents the element in row i and column l of the first linear mapping weight matrix of the sth encoding module, b 2,s (t) represents the first linear mapping bias vector of the s-th encoding module, Represents the lth element of the first linear mapping bias vector of the sth encoding module, W 3,s (t) represents the second linear mapping weight matrix of the s-th encoding module, represents the i-th row vector of the second linear mapping weight matrix of the s-th encoding module, represents the element in the i-th row and l-th column of the second linear mapping weight matrix of the s-th encoding module, b 3,s (t) represents the second linear mapping bias vector of the s-th encoding module, Represents the lth element of the first linear mapping bias vector of the sth encoding module, and the feedforward network output matrix of the sth encoding module Perform residual connection operation and layer normalization operation to obtain the word unit expression matrix Z output by the s-th encoding module at time t s+1 (t), represents the mth row vector of the output matrix of the word unit expression of the sth encoding module, Represents the element in the mth row and ith column of the word unit expression output matrix of the sth encoding module;

[0035] Output layer: A linear regression structure is used to map the word-unit expression matrix output by the third encoding module to the final output of the model. The calculation formula of the model output is as follows:

[0036]

[0037] in, represents the output value of the main motor bearing temperature predicted by the model at time t, in degrees Celsius, and W4(t) represents the weight vector of the linear regression of the output layer at time t. represents the i-th item of the linear regression weight vector of the output layer at time t, and b4(t) represents the bias term of linear regression;

[0038] 3. Training the blower bearing temperature prediction model parameters:

[0039] ①Define the loss function of the model as:

[0040]

[0041] Where J(t) represents the loss of the model at time t, and y(t) represents the actual measured main motor bearing temperature value at time t, in degrees Celsius, which is used as a supervisory signal in model training.

[0042] ② Set the current training time to t, initialize the number of training rounds τ = 1, and set the number of training iterations to 20; initialize the weight parameters and bias parameters of the model, with the weight parameters randomly taking values in the interval [-0.2, 0.2] and the bias parameters taking values of 0;

[0043] ③ Use formulas (1)-(15) to calculate the actual output of the model trained for the τth time at time t Use formula (16) to calculate the loss J of the model trained for the τth time at time t τ (t), use the gradient descent method to update the weight matrix and bias matrix, the calculation formula is:

[0044]

[0045] in, Represents the coefficient of the nth row and ith column of the frequency embedding weight matrix for the τth training at time t, represents the coefficient of the ith row and lth column of the sth query weight matrix of the τth training at time t, Represents the coefficient of the mth row and ith column of the sth key weight matrix of the τth training at time t, Represents the coefficient of the mth row and ith column of the sth value weight matrix of the τth training at time t, represents the coefficient of the first linear mapping weight matrix of the s-th encoding module trained for the τth time at time t, represents the coefficient of the second linear mapping weight matrix of the ith row and lth column of the sth encoding module trained for the τth time at time t, represents the i-th coefficient of the output layer linear regression weight vector of the τ-th training at time t, Represents the coefficient of the nth row and ith column of the frequency embedding weight matrix of the τ+1th training at time t, represents the coefficient of the ith row and lth column of the sth query weight matrix of the τ+1th training at time t, Represents the coefficient of the mth row and ith column of the sth key weight matrix of the τ+1th training at time t, Represents the coefficient of the mth row and ith column of the sth value weight matrix of the τ+1th training at time t, represents the coefficient of the first linear mapping weight matrix of the s-th encoding module in the τ+1-th training at time t, represents the coefficient of the second linear mapping weight matrix of the s-th encoding module in the τ+1-th training at time t, represents the i-th coefficient of the output layer linear regression weight vector of the τ+1-th training at time t, Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of represents the i-th frequency embedding bias term of the τ-th training at time t, represents the lth bias term of the first linear mapping of the sth encoding module trained for the τth time, represents the lth bias term of the second linear mapping of the sth encoding module trained for the τth time at time t, represents the bias term of the linear regression of the τth training at time t, represents the i-th frequency embedding bias term of the τ+1-th training at time t, represents the lth bias term of the first linear mapping of the sth encoding module trained at time t, represents the lth bias term of the second linear mapping of the sth encoding module trained at time t, represents the bias term of the linear regression of the τ+1th training at time t, Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes Partial derivatives of τ = 1, 2, …, 20;

[0046] ④ If the number of training rounds τ < 20, τ is increased by 1, and the training is continued in step ③; if the number of training rounds τ ≥ 20, the model parameter training and update are terminated, and the prediction is performed in step 4;

[0047] 4. Predict blower bearing temperature:

[0048] Using the trained blower bearing temperature prediction model, the outlet air pressure data, high-speed shaft drive end bearing temperature data, high-speed shaft vibration value data, inlet air temperature data, main motor current data and historical blower main motor bearing temperature data collected at time t are used as the model input, and the model output value at time t is obtained according to formulas (1)-(15) Model output value The blower bearing temperature at time t as the prediction for the next moment, in degrees Celsius.

