Weighing sensor fault diagnosis method based on improved PSO-GRNN neural network

Abstract

The present application relates to a kind of based on improved PSO-GRNN neural network weighing sensor fault diagnosis method, comprising the following steps: step 1: collect the vibration signal under the normal state of weighing sensor and the vibration signal under different fault states, form first data set;Step 2: the data sample of first data set is labeled and handled, forms second data set;Step 3: the signal characteristics of output vibration signal under different states are calculated and extracted, as the input feature vector of model, form third data set;Step 4: introduce Lévy flight and complete the optimization of PSO algorithm;Step 5: the smoothing factor of GRNN neural network is optimized using improved PSO algorithm, and GRNN neural network fault diagnosis model is constructed;Step 6: based on third data set, GRNN neural network fault diagnosis model is trained, and the final GRNN neural network fault diagnosis model is obtained.The present application effectively improves the reliability of weighing sensor fault diagnosis efficiency.

Description

Technical Field

[0001] This invention relates to the field of weighing sensor fault diagnosis technology, and specifically to a weighing sensor fault diagnosis method based on an improved PSO-GRNN neural network. Background Technology

[0002] Load cells are widely used sensing and measuring devices in force measurement and weighing, and are important components for information acquisition in modern industry. They convert load values ​​into measurable signals, and the accuracy of the output physical signal affects the system's measurement results. Load cells can be used in both static and dynamic environments. Static applications include static weighing and force measurement value transfer; dynamic applications include dynamic weighing of vehicles and goods, condition monitoring of mechanical equipment, and force measurement and control in material testing machines. In complex industrial environments, prolonged adverse external conditions (such as extremely high temperatures, excessive pressure, and particularly humid environments) and load cell malfunctions often cause abnormal data output from the sensors. The output signal will deviate from the true signal, leading to erroneous results in subsequent analysis, and in severe cases, even causing industrial accidents.

[0003] To prevent sensor malfunctions or failures from affecting the normal operation of the system, it is necessary to identify and diagnose faults promptly. Traditional methods involve periodic manual calibration. While this can identify some common physical damage to sensors, it cannot pinpoint internal faults. Furthermore, manual calibration is extremely wasteful of manpower and resources. Many sensors operate in complex environments, making it difficult for on-site personnel to handle sensor faults promptly and efficiently. More seriously, the noise and toxic gases in the production environment can harm their physical and mental health. Therefore, researching fault diagnosis methods for weighing sensors is crucial for ensuring the safe operation of equipment.

[0004] Traditional fault diagnosis techniques based on manual and physical models cannot provide timely feedback on the type of fault. With the development of big data and artificial intelligence technologies, more and more people are applying big data information, machine learning, and deep learning to fault diagnosis to improve the rapid fault detection, diagnosis, and repair capabilities of equipment. GRNN is a feedforward supervised neural network. Due to its feedforward structure, it possesses strong nonlinear mapping capabilities, a flexible network structure, and high fault tolerance and robustness, achieving good regression results even with a small number of training samples. However, its accuracy is highly dependent on the selection of the structural parameter smoothing factor. PSO algorithm theory is relatively mature and is commonly used to optimize neural networks for application in fault diagnosis. However, its drawback is that it is prone to getting trapped in local optima when optimizing parameters. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a fault diagnosis method for weighing sensors based on an improved PSO-GRNN neural network, which aims to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A fault diagnosis method for weighing sensors based on an improved PSO-GRNN neural network includes the following steps:

[0008] Step 1: Collect vibration signals from the load cell under normal conditions and vibration signals under different fault conditions to form the first dataset;

[0009] Step 2: Label the data samples in the first dataset to form the second dataset;

[0010] Step 3: Calculate and extract the signal features of the output vibration signal under different states, and use them together as the input feature vector of the model to form a third dataset;

[0011] Step 4: Introduce Lévy flight to improve the PSO algorithm;

[0012] Step 5: Optimize the smoothness factor of the GRNN neural network using the improved PSO algorithm, and construct a GRNN neural network fault diagnosis model;

[0013] Step 6: Train the GRNN neural network fault diagnosis model based on the third dataset to obtain the final GRNN neural network fault diagnosis model.

[0014] Furthermore, the first dataset is an n×m matrix, where n is the number of samples of the measured data and m is the sample dimension, i.e. the number of sampling points of the sample;

[0015] Furthermore, the different states include six states: normal, deviation fault, impact fault, drift fault, jamming fault, and accuracy reduction fault, which are represented by 1, 2, 3, 4, 5, and 6 respectively, and are labeled accordingly.

