Chemical process fault diagnosis and detection method and system

Abstract

The invention discloses a chemical process fault diagnosis and detection method and system. Belongs to the field of mechanical equipment fault diagnosis, and comprises the following detection steps: de-noising collected data by using a de-noising diffusion probability model (DDPM), and reducing noise interference in signals, thereby providing a clearer data basis for subsequent fault diagnosis; then, a fault diagnosis model (AResNet-Linformer) is built, the AResNet model is used for carrying out feature extraction on the denoised signals, and then feature information is sent to a Linformer neural network model for fault diagnosis; in order to improve the accuracy of the model, an improved star-sparrow optimization algorithm is used to carry out parameter optimization on the established fault diagnosis model. According to the invention, faults in the chemical process can be effectively identified, fault types can be rapidly determined, and a powerful data basis is provided for prevention and maintenance work.

Description

Technical Field

[0001] The present invention belongs to the technical field of engineering machinery transportation, and relates to a chemical process fault diagnosis and detection method and system; specifically, it relates to a centrifugal pump fault diagnosis method and system based on multi-sensor data fusion and deep learning model. Background Art

[0002] In the chemical production process, centrifugal pumps are one of the key equipment and are widely used in multiple links such as fluid transportation, cooling and pressurization. Due to the complex working conditions and large load fluctuations in the chemical process, centrifugal pumps are prone to failures during long-term operation, such as impeller damage, bearing wear, seal failure, etc. These failures may not only lead to production stagnation and affect the normal operation of the process flow, but also increase maintenance costs and even pose a threat to production safety. Therefore, timely and accurate fault diagnosis and detection of centrifugal pumps is the key to ensuring the stable operation of the chemical process and improving the operating efficiency of the equipment. Traditional fault diagnosis methods mostly rely on manual experience, regular inspection of equipment, and traditional monitoring methods such as vibration monitoring and temperature monitoring. However, these methods are usually limited by the interference of monitoring equipment and the environment, and cannot identify the fault type in real time and accurately, and it is difficult to detect potential fault problems in the early stage. In addition, traditional diagnostic methods rely heavily on operators and lack intelligent fault identification capabilities, which can easily lead to missed diagnosis or misdiagnosis.

[0003] With the rapid development of information technology, data-driven fault diagnosis methods have gradually become a hot topic of research. By collecting multidimensional data during the operation of centrifugal pumps, such as vibration signals, flow, pressure, etc., combined with advanced signal processing technology and intelligent algorithms, the accuracy and real-time performance of fault diagnosis can be improved. In particular, the advantages of deep learning technology in feature extraction and pattern recognition have significantly improved the accuracy and robustness of fault diagnosis. However, equipment data in chemical processes are often accompanied by strong noise. Directly using raw data for fault diagnosis will affect the performance of the model and lead to lower diagnostic accuracy. Summary of the invention

[0004] In view of the above problems, the purpose of the present invention is to propose a chemical process fault diagnosis and detection method and system, which can effectively identify the fault of the centrifugal pump in the chemical process, and quickly and accurately diagnose the fault type; by using the denoising diffusion probability model (DDPM) for data denoising, combining the adaptive residual network (AResNet) for deep feature extraction, and then using the Linformer neural network model for fault diagnosis, while using the improved starfinch optimization algorithm to optimize the model parameters, finally achieving accurate identification and classification of centrifugal pump faults. The present invention solves the problems of noise interference, insufficient feature extraction and insufficient model optimization in the prior art, thereby improving the accuracy and efficiency of fault diagnosis.

[0005] The technical solution of the present invention is: a chemical process fault diagnosis and detection method according to the present invention, the operation steps of which are as follows:

[0006] Step (1): collect the original signal data of the centrifugal pump through a variety of sensors, store the data in the form of time series, and classify the data to ensure that each piece of data has a correct label indicating the corresponding fault type;

[0007] Step (2): Use the denoising diffusion probability model to perform noise reduction on the collected signal data to ensure the accuracy of subsequent fault diagnosis;

[0008] Step (3): Build a fault diagnosis system that integrates AResNet and Linformer models, and use the improved starfinch optimization algorithm to optimize the parameters of the Linformer model;

[0009] Step (4): Determine the fault type based on the diagnosis result obtained in step (3); when a fault occurs, promptly feedback the fault information and perform on-site processing to ensure the normal operation and timely maintenance of the equipment.

[0010] Furthermore, in step (1), the original signal data includes vibration, temperature, current and pressure of the centrifugal pump.

[0011] Furthermore, in step (1), the steps of data collection are as follows:

[0012] Data collection from multiple sensors: Multiple sensors are installed at key locations of the centrifugal pump. The installed sensors collect multi-dimensional signal data in real time under the operating status of the equipment, including vibration signals x V (t), temperature signal x T (t), current signal x I (t) and pressure signal x P (t); the above data is stored in the form of a time series matrix to form a multi-channel input;

[0013] Construct the collected data into fault data H R =[H1,H2,...,H n ] is the input feature; where n is the number of variables, and R is the number of fault samples;

[0014] Fault types include: excessive vibration of the pump body, insufficient flow or reduced head, overheating of the motor, abnormal noise;

[0015] Add fault type label K = [K1, K2, ..., K m ] as the output feature; where m is the number of label types; construct a fault sample data set (H R , K m ).

