Bearing early fault prediction method and system based on BP neural network

Abstract

The invention relates to the field of mechanical engineering, in particular to a bearing early-stage fault prediction method and system based on a BP neural network, and the method comprises the steps: collecting multi-source data; extracting characteristic parameters; constructing and training a BP neural network model; inputting the input feature vector into a pre-trained BP neural network model; the output of the BP neural network model is a bearing fault characteristic frequency amplitude and a vibration energy prediction value at a future moment; fault prediction and early warning: processing data collected in real time and inputting the processed data into the trained BP neural network model to obtain a predicted future state value; the method has the advantages that the output of the BP neural network model is defined as the fault feature state at the future moment, the model learns the mapping relation from the current feature to the future state, early warning can be given out at the early stage of fault germination and when serious damage is not caused by monitoring the deterioration trend of the predicted value, the hysteresis of traditional threshold value alarm is broken through, and the fault detection accuracy is improved. And a sufficient time window is provided for preventive maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical engineering, and in particular to a bearing early fault prediction method and system based on a BP neural network. BACKGROUND

[0002] Rolling bearings are one of the most critical and easily damaged components in rotating machinery. If early faults such as micro-cracks and pitting corrosion are not detected in time, they may eventually lead to a series of chain damages in the equipment, causing huge economic losses and safety accidents.

[0003] Currently, traditional fault diagnosis methods mainly rely on vibration signal analysis, identifying the amplitude changes of bearing fault characteristic frequencies (such as inner ring fault characteristic frequency BPFI, outer ring fault characteristic frequency BPFO, and rolling element fault characteristic frequency BSF) through spectral analysis techniques. However, this method has obvious limitations:

[0004] 1. Hysteresis:

[0005] Traditional vibration analysis can only be effectively detected when the bearing fault has developed to a certain extent and produced significant vibration energy, making it impossible to predict the evolution trend of the fault and achieve true early warning.

[0006] 2. Dependence on expert experience:

[0007] Often requires professional interpretation of spectral graphs, with low automation and difficulty in meeting the needs of automated and intelligent monitoring of equipment health status in large-scale industrial Internet of Things scenarios.

[0008] 3. Unreliability of single signal source:

[0009] Relying solely on vibration signals or current signals for judgment can lead to misjudgment or missed judgment under complex working conditions and noise interference. For example, current signals can reflect the overall load changes of the motor, but are not sensitive to early micro-faults of bearings; while high-frequency vibration signals are sensitive to early faults, they are easily affected by sensor installation location, signal transmission path, and environmental noise. SUMMARY

[0010] The purpose of the present application is to provide a bearing early fault prediction method and system based on a BP neural network, which can achieve early, automatic, and accurate prediction of bearing faults. By fusing the overall trend of motor current and the detailed features of high-frequency vibration of edge-side bearings, the BP neural network model learns the evolution law of bearing faults, thereby issuing an early warning before the fault occurs.

[0011] To achieve the above purpose, the present application realizes the following technical solutions:

[0012] A bearing early fault prediction method based on BP neural network, comprising:

[0013] S1, multi-source data acquisition:

[0014] Synchronously collecting vibration signals of the motor bearing and current signals of the motor;

[0015] S2, feature parameter extraction:

[0016] At least the characteristic parameters including root mean square value, kurtosis and high harmonic are extracted from the vibration signal, and the effective value of the current signal is used to form an input feature vector;

[0017] S3, constructing and training a BP neural network model;

[0018] The input feature vector is input into the pre-trained BP neural network model;

[0019] The output of the BP neural network model is the bearing fault characteristic frequency amplitude and vibration energy prediction value at the future time:

[0020] S4, fault prediction and early warning:

[0021] The real-time collected data is input into the trained BP neural network model after S2 processing to obtain the predicted future state value;

[0022] The early warning is triggered by analyzing the change trend of the future state value.

[0023] In S2, the input feature vector is as follows:

[0024] (1);

[0025] (2);

[0026] (3);

[0027] (4);

[0028] In formula (1), The root mean square value of the vibration signal is represented, and the unit is m / s 2 ; The kurtosis of the vibration signal is represented, The high harmonic energy of the vibration signal is represented, and the unit is m / s 2 ; The effective value of the motor current is represented, and the unit is A; The input feature vector is represented;

[0029] In formula (2), denotes the discrete vibration acceleration signal value, unit: m / s 2 ; N denotes the number of data points in the window;

[0030] In formula (3), denotes the signal mean value, unit: m / s 2 ;

[0031] In formula (4),

[0032] A(f) denotes the amplitude at frequency f in the spectrum, f is the bearing inner ring fault characteristic frequency , outer ring fault characteristic frequency , rolling element fault characteristic frequency , and M is the number of harmonics considered;

[0033] The effective value of motor current in the same time window is taken as one of the input features to introduce load information.