[0049] The system implementation structure of the method of the present invention is as follows:

[0050] In order to realize the above-mentioned blower bearing temperature prediction method based on spectrum reconstruction neural network, a modular integrated prediction system is constructed, which includes: industrial sensor module, programmable logic controller (PLC) module, database module, neural network processing module and display module; wherein the industrial sensor module includes a pressure transmitter, a three-axis acceleration sensor, a temperature transmitter, an AC current transformer and two resistance temperature sensors, which are used to collect data such as outlet air pressure, high-speed shaft vibration, inlet air temperature, main motor current, high-speed shaft drive end bearing temperature and main motor bearing temperature during the operation of the blower, and is connected to the PLC module through a signal line; the PLC module is used to receive the incoming The signal from the industrial sensor module is collected and the collected data is sent to the database module and the neural network processing module through the Ethernet interface; the database module and the neural network processing module are communicated and connected through the Ethernet interface to provide historical operation data to the neural network processing module; the neural network processing module is constructed using an industrial control computer, and the industrial control computer is configured with an Ethernet port and a display output interface, which are used to connect the PLC module, the database module and the display module respectively, deploy the spectrum reconstruction neural network model, receive data and perform temperature prediction; the display module is an industrial touch display terminal, which is connected to the neural network processing module using a display output interface to display the temperature prediction results.

[0051] The creativity of the present invention is mainly reflected in:

[0052] 1. This paper addresses the problem of concept drift in input variables during blower operation, which can lead to reduced prediction accuracy. This paper proposes a blower bearing temperature prediction method based on a spectrum reconstruction neural network. This method uses a spectrum reconstruction layer to extract and smooth the input data in the frequency domain, alleviating the impact of concept drift on the model's prediction performance and enabling accurate prediction of the future state of the blower bearing temperature.

[0053] 2. The present invention designs a spectral projection attention mechanism, which uses an approximate orthogonal projection matrix to decompose the query matrix and key matrix in the attention mechanism, adaptively adjusts the weights of different frequency features, enhances the model's ability to recognize new and old operating modes under concept drift, and improves the accuracy and stability of the prediction of the future state of the blower bearing temperature. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a blower bearing temperature prediction model diagram of the present invention;

[0055] Figure 2 This is a graph showing the results of the blower bearing temperature prediction according to the present invention;

[0056] Figure 3 It is an error diagram of the blower bearing temperature prediction according to the present invention. DETAILED DESCRIPTION

[0057] The present invention obtains a blower bearing temperature prediction method based on spectrum reconstruction neural network, which selects outlet air pressure, high-speed shaft drive end bearing temperature, high-speed shaft vibration value, inlet air temperature, main motor current and historical blower main motor bearing temperature as model input variables, and blower main motor bearing temperature as model output variable;

[0058] The experimental data comes from a Howden-Hua blower actually operating in a city sewage treatment plant. Using a Rosemount 3051 pressure transmitter, a WZPK-187 PT100 platinum resistance temperature sensor, an Enerpac VIB-100 triaxial accelerometer, a Honeywell T775U temperature transmitter, and a Schneider LC1D95 AC current transformer, the outlet air pressure, high-speed shaft drive end bearing temperature, high-speed shaft vibration value, inlet air temperature, main motor current data, and main motor bearing temperature data were collected with a sampling period of 1 minute. A total of 14 hours of data, totaling 840 groups of samples, were collected. The data were divided into training and test sets in chronological order, with 420 groups of data in the first 7 hours as training samples and 420 groups of data in the last 7 hours as test samples.

[0059] The present invention adopts the following technical solutions and implementation steps:

[0060] A blower bearing temperature prediction method based on spectrum reconstruction neural network is characterized by: collecting blower data, establishing a blower bearing temperature prediction model, training blower bearing temperature prediction model parameters, and predicting the blower bearing temperature, including the following steps:

[0061] 1. Collect blower data:

[0062] Taking the blower as the research object, the outlet air pressure, high-speed shaft drive end bearing temperature, high-speed shaft vibration value, inlet air temperature, main motor current, and historical blower main motor bearing temperature are selected as the input variables of the blower bearing temperature prediction model, and the blower main motor bearing temperature is used as the output variable of the blower bearing temperature prediction model; among them, the main motor refers to the three-phase asynchronous motor that provides direct driving power for the blower, and its output shaft is connected to the high-speed shaft to drive the entire blower system; the high-speed shaft refers to the high-speed rotating shaft section connecting the main motor output shaft and the impeller transmission system, which is mainly used to transmit the rotational power of the main motor; the outlet air pressure data at time t is collected by using a pressure transmitter installed in the outlet pipe as a vector x 1 (t) = [x 1 (t-L+1),…,x 1 (t-1),x 1 (t)], the resistance temperature sensor arranged on the bearing seat is used to collect the bearing temperature data of the high-speed shaft drive end at time t as vector x2 (t) = [x 2 (t-L+1),…,x 2 (t-1),x 2 (t)], the high-speed shaft vibration value data at time t is collected using a three-axis acceleration sensor as vector x 3 (t) = [x 3 (t-L+1),…,x 3 (t-1),x 3 (t)], the temperature transmitter is used to collect the inlet air temperature data at time t as vector x 4 (t) = [x 4 (t-L+1),…,x 4 (t-1),x 4 (t)], the AC current transformer is used to collect the main motor current data at time t as vector x 5 (t) = [x 5 (t-L+1),…,x 5 (t-1),x 5 (t)], the resistance temperature sensor is used to collect the historical blower main motor bearing temperature data at time t as vector x 6 (t) = [x 6 (t-L+1),…,x 6 (t-1),x 6 (t)], construct the input sample matrix as X(t)=[x 1 (t) T ,…,x n (t) T ,…,x 6 (t) T ] T , x n (t) = [x n (t-L+1),…,x n (t-1),x n (t)], n=1,2,…,6, t=L+1,L+2,…,N+L+1, N is the number of sample sequences, L is the number of samples of the input variable, and T represents the transpose of the matrix;

[0063] 2. Establish a blower bearing temperature prediction model:

[0064] A blower bearing temperature prediction model based on spectrum reconstruction neural network is constructed. The spectrum reconstruction neural network consists of an input layer, a spectrum reconstruction layer, a spectrum projection attention layer, and an output layer.