[0016] Furthermore, the signal features include the root mean square value y. rms Variance σ², peak value C, impulse rate I, mean kurtosis index y q Maximum value y max Minimum value y min Peak value y pp Root amplitude y r Average amplitude y * The waveform index K and the margin index L are calculated using the following formulas:

[0017]

[0018] In the formula, y i This represents the value of a collection point in a single sample; m represents the sample dimension.

[0019]

[0020]

[0021]

[0022] y max =max{y i} (7)

[0023] y min =min{y i} (8)

[0024] y pp =y max -y min (9)

[0025]

[0026]

[0027]

[0028]

[0029] Furthermore, step S4 specifically involves: improving the PSO algorithm by introducing Lévy flight, and avoiding the PSO algorithm from getting trapped in local optima by generating a random step size, as detailed below:

[0030] After introducing Lévy flight, the formula for updating the particle position is:

[0031]

[0032] Among them, S i_(t+1) This represents the position of the (i+1)th particle after the Lévy flight and the position after the t-th update; α represents the step size control variable. This is a point-to-point multiplication; L(u,v) is a random step-size path, and the Mantegna algorithm is used to perform Lévy flight. The step-size calculation formula is as follows:

[0033]

[0034] Wherein, the range of β is generally 1 < β < 3; u and v both follow a normal distribution, as shown in the following formula.

[0035] u~N(0,σ u 2 (16)

[0036] v~N(0,σ v 2 (17)

[0037] σ u σ v The calculation method is as follows:

[0038]

[0039] Furthermore, step S5 specifically includes:

[0040] Step 501: Determine the topology of the GRNN neural network.

[0041] Step 502: Establish the initial particle population S = {s1, s2, ..., s} m} T Set the following parameters for the particle swarm: initial parameters c1 and c2, initial velocity matrix V, and initial optimal position p of each individual particle. i and the global optimal position p g ;

[0042] Step 503: Evaluate each particle in the particle population and obtain the fitness value of each particle;

[0043] Step 504: Using the particle update formula introduced by Lévy flight, update the particles in the particle swarm, including updating the state of each particle in the particle swarm and the optimal position p of each individual particle. i The update, and the globally optimal position p g Update;

[0044] Step 505: When the number of updates reaches the set maximum number of training iterations, or when the particle fitness value of the updated result meets the requirements, training terminates and the final result is output. Otherwise, repeat steps 503 and 504 to continue evaluation and updates until the maximum number of training iterations is reached or the global optimum is found;

[0045] Step 506: Global optimal position p g This is the global optimal value, which is also the optimal solution for the smoothness factor σ, a structural parameter of the GRNN neural network.

[0046] Furthermore, the topology includes: the number of nodes j in the input layer of the neural network, the number of neurons k in the pattern layer, and the number of nodes l in the output layer; the number of nodes j in the input layer depends on the vector dimension of the input samples; the number of neurons k in the pattern layer is the same as the number of learning samples; and the number of nodes l in the output layer is equal to the dimension of the output vector.

[0047] Furthermore, particle s i The fitness value refers to the number of particles s in n support vector machines. i The sum of squared errors between predicted and actual values ​​is used to quantify the individual particles based on their fitness, and the fitness F(s) of each particle is calculated according to the formula. i ):

[0048]

[0049] A fault diagnosis system for weighing sensors based on an improved PSO-GRNN neural network, characterized in that the system comprises:

[0050] Data acquisition module: used to acquire the output signal of the weighing sensor to obtain the first dataset;

[0051] Data processing module: used to perform data processing operations based on the first data information obtained by the data acquisition module, specifically including extracting labels and various features of vibration signals from the data, and finally obtaining a third dataset;

[0052] Training module: Used to establish a fault diagnosis model based on the improved PSO-GRNN neural network. The improved PSO-GRNN neural network model is trained using the training set in the third dataset, and the accuracy of the fault diagnosis model based on the improved PSO-GRNN neural network is verified using the validation set in the third training set.

[0053] Fault diagnosis module: Real-time detection of the weighing sensor, processing of the data to obtain data samples, inputting the data samples into the trained fault diagnosis model, and finally obtaining the result of the weighing sensor's structure.