[0034] In S3, a three-layer BP neural network model including input layer, hidden layer and output layer is constructed;

[0035] The input layer includes four feature parameters, respectively the root mean square value of the vibration signal , the kurtosis of the vibration signal , the high harmonic energy of the vibration signal , and the effective value of the motor current ;

[0036] The output of the jth node of the hidden layer is calculated as follows:

[0037] (5);

[0038] (6);

[0039] In formula (5), formula (6), denotes the input of the ith node of the input layer, denotes the connection weight from the ith node of the input layer to the jth node of the hidden layer, denotes the bias of the jth node of the hidden layer, denotes the net input of the jth node of the hidden layer;

[0040] The hidden layer adopts Sigmoid activation function for nonlinear transformation, and the formula is as follows:

[0041] (7);

[0042] In formula (7), denotes the activation function, This represents the input to the activation function;

[0043] The output layer predicts the future fault characteristic frequency amplitude and vibration energy using the following formula:

[0044] (8);

[0045] The output of the kth node in the output layer The calculation is as follows:

[0046] (9);

[0047] (10);

[0048] In formula (9) and formula (10), =1,2,3,4; This represents the connection weight from the j-th node in the hidden layer to the k-th node in the output layer. This represents the bias of the k-th node in the output layer. This represents the net input of the k-th node in the output layer;

[0049] In formula (8), the output of the first node of the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing inner ring failure, in m / s. 2 Output of the second node in the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing outer ring failure, in m / s. 2 Output of the 3rd node in the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing rolling element failure, in m / s. 2 Output of the 4th node in the output layer ,Right now This represents the predicted root mean square value of vibration acceleration, which is the predicted vibration energy value in m / s². 2 .

[0050] In S4, the trend of future state values ​​is analyzed, including:

[0051] like It shows a continuous upward trend. and If the trend is stable, the prediction is a failure in the inner ring of the bearing.

[0052] like It shows a continuous upward trend. and If the trend is stable, the prediction is a failure in the outer ring of the bearing;

[0053] like It shows a continuous upward trend. and If the trend is stable, it is predicted to be a bearing rolling element failure;

[0054] When the trend slope increases or the predicted vibration energy exceeds the safety threshold, an early fault warning notification will be automatically issued.

[0055] In S3, the BP neural network model is trained using historical data of the bearing's life cycle from normal to failure, and the prediction error is minimized through the error backpropagation algorithm.

[0056] A bearing early fault prediction system based on BP neural network includes a data acquisition module, an edge early warning terminal, and an early warning system platform server;

[0057] The data acquisition module includes a vibration sensor for acquiring bearing vibration signals and a current transmitter for acquiring motor current signals.

[0058] The edge warning terminal is connected to the data acquisition module to receive vibration and current signals and perform the following operations:

[0059] Extract input feature vectors, call a pre-trained BP neural network model for fault prediction, and trigger early warning based on trend analysis of the prediction results;

[0060] The early warning system platform server communicates with the edge early warning terminal to receive and display early warning information.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] 1. The output of the BP neural network model is defined as the fault characteristic state at a future time, so that the model learns the mapping relationship from "current characteristics" to "future state". By monitoring the deterioration trend of the predicted value, it can issue an early warning in the early stage of fault occurrence and before serious damage occurs, which breaks through the lag of traditional threshold alarms and provides a sufficient time window for preventive maintenance.

[0063] 2. By integrating the macroscopic load information of motor current with the microscopic details of bearing high-frequency vibration, a multi-dimensional sensing system was constructed. This multi-source information fusion mechanism effectively overcomes the defect of a single signal source being susceptible to interference under complex working conditions. By comprehensively utilizing the complementarity of different signals, the false alarm and false alarm rates were significantly reduced.

[0064] 3. The entire process, from feature extraction and trend prediction to fault diagnosis, is automatically completed by the BP neural network model and algorithm, without relying on manual spectrum analysis by professionals. This not only reduces the dependence on expert experience, but also enables the solution to be efficiently deployed in large-scale industrial IoT scenarios, realizing the automation and intelligence of equipment health management.

[0065] 4. By monitoring the independent prediction trends of different fault characteristic frequencies (such as BPFI, BPFO, BSF) output by the model, it is possible to initially identify and distinguish potential fault locations (bearing inner ring, bearing outer ring, or bearing rolling elements). This provides maintenance personnel with richer and more accurate decision-making information, realizing a leap from simple judgment of "whether there is a fault" to fine diagnosis of "what kind of fault".

[0066] 5. By introducing nonlinear activation functions such as Sigmoid into the hidden and output layers of the neural network, the model's ability to fit complex nonlinear relationships is enhanced. At the same time, this structure helps stabilize the training process and improves the model's adaptability to unseen working condition data, thereby ensuring the robustness and generalization of the system in practical applications. Attached Figure Description

[0067] Figure 1 This is a structural diagram of the BP neural network model for an early bearing failure prediction system. Detailed Implementation

[0068] The present invention will now be described in detail with reference to the accompanying drawings, but it should be noted that the implementation of the present invention is not limited to the following embodiments.

[0069] The following embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified, the methods used in the following embodiments are conventional methods.