[0065] Input layer: Input the input sample matrix into the model row by row. The output of the input layer at time t is X(t) = [x 1 (t) T,x 2 (t) T ,…,x 6 (t) T ] T ;

[0066] Spectrum reconstruction layer: Perform real cosine transform on the input sample matrix. Each row of input samples is mapped to a real vector consisting of M frequency projection values. The calculation formula is as follows:

[0067]

[0068] in, represents the mth frequency projection value of the nth row input sample at time t, cos(·) represents the cosine function, F n (t) represents the nth frequency projection vector of the input sample matrix at time t, m = 1, 2, ..., M; the frequency projection vector is smoothed, and the calculation formula of the nth smoothed frequency projection vector is as follows:

[0069]

[0070] Among them, a n (t) represents the weighted cumulative value of the change amplitude of the nth frequency projection vector at time t. The change amplitude refers to the Euclidean distance between the frequency projection vector at the current moment and the previous moment. n (t-1) represents the weighted cumulative value of the change amplitude of the nth frequency projection vector at the moment before time t, F n (t-1) represents the nth frequency projection vector of the input sample matrix at the moment before time t, represents the mth frequency projection value of the nth row of the input sample matrix at the moment before time t, ||·|| represents the Euclidean norm operation of the vector, b n (t) represents the weighted cumulative value of the change stability of the nth frequency projection vector at time t. The change stability refers to the difference between the Euclidean distance between the current frequency projection vector and the previous moment and 1. b n (t-1) represents the weighted cumulative value of the stability of the n-th frequency projection vector change at the moment before time t, represents the smoothed nth frequency projection vector, represents the mth frequency projection value of the nth row input sample at time t after smoothing, α n (t) represents the smoothing coefficient corresponding to the nth frequency projection vector, α(t) = [α 1 (t),α 2 (t),…,α 6 (t)] T, α(t) represents the smoothing coefficient vector at time t; the smoothed frequency projection vectors are combined into a frequency embedding matrix, and a linear transformation operation is performed on the frequency embedding matrix to obtain the word unit expression matrix. The calculation formula is as follows:

[0071] Z1(t)=E(t)W1(t)+b1(t) (24)

[0072] Among them, E(t) represents the frequency embedding matrix at time t, Z1(t) represents the word expression matrix at time t, represents the mth row vector of the word expression matrix, represents the element in the mth row and ith column of the word expression matrix, W1(t) represents the frequency embedding weight matrix at time t, represents the nth row vector of the frequency embedding weight matrix, represents the element in the nth row and ith column of the frequency embedding weight matrix, b1(t) represents the frequency embedding bias vector at time t, represents the i-th frequency embedding bias term, D is the embedding dimension of the word unit expression, i = 1, 2, ..., D;

[0073] Encoding layer: It consists of three stacked encoding modules, where s = 1, 2, 3 represents the number of the sth encoding module. Each encoding module performs frequency projection attention operation, residual connection operation, layer normalization operation, feedforward operation, residual connection operation, and layer normalization operation on the input word expression matrix in sequence. The frequency projection attention operation constructs the frequency projection matrix to achieve the division of attention representation in the new and old mode subspaces and enhance the model's adaptability to concept drift. The calculation formula of the frequency projection matrix is as follows:

[0074] P(t)=E(t)α(t)diag(Z1(t) T Z1(t) T (25)

[0075] Wherein, P(t) represents the frequency projection matrix at time t, P(t)=[P 1 (t) T ,...,P m (t) T ,...,P M (t) T ] T , P m (t) represents the mth row vector of the first frequency projection matrix at time t, P m (t)=[P m,1 (t),...,P m,j (t),...,P m,M (t)],P m,j(t) represents the element in the mth row and jth column of the frequency projection matrix at time t, diag(·) represents extracting the diagonal elements of the matrix and outputting them as column vectors, j = 1, 2, …, M. The calculation formulas for the query matrix, key matrix, and value matrix of the sth encoding module are:

[0076] Q s (t) = Z s (t)W Q,s (t) (26)

[0077] K s (t) = Z s (t)W K,s (t) (27)

[0078] V s (t) = Z s (t)W V,s (t) (28)