[0016] Furthermore, in step (2), the steps of performing noise reduction processing on the collected data are as follows:

[0017] De-noising diffusion probability model processing: The original signal data is accompanied by noise and needs to be processed by noise reduction;

[0018] Adopting the denoising diffusion probability model, efficient noise reduction is achieved by simulating the forward noise addition process and the reverse denoising process of the signal;

[0019] Forward noise addition process: In the process of adding noise to the original signal, forward Gaussian noise will be added, with a total of T steps of noise addition. After multiple noise additions, the data will eventually become random noise;

[0020] Noise the data X at step t t The distribution of is in accordance with the following formula:

[0021]

[0022] The noise addition process generates a new Markov chain:

[0023]

[0024] In the formula, x t-1 represents the data generated by adding noise at the previous moment, x t represents the data generated by adding noise at the current moment, I represents the unit matrix, θ t represents the variance used at time t, X 1：T Represents the data of X from time 1 to time T;

[0025] Reverse denoising process: In the forward denoising process, the data Xt at time t is known. The reverse denoising process is to reversely solve the real distribution of the previous step step by step through the known noise signal Xt, but the reverse distribution b(x t |x t-1 ) is unknown; therefore, the model f θThe estimation process also follows the Markov chain:

[0026] f θ (x t-1 |x t )=N(x t-1 ; γ θ (x t ,t),∑ θ (X t ,t))

[0027] Where: θ represents the model parameters; γ θ (x t ,t) represents the learnable mean;∑ θ (X t ,t) represents the learnable variance;

[0028] Finally, the noise-reduced signal X is obtained. d , providing accurate data input for feature extraction.

[0029] Furthermore, in step (3), AResNet is used to extract features from the denoised signal in step (2), and then the feature information is divided into a training set and a test set;

[0030] Use the training set to train the Linformer model to obtain the optimal model parameters;

[0031] The trained AResNet-Linformer model is used to perform fault diagnosis on the divided test set, and the accuracy of the model on the test set is calculated to evaluate its diagnostic performance.

[0032] Furthermore, in step (3), the steps of constructing the AResNet-Linformer model are as follows:

[0033] The denoised feature vector is input into AResNet, and adaptive feature extraction of the signal is performed through Conv and three ARBUs to obtain adaptively amplified fault-sensitive features and suppressed irrelevant features;

[0034] Among them, Conv is the convolution operation, and its formula is as follows:

[0035]

[0036] In the formula, C i Represents the input of the i-th channel; Z j represents the output of the jth channel; J represents the convolution kernel; b represents the bias; N j Represents a method for calculating Z j The channel set of

[0037] ARBU adds an adaptive module to the original RBU. The adaptive module contains BN, ReLU, Conv and Sigmoid functions. The final output A is obtained by multiplying the adaptive coefficient and the feature b1 output by the original RBU branch. The output formula of the adaptive module is as follows:

[0038]

[0039] In the formula, μ i represents the i-th adaptive coefficient; C i Represents the characteristics of the i-th channel;

[0040] The output formula is as follows:

[0041] A=b1·μ i

[0042] As an adaptive residual network model, AResNet captures key feature information in the denoised signal. Its residual structure alleviates the gradient vanishing problem in deep networks through cross-layer connections, improving the accuracy and efficiency of feature extraction.

[0043] The Linformer model includes a sequence feature layer, an attention layer, and a neural network layer. The Linformer model takes the output matrix A of the AResNet model as input and outputs the fault feature sequence X through the position encoding layer and the embedding sublayer. q ={x q1 ,x q2 ,...,x qn};

[0044] The attention layer includes self-attention mechanism and multi-head attention mechanism. The sub-attention mechanism transforms the fault feature sequence into query vector Q, key vector W, and value vector E through a matrix. The conversion formula is as follows:

[0045] Q=X q R Q , W = X q R W , E = X q R E

[0046] In the formula, R Q , R W and R E Represent the transformation matrices of the three vectors Q, W, and E respectively;

[0047] Q and W represent the attributes of the feature to be processed and the target feature respectively. The self-attention mechanism learns the fault features by calculating the similarity between the two. The self-attention calculation formula is as follows:

[0048]

[0049] Where, d w Represents the scaling factor; the softmax function converts the similarity between the two into a weight value;

[0050] The multi-head attention mechanism speeds up feature learning by using multiple self-attention calculation modules in multiple spaces; its formula is as follows:

[0051] Thead(Q,W,E)=[head1,head2,head3]R O

[0052] In the formula, head1, head2 and head3 represent the calculation results of the self-attention mechanism in three different spaces; R O Represents the weight matrix connected to the calculation results of each self-attention module;

[0053] The neural network layer obtains feature expressions by performing nonlinear transformations on features in different spaces to achieve accurate fault diagnosis. The neural network calculates the corresponding attention weights, calculates the feature information of the self-attention sum, and establishes a mapping between fault features and fault categories.

[0054] The output formula of the neural network layer is:

[0055] FF(P)=max(0,PW1+z1)W2+z2

[0056] Where P represents the calculation result of multi-head attention; W1 and W2 represent the weight matrix of the network; max represents the activation function; z1 and z2 represent two bias parameters.