[0070] Example 1

[0071] A method and system for predicting early bearing failures based on BP neural networks, see [link to relevant documentation]. Figure 1 Using readily available motor current signals as the macroscopic background and sensitive features extracted from high-frequency vibration signals collected from the edge side as microscopic fault evidence, a BP neural network model is trained together. This model outputs not a simple "normal / fault" classification, but rather the future development trend of fault feature frequency components and vibration energy. By continuously monitoring the output of the BP neural network model, the degree of trend deterioration is monitored, enabling early fault prediction. Specifically, this includes:

[0072] S1. Multi-source data acquisition:

[0073] Simultaneously acquire vibration signals from the motor bearings and current signals from the motor;

[0074] Vibration signal acquisition: A high-frequency acceleration vibration sensor is installed on the bearing housing and connected to the high-speed data acquisition module of the edge early warning terminal; the sampling rate of the high-speed data acquisition module is not less than 25.6kHz to meet the requirements of high-frequency resonance band analysis and continuously acquire the original vibration acceleration signal;

[0075] A current transmitter is installed in the motor control box to collect the three-phase current signal of the motor, calculate its effective value, and connect it to the analog quantity acquisition module of the edge early warning terminal as a background reference for the overall operating load of the motor.

[0076] S2. Feature parameter extraction:

[0077] A set of feature parameters characterizing early bearing failure is extracted from the vibration signal of each time window (each window is 10 seconds) and used as the input vector of the BP neural network. The feature parameters include:

[0078] 1) Root Mean Square (RMS):

[0079] The average energy characterizing the vibration signal is highly sensitive to localized damage in bearings, and the formula is as follows:

[0080] (2);

[0081] 2) Kurtosis:

[0082] It is extremely sensitive to the impact component in the signal, making it ideal for detecting early surface damage faults such as early pitting and minor cracks. The formula is as follows:

[0083] (3);

[0084] The kurtosis value of a normal bearing is close to 3. In the early stages of failure, the kurtosis value will increase. The larger the kurtosis value, the further the surface bearing condition deviates from the normal operating condition.

[0085] 3), High harmonic energy (HHE):

[0086] The weak impact caused by early failures can lead to high-frequency resonance in the system, generating higher-order harmonics in the spectrum at intervals of the fault characteristic frequency. The total harmonic energy of the bearing resonance frequency band is calculated as a characteristic, as shown in the following formula:

[0087] (4);

[0088] At the same time, the effective value of motor current within the same time window It is also used as one of the input features to introduce load information;

[0089] An input sample of a neural network can be represented as a vector:

[0090] (1);

[0091] In formula (1), The root mean square value of the vibration signal, in m / s. 2 ; Indicates the kurtosis of the vibration signal; This represents the higher harmonic energy of a vibration signal, measured in m / s². 2 ; This represents the effective value of the motor current, expressed in amperes (A). This represents the input feature vector;

[0092] In formula (2), This represents the discrete vibration acceleration signal value, with units of m / s². 2 N represents the number of data points within the window;

[0093] In formula (3), This represents the signal mean, measured in m / s. 2 ;

[0094] In formula (4),

[0095] A(f) represents the amplitude at frequency f in the spectrum, where f is the characteristic frequency of the bearing inner ring fault. Outer ring fault characteristic frequency characteristic frequency of rolling element failure M represents the harmonic order being considered (e.g., for the 5th harmonic, M is 5).

[0096] S3. Construct and train the BP neural network model;

[0097] The input feature vector is fed into a pre-trained BP neural network model;

[0098] The output of the BP neural network model is the predicted amplitude of the bearing fault characteristic frequency and the vibration energy at future time points:

[0099] Construct a three-layer backpropagation (BP) neural network model consisting of an input layer, a hidden layer, and an output layer;

[0100] 1) The input layer has 4 nodes, corresponding to four feature parameters, namely the root mean square value of the vibration signal. Kurtosis of vibration signals High-order harmonic energy of vibration signals Effective value of motor current ;

[0101] 2) The hidden layer considers the correlation of feature parameters, and the number of hidden layer nodes is set to 3 (the number of hidden layer nodes can be optimized based on experiments and experience). The hidden layer is the intermediate processing unit connecting the input layer and the output layer. Through the data processing of the hidden layer, the relationship between the four feature parameters of the input layer is constructed. Each output in the hidden layer node corresponds to the influence of the four vibration signal feature parameters of the input layer and the effective value of the motor current on the fault characteristic frequency and bearing vibration energy, thereby completing the information transmission from the input layer to the output layer.