[0079] Among them, Q s (t) represents the query matrix of the s-th encoding module at time t, represents the mth row vector of the query matrix of the sth encoding module, represents the element in the mth row and ith column of the query matrix of the sth encoding module, W Q,s (t) represents the query weight matrix of the s-th encoding module at time t, represents the i-th row vector of the query weight matrix of the s-th encoding module, K represents the element in row i and column l of the query weight matrix of the s-th encoding module, s (t) represents the key matrix of the s-th encoding module at time t, represents the mth row vector of the sth encoding module key matrix, represents the element in the mth row and ith column of the sth encoding module key matrix, W K,s (t) represents the key weight matrix of the s-th encoding module, represents the i-th row vector of the key weight matrix of the s-th encoding module, Represents the element in row i and column l of the key weight matrix of the sth encoding module, V s (t) represents the value matrix of the s-th coding module at time t, represents the mth row vector of the sth encoding module value matrix, Represents the element in the mth row and ith column of the sth encoding module value matrix, W V,s (t) represents the value weight matrix of the s-th coding module at time t, represents the i-th row vector of the weight matrix of the s-th encoding module value, Represents the element in the i-th row and l-th column of the weight matrix of the s-th encoding module value, l = 1, 2, .., D; the calculation formula for the frequency projection attention output is as follows:

[0080]

[0081] Among them, A 1,s (t) represents the first attention matrix of the s-th encoding module at time t, represents the mth row vector of the first attention matrix of the sth encoding module at time t, represents the element of the mth row and jth column of the first attention matrix of the sth encoding module at time t, I represents the identity matrix, A 2,s (t) represents the second attention matrix of the s-th encoding module at time t, represents the mth row vector of the second attention matrix of the sth encoding module at time t, represents the element of the mth row and jth column of the second attention matrix of the sth encoding module at time t. Softmax(·) represents the normalization of the input matrix in the row dimension so that the sum of the elements in each row is 1. s (t) represents the attention output matrix of the s-th encoding module at time t, represents the mth row vector of the attention output matrix of the sth encoding module, Represents the element of the mth row and ith column of the attention output matrix of the sth encoding module; the attention output matrix O of the sth encoding module s (t) Perform residual connection operation and layer normalization operation to obtain the residual update output matrix R of the sth encoding module s (t), represents the mth row vector of the residual update output matrix of the sth encoding module,

[0082] Represents the element in the mth row and ith column of the residual update output matrix of the sth encoding module; the matrix after the residual update is input into the feedforward neural network. The feedforward neural network structure consists of a first linear mapping layer with an input dimension of D and an output dimension of D, and a second linear mapping layer with an input dimension of D and an output dimension of D. A nonlinear activation layer is connected between the two. The nonlinear activation function is a linear rectification function, which is used to reconstruct the dimension and map the word unit expression vector, where the dimension D is equal to the embedding dimension of the word unit expression; the calculation formula of the feedforward neural network output of the sth encoding module is as follows:

[0083]

[0084] in, represents the feedforward network output matrix of the s-th encoding module, represents the mth row vector of the feedforward network output matrix of the sth encoding module,

[0085] represents the element of the mth row and ith column of the feedforward network output matrix of the sth encoding module, ReLU(·) represents the linear rectification function, W 2,s (t) represents the first linear mapping weight matrix of the s-th encoding module, represents the i-th row vector of the first linear mapping weight matrix of the s-th encoding module, represents the element in row i and column l of the first linear mapping weight matrix of the sth encoding module, b 2,s (t) represents the first linear mapping bias vector of the s-th encoding module, Represents the lth element of the first linear mapping bias vector of the sth encoding module, W 3,s (t) represents the second linear mapping weight matrix of the s-th encoding module,

[0086] represents the i-th row vector of the second linear mapping weight matrix of the s-th encoding module, represents the element in the i-th row and l-th column of the second linear mapping weight matrix of the s-th encoding module, b 3,s (t) represents the second linear mapping bias vector of the s-th encoding module,

[0087] Represents the lth element of the first linear mapping bias vector of the sth encoding module, and the feedforward network output matrix of the sth encoding module Perform residual connection operation and layer normalization operation to obtain the word unit expression matrix Z output by the s-th encoding module at time t s+1 (t),

[0088] represents the mth row vector of the output matrix of the word unit expression of the sth encoding module, Represents the element in the mth row and ith column of the word unit expression output matrix of the sth encoding module;

[0089] Output layer: A linear regression structure is used to map the word-unit expression matrix output by the third encoding module to the final output of the model. The calculation formula of the model output is as follows:

[0090]

[0091] in, represents the output value of the main motor bearing temperature predicted by the model at time t, in degrees Celsius, and W4(t) represents the weight vector of the linear regression of the output layer at time t. represents the i-th item of the linear regression weight vector of the output layer at time t, and b4(t) represents the bias term of linear regression;

[0092] 3. Training the blower bearing temperature prediction model parameters:

[0093] ①Define the loss function of the model as:

[0094]

[0095] Where J(t) represents the loss of the model at time t, and y(t) represents the actual measured main motor bearing temperature value at time t, in degrees Celsius, which is used as a supervisory signal in model training.