[0057] Furthermore, in step (3), the specific steps of the star sparrow optimization algorithm are as follows:

[0058] (1) Initialization: First, assume that the size of the population is N and the dimension of the problem is D; the initialization formula is as follows:

[0059]

[0060] Where, t represents the number of iterations; n = 1, 2, ...; N; m = 1, 2, ..., D; represents the m-dimensional vector of the nth nutcracker at the current iteration number; U m and L m Respectively represent the upper and lower bounds of the m-dimensional variable; represents a random vector between [0,1];

[0061] (2) Position update: divided into exploration behavior and development behavior;

[0062] The exploration behavior is to randomly search new areas with a relatively large step size. The formula is as follows:

[0063] X i (t+1)=X i (t)+r1·(X best -X i (t))

[0064] Where, X i (t) represents the position of the ith starfinch in the tth generation, X best represents the position of the starfinch with the best fitness in the group, r1 represents a random number, and its value is between 0 and 1;

[0065] The development behavior is to perform local optimization in the current neighborhood with a smaller step size, and its formula is as follows:

[0066] X i (t+1)=X i (t)+r2·(X leader -X i (t))

[0067] Where, X leader is the current best star position, r2 is a random number between 0 and 1;

[0068] (3) Fitness evaluation: After each position update, the position of the starfinch is evaluated for its fitness according to the objective function, and the optimal solution in the group is updated; the fitness function F(X i ) is usually an objective function associated with an optimization problem, evaluating the goodness of each new position of the starfinch;

[0069] (4) Termination condition: The Starfinch optimization algorithm will continue to iterate until the preset termination condition is reached; when the maximum number of iterations T is reached max Or the fitness is less than the set threshold ∈; the termination condition formula is as follows:

[0070] Termination Condition

[0071] In the formula, t represents the current iteration number, and ε represents the allowable error;

[0072] (5) Return the optimal solution: When the algorithm meets the termination condition, the optimal solution is output.

[0073] Furthermore, the improvement steps of the starfinch optimization algorithm are as follows:

[0074] (1) Good point set strategy: The initial position is randomly generated during initialization, and its traversability and diversity cannot be guaranteed. The low quality of the population affects the convergence speed of the algorithm, so the good point set is introduced to uniformly select points; the point set formula is as follows:

[0075]

[0076] Based on the good point set theory, the new initialization strategy formula is as follows:

[0077] x i (k)=(upper j -low j ){P n (k)}+low j

[0078] In the formula, upper represents the upper bound, and low represents the lower bound;

[0079] (2) Adaptive adjustment of learning factor strategy: According to the current fitness and progress of the group, dynamically adjust the values ​​of c1 and c2 to make the balance between the exploration phase and the development phase more appropriate; define learning factors c1 and c2 as variables that change over time, and the learning factor formula is as follows:

[0080]

[0081] In the formula, t represents the current iteration number; c 1_max and c 1_min represents the maximum and minimum value of the learning factor c1; c 2_max and c 2_min Represents the maximum and minimum values ​​of the learning factor c2;

[0082] The improved starfinch optimization algorithm is used to optimize the learning rate and weight of the Linformer model. The model is trained using the training set to obtain the optimal model parameters, and the trained model is used to perform fault diagnosis on the test set data.

[0083] Furthermore, in step (4), the specific steps of fault judgment and feedback are as follows:

[0084] (1) Fault judgment: After the model optimized by the improved Starbird optimization algorithm outputs the fault type label, the system determines whether a fault has occurred based on the output result;

[0085] Specifically, if the fault type label output by the model is an abnormal category of Fault 1 or Fault 2, the system determines that the corresponding fault has occurred and further analyzes the severity and impact of the fault;

[0086] If the output result is labeled "Normal", it means that the system is in normal operation and no further processing is required;

[0087] (2) Fault information feedback: Once a fault is identified and confirmed, the system will automatically generate a detailed fault report, which includes the fault type, occurrence time, possible cause of the fault, and impact on production. The fault report is sent promptly through the real-time communication network to ensure that the fault information is received in a timely manner;

[0088] (3) On-site processing suggestions: Based on the fault type and the output results of the model, the system provides targeted maintenance suggestions; the fault report contains the recommended maintenance steps, required spare parts and technical documents;

[0089] (4) Fault handling tracking: After the fault information is fed back, the system continues to track the progress of fault handling to ensure that the fault is detected and repaired in a timely manner based on the information provided.

[0090] Further, a chemical process fault diagnosis and detection system includes a data acquisition and classification unit, a data denoising unit, a feature extraction unit, a fault diagnosis unit, a model optimization unit and a fault diagnosis result output unit;

[0091] The system determines the fault type based on the final diagnosis results and feeds back the fault information through the interface to provide support for subsequent equipment maintenance and fault warning.

