A method and device for intelligent diagnosis of important plant water pumps

Abstract

The present invention relates to a method and device for intelligent diagnosis of important plant water pumps, the method comprising: S1: collecting vibration signals of important plant water pumps under different working conditions, and obtaining sample points of vibration signals containing different fault states; S2: performing wavelet threshold denoising preprocessing on the sample points; S3: establishing a fault game model to provide an interactive environment for observation, action and reward acquisition for the fault diagnosis agent; S4: reducing the dimension and extracting features of multiple hidden layers by stacking autoencoder neural networks, optimizing the initial parameters by using BP neural networks, and obtaining a deep neural network model with feature extraction and pattern recognition functions; S5: outputting the diagnosis result. The present invention can effectively establish a deep neural network model by designing a fault diagnosis game environment according to vibration signals in different states, combining reinforcement learning with deep learning, and realizing intelligent diagnosis, and has a good diagnostic effect.

Description

Technical Field

[0001] The invention relates to the fields of mechanical fault diagnosis and computer artificial intelligence, and in particular to a method and a device for intelligent diagnosis of an important plant water pump. Background Art

[0002] Important plant water pumps are important nuclear safety level 3 equipment in nuclear power plants. They serve the heat exchangers in the cooling water system of nuclear power plant equipment. The seawater transported by important plant water pumps transfers the heat in the heat exchanger to nature (the sea) to ensure the safe and reliable operation of various equipment in nuclear power plants. Failure of the pump will lead to performance degradation, which may cause huge economic losses or even catastrophic accidents.

[0003] In the prior art, the commonly used rotating machinery fault diagnosis method collects signal data and preprocesses the data by methods such as empirical mode decomposition, local mean decomposition, and wavelet packet decomposition. After the raw data is preprocessed, feature extraction and fault identification are required. Deep learning methods have been successfully applied in the field of mechanical fault diagnosis because they can automatically learn representative features from data. However, deep learning models still have shortcomings: (1) It is impossible to establish a direct linear or nonlinear mapping between the raw data and the corresponding fault mode, and the performance of these fault diagnosis methods depends on the quality of building and training deep models. (2) The training mechanism of this method is mainly based on supervised learning or semi-supervised learning, which means that the diagnosis algorithm requires an expert system to specifically learn different fault modes.

[0004] By designing a fault diagnosis game environment, combining reinforcement learning with deep learning, and allowing artificial agents to learn their knowledge and experience directly from data, this method can successfully use DRL (deep reinforcement learning) to obtain intelligent fault diagnosis agents. It can effectively establish an unsupervised diagnosis model, realize intelligent diagnosis, and have a good diagnostic effect. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies in the prior art and to provide a method and device for intelligent diagnosis of important plant water pumps that are simple, efficient, stable and accurate.

[0006] The technical solution of the present invention is as follows: A method for intelligent diagnosis of important plant water pumps comprises the following steps:

[0007] (S1) collecting vibration signals of important plant water pumps under different working conditions to obtain sample points of vibration signals containing different fault states;

[0008] (S2) performing wavelet threshold denoising preprocessing on the sample points;

[0009] (S3) Establish a fault game model to provide an interactive environment for the fault diagnosis agent to observe, act and obtain rewards;

[0010] (S4) reducing the dimensions and extracting features of multiple hidden layers by stacking autoencoder neural networks, and optimizing initial parameters by using BP neural networks to obtain a deep neural network model with feature extraction and pattern recognition functions;

[0011] (S5) Output the diagnosis result.

[0012] Furthermore, in the above-mentioned method for intelligent diagnosis of important plant water pumps, in step (S1), a measurement system composed of an acceleration sensor and LabVIEW software is used to collect vibration signals under different fault states.

[0013] Furthermore, in step (S1), the causes of the fault include: severe wear and cavitation at the mouth ring, the center of the rotor does not coincide with the center of the volute, the blades are not properly aligned with the diffuser, the inlet and outlet pipelines are not designed reasonably, the inlet and outlet pipelines have no straight pipe section or the straight pipe section is too short (especially the inlet has a large impact), causing turbulence before the fluid enters the impeller, the blades do not operate under the design conditions (such as operating at ultra-small flow or ultra-large flow), the fluid is prone to turbulence, flow separation, impact, cavitation, random and uneven impact on the blades, etc.

[0014] Further, in the above-mentioned method for intelligent diagnosis of important plant water pumps, the method for performing denoising preprocessing on the sample points in step (S2) comprises the following steps:

[0015] (S201) De-noising the noise data by improving the threshold function, and selecting the improved threshold th as follows:

[0016]

[0017] Among them, σ 2 is the noise variance, σ=median, median is the median of the absolute values ​​of high-frequency wavelet coefficients, n is the signal length, j is the number of wavelet decomposition layers, and β is the control factor.