[0054] Furthermore, the data acquisition module connects to a data acquisition card connected to the weighing sensor to acquire the sensor's vibration signal. After acquiring signals from multiple different states, a first dataset is obtained.

[0055] Compared with the prior art, the present invention has the following advantages:

[0056] This invention solves the problem that the PSO algorithm's weak ability to find the global optimum leads to convergence into local optima, which in turn causes the GRNN neural network to fail to match the optimal parameter smoothing factor, resulting in low fault diagnosis accuracy. It provides a new artificial intelligence-based method for fault diagnosis of weighing sensors, improving safe production efficiency and reducing economic losses. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 The flowchart of the improved PSO-GRNN algorithm of this invention is shown below;

[0059] Figure 3 This is a network structure diagram of the GRNN neural network of the present invention;

[0060] Figures 4-9 These are signal characteristic diagrams of different fault types of the weighing sensor in this invention;

[0061] Figure 10 The diagram shows the optimization process before and after the algorithm improvement of this invention;

[0062] Figure 11 This is a physical connection diagram of the present invention;

[0063] In the diagram, 1 represents the weighing sensor fault diagnosis system, which includes a data processing module, a training module, and a fault diagnosis module; 2 represents the data acquisition card; and 3 represents the weighing sensor. Together, 2 and 3 form the data acquisition module. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0065] Please refer to Figure 1 This invention provides a fault diagnosis method for weighing sensors based on an improved PSO-GRNN neural network, comprising the following steps:

[0066] Step 1: Collect vibration signals from the load cell under normal conditions and vibration signals under different fault conditions to form the first dataset;

[0067] Step 2: Label the data samples in the first dataset to form the second dataset;

[0068] Step 3: Calculate and extract the signal features of the output vibration signal under different states, and use them together as the input feature vector of the model to form a third dataset;

[0069] Step 4: Introduce Lévy flight to improve the PSO algorithm;

[0070] Step 5: Optimize the smoothness factor of the GRNN neural network using the improved PSO algorithm, and construct a GRNN neural network fault diagnosis model;

[0071] Step 6: Train the GRNN neural network fault diagnosis model based on the third dataset to obtain the final GRNN neural network fault diagnosis model.

[0072] In this embodiment, the first dataset is an n×m matrix, where n is the number of samples of the measured data, and m is the sample dimension, i.e., the number of sampling points of the samples, as shown in the following formula:

[0073]

[0074] In this embodiment, numbers 1, 2, 3, 4, 5, and 6 are used to label different states of the weighing sensor, as shown in the table below:

[0075] Table 1. Different Fault Types and Labels

[0076]

[0077]

[0078] In this embodiment, the signal features include the root mean square value y. rms Variance σ², peak value C, impulse rate I, mean kurtosis index y q Maximum value y max Minimum value y min Peak value y pp Root amplitude y r Average amplitude y * The waveform index K and the margin index L are calculated using the following formulas:

[0079]

[0080] In the formula, y i This represents the value of a collection point in a single sample; m represents the sample dimension.

[0081]

[0082]

[0083]

[0084] y max =max{y i} (7)

[0085] y min =min{y i} (8)

[0086] y pp =y max -y min (9)

[0087]

[0088]

[0089]

[0090]

[0091] The third dataset obtained is:

[0092]

[0093] In this embodiment, step S4 specifically involves improving the PSO algorithm by introducing Lévy flight. Figure 2 This is a flowchart of the improved PSO-GRNN algorithm of this invention. It mainly avoids the PSO algorithm from getting trapped in local optima by generating random step sizes. After introducing Lévy flight, the particle position update formula is:

[0094]

[0095] Among them, S i_(t+1) This represents the position of the (i+1)th particle after the Lévy flight and the position after the t-th update; α represents the step size control variable. This is a point-to-point multiplication; L(u,v) is a random step-size path, and the Mantegna algorithm is used to perform Lévy flight. The step-size calculation formula is as follows:

[0096]

[0097] Wherein, the range of β is generally 1 < β < 3; u and v both follow a normal distribution, as shown in the following formula.