[0092] The beneficial effects of the present invention are as follows: firstly, by introducing a denoising diffusion probability model (DDPM) to denoise the collected signal, noise interference is significantly reduced, the quality of the signal is improved, and more accurate input data is provided for subsequent fault diagnosis; secondly, by combining the AResNet-Linformer model, not only can deep features be effectively extracted, but also efficient fault diagnosis can be performed through the Linformer model, thereby improving the accuracy and real-time performance of fault classification; in addition, by optimizing the model parameters in combination with the improved starbird optimization algorithm, the performance of the diagnosis model is further improved, so that the method has stronger robustness and accuracy when dealing with fault diagnosis tasks in complex chemical processes; therefore, the present invention provides an efficient and reliable chemical process fault diagnosis method, which can provide strong data support for preventive maintenance and fault detection of equipment, and reduce human misjudgment and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 It is the overall operation flow chart of the present invention;

[0094] Figure 2 It is a specific flow chart of DDPM in the present invention;

[0095] Figure 3 It is a flow chart of the overall structure of the system of the present invention. DETAILED DESCRIPTION

[0096] The specific technical scheme of the present invention is further described in detail below with reference to specific examples.

[0097] As shown in the figure, the present invention discloses a chemical process fault diagnosis and detection method, comprising the following steps:

[0098] Step (1), collecting the original signal data of the centrifugal pump through a variety of sensors, including vibration, temperature, current, pressure, etc., and storing the data in a time series form, and classifying the data to ensure that each piece of data has a correct label indicating the corresponding fault type;

[0099] The data collection steps are as follows:

[0100] Multi-sensor data acquisition: Multiple sensors are installed at key parts of the centrifugal pump, such as vibration sensors, temperature sensors, current sensors and pressure sensors; these sensors can collect multi-dimensional signal data under the operation status of the equipment in real time, including vibration signal x V (t), temperature signal x T (t), current signal x I (t) and pressure signal x P (t); These data are stored in the form of a time series matrix to form a multi-channel input, laying the foundation for subsequent analysis; the collected data is constructed into fault data H R =[H1,H2,...,H n ] is the input feature; where n is the number of variables and R is the number of fault samples;

[0101] Fault types include: excessive vibration of the pump body, insufficient flow or head drop, motor overheating, abnormal noise; add fault type label K = [K1, K2, ..., K m ] as the output feature, where m is the number of label types; construct a fault sample data set (H R , K m );

[0102] Step (2), using a denoising diffusion probability model (DDPM) to perform denoising on the signal data collected in step (1) to reduce noise interference in the signal and ensure the accuracy of subsequent fault diagnosis;

[0103] The steps of collecting data denoising are as follows:

[0104] De-noising Diffusion Probability Model (DDPM) processing: The original signal data is usually accompanied by noise and needs to be de-noised. The De-noising Diffusion Probability Model (DDPM) is used to achieve efficient noise reduction by simulating the forward noise addition process and the reverse noise removal process of the signal.

[0105] Forward noise addition process: In the process of adding noise to the original signal, forward Gaussian noise will be continuously added, with a total of T steps of noise addition. After enough noise additions, the data will eventually become random noise;

[0106] Noise the data X at step t t The distribution of is in accordance with the following formula:

[0107]

[0108] The noise addition process generates a Markov chain:

[0109]

[0110] In the formula, x t-1 Indicates the data generated by adding noise at the previous moment; x t represents the data generated by adding noise at the current moment; I represents the unit matrix; θ t represents the variance used at time t; X 1：T Represents the data of X from time 1 to time T;

[0111] Here the generated X t The same dimension as the original data;

[0112] Reverse denoising process: In the forward denoising process, the data Xt at time t is known. The reverse denoising process is to reversely solve the real distribution of the previous step step by step through the known noise signal Xt, but the reverse distribution b(x t |x t-1 ) is unknown;

[0113] So we build the model f θ The estimation process also follows the Markov chain:

[0114] f θ (x t-1 |x t )=N(x t-1 ; γ θ (x t ,t),∑ θ (X t ,t))

[0115] Where: θ is the model parameter; γ θ (x t ,t) is the learnable mean;∑ θ (X t ,t) is the learnable variance;

[0116] Finally, the noise-reduced signal X is obtained. d , providing more accurate data input for feature extraction;

[0117] Step (3), constructing a fault diagnosis system integrating an adaptive residual network (AResNet) and a Linformer model, and optimizing the parameters of the Linformer model using an improved star-sparrow optimization algorithm (NOA);

[0118] AResNet is used to extract features from the denoised signal in step (2), and then the feature information is divided into a training set and a test set; the Linformer model is trained using the training set to obtain the optimal model parameters; the trained AResNet-Linformer model is used to perform fault diagnosis on the divided test set, the accuracy of the model on the test set is calculated, and its diagnostic performance is evaluated;

[0119] The steps for building the AResNet-Linformer model are as follows:

[0120] The denoised feature vector is input into AResNet, and the adaptive feature extraction of the signal is performed through the convolution layer (Conv) and three adaptive residual unit (ARBU) layers to obtain adaptive amplification of fault sensitive features and suppression of irrelevant features; Conv is mainly a convolution operation, and the formula is as follows:

[0121]

[0122] In the formula, C i is the input of the ith channel; Z j is the output of the jth channel; J is the convolution kernel; b is the bias; N j is a method for calculating Z j The channel set of

[0123] ARBU adds an adaptive module to the original residual unit (RBU). The adaptive module includes BN, ReLU, Conv and Sigmoid functions. The final output A is obtained by multiplying the adaptive coefficient and the feature b1 output by the original RBU branch. The output formula of the adaptive module is as follows:

[0124]

[0125] In the formula, μ i is the i-th adaptive coefficient; C i is the feature of the i-th channel;

[0126] The output formula is as follows:

[0127] A=b1·μ i

[0128] As an adaptive residual network model, AResNet can effectively capture the key feature information in the denoised signal; its residual structure alleviates the gradient vanishing problem in deep networks through cross-layer connections, thereby improving the accuracy and efficiency of feature extraction;

[0129] The Linformer model includes a sequence feature layer, an attention layer, and a neural network layer. The Linformer model takes the output matrix A of the AResNet model as input, and outputs the fault feature sequence X through the position encoding layer and the embedding sublayer. q ={x q1 ,x q2 ,...,x qn}; The attention layer includes self-attention mechanism and multi-head attention mechanism, in which the sub-attention mechanism transforms the fault feature sequence into query vector Q, key vector W, and value vector E through a matrix; the conversion formula is as follows:

[0130] Q=X q R Q , W = X q R W , E = X q R E

[0131] In the formula, R Q , R W and R E They are the transformation matrices of the three vectors Q, W, and E;

[0132] Q and W represent the attributes of the feature to be processed and the target feature respectively. The self-attention mechanism mainly learns the fault features by calculating the similarity between the two. The self-attention calculation formula is as follows:

[0133]

[0134] Where, d w is the scaling factor; the softmax function converts the similarity between the two into a weight value;

[0135] The multi-head attention mechanism speeds up feature learning by using multiple self-attention calculation modules in multiple spaces; the formula is as follows:

[0136] Thead(Q,W,E)=[head1,head2,head3]R O

[0137] Where head1, head2 and head3 are the calculation results of the self-attention mechanism in three different spaces; R O It is the weight matrix connected to the calculation results of each self-attention module;

[0138] The neural network layer obtains a more advanced feature expression by performing nonlinear transformation on the features in different spaces, thus achieving accurate fault diagnosis. The neural network calculates the corresponding attention weights and the feature information of the total self-attention, thus establishing a mapping between fault features and fault categories.

[0139] Here is the output formula of the neural network layer:

[0140] FF(P)=max(0,PW1+z1)W2+z2

[0141] Where P is the calculation result of multi-head attention; W1 and W2 are the weight matrices of the network; max is the activation function; z1 and z2 are two bias parameters;

[0142] As an improved neural network structure, the Linformer model significantly reduces the computational complexity of the traditional attention mechanism through the linearized self-attention mechanism; this optimization enables the model to maintain diagnostic accuracy while significantly improving computational efficiency when processing long time series data; the construction of the AResNet-Linformer model closely combines feature extraction with fault classification to form an efficient and accurate diagnostic framework; this framework can adapt to various types of signal inputs, providing an important guarantee for the stable operation of chemical process equipment;

[0143] In order to further improve the performance of the fault diagnosis model, the improved starfinch optimization algorithm (NOA) is used to optimize the key parameters of the Linformer model; these parameters include the number of iterations, model dimension and batch size, which play a vital role in the model's convergence speed, accuracy and consumption of computing resources during the model training process;

[0144] The steps of the starfinch optimization algorithm in step (3) are as follows:

[0145] The process of the Noah Optimization Algorithm (NOA) is as follows:

[0146] (1) Initialization: First, assume that the size of the population is N and the dimension of the problem is D. The initialization formula is as follows:

[0147]

[0148] Where, t represents the number of iterations; n = 1, 2, ...; N; m = 1, 2, ..., D; represents the m-dimensional vector of the nth nutcracker at the current iteration number; U m and L m Respectively represent the upper and lower bounds of the m-dimensional variable; represents a random vector between [0,1];

[0149] (2) Position update: divided into exploration behavior and development behavior;

[0150] The exploration behavior is to randomly search new areas with a relatively large step size. The formula is as follows:

[0151] X i (t+1)=X i (t)+r1·(X best -X i (t))

[0152] In the formula, X i (t) is the position of the ith starfinch in the tth generation, X best is the position of the starfinch with the best fitness in the group, r1 is a random number (between 0 and 1);

[0153] The development behavior is to perform local optimization in the current neighborhood with a smaller step size, and the formula is as follows:

[0154] X i (t+1)=X i (t)+r2·(X leader -X i (t))

[0155] In the formula, X leader is the position of the current best star bird, r2 is a random number between (0,1);

[0156] (3) Fitness evaluation: After each position update, the position of the starfinch will evaluate its fitness according to the objective function and update the optimal solution in the group; the fitness function F(X i ) is usually an objective function associated with an optimization problem, evaluating the goodness of each new position of the starfinch;

[0157] (4) Termination condition: The Starfinch optimization algorithm will continue to iterate until the preset termination condition is reached; when the maximum number of iterations T is reached max Or the fitness is less than the set threshold ε; the termination condition formula is as follows:

[0158] Termination Condition

[0159] Where t is the current iteration number, ∈ is the allowable error;

[0160] Return the optimal solution: When the algorithm meets the termination condition, the optimal solution is output;