[0102] The output of the j-th node in the hidden layer The calculation is as follows:

[0103] (5);

[0104] (6);

[0105] In formula (5) and formula (6), This represents the input of the i-th node in the input layer. This represents the connection weight from the i-th node in the input layer to the j-th node in the hidden layer. This represents the bias of the j-th node in the hidden layer. This represents the net input of the j-th node in the hidden layer;

[0106] The hidden layer uses the Sigmoid activation function for non-linear transformation, as shown in the following formula:

[0107] (7);

[0108] In formula (7), This represents the activation function. This represents the input to the activation function;

[0109] 3) The output layer predicts the fault characteristic frequency amplitude and vibration energy within a future period (within 1 minute). The output layer uses the Sigmoid activation function (formula (7)) to obtain the nonlinear transformation output, as shown in the following formula:

[0110] (8);

[0111] The output of the kth node in the output layer The calculation is as follows:

[0112] (9);

[0113] (10);

[0114] In formula (9) and formula (10), =1,2,3,4; This represents the connection weight from the j-th node in the hidden layer to the k-th node in the output layer. This represents the bias of the k-th node in the output layer. This represents the net input of the k-th node in the output layer;

[0115] In formula (8), the output of the first node of the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing inner ring failure, in m / s. 2 Output of the second node in the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing outer ring failure, in m / s. 2 Output of the 3rd node in the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing rolling element failure, in m / s. 2 Output of the 4th node in the output layer ,Right now This represents the predicted root mean square value of vibration acceleration, which is a predicted value of vibration energy, and its unit is m / s². 2 The root mean square (RMS) value of vibration acceleration characterizes the energy generated by vibration. This parameter can be obtained by measuring the vibration acceleration of the bearing using sensors and then processing the data. When the bearing is worn, the high-frequency components of the vibration will increase, and the RMS value of the bearing vibration acceleration will also increase. This parameter can be used to determine whether the bearing vibration exceeds the standard.

[0116] The BP neural network model is trained using historical data of the bearing's life cycle from normal operation to failure, and the prediction error is minimized through the error backpropagation algorithm.

[0117] 4) Construct a BP neural network fault prediction model:

[0118] Hidden layer net input:

[0119] = + ;

[0120] = + ;

[0121] = + (7);

[0122] Hidden layer output:

[0123] ;

[0124] ;

[0125] ;

[0126] In formula (7), Weight parameters representing the root mean square value; Weight parameters representing kurtosis; Weighting parameters representing higher harmonic energy; Weighting parameters representing the effective value of motor current; This indicates the bias of the hidden layer output; ~ This indicates the influence of the input vibration signal characteristic parameters and the effective value of the motor current on the fault characteristic frequency and bearing vibration energy.

[0127] That is, the weights from the input layer to the hidden layer. :

[0128] (8);

[0129] That is, the bias from the input layer to the hidden layer. :

[0130] (9);

[0131] The hidden layers undergo nonlinear transformations using the Sigmoid activation function:

[0132] (10);

[0133] In formula (10), This represents the activation function. This represents the input to the activation function;

[0134] Output of the BP neural network fault prediction model:

[0135] Output layer net input:

[0136] = + ;

[0137] = + ;

[0138] = + ;

[0139] = + (11);

[0140] Output layer output:

[0141] ;

[0142] ;

[0143] ;

[0144] ;

[0145] In formula (11), Weighted parameters representing the amplitude of characteristic frequencies of bearing inner ring failures. Weighted parameters representing the amplitude of characteristic frequencies of bearing outer ring failures. Weighted parameters representing the amplitude of characteristic frequencies of bearing rolling element failures. The weighting parameter represents the root mean square value of bearing vibration acceleration. This indicates the bias of the model output.

[0146] That is, the weights from the hidden layer to the output layer. :

[0147] (12);

[0148] That is, the bias from the hidden layer to the output layer. :

[0149] (13).

[0150] S4. Model Training:

[0151] 1) Training data preparation:

[0152] Collect a large amount of historical data throughout the entire life cycle of the bearing, from normal operation to complete failure, or through accelerated life testing. For each historical time point t, its feature vector is... As input, future time points ( The actual calculated fault frequency amplitude and RMS value (using the prediction step size) are used as the target output, as shown in the following formula:

[0153] T= [ , , , (14);

[0154] In formula (14), T represents the target output of the sample, including: Indicates a future point in time. Target output of the characteristic frequency amplitude of bearing inner ring fault, in m / s 2 ; Future Time Point Target output of the characteristic frequency amplitude of bearing outer ring fault, in m / s 2 ; Future Time Point Target output of the characteristic frequency amplitude of bearing rolling element failure, in m / s 2 ; Future Time Point The target output of the root mean square value of bearing vibration acceleration, in m / s². 2 ;

[0155] 2. Training process: The network is trained using the backpropagation algorithm.

[0156] 1) Forward propagation: Computing the actual output of the network ;

[0157] 2) Calculation error: ;

[0158] Backpropagation:

[0159] According to error Starting from the output layer, the error gradient is calculated layer by layer, and the gradient descent method is used to update all weights in the network. and bias until the corresponding deviation function value is minimized:

[0160] (15);

[0161] (16);

[0162] In formulas (15) and (16), For learning rate, The value range is 0.0001 to 0.5; Indicates the updated weight; Indicates the weights before the update; Indicates the updated bias; Indicates the bias before the update; This represents the gradient of the output layer weights. This represents the output layer bias gradient.

[0163] 3. Training objectives:

[0164] The network iterates continuously to minimize the average error of the entire training set until it can accurately predict the future development of fault characteristics from the current state features.