[0098] u~N(0,σ u 2 (16)

[0099] v~N(0,σ v 2 (17)

[0100] σ u σ v The calculation method is as follows:

[0101]

[0102] In this embodiment, step S5 specifically includes:

[0103] Step 501: Determine the topology of the GRNN neural network, which includes: the number of nodes j in the input layer, the number of neurons k in the pattern layer, and the number of nodes l in the output layer. The number of nodes j in the input layer depends on the dimension of the input sample vector; the number of neurons k in the pattern layer is the same as the number of training samples; the number of nodes l in the output layer is equal to the dimension of the output vector. Figure 3 This is a network structure diagram of the GRNN neural network of the present invention;

[0104] Step 502: Initialize the parameters and establish the initial particle population S = {s1, s2, ..., s} m} T The particle size is m. And for particle s... i The position of its particle is s i ={s i1 ,s i2 ,...,s iD} T ;

[0105] The particle's velocity is v i ={v i1 ,v i2 ,...,v iD} T The optimal position of the particle is p. i ={p i1 ,p i2 ,...,p iD} T ;

[0106] The optimal position of the particle swarm is p g ={p g1 ,p g2 ,...,p gm} T Each particle possesses a different individual fitness. The following parameters of the particle swarm are set: initial parameters c1 and c2, initial velocity matrix V, and initial optimal position p of each individual particle. i and the global optimal position p g ;

[0107] Step 503: Evaluate each particle in the particle population and obtain the fitness value of each particle. This value is the main indicator describing particle performance. i The fitness value refers to the number of particles s in n support vector machines. i The sum of squares of the errors between predicted and actual values. Based on fitness, individual particles are eliminated through a survival-of-the-fittest process. The fitness F(s) of each particle is calculated using the formula. i ):

[0108]

[0109] Step 504: Using the particle update formula introduced by Lévy flight, update the particles in the particle swarm, including updating the state of each particle in the particle swarm and the optimal position p of each individual particle. i The update, and the globally optimal position p g Update;

[0110] Step 505: When the number of updates reaches the set maximum number of training iterations, or when the particle fitness value of the updated result meets the requirements, the training terminates, and the final result is output as shown in the following formula. Otherwise, repeat steps 503 and 504 to continue evaluation and updating until the maximum number of training iterations is reached or the global optimum is found;

[0111]

[0112] t = t max

[0113] Step 506: Global optimal position p g This is the global optimal value, which is also the optimal solution for the smoothness factor σ, a structural parameter of the GRNN neural network.

[0114] σ=p g

[0115] In this embodiment, step S6 involves training the optimized GRNN neural network fault diagnosis model to obtain the final GRNN neural network fault diagnosis model. The specific steps are as follows:

[0116] Step 601: Assumption The input samples for the GRNN network are n, and the total number of samples in the training and learning sample set is n; S = [s1, s2, s3, ..., s j ] T d represents the output sample corresponding to the network, where d is the sample dimension of the output sample; This represents the network's predicted output when the input is Y. The number of neurons in the input layer is equal to the dimension of the input sample Y vector, and each neuron directly transmits information from the Y vector to the pattern layer.

[0117] Step 602: In the pattern layer, the number of neurons is equal to the number of training samples, and a Gaussian function is used as the basis function. The neuron transfer function is:

[0118]

[0119] In the formula, Y i σ represents the input learning sample corresponding to the i-th neuron; σ is the smoothing factor.

[0120] Step 603: The summation layer mainly sums the transfer functions in the pattern layer, and there are two main types of summation. The first type of summation calculates the algebraic sum of the transfer functions of each neuron in the pattern layer, with a connection weight of 1, also known as the denominator unit S. D The transfer function for this summation type is:

[0121]

[0122] The second type of summation calculates the weighted sum of the transfer functions of each neuron in the pattern layer. The connection weights are the expected output values ​​from the training samples, also known as the molecular unit S. Nj Its transfer function is:

[0123]

[0124] In the formula, y ij Let be the expected output value of the i-th sample.

[0125] Step 604: The number of neurons in the output layer is equal to the dimension d of the output vector. The output layer mainly obtains the output result by dividing the output of the numerator unit and the output of the denominator unit obtained from the summation layer.

[0126]

[0127] The predicted output of the improved LPO-GRNN neural network model after obtaining the i-th sample is:

[0128]

[0129] In this embodiment, a fault diagnosis system for weighing sensors based on an improved PSO-GRNN neural network is also provided, characterized in that the system includes:

[0130] Data acquisition module: used to acquire the output signal of the weighing sensor to obtain the first dataset;

[0131] Data processing module: used to perform data processing operations based on the first data information obtained by the data acquisition module, specifically including extracting labels and various features of vibration signals from the data, and finally obtaining a third dataset;

[0132] Training module: Used to establish a fault diagnosis model based on the improved PSO-GRNN neural network. The improved PSO-GRNN neural network model is trained using the training set in the third dataset, and the accuracy of the fault diagnosis model based on the improved PSO-GRNN neural network is verified using the validation set in the third training set.