[0161] Furthermore, the improvement steps of the starfinch optimization algorithm are as follows: Although the starfinch optimization algorithm performs well in solving many problems, it still has shortcomings. In order to avoid the low efficiency of the algorithm search or falling into the local optimum, a hybrid strategy and an adaptive adjustment learning factor method are introduced to enhance its exploration ability and local search accuracy in the solution space, thereby improving the optimization effect and the overall performance of the algorithm;

[0162] 1. Good point collection strategy:

[0163] The initial position is randomly generated during initialization, and its traversability and diversity cannot be guaranteed. The low quality of the population affects the convergence speed of the algorithm, so the introduction of the good point set can effectively select points evenly; the point set formula is as follows:

[0164]

[0165] Based on the good point set theory, the new initialization strategy formula is as follows:

[0166] x i (k)=(upper j -low j ){P n (k)}+low j

[0167] In the formula, upper is the upper bound and low is the lower bound;

[0168] 2. Adaptive adjustment of learning factor strategy:

[0169] According to the current fitness and progress of the group, the values ​​of c1 and c2 are dynamically adjusted to make the balance between the exploration phase and the development phase more appropriate; the learning factors c1 and c2 are defined as variables that change over time. The learning factor formula is as follows:

[0170]

[0171] Where t is the current iteration number; c 1_max and c 1_min is the maximum and minimum value of the learning factor c1; c 2_max and c 2_min is the maximum and minimum value of the learning factor c2;

[0172] Use the improved Starfinch optimization algorithm to optimize the learning rate and weight of the Linformer model, use the training set to fully train the model to obtain the optimal model parameters, and use the trained model to perform fault diagnosis on the test set data;

[0173] The improved Starfinch Optimization Algorithm (NOA) balances the capabilities of global search and local search by enhancing the position update mechanism, effectively avoiding the problem that traditional optimization algorithms are prone to fall into local optimality; the improved Starfinch Algorithm significantly reduces the number of iterations during the optimization process, effectively shortening the training time of the model. Combined with the efficient structure of the AResNet-Linformer model, the real-time performance of the entire system is excellent;

[0174] Step (4), judging the fault type according to the diagnosis result obtained in step (3); when a fault occurs, promptly feedback the fault information and perform on-site processing to ensure the normal operation and timely maintenance of the equipment;

[0175] Among them, fault judgment and feedback are as follows:

[0176] 1. Fault judgment: After the model optimized by the improved Starbird optimization algorithm outputs the fault type label, the system determines whether a fault has occurred based on the output result; specifically, if the fault type label output by the model is an abnormal category such as Fault 1 or Fault 2, the system will determine that a corresponding fault has occurred, and further analyze the severity and impact of the fault; if the output result is a "normal" label, it means that the system is in normal operation and no further processing is required;

[0177] 2. Fault information feedback: Once a fault is identified and confirmed, the system will automatically generate a detailed fault report, which includes the fault type, occurrence time, possible cause of the fault, impact on production, etc. The fault report is sent to relevant technicians or maintenance teams through a real-time communication network to ensure that they can receive fault information in a timely manner;

[0178] 3. On-site processing suggestions: Based on the fault type and the output results of the model, the system can also provide targeted maintenance suggestions. The fault report can include recommended maintenance steps, required spare parts, technical documents, etc., to help technicians quickly locate the problem and take effective on-site processing measures;

[0179] 4. Fault handling tracking: After the fault information is fed back, the system continues to track the progress of fault handling to ensure that the technicians conduct timely troubleshooting and repairs based on the information provided.

[0180] Further, a chemical process fault diagnosis and detection system includes: a data acquisition and classification unit for acquiring raw signal data of a centrifugal pump through a variety of sensors, including information such as vibration, temperature, current, pressure, etc., and storing the acquired data in time series;

[0181] Classify the collected data to ensure that each data point corresponds to a different fault type label and build a fault data set with labels;

[0182] A data denoising unit is used to denoise the original signal collected in step (1) by using a denoising diffusion probability model (DDPM); by effectively removing the noise in the signal, the reliability of the data is enhanced, and a clear data basis is provided for subsequent feature extraction and fault diagnosis;

[0183] The feature extraction unit uses the adaptive residual network (AResNet) to perform deep feature extraction on the denoised signal. The AResNet model can extract meaningful fault feature information from the noise-removed data through the residual learning mechanism, providing effective input for fault diagnosis.

[0184] The fault diagnosis unit uses the Linformer model to perform fault diagnosis on the extracted features. The Linformer model uses a self-attention mechanism to effectively classify and diagnose fault types by processing long time series data.

[0185] The model optimization unit uses the improved star-sparrow optimization algorithm (NOA) to optimize the parameters of the fault diagnosis model (Linformer). Through the star-sparrow optimization algorithm, a balance is maintained between global search and local search, the network parameters of Linformer are optimized, and the accuracy and stability of the diagnosis model are improved.

[0186] The fault diagnosis result output unit combines the diagnosis result with the optimized model result to obtain the final fault diagnosis output.

[0187] The system determines the fault type based on the final diagnosis results and feeds back the fault information to the technicians through the interface to provide support for subsequent equipment maintenance and fault warning.