[0096] ② Set the current training time to t, initialize the number of training rounds τ = 1, and set the number of training iterations to 20; initialize the weight parameters and bias parameters of the model, with the weight parameters randomly taking values in the interval [-0.2, 0.2] and the bias parameters taking values of 0;

[0097] ③ Use formulas (19)-(33) to calculate the actual output of the model trained for the τth time at time t Use formula (34) to calculate the loss J of the model trained for the τth time at time t τ (t), use the gradient descent method to update the weight matrix and bias matrix, the calculation formula is:

[0098]

[0099] in, Represents the coefficient of the nth row and ith column of the frequency embedding weight matrix for the τth training at time t, represents the coefficient of the ith row and lth column of the sth query weight matrix of the τth training at time t, Represents the coefficient of the mth row and ith column of the sth key weight matrix of the τth training at time t, Represents the coefficient of the mth row and ith column of the sth value weight matrix of the τth training at time t, represents the coefficient of the first linear mapping weight matrix of the s-th encoding module trained for the τth time at time t, represents the coefficient of the second linear mapping weight matrix of the ith row and lth column of the sth encoding module trained for the τth time at time t, represents the i-th coefficient of the output layer linear regression weight vector of the τ-th training at time t, Represents the coefficient of the nth row and ith column of the frequency embedding weight matrix of the τ+1th training at time t, represents the coefficient of the ith row and lth column of the sth query weight matrix of the τ+1th training at time t, Represents the coefficient of the mth row and ith column of the sth key weight matrix of the τ+1th training at time t, Represents the coefficient of the mth row and ith column of the sth value weight matrix of the τ+1th training at time t, represents the coefficient of the first linear mapping weight matrix of the s-th encoding module in the τ+1-th training at time t, represents the coefficient of the second linear mapping weight matrix of the s-th encoding module in the τ+1-th training at time t, represents the i-th coefficient of the output layer linear regression weight vector of the τ+1-th training at time t, Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of represents the i-th frequency embedding bias term of the τ-th training at time t, represents the lth bias term of the first linear mapping of the sth encoding module trained for the τth time, represents the lth bias term of the second linear mapping of the sth encoding module trained for the τth time at time t, represents the bias term of the linear regression of the τth training at time t, represents the i-th frequency embedding bias term of the τ+1-th training at time t, represents the lth bias term of the first linear mapping of the sth encoding module trained at time t, represents the lth bias term of the second linear mapping of the sth encoding module trained at time t, represents the bias term of the linear regression of the τ+1th training at time t, Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes Partial derivatives of τ = 1, 2, …, 20;

[0100] ④ If the number of training rounds τ < 20, τ is increased by 1, and the training is continued in step ③; if the number of training rounds τ ≥ 20, the model parameter training and update are terminated, and the prediction is performed in step 4;

[0101] 4. Predict blower bearing temperature:

[0102] Using the trained blower bearing temperature prediction model, the outlet air pressure data, high-speed shaft drive end bearing temperature data, high-speed shaft vibration value data, inlet air temperature data, main motor current data and historical blower main motor bearing temperature data collected at time t are used as the model input, and the model output value at time t is obtained according to formulas (19)-(33) Model output value The blower bearing temperature at time t as the prediction for the next moment, in degrees Celsius.

[0103] The specific implementation examples of the system deployment structure of the present invention are as follows:

[0104] In actual engineering applications, a modular integrated prediction system can be constructed to deploy the blower bearing temperature prediction method based on spectrum reconstruction neural network described in the present invention. The system includes multiple functional modules that work together: an industrial sensor module, a PLC module, a database module, a neural network processing module, and a display module. The industrial sensor module consists of a pressure transmitter, a triaxial accelerometer, a temperature transmitter, an AC current transformer, and two resistance temperature sensors. It is used to collect real-time data on multiple variables during blower operation, including outlet air pressure, high-speed shaft vibration, inlet air temperature, main motor current, and high-speed shaft drive-end bearing temperature and main motor bearing temperature. These sensors are connected to the PLC module via signal cables to complete standardized data collection and initial transmission. The PLC module synchronously transmits the collected operating data to the database module and the neural network processing module via an Ethernet interface. The database module is used to construct a time series historical dataset, providing a data foundation for subsequent predictions. The neural network processing module is built using an industrial control computer and integrates a spectrum reconstruction neural network model to perform intelligent predictions based on combined database and PLC data. The industrial control computer is equipped with a standard Ethernet port and display output interface, which are connected to the database module, PLC module, and display module respectively. Finally, the prediction results are displayed on an industrial-grade touch display terminal, enabling visual presentation and operation and maintenance support.