[0165] Based on the number of output nodes (4), the formula for calculating the average error is as follows:

[0166] (17);

[0167] In formula (17), Output to the target; This is the actual output of the network; This represents the mean squared error between the network's predicted value and the actual target value. =1~4.

[0168] S5. Fault Prediction and Early Warning:

[0169] The trained model is deployed to the edge warning terminal box. The edge warning terminal uses the BL410 model, which performs high-speed calculations and judgments locally, and only needs to upload key warning information and other result data to the upper-level platform warning system server.

[0170] 1) Real-time data input:

[0171] The system acquires current and vibration signals in real time, and extracts feature vectors according to step S2, as shown in the following formula:

[0172] (18);

[0173] In formula (18), The feature vector representing the real-time acquired input signal includes: This represents the root mean square value of the vibration signal acquired in real time, in m / s. 2 ; Indicates the kurtosis of the vibration signal acquired in real time; This represents the higher harmonic energy of the vibration signal acquired in real time, with units of m / s. 2 ; This represents the effective value of the motor current collected in real time, in amperes (A).

[0174] 2) Model prediction:

[0175] Will The input is fed into a pre-trained BP neural network model to obtain the predicted output, as shown in the following formula:

[0176] ;

[0177] In formula (19), The representation indicates the model's predicted output, including This represents the predicted characteristic frequency value of bearing inner ring failure, in m / s. 2 ; This represents the predicted characteristic frequency amplitude of bearing outer ring failure, in m / s. 2 ; This represents the predicted characteristic frequency amplitude of bearing rolling element failure, in m / s. 2 ; This represents the predicted root mean square value of vibration acceleration, in m / s². 2 ;

[0178] 3) Trend Analysis and Fault Prediction:

[0179] The system will provide the predicted values ​​for this test. Compare with historical forecasts over a previous period (such as the past 30 minutes).

[0180] Fault type determination:

[0181] if If the frequency of a bearing exhibits a continuous and significant upward trend, while other characteristic frequencies remain stable, it is predicted that the bearing will experience inner ring failure (such as inner ring spalling or cracking).

[0182] Similarly, the continuously rising Trend prediction of bearing outer ring failure;

[0183] Continuing to rise Trend prediction of bearing rolling element failure.

[0184] Fault severity warning: Simultaneously, if the predicted vibration energy... If the value also shows an upward trend and approaches or exceeds the safety threshold, it indicates that the fault is worsening.

[0185] 4) Triggering an alert:

[0186] When the predicted frequency of any fault characteristic shows an upward trend, or the predicted value of vibration energy shows an upward trend, the system automatically triggers an early fault warning to notify maintenance personnel.

[0187] A bearing early fault prediction system based on BP neural network includes a data acquisition module, an edge early warning terminal, and an early warning system platform server;

[0188] The data acquisition module includes a vibration sensor for acquiring bearing vibration signals and a current transmitter for acquiring motor current signals.

[0189] The edge warning terminal is connected to the data acquisition module to receive vibration and current signals and perform the following operations:

[0190] Extract input feature vectors, call a pre-trained BP neural network model for fault prediction, and trigger early warning based on trend analysis of the prediction results;

[0191] The early warning system platform server communicates with the edge early warning terminal to receive and display early warning information.

[0192] Example 2

[0193] In this embodiment, the bearing early failure prediction method and system based on BP neural network are the same as in Embodiment 1, with the following additional working process:

[0194] 1. Model training;

[0195] (1) Forward Propagation;

[0196] Step 1: Initialize weights and biases;

[0197] Input layer to hidden layer weights:

[0198] ;

[0199] Hidden layer bias:

[0200] Input layer to hidden layer bias :

[0201] ;

[0202] Hidden layer to output layer weights:

[0203] Hidden layer to output layer weights :

[0204] ;

[0205] Output layer bias:

[0206] Hidden layer to output layer bias :

[0207] ;

[0208] Step 2: Define the input and target output;

[0209] enter :

[0210] [ , , , = [0.5, 0.6, 0.7, 0.8];

[0211] Target output :

[0212] = [ , , , ] = [0.9, 0.8, 0.7, 0.6] ;

[0213] Step 3: Calculate the hidden layer inputs and outputs;

[0214] Hidden layer net input:

[0215] (1);

[0216] In formula (1), This represents the net input of the j-th node in the hidden layer; This represents the input of the i-th node in the input layer; This represents the connection weight from the i-th node in the input layer to the j-th node in the hidden layer; This represents the bias of the j-th node in the hidden layer;

[0217] calculate:

[0218] 0.5*0.1 + 0.6*0.4 + 0.7*0.7 + 0.8*1.0 + 0.1 = 0.05 + 0.24 + 0.49 + 0.8 + 0.1 = 1.68;

[0219] 0.5*0.2 + 0.6*0.5 + 0.7*0.8 + 0.8*1.1 + 0.2 = 0.1 + 0.3 + 0.56 + 0.88 + 0.2 = 2.04;