Claims

1. A chemical process fault diagnosis and detection method, characterized in that: The operation steps are as follows: Step (1): collect the original signal data of the centrifugal pump through a variety of sensors, store the data in the form of time series, and classify the data to ensure that each piece of data has a correct label indicating the corresponding fault type; Step (2): Use the denoising diffusion probability model to perform noise reduction on the collected signal data to ensure the accuracy of subsequent fault diagnosis; Step (3): Build a fault diagnosis system that integrates AResNet and Linformer models, and use the improved starfinch optimization algorithm to optimize the parameters of the Linformer model; Step (4): Determine the fault type based on the diagnosis result obtained in step (3); when a fault occurs, promptly feedback the fault information and perform on-site processing to ensure the normal operation and timely maintenance of the equipment.

2. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (1), the original signal data includes vibration, temperature, current and pressure of the centrifugal pump.

3. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (1), the steps of data collection are as follows: Data collection from multiple sensors: Multiple sensors are installed at key locations of the centrifugal pump. The installed sensors collect multi-dimensional signal data in real time under the operating status of the equipment, including vibration signals x V (t), temperature signal x T (t), current signal x I (t) and pressure signal x P (t); the above data is stored in the form of a time series matrix to form a multi-channel input; Construct the collected data into fault data H R =[H1,H2,...,H n ] is the input feature; where n is the number of variables, and R is the number of fault samples; Fault types include: excessive vibration of the pump body, insufficient flow or reduced head, overheating of the motor, abnormal noise; Add fault type label K = [K1, K2, ..., K m ] as the output feature; where m is the number of label types; construct a fault sample data set (H R , K m ).

4. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (2), the steps of performing noise reduction processing on the collected data are as follows: De-noising diffusion probability model processing: The original signal data is accompanied by noise and needs to be processed by noise reduction; Adopting the denoising diffusion probability model, efficient noise reduction is achieved by simulating the forward noise addition process and the reverse denoising process of the signal; Forward noise addition process: In the process of adding noise to the original signal, forward Gaussian noise will be added, with a total of T steps of noise addition. After multiple noise additions, the data will eventually become random noise; Noise the data X at step t t The distribution of is in accordance with the following formula: The noise adding process generates a new Markov chain, namely: In the formula, x t-1 represents the data generated by adding noise at the previous moment, x t represents the data generated by adding noise at the current moment, I represents the unit matrix, θ t represents the variance used at time t, X 1：T Represents the data of X from time 1 to time T; Reverse denoising process: In the forward denoising process, the data Xt at time t is known. The reverse denoising process is to reversely solve the real distribution of the previous step step by step through the known noise signal Xt, but the reverse distribution b(x t |x t-1 ) is unknown; therefore, the model f θ The estimation process also follows the Markov chain: f θ (x t-1 |x t )=N(x t-1 ;γ θ (x t ,t),∑ θ (X t ,t)) Where: θ represents the model parameters; γ θ (x t ,t) represents the learnable mean;∑ θ (X t ,t) represents the learnable variance; Finally, the noise-reduced signal X is obtained. d , providing accurate data input for feature extraction.

5. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (3), AResNet is used to extract features from the denoised signal in step (2), and then the feature information is divided into a training set and a test set; Use the training set to train the Linformer model to obtain the optimal model parameters; The trained AResNet-Linformer model is used to perform fault diagnosis on the divided test set, and the accuracy of the model on the test set is calculated to evaluate its diagnostic performance.

6. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (3), the steps of constructing the AResNet-Linformer model are as follows: The denoised feature vector is input into AResNet, and adaptive feature extraction of the signal is performed through Conv and three ARBUs to obtain adaptively amplified fault-sensitive features and suppressed irrelevant features; Among them, Conv is the convolution operation, and its formula is as follows: In the formula, C i Represents the input of the i-th channel; Z j represents the output of the jth channel; J represents the convolution kernel; b represents the bias; N j Represents a method for calculating Z j The channel set of ARBU adds an adaptive module to the original RBU. The adaptive module contains BN, ReLU, Conv and Sigmoid functions. The final output A is obtained by multiplying the adaptive coefficient and the feature b1 output by the original RBU branch. The output formula of the adaptive module is as follows: In the formula, μ i represents the i-th adaptive coefficient; C i Represents the characteristics of the i-th channel; The output formula is as follows: A=b1·μ i As an adaptive residual network model, AResNet captures key feature information in the denoised signal. Its residual structure alleviates the gradient vanishing problem in deep networks through cross-layer connections, improving the accuracy and efficiency of feature extraction. The Linformer model includes a sequence feature layer, an attention layer, and a neural network layer. The Linformer model takes the output matrix A of the AResNet model as input and outputs the fault feature sequence X through the position encoding layer and the embedding sublayer. q ={x q1 ,x q2 ,...,x qn }; The attention layer includes self-attention mechanism and multi-head attention mechanism. The sub-attention mechanism transforms the fault feature sequence into query vector Q, key vector W, and value vector E through a matrix. The conversion formula is as follows: Q=X q R Q ,W=X q R W ,E=X q R E In the formula, R Q , R W and R E Represent the transformation matrices of the three vectors Q, W, and E respectively; Q and W represent the attributes of the feature to be processed and the target feature respectively. The self-attention mechanism learns the fault features by calculating the similarity between the two. The self-attention calculation formula is as follows: Where, d w Represents the scaling factor; the softmax function converts the similarity between the two into a weight value; The multi-head attention mechanism speeds up feature learning by using multiple self-attention calculation modules in multiple spaces; its formula is as follows: Thead(Q,W,E)=[head1,head2,head3]R O In the formula, head1, head2 and head3 represent the calculation results of the self-attention mechanism in three different spaces; R O Represents the weight matrix connected to the calculation results of each self-attention module; The neural network layer obtains feature expressions by performing nonlinear transformations on features in different spaces to achieve accurate fault diagnosis. The neural network calculates the corresponding attention weights, calculates the feature information of the self-attention sum, and establishes a mapping between fault features and fault categories. The output formula of the neural network layer is: FF(P)=max(0,PW1+z1)W2+z2 Where P represents the calculation result of multi-head attention; W1 and W2 represent the weight matrix of the network; max represents the activation function; z1 and z2 represent two bias parameters.

7. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (3), the specific steps of the star sparrow optimization algorithm are as follows: (1) Initialization: First, assume that the size of the population is N and the dimension of the problem is D; the initialization formula is as follows: Where, t represents the number of iterations; n = 1, 2, ...; N; m = 1, 2, ..., D; represents the m-dimensional vector of the nth nutcracker at the current iteration number; U m and L m Respectively represent the upper and lower bounds of the m-dimensional variable; represents a random vector between [0,1]; (2) Position update: divided into exploration behavior and development behavior; The exploration behavior is to randomly search new areas with a relatively large step size. The formula is as follows: X i (t+1)=X i (t)+r1·(X best -X i (t)) In the formula, X i (t) represents the position of the ith starfinch in the tth generation, X best represents the position of the starfinch with the best fitness in the group, r1 represents a random number, and its value is between 0 and 1; The development behavior is to perform local optimization in the current neighborhood with a smaller step size, and its formula is as follows: X i (t+1)=X i (t)+r2·(X leader -X i (t)) In the formula, X leader is the current best star position, r2 is a random number between 0 and 1; (3) Fitness evaluation: After each position update, the position of the starfinch is evaluated for its fitness according to the objective function, and the optimal solution in the group is updated; the fitness function F(X i ) is usually an objective function associated with an optimization problem, evaluating the goodness of each new position of the starfinch; (4) Termination condition: The Starfinch optimization algorithm will continue to iterate until the preset termination condition is reached; when the maximum number of iterations T is reached max Or the fitness is less than the set threshold ∈; The termination condition formula is as follows: Termination Condition In the formula, t represents the current iteration number, ∈ represents the allowable error; (5) Return the optimal solution: When the algorithm meets the termination condition, the optimal solution is output.

8. A chemical process fault diagnosis and detection method according to claim 7, characterized in that: The improvement steps of the starfinch optimization algorithm are as follows: (1) Good point set strategy: The initial position is randomly generated during initialization, and its traversability and diversity cannot be guaranteed. The low quality of the population affects the convergence speed of the algorithm, so the good point set is introduced to uniformly select points; the point set formula is as follows: Based on the good point set theory, the new initialization strategy formula is as follows: x i (k)=(upper j -low j ){P n (k)}+low j In the formula, upper represents the upper bound, and low represents the lower bound; (2) Adaptive adjustment of learning factor strategy: According to the current fitness and progress of the group, dynamically adjust the values ​​of c1 and c2 to make the balance between the exploration phase and the development phase more appropriate; define learning factors c1 and c2 as variables that change over time, and the learning factor formula is as follows: In the formula, t represents the current iteration number; c 1_max and c 1_min represents the maximum and minimum value of the learning factor c1; c 2_max and c 2_min Represents the maximum and minimum values ​​of the learning factor c2; Use the improved Starfinch optimization algorithm to optimize the learning rate and weight of the Linformer model; Use the training set to train the model to obtain the optimal model parameters, and use the trained model to perform fault diagnosis on the test set data.

9. A chemical process fault diagnosis and detection method according to claim 1, characterized in that: In step (4), the specific steps of fault judgment and feedback are as follows: (1) Fault judgment: After the model optimized by the improved Starbird optimization algorithm outputs the fault type label, the system determines whether a fault has occurred based on the output result; Specifically, if the fault type label output by the model is an abnormal category of Fault 1 or Fault 2, the system determines that the corresponding fault has occurred and further analyzes the severity and impact of the fault; If the output result is labeled "Normal", it means that the system is in normal operation and no further processing is required; (2) Fault information feedback: Once a fault is identified and confirmed, the system will automatically generate a detailed fault report, which includes the fault type, occurrence time, possible cause of the fault, and impact on production. The fault report is sent promptly through the real-time communication network to ensure that the fault information is received in a timely manner; (3) On-site processing suggestions: Based on the fault type and the output results of the model, the system provides targeted maintenance suggestions; The fault report contains recommended repair steps, required spare parts and technical documentation; (4) Fault handling tracking: After the fault information is fed back, the system continues to track the progress of fault handling to ensure that the fault is detected and repaired in a timely manner based on the information provided.

10. A chemical process fault diagnosis and detection system, characterized in that: It includes a data acquisition and classification unit, a data denoising unit, a feature extraction unit, a fault diagnosis unit, a model optimization unit and a fault diagnosis result output unit.

Images