Claims

1. A blower bearing temperature prediction method based on spectrum reconstruction neural network, characterized by: Collecting blower data, establishing a blower bearing temperature prediction model, training blower bearing temperature prediction model parameters, and predicting the blower bearing temperature include the following steps: (1) Collect blower data: Taking the blower as the research object, the outlet air pressure, high-speed shaft drive end bearing temperature, high-speed shaft vibration value, inlet air temperature, main motor current, and historical blower main motor bearing temperature are selected as the input variables of the blower bearing temperature prediction model, and the blower main motor bearing temperature is used as the output variable of the blower bearing temperature prediction model; among them, the main motor refers to the three-phase asynchronous motor that provides direct driving power for the blower, and its output shaft is connected to the high-speed shaft to drive the entire blower system; the high-speed shaft refers to the high-speed rotating shaft section connecting the main motor output shaft and the impeller transmission system, which is mainly used to transmit the rotational power of the main motor; the outlet air pressure data at time t is collected by using a pressure transmitter installed in the outlet pipe as a vector x 1 (t) = [x 1 (t-L+1),…,x 1 (t-1),x 1 (t)], the resistance temperature sensor arranged on the bearing seat is used to collect the bearing temperature data of the high-speed shaft drive end at time t as vector x 2 (t) = [x 2 (t-L+1),…,x 2 (t-1),x 2 (t)], the high-speed shaft vibration value data at time t is collected using a three-axis acceleration sensor as vector x 3 (t) = [x 3 (t-L+1),…,x 3 (t-1),x 3 (t)], the temperature transmitter is used to collect the inlet air temperature data at time t as vector x 4 (t) = [x 4 (t-L+1),…,x 4 (t-1),x 4 (t)], the AC current transformer is used to collect the main motor current data at time t as vector x 5 (t) = [x 5 (t-L+1),…,x 5 (t-1),x 5 (t)], the resistance temperature sensor is used to collect the historical blower main motor bearing temperature data at time t as vector x 6 (t) = [x 6 (t-L+1),…,x 6 (t-1),x 6 (t)], construct the input sample matrix as X(t)=[x 1 (t) T ,…,x n (t) T ,…,x 6 (t) T ] T , x n (t) = [x n (t-L+1),…,x n (t-1),x n (t)], n=1,2,…,6, t=L+1,L+2,…,N+L+1, N is the number of sample sequences, L is the number of samples of the input variable, and T represents the transpose of the matrix; (2) Establish a blower bearing temperature prediction model: A blower bearing temperature prediction model based on spectrum reconstruction neural network is constructed. The spectrum reconstruction neural network consists of an input layer, a spectrum reconstruction layer, a spectrum projection attention layer, and an output layer. Input layer: Input the input sample matrix into the model row by row. The output of the input layer at time t is X(t) = [x 1 (t) T ,x 2 (t) T ,…,x 6 (t) T ] T ; Spectrum reconstruction layer: Perform real cosine transform on the input sample matrix. Each row of input samples is mapped to a real vector consisting of M frequency projection values. The calculation formula is as follows: in, represents the mth frequency projection value of the nth row input sample at time t, cos(·) represents the cosine function, F n (t) represents the nth frequency projection vector of the input sample matrix at time t, m = 1, 2, ..., M; the frequency projection vector is smoothed, and the calculation formula of the nth smoothed frequency projection vector is as follows: a n (t)=0.9a n (t-1)+0.1||F n (t)-F n (t-1)|| (2) b n (t)=0.9b n (t-1)+0.1(1-||F n (t)-F n (t-1)||) (3) Among them, a n (t) represents the weighted cumulative value of the change amplitude of the nth frequency projection vector at time t. The change amplitude refers to the Euclidean distance between the frequency projection vector at the current moment and the previous moment. n (t-1) represents the weighted cumulative value of the change amplitude of the nth frequency projection vector at the moment before time t, F n (t-1) represents the nth frequency projection vector of the input sample matrix at the moment before time t, represents the mth frequency projection value of the nth row of the input sample matrix at the moment before time t, ||·|| represents the Euclidean norm operation of the vector, b n (t) represents the weighted cumulative value of the change stability of the nth frequency projection vector at time t. The change stability refers to the difference between the Euclidean distance between the current frequency projection vector and the previous moment and 1. b n (t-1) represents the weighted cumulative value of the stability of the n-th frequency projection vector change at the moment before time t, represents the smoothed nth frequency projection vector, represents the mth frequency projection value of the nth row input sample at time t after smoothing, α n (t) represents the smoothing coefficient corresponding to the nth frequency projection vector, α(t) = [α 1 (t),α 2 (t),…,α 6 (t)] T , α(t) represents the smoothing coefficient vector at time t; the smoothed frequency projection vectors are combined into a frequency embedding matrix, and a linear transformation operation is performed on the frequency embedding matrix to obtain the word unit expression matrix. The calculation formula is as follows: Z1(t)=E(t)W1(t)+b1(t) (6) Among them, E(t) represents the frequency embedding matrix at time t, Z1(t) represents the word expression matrix at time t, represents the mth row vector of the word expression matrix, represents the element in the mth row and ith column of the word expression matrix, W1(t) represents the frequency embedding weight matrix at time t, represents the nth row vector of the frequency embedding weight matrix, W1 n,i (t) represents the element in the nth row and ith column of the frequency embedding weight matrix, b1(t) represents the frequency embedding bias vector at time t, represents the i-th frequency embedding bias term, D is the embedding dimension of the word unit expression, i = 1, 2, ..., D; Encoding layer: It consists of three stacked encoding modules, where s = 1, 2, 3 represents the number of the sth encoding module. Each encoding module performs frequency projection attention operation, residual connection operation, layer normalization operation, feedforward operation, residual connection operation, and layer normalization operation on the input word expression matrix in sequence. The frequency projection attention operation constructs the frequency projection matrix to achieve the division of attention representation in the new and old mode subspaces and enhance the model's adaptability to concept drift. The calculation formula of the frequency projection matrix is as follows: P(t)=E(t)α(t)diag(Z1(t) T Z1(t)) T (7) Wherein, P(t) represents the frequency projection matrix at time t, P(t)=[P 1 (t) T ,...,P m (t) T ,...,P M (t) T ] T , P m (t) represents the mth row vector of the first frequency projection matrix at time t, P m (t)=[P m,1 (t),...,P m,j (t),...,P m,M (t)],P m,j (t) represents the element in the mth row and jth column of the frequency projection matrix at time t, diag(·) represents extracting the diagonal elements of the matrix and outputting them as column vectors, j = 1, 2, …, M. The calculation formulas for the query matrix, key matrix, and value matrix of the sth encoding module are: Q s (t)=Z s (t)W Q,s (t) (8) K s (t)=Z s (t)W K,s (t) (9) V s (t)=Z s (t)W V,s (t) (10) Among them, Q s (t) represents the query matrix of the s-th encoding module at time t, represents the mth row vector of the query matrix of the sth encoding module, represents the element in the mth row and ith column of the query matrix of the sth encoding module, W Q,s (t) represents the query weight matrix of the s-th encoding module at time t, represents the i-th row vector of the query weight matrix of the s-th encoding module, represents the element in row i and column l of the query weight matrix of the s-th encoding module, K s (t) represents the key matrix of the s-th encoding module at time t, represents the mth row vector of the sth encoding module key matrix, represents the element in the mth row and ith column of the sth encoding module key matrix, W K,s (t) represents the key weight matrix of the s-th encoding module, represents the i-th row vector of the key weight matrix of the s-th encoding module, Represents the element in row i and column l of the key weight matrix of the sth encoding module, V s (t) represents the value matrix of the s-th coding module at time t, represents the mth row vector of the sth encoding module value matrix, V s m,i (t) represents the element of the mth row and ith column of the sth coding module value matrix, W V,s (t) represents the value weight matrix of the s-th coding module at time t, represents the i-th row vector of the weight matrix of the s-th encoding module value, Represents the element in the i-th row and l-th column of the weight matrix of the s-th encoding module value, l = 1, 2, .., D; the calculation formula for the frequency projection attention output is as follows: A 1,s (t)=softmax(P(t)Q s (t)K s (t) T P(t) T ) (11) A 2,s (t)=softmax((I-P(t))Q s (t)K s (t) T (I-P(t)) T ) (12) O s (t)=(A 1,s (t)+A 2,s (t))V s (t) (13) Among them, A 1,s (t) represents the first attention matrix of the s-th encoding module at time t, Represents the mth row vector of the first attention matrix of the sth encoding module at time t represents the element of the mth row and jth column of the first attention matrix of the sth encoding module at time t, I represents the identity matrix, A 2,s (t) represents the second attention matrix of the s-th encoding module at time t, represents the mth row vector of the second attention matrix of the sth encoding module at time t, represents the element of the mth row and jth column of the second attention matrix of the sth encoding module at time t. Softmax(·) represents the normalization of the input matrix in the row dimension so that the sum of the elements in each row is 1. s (t) represents the attention output matrix of the s-th encoding module at time t, represents the mth row vector of the attention output matrix of the sth encoding module, Represents the element of the mth row and ith column of the attention output matrix of the sth encoding module; the attention output matrix O of the sth encoding module s (t) Perform residual connection operation and layer normalization operation to obtain the residual update output matrix R of the sth encoding module s (t), represents the mth row vector of the residual update output matrix of the sth encoding module, Represents the element in the mth row and ith column of the residual update output matrix of the sth encoding module; the matrix after the residual update is input into the feedforward neural network. The feedforward neural network structure consists of a first linear mapping layer with an input dimension of D and an output dimension of D, and a second linear mapping layer with an input dimension of D and an output dimension of D. A nonlinear activation layer is connected between the two. The nonlinear activation function is a linear rectification function, which is used to reconstruct the dimension and map the word unit expression vector, where the dimension D is equal to the embedding dimension of the word unit expression; the calculation formula of the feedforward neural network output of the sth encoding module is as follows: in, represents the feedforward network output matrix of the s-th encoding module, represents the mth row vector of the feedforward network output matrix of the sth encoding module, represents the element of the mth row and ith column of the feedforward network output matrix of the sth encoding module, ReLU(·) represents the linear rectification function, W 2,s (t) represents the first linear mapping weight matrix of the s-th encoding module, represents the i-th row vector of the first linear mapping weight matrix of the s-th encoding module, represents the element in the i-th row and l-th column of the first linear mapping weight matrix of the s-th encoding module, b 2,s (t) represents the first linear mapping bias vector of the s-th encoding module, represents the lth element of the first linear mapping bias vector of the sth encoding module, W 3,s (t) represents the second linear mapping weight matrix of the s-th encoding module, represents the i-th row vector of the second linear mapping weight matrix of the s-th encoding module, represents the element in the i-th row and l-th column of the second linear mapping weight matrix of the s-th encoding module, b 3,s (t) represents the second linear mapping bias vector of the s-th encoding module, Represents the lth element of the first linear mapping bias vector of the sth encoding module, and the feedforward network output matrix of the sth encoding module Perform residual connection operation and layer normalization operation to obtain the word unit expression matrix Z output by the s-th encoding module at time t s+1 (t), represents the mth row vector of the output matrix of the word unit expression of the sth encoding module, Represents the element in the mth row and ith column of the word unit expression output matrix of the sth encoding module; Output layer: A linear regression structure is used to map the word-unit expression matrix output by the third encoding module to the final output of the model. The calculation formula of the model output is as follows: in, represents the output value of the main motor bearing temperature predicted by the model at time t, in degrees Celsius, and W4(t) represents the weight vector of the linear regression of the output layer at time t. represents the i-th item of the linear regression weight vector of the output layer at time t, and b4(t) represents the bias term of linear regression; (3) Training the blower bearing temperature prediction model parameters: (4) Predicting blower bearing temperature: Using the trained blower bearing temperature prediction model, the outlet air pressure data, high-speed shaft drive end bearing temperature data, high-speed shaft vibration value data, inlet air temperature data, main motor current data and historical blower main motor bearing temperature data collected at time t are used as the input of the model to obtain the model output value at time t. Model output value The blower bearing temperature at time t as the prediction for the next moment, in degrees Celsius.