[0220] = 0.5*0.3 + 0.6*0.6 + 0.7*0.9 + 0.8*1.2 + 0.3 = 0.15 + 0.36 + 0.63 + 0.96 + 0.3 = 2.40(6);

[0221] Hidden layer output (using Sigmoid):

[0222] (2);

[0223] In formula (2), the activation function of the hidden layer is the sigmoid function to obtain the hidden layer output. This represents the output of the j-th hidden layer node; This represents the net input of the j-th node in the hidden layer;

[0224] calculate:

[0225] 0.843;

[0226] 0.885;

[0227] 0.917;

[0228] Hidden layer output:

[0229] = [0.843, 0.885, 0.917] ;

[0230] Step 4: Calculate the input and output of the output layer;

[0231] Output layer net input The formula is as follows:

[0232] (3);

[0233] In formula (3), This represents the connection weight from the j-th node in the hidden layer to the k-th node in the output layer; This represents the bias of the k-th node in the output layer;

[0234] Calculate the output of the first node of the output layer :

[0235] = 0.843*0.13 + 0.885*0.17 + 0.917*0.21 + 0.25 0.1096 + 0.1505 + 0.1926 + 0.25 = 0.7027;

[0236] in This represents the net input of the first node in the output layer;

[0237] Applying the Sigmoid function, the formula is as follows:

[0238] = 0.668;

[0239] Calculate the output of the second node in the output layer. :

[0240] = 0.843*0.14 + 0.885*0.18 + 0.917*0.22 + 0.26 0.1180 + 0.1593 + 0.2017 + 0.26 = 0.7390;

[0241] in, This represents the net input of the second node in the output layer;

[0242] Applying Sigmoid:

[0243] = 0.677 (13);

[0244] Calculate the output of the 3rd node in the output layer. :

[0245] = 0.843*0.15 + 0.885*0.19 + 0.917*0.23 + 0.27 0.1265 + 0.1682 + 0.2109 + 0.27 = 0.7756

[0246] in, This represents the net input of the 3rd node in the output layer;

[0247] Applying Sigmoid:

[0248] = 0.685;

[0249] Calculate the output of the 4th node in the output layer. :

[0250] = 0.843*0.16 + 0.885*0.20 + 0.917*0.24 + 0.28 0.1349 + 0.1770 + 0.2201 + 0.28 = 0.8120;

[0251] in, This represents the net input of the 4th node in the output layer;

[0252] Applying Sigmoid:

[0253] = 0.692;

[0254] Therefore, the output [0.668, 0.677, 0.685, 0.692] ;

[0255] in This represents the model's predicted output;

[0256] Step 5: Calculate the total error;

[0257] (1) The total error formula is as follows:

[0258]

[0259]

[0260] = [0.0538 + 0.0151 + 0.000225 + 0.00846] 0.0388

[0261] (2) Backpropagation;

[0262] Learning rate = 0.5

[0263] Step 1: Calculate the output layer error term ;

[0264] For the Sigmoid output layer:

[0265] (14);

[0266] In formula (14), This represents the error term of the k-th node in the output layer; This represents the target output of the k-th node in the output layer; This represents the actual output of the k-th node in the output layer; k = 1~4;

[0267] calculate:

[0268] = =(0.9 - 0.668) * 0.668 * (1-0.668) 0.0514;

[0269] = = (0.8 - 0.677) * 0.677 * (1-0.677) 0.0269;

[0270] = = (0.7 - 0.685) * 0.685 * (1-0.685) 0.00324;

[0271] = =(0.6 - 0.692) * 0.692 * (1-0.692) -0.0196;

[0272] get:

[0273] [0.0514, 0.0269, 0.00324, -0.0196];

[0274] in express The output layer error vector is formed.

[0275] Step 2: Update the hidden layer weights to the output layer weights. and bias ;

[0276] (15);

[0277] In formula (15), This represents the update amount of the connection weight from the j-th node in the hidden layer to the k-th node in the output layer; This represents the error term of the k-th node in the output layer;

[0278]

[0279] (16);

[0280] In formula (16), This represents the bias update amount of the k-th node in the output layer;

[0281]

[0282] These are the weights before the hidden layer is updated to the output layer.

[0283] The connection weights from the first node in the hidden layer to the first node in the output layer For example:

[0284] = =0.5 * 0.0514 * 0.843 0.0217;

[0285] new = 0.13 + 0.0217 = 0.1517;

[0286] The bias of the first node of the output layer For example:

[0287] = =0.5 * 0.0514 0.0257;

[0288] new = = 0.25 + 0.0257 = 0.2757;

[0289] Other weights and Other bias updates are similar.