[0018] (S202) extracting wavelet packet energy of the denoised data using wavelet packet transform;

[0019] (S203) performing wavelet packet decomposition on the vibration signal of the fault state, selecting the wavelet basis as db3 and the decomposition level as 4;

[0020] (S204) After decomposition, the wavelet packet energy of each node is obtained by the following formula:

[0021] E jn =1M∑t=1M((X jn (t)) 2

[0022] Among them, X jn (t) represents the time domain vibration signal, t represents time, and M represents the node X jn The number of samples in .

[0023] Further, in the above-mentioned method for intelligent diagnosis of important plant water pumps, the specific process of step (S3) includes:

[0024] (S301) generating a fault diagnosis question;

[0025] (S302) accepting diagnostic questions and providing result feedback via an agent;

[0026] (S303) The game model determines whether the diagnosis result is correct. If it is correct, the total reward is increased by 1 and the game reward value is output. Otherwise, the total reward is reduced by 1 and the game model is returned to (S301) to regenerate the fault diagnosis problem.

[0027] Further, in the above-mentioned method for intelligent diagnosis of important plant water pumps, the specific process of step (S4) includes:

[0028] (S401) adding damage noise to the preprocessed data, adding noise reduction restriction and sparsity restriction to the autoencoder to obtain a sparse noise reduction autoencoder;

[0029] (S402) pre-training a sparse denoising autoencoder;

[0030] (S403) taking the hidden layer output value of the previous sparse denoising autoencoder as the input value of the next sparse denoising autoencoder, and repeating step (S402) using a layer-by-layer greedy algorithm until all sparse denoising autoencoders are trained;

[0031] (S404) adding a softmax classifier to the end of the sparse denoising autoencoder for classification;

[0032] (S405) using the BP back propagation algorithm to update the weights and biases in each iteration and fine-tune the parameters of the entire deep network;

[0033] (S406) testing the classification accuracy of the algorithm using a test set;

[0034] (S407) Complete the deep neural network model.

[0035] Furthermore, in the above-mentioned method for intelligent diagnosis of important plant water pumps, in step (S401), the noise reduction restriction uses masknoise as a correction, that is, the input data is randomly set to 0, and the corrected input vector is denoted as x′, and the expression is:

[0036] h=f(x′)=S f(Wx′+b)

[0037] Where: x′ is the corrected input vector; S f represents the nonlinear activation function; W is the encoder weight matrix; b is the encoder bias.

[0038] The sparsity constraint is achieved by controlling the average activation of neurons in the hidden layer. j (x) represents the jth activation unit in the hidden layer, then the average activation of the jth unit in the hidden layer is:

[0039]

[0040] Where: N represents the total number of samples, j represents the hidden unit.

[0041] In order to ensure that j In order not to deviate from ρ, a sparse term penalty factor needs to be added to the cost function.

[0042]

[0043] Where: S2 represents the number of neurons in the hidden layer; ρ is the sparsity parameter; KL is the Kullback-Leibler (induced sparsity) divergence; ρ j represents the average activation of the jth unit in the hidden layer.

[0044] After denoising the stacked autoencoder (AE) and adding a sparse penalty factor, a sparse denoising autoencoder (SDAE) is obtained. The original cost function can be expressed as:

[0045] J sDAE =J AE +γPN

[0046] In the formula, γ represents the weight of the sparse penalty factor; J AE is the stacked autoencoder cost function; PN represents the sparse term penalty factor.

[0047] Furthermore, in the above-mentioned method for intelligent diagnosis of important plant water pumps, in step (S402), the sparse denoising autoencoder updates the connection weights and biases according to the following formula during the pre-training process:

[0048]

[0049]

[0050] Where: ε represents the learning rate, J SDAE represents the sparse denoising autoencoder cost function, W ij (l) represents the weight between the i-th neuron and the j-th neuron in the l-th layer, bi (l) represents the bias between the lth layer and the ith neuron, Indicates the difference.

[0051] Furthermore, in the above-mentioned method for intelligent diagnosis of important plant water pumps, the specific process of step (S5) includes: the display unit in the computer device warns of the fault type corresponding to the vibration signal and stores it in the storage unit in a timely manner.

[0052] The present invention further provides a device for implementing the above-mentioned important plant water pump intelligent diagnosis method, comprising:

[0053] Acceleration sensor, used to obtain vibration signals of different fault types;

[0054] Data acquisition unit, used to isolate and regulate input voltage, and convert analog input quantity into digital quantity required by microcomputer system so as to interface with CPU;

[0055] A signal processing unit, used for performing wavelet threshold denoising preprocessing on the acquired signal;

[0056] Intelligent diagnosis unit, used to establish fault game model and deep neural network model, intelligently identify fault signals, and use deep neural network to automatically determine the fault type;

[0057] The display unit is used to label the classified fault type on the human-computer interaction interface, and when a fault occurs, timely warn the corresponding fault type;

[0058] Storage unit, used to store data and read and write data in real time;

[0059] The reset unit is used to reset the current CPU and running data to zero and then start up without power failure.