2. The blower bearing temperature prediction method according to claim 1, wherein: ①Define the loss function of the model as: Where J(t) represents the loss of the model at time t, and y(t) represents the actual measured main motor bearing temperature value at time t, in degrees Celsius, which is used as a supervisory signal in model training. ② Set the current training time to t, initialize the number of training rounds τ = 1, and set the number of training iterations to 20; initialize the weight parameters and bias parameters of the model, with the weight parameters randomly taking values in the interval [-0.2, 0.2] and the bias parameters taking values of 0; ③ Use formulas (1)-(15) to calculate the actual output of the model trained for the τth time at time t Use formula (16) to calculate the loss J of the model trained for the τth time at time t τ (t), use the gradient descent method to update the weight matrix and bias matrix, the calculation formula is: Among them, W1 n,i,τ (t) represents the coefficient of the nth row and ith column of the frequency embedding weight matrix of the τth training at time t, represents the coefficient of the ith row and lth column of the sth query weight matrix of the τth training at time t, Represents the coefficient of the mth row and ith column of the sth key weight matrix of the τth training at time t, Represents the coefficient of the mth row and ith column of the sth value weight matrix of the τth training at time t, represents the coefficient of the first linear mapping weight matrix of the s-th encoding module trained for the τth time at time t, represents the coefficient of the second linear mapping weight matrix of the ith row and lth column of the sth encoding module trained for the τth time at time t, Represents the i-th coefficient of the output layer linear regression weight vector of the τ-th training at time t, W1 n,i,τ+1 (t) represents the coefficient of the nth row and ith column of the frequency embedding weight matrix of the τ+1th training at time t, represents the coefficient of the ith row and lth column of the sth query weight matrix of the τ+1th training at time t, Represents the coefficient of the mth row and ith column of the sth key weight matrix of the τ+1th training at time t, Represents the coefficient of the mth row and ith column of the sth value weight matrix of the τ+1th training at time t, represents the coefficient of the first linear mapping weight matrix of the s-th encoding module in the τ+1-th training at time t, represents the coefficient of the second linear mapping weight matrix of the s-th encoding module in the τ+1-th training at time t, represents the i-th coefficient of the output layer linear regression weight vector of the τ+1-th training at time t, Indicates J τ (t) for W1 n,i,τ The partial derivative of (t), Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of represents the i-th frequency embedding bias term of the τ-th training at time t, represents the lth bias term of the first linear mapping of the sth encoding module trained for the τth time, represents the lth bias term of the second linear mapping of the sth encoding module trained for the τth time at time t, represents the bias term of the linear regression of the τth training at time t, represents the i-th frequency embedding bias term of the τ+1-th training at time t, represents the lth bias term of the first linear mapping of the sth encoding module trained at time t, represents the lth bias term of the second linear mapping of the sth encoding module trained at time t, represents the bias term of the linear regression of the τ+1th training at time t, Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes The partial derivative of Indicates J τ (t)Yes Partial derivatives of τ = 1, 2, …, 20; ④ If the number of training rounds τ<20, τ is increased by 1, and the training is continued in step ③; if the number of training rounds τ≥20, the model parameter training and update are terminated, and the prediction is performed in step (4).