[0290] Step 3: Calculate the hidden layer error term ;

[0291] (17);

[0292] In formula (17), This represents the error term of the j-th node in the hidden layer; This represents the error term of the k-th node in the output layer; This represents the weights from the hidden layer to the output layer before the update. This represents the output of the j-th hidden layer node;

[0293] Calculate the error term of the first node in the hidden layer. :

[0294] Summation part:

[0295] 0.0514*0.13 + 0.0269*0.14 + 0.00324*0.15 + (-0.0196)*0.16

[0296] 0.006682 + 0.003766 + 0.000486 + (-0.003136) 0.007798;

[0297] 0.007798 * 0.843 * (1-0.843) 0.001032;

[0298] Calculate the error term of the second node in the hidden layer. :

[0299] Summation part:

[0300] 0.0514*0.17 + 0.0269*0.18 + 0.00324*0.19 + (-0.0196)*0.20

[0301] 0.0102756;

[0302] 0.0102756 * 0.885 * (1-0.885) 0.001046;

[0303] Calculate the error term of the third node in the hidden layer. :

[0304] Summation part:

[0305] 0.0514*0.21 + 0.0269*0.22 + 0.00324*0.23 + (-0.0196)*0.24 0.0127532;

[0306] 0.0127532 * 0.917 * (1-0.917) 0.000970;

[0307] Step 4: Update the input layer weights to the hidden layer weights. and bias ;

[0308] ;

[0309] in, This represents the update amount of the connection weight from the i-th node in the input layer to the j-th node in the hidden layer;

[0310] ;

[0311] ;

[0312] in, This represents the bias update amount of the j-th node in the hidden layer;

[0313] ;

[0314] The connection weights from the first node in the input layer to the first node in the hidden layer For example:

[0315] = 0.5 * 0.001032 * 0.5 0.000258

[0316] new = 0.1 + 0.000258 = 0.100258

[0317] The bias of the first node in the hidden layer For example:

[0318] = 0.5 * 0.001032 0.000516

[0319] new = 0.1 + 0.000516 = 0.100516

[0320] Other weights and Other bias updates are similar.

[0321] At this point, we have completed using a sample (input sample) and target sample This represents a complete forward and backward propagation of the network. All weights and biases of the network are fine-tuned based on the error.

[0322] In actual training, this process is repeated thousands of times on the entire training set until the error converges to an acceptable level, thus obtaining the neural network model.

[0323] 1. System Deployment: Install a vibration acceleration sensor on the bearing housing of the drive motor of the wind turbine to collect vibration signals, and install a current transmitter in the motor housing to collect the effective value of the current.

[0324] 2. Data Acquisition and Feature Extraction: A data point is generated every 10 seconds. For each data point, the following are calculated from the 10 seconds of vibration data: RMS value, kurtosis, and HHE value (calculating the energy of the first 5 harmonics). The current average current is also recorded. This forms the input vector.

[0325] 3. Model Application: Input the vector into the trained model, and the model will output the predicted value.

[0326] 4. Prediction Results and Early Warning: The system queries the prediction records for this bearing within the past 30 minutes. If any are found... , , , If the slope of the amplitude of a certain fault characteristic frequency or the root mean square value of vibration acceleration continuously increases, exhibiting a clear monotonically increasing trend, the system will make a fault judgment. For example, if the model outputs a predicted value: = 0.08, the system queries the prediction records for this bearing over the past 30 minutes and finds The value sequence is [0.01, 0.02, 0.03, 0.05, 0.06], showing a clear monotonically increasing trend, and the slope of the trend is constantly increasing. and The trend is stable. The value also increased from 0.3 to 0.7. The system determined that the bearing had a significant risk of early inner ring failure (such as pitting corrosion), and that the failure was accelerating. The system immediately issued an early failure warning, suggesting that "bearing inner ring failure is predicted, and inspection and maintenance should be arranged within 2 weeks."

[0327] This invention defines the output of a BP neural network model as the fault characteristic state at a future time, enabling the model to learn the mapping relationship from "current features" to "future states." By monitoring the deterioration trend of predicted values, it can issue early warnings in the early stages of fault occurrence, before serious damage occurs, overcoming the lag of traditional threshold alarms and providing sufficient time windows for preventative maintenance. By fusing macroscopic load information of motor current with microscopic details of high-frequency bearing vibration, a multi-dimensional sensing system is constructed. This multi-source information fusion mechanism effectively overcomes the susceptibility of single signal sources to interference under complex operating conditions, comprehensively utilizing the complementarity of different signals to significantly reduce false alarms and missed alarms. The entire process, from feature extraction and trend prediction to fault judgment, is automatically completed by the BP neural network model and algorithm, eliminating the need for manual spectrum analysis by professionals. This not only reduces reliance on expert experience but also allows the solution to be efficiently deployed in large-scale industrial IoT scenarios, achieving automation and intelligence in equipment health management. By monitoring different fault characteristic frequencies output by the model (such as BPFI, BPFO, ...), ... The independent predictive trend of the bearing finite element (BSF) can initially identify and differentiate potential fault locations (bearing inner ring, bearing outer ring, or bearing rolling elements), providing maintenance personnel with richer and more accurate decision-making information. This achieves a leap from simple judgment of "whether there is a fault" to precise diagnosis of "what kind of fault." By introducing nonlinear activation functions such as Sigmoid into the hidden and output layers of the neural network, the model's ability to fit complex nonlinear relationships is enhanced. At the same time, this structure helps stabilize the training process and improves the model's adaptability to unseen operating condition data, thereby ensuring the robustness and generalization of the system in practical applications.