[0060] The beneficial effects of the present invention are as follows:

[0061] 1. An improved threshold function is used to denoise the fault samples of important plant water pumps, extract the energy of the denoised wavelet packet, and retain the fault information to the maximum extent. After the sparse denoising autoencoder (SDAE), accurate diagnosis of the fault data can be achieved.

[0062] 2. By designing a fault diagnosis game environment, this method can successfully use DRL to obtain an intelligent fault diagnosis agent and can effectively establish a diagnostic model. On the premise of ensuring diagnostic accuracy and reliability, this method greatly reduces the reliance on prior knowledge and expert experience of fault diagnosis and has a better diagnostic effect.

[0063] 3. The present invention integrates sparse denoising autoencoder deep neural network for intelligent fault diagnosis, which significantly improves the classification accuracy. The sparse denoising autoencoder network mines the features of the original data in an unsupervised learning manner, which improves the robustness and generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0065] Figure 1 A flow chart of a method for intelligent diagnosis of an important plant water pump provided by the present invention;

[0066] Figure 2 A flow chart of a wavelet improved threshold denoising method provided by the present invention;

[0067] Figure 3 A flow chart of a method for establishing a fault game model provided by the present invention;

[0068] Figure 4 A flow chart for establishing a deep neural network model provided by the present invention;

[0069] Figure 5 This is a schematic diagram of a detailed deconstruction of an important plant water pump intelligent diagnosis device provided by the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0071] The present invention takes into account that there are many types of important plant water pump failures. The traditional supervision method is not only time-consuming and labor-intensive, but also has low diagnostic accuracy. Therefore, a fault game model is designed to combine reinforcement learning with deep learning. The fault signal is first preprocessed, and then multiple hidden layers are reduced in dimension and feature extracted by stacking autoencoder neural networks. The initial parameters are optimized using BP neural networks to obtain a deep neural network model with feature extraction and pattern recognition functions. This method can effectively establish an unsupervised diagnosis model, realize intelligent diagnosis, and has a good diagnostic effect.

[0072] The embodiment of the present invention provides a method for intelligent diagnosis of an important plant water pump, such as Figure 1 As shown, the following steps are included:

[0073] S1) collecting vibration signals of important plant water pumps under different working conditions to obtain sample points of vibration signals containing different fault states;

[0074] S2) preprocessing the sample points using wavelet improved threshold denoising method;

[0075] S3) Establish a fault game model to provide an interactive environment for the fault diagnosis agent to observe, act and obtain rewards;

[0076] S4) reducing the dimensions and extracting features of multiple hidden layers by stacking autoencoder neural networks, and optimizing initial parameters by using BP neural networks to obtain a deep neural network model with feature extraction and pattern recognition functions;

[0077] S5) Output the diagnosis result.

[0078] In the above step S1, the vibration signals under different fault states are obtained by a measurement system composed of an acceleration sensor and LabVIEW; the causes of the faults include: severe wear and cavitation at the mouth ring, the rotor center does not coincide with the volute center, the blades and the diffuser are not properly aligned, the inlet and outlet pipelines are not designed reasonably, the inlet and outlet pipelines have no straight pipe section or the straight pipe section is too short (especially the inlet has a large impact), causing turbulence before the fluid enters the impeller, the blades do not operate under the design conditions (such as operating at ultra-small flow or ultra-large flow), the fluid is prone to turbulence, flow separation, impact, cavitation, random and uneven impact on the blades, etc.

[0079] In an optional implementation, the preprocessing may adopt an improved wavelet threshold method; Figure 2 As shown, the following steps are included:

[0080] S201) denoising the noise data by improving the threshold function;

[0081] In order to further solve the problem that the hard threshold function is discontinuous and the soft threshold function has a certain deviation, the present invention proposes a new threshold function between the soft and hard thresholds, which combines the advantages of the soft and hard threshold functions.

[0082] In view of the shortcomings of the traditional fixed threshold, the present invention proposes an improved threshold:

[0083]

[0084] Among them, σ 2 is the noise variance, σ=median, median is the median of the absolute value of the high-frequency wavelet coefficients, n is the signal length, j is the number of wavelet decomposition layers, and β is the control factor. As the number of decomposition layers increases, the threshold will gradually decrease.

[0085] S202) extracting wavelet packet energy of denoised data using wavelet packet transform;

[0086] S203) performing wavelet packet decomposition on the fault signal, selecting the wavelet basis as db3 and the decomposition level as 4;

[0087] S204) After decomposition, the wavelet packet energy of