3. The blower bearing temperature prediction method according to claim 1, wherein: The method is implemented through a modular integrated prediction system, which includes an industrial sensor module, a programmable logic controller module, a database module, a neural network processing module and a display module; wherein the industrial sensor module includes a pressure transmitter, a three-axis acceleration sensor, a temperature transmitter, an AC current transformer and two resistance temperature sensors, which are used to collect the outlet air pressure, high-speed shaft vibration, inlet air temperature, main motor current, high-speed shaft drive end bearing temperature and main motor bearing temperature during the operation of the blower, and is connected to the programmable logic controller module through a signal line; the programmable logic controller module is used to receive signals from the industrial sensor module and transmit the collected signals through an Ethernet interface The data is sent to the database module and the neural network processing module; the database module and the neural network processing module are communicatively connected via an Ethernet interface, and are used to provide historical operation data to the neural network processing module; the neural network processing module is constructed using an industrial control computer, and the industrial control computer is configured with an Ethernet interface and a display output interface, which are used to connect to the programmable logic controller module, the database module and the display module respectively, deploy a spectrum reconstruction neural network model, receive data provided by the database module and the programmable logic controller module and perform temperature prediction; the display module is an industrial touch display terminal, which is connected to the neural network processing module via a display output interface and is used to display temperature prediction results.

Images