Claims

1. A method for predicting early bearing failures based on a BP neural network, characterized in that, include: S1. Multi-source data acquisition: Simultaneously acquire vibration signals from the motor bearings and current signals from the motor; S2. Feature parameter extraction: Extract feature parameters from the vibration signal, including at least the root mean square value, kurtosis and higher harmonics, and combine them with the effective value of the current signal to form an input feature vector. S3. Construct and train the BP neural network model; The input feature vector is fed into a pre-trained BP neural network model; The output of the BP neural network model is the predicted amplitude of the bearing fault characteristic frequency and the vibration energy at future time points: S4. Fault Prediction and Early Warning: The real-time collected data is processed by S2 and then input into the trained BP neural network model to obtain the predicted future state value. Early warnings are triggered by analyzing the changing trends of future state values.

2. The bearing early failure prediction method based on BP neural network according to claim 1, characterized in that, In S2, the input feature vector is defined by the following formula: (1)； (2)； (3)； (4)； In formula (1), The root mean square value of the vibration signal, in m / s. 2 ; Indicates the kurtosis of the vibration signal; This represents the higher harmonic energy of a vibration signal, measured in m / s². 2 ; This represents the effective value of the motor current, expressed in amperes (A). This represents the input feature vector; In formula (2), This represents the discrete vibration acceleration signal value, with units of m / s². 2 N represents the number of data points within the window; In formula (3), This represents the signal mean, measured in m / s. 2 ; In formula (4), A(f) represents the amplitude at frequency f in the spectrum, where f is the characteristic frequency of the bearing inner ring fault. Outer ring fault characteristic frequency characteristic frequency of rolling element failure M represents the harmonic order being considered; The effective value of motor current within the same time window As one of the input features, it is used to introduce load information.

3. The bearing early failure prediction method based on BP neural network according to claim 1, characterized in that, In S3, a three-layer BP neural network model containing an input layer, a hidden layer, and an output layer is constructed. The input layer includes four feature parameters, which are the root mean square values ​​of the vibration signal. Kurtosis of vibration signals High-order harmonic energy of vibration signals Effective value of motor current ; The output of the j-th node in the hidden layer The calculation is as follows: (5)； (6)； In formula (5) and formula (6), This represents the input of the i-th node in the input layer. This represents the connection weight from the i-th node in the input layer to the j-th node in the hidden layer. This represents the bias of the j-th node in the hidden layer. This represents the net input of the j-th node in the hidden layer; The hidden layer uses the Sigmoid activation function for non-linear transformation, as shown in the following formula: (7)； In formula (7), This represents the activation function. This represents the input to the activation function; The output layer predicts the future fault characteristic frequency amplitude and vibration energy using the following formula: (8)； The output of the kth node in the output layer The calculation is as follows: (9); (10)； In formula (9) and formula (10), =1,2,3,4; This represents the connection weight from the j-th node in the hidden layer to the k-th node in the output layer. This represents the bias of the k-th node in the output layer. This represents the net input of the k-th node in the output layer; In formula (8), the output of the first node of the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing inner ring failure, in m / s. 2 Output of the second node in the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing outer ring failure, in m / s. 2 Output of the 3rd node in the output layer ,Right now This represents the predicted characteristic frequency amplitude of bearing rolling element failure, in m / s. 2 Output of the 4th node in the output layer ,Right now This represents the predicted root mean square value of vibration acceleration, which is the predicted vibration energy value in m / s². 2 .

4. The bearing early failure prediction method based on BP neural network according to claim 1, characterized in that, In S4, the analysis of the changing trend of future state values ​​includes: like It shows a continuous upward trend. and If the trend is stable, the prediction is a failure in the inner ring of the bearing. like It shows a continuous upward trend. and If the trend is stable, the prediction is a failure in the outer ring of the bearing; like It shows a continuous upward trend. and If the trend is stable, it is predicted to be a bearing rolling element failure; When the trend slope increases or the predicted vibration energy exceeds the safety threshold, an early fault warning notification will be automatically issued.

5. The bearing early failure prediction method based on BP neural network according to claim 1, characterized in that, In S3, the BP neural network model is trained using historical data of the bearing's life cycle from normal to failure, and the prediction error is minimized through the error backpropagation algorithm.

6. A bearing early fault prediction system based on a BP neural network for implementing the method of any one of claims 1-5, characterized in that, Includes a data acquisition module, an edge early warning terminal, and an early warning system platform server; The data acquisition module includes a vibration sensor for acquiring bearing vibration signals and a current transmitter for acquiring motor current signals. The edge warning terminal is connected to the data acquisition module to receive vibration and current signals and perform the following operations: Extract input feature vectors, call a pre-trained BP neural network model for fault prediction, and trigger early warning based on trend analysis of the prediction results; The early warning system platform server communicates with the edge early warning terminal to receive and display early warning information.

Images