[0133] Fault diagnosis module: Real-time detection of the weighing sensor, processing of the data to obtain data samples, inputting the data samples into the trained fault diagnosis model, and finally obtaining the result of the weighing sensor's structure.

[0134] Preferably, the data acquisition module is connected to a data acquisition card to the weighing sensor. The data acquisition card acquires the sensor's vibration signal, and after acquiring signals from multiple different states, a first dataset is obtained.

[0135] The present invention provides another embodiment, which uses a T070B four-channel strain acquisition card to acquire vibration signals of a weighing sensor in different states, with a sampling rate of 30S / s and a refresh period of 20ms.

[0136] The six types of load cell faults diagnosed in this example (including normal mode) include: normal mode, deviation fault, impact fault, drift fault, jamming fault, and accuracy reduction fault.

[0137] When training a fault diagnosis system for a weighing sensor, it is necessary to combine the datasets of six states—normal, deviation, impact, drift, jamming, and accuracy degradation—into a matrix. The total sample size is 180, and each sample has 4000 data points, meaning the sample dimension is 4000. This can be represented by a data matrix of 180 rows and 4000 columns, forming the first dataset.

[0138] According to step 4, the first dataset is labeled to identify the fault types mentioned in the data, forming the second dataset: In this example, 1, 2, 3, 4, 5, and 6 are used to represent six different faults, and the labels corresponding to the fault types are shown in Table 1.

[0139] According to step 3, vibration signal feature extraction processing is performed on the second dataset to obtain the third dataset. Since 13 features are extracted for each fault signal, the total matrix of the third dataset is 180×13. It is divided into training and testing samples in a 2:1 ratio, that is, the training set matrix of the diagnostic system is 120×13 and the testing set matrix is ​​60×13.

[0140] Based on step 4, the improvement of the PSO algorithm is completed;

[0141] According to step 5, the parameter optimization of the GRNN neural network is completed. The optimal position of the particle swarm, i.e., the optimal value of the smoothness factor σ, a core structural parameter of the GRNN neural network, changes with the number of training iterations as follows: Figure 10 As shown in Table 2, the smoothing factor σ and diagnostic accuracy of the PSO algorithm before and after optimization of the GRNN neural network are as follows:

[0142] Table 2. σ values ​​for GRNN, PSO-GRNN, and improved PSO-GRNN

[0143]

[0144]

[0145] The diagnostic performance of three fault diagnosis models—the GRNN network model, the PSO-GRNN network model, and the improved PSO-GRNN network model—was compared. In terms of diagnostic accuracy, the unoptimized GRNN network model achieved 81.67%, the PSO-GRNN network model achieved 90%, and the improved PSO-GRNN network model achieved a remarkable 96.67%, demonstrating a significant improvement in accuracy compared to the other two. Therefore, the fault diagnosis model proposed in this invention significantly improves the fault classification accuracy of the system compared to previous algorithms.

[0146] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A fault diagnosis method for weighing sensors based on an improved PSO-GRNN neural network, characterized in that, Includes the following steps: Step 1: Collect vibration signals from the load cell under normal conditions and vibration signals under different fault conditions to form the first dataset; Step 2: Label the data samples in the first dataset to form the second dataset; Step 3: Calculate and extract the signal features of the output vibration signal under different states, and use them together as the input feature vector of the model to form a third dataset; Step 4: Introduce Lévy flight to improve the PSO algorithm. By generating a random step size, the PSO algorithm is prevented from getting trapped in local optima. The details are as follows: After introducing Lévy flight, the formula for updating the particle position is: （14） in, Indicates the first After the first particle was introduced into Lévy's flight, it was at the... The position after the next update; Represents the step size control quantity; This is point-to-point multiplication; For a random step size path, the Mantegna algorithm is used to perform Lévy flight. The step size calculation formula is as follows: （15） in, The range of values ​​for is generally as follows: ; , All follow a normal distribution, as shown in equation (17). （16） （17） , The calculation method is as follows: （18）； Step 5: Optimize the smoothness factor of the GRNN neural network using the improved PSO algorithm, and construct a GRNN neural network fault diagnosis model; Step 6: Train the GRNN neural network fault diagnosis model based on the third dataset to obtain the final GRNN neural network fault diagnosis model.

2. The fault diagnosis method for weighing sensors based on the improved PSO-GRNN neural network according to claim 1, characterized in that, The first dataset is one The matrix, where, The number of samples in the measured data. The dimension of the sample is the number of data collection points.

3. The fault diagnosis method for weighing sensors based on the improved PSO-GRNN neural network according to claim 1, characterized in that, The different states include six states: normal, deviation fault, impact fault, drift fault, jamming fault, and accuracy reduction fault. They are represented by 1, 2, 3, 4, 5, and 6 respectively, and are labeled accordingly.

4. The fault diagnosis method for weighing sensors based on the improved PSO-GRNN neural network according to claim 1, characterized in that, The signal features include the root mean square value. variance Peak indicators Pulse Indicators mean kurtosis index Maximum value Minimum value Peak-to-peak value Root amplitude Average amplitude Waveform Indicators and margin indicators The calculation formulas are as follows: （1） In formula (1), This represents the value of the collection point in a single sample; Indicates the sample dimension; （2） （3） （4） （5） （6） （7） （8） （9） （10） （11） （12） （13）。 5. The fault diagnosis method for weighing sensors based on the improved PSO-GRNN neural network according to claim 1, characterized in that, Step S5 specifically involves: Step 501: Determine the topology of the GRNN neural network. Step 502: Establish the initial particle population Set the relevant parameters of the particle swarm separately: initial parameters and Initial velocity matrix Initial optimal position of individual particles and global optimal position ; Step 503: Evaluate each particle in the particle population and obtain the fitness value of each particle; Step 504: Using the particle update formula (14) after introducing Lévy flight, complete the update of particles in the particle swarm, including the update of the state of each particle in the particle swarm and the optimal position of each individual particle. Updates and the globally optimal position Update; Step 505: When the number of updates reaches the set maximum number of training iterations, or when the particle fitness value of the updated result meets the requirements, training terminates and the final result is output. Otherwise, repeat steps 503 and 504 to continue evaluation and updates until the maximum number of training iterations is reached or the global optimum is found; Step 506: Global Optimal Position This is the global optimum, which is also the smoothing factor of the GRNN neural network structure parameters. The optimal solution.

6. The fault diagnosis method for weighing sensors based on the improved PSO-GRNN neural network according to claim 5, characterized in that, The topology includes: the number of nodes in the neural network input layer. Number of neurons in the pattern layer and the number of output layer nodes Number of nodes in the input layer Depends on the vector dimension of the input samples; number of neurons in the pattern layer The same as the number of training samples; the number of nodes in the output layer. It is equal to the dimension of the output vector.

7. The fault diagnosis method for weighing sensors based on the improved PSO-GRNN neural network according to claim 5, characterized in that, particle The fitness value refers to the value of fitness in the context of fitness. Particles in each support vector machine The sum of squares of the errors between predicted and actual values ​​is used to quantify the fitness of individual particles, allowing for natural selection. The fitness of each particle is then calculated using the formula. : （19）。 8. A fault diagnosis system implementing the fault diagnosis method for weighing sensors based on an improved PSO-GRNN neural network as described in any one of claims 1-7, characterized in that, The system includes: Data acquisition module: used to acquire the output signal of the weighing sensor to obtain the first dataset; Data processing module: used to perform data processing operations based on the first data information obtained by the data acquisition module, specifically including extracting labels and various features of vibration signals from the data, and finally obtaining a third dataset; Training module: Used to establish a fault diagnosis model based on the improved PSO-GRNN neural network. The improved PSO-GRNN neural network model is trained using the training set in the third dataset, and the accuracy of the fault diagnosis model based on the improved PSO-GRNN neural network is verified using the validation set in the third dataset. Fault diagnosis module: Real-time detection of the weighing sensor, processing of the data to obtain data samples, inputting the data samples into the trained fault diagnosis model, and finally obtaining the fault diagnosis result of the weighing sensor.

9. The fault diagnosis system according to claim 8, characterized in that, The data acquisition module connects to a data acquisition card connected to the weighing sensor. The data acquisition card acquires the sensor's vibration signal. After acquiring signals from multiple different states, a first dataset is obtained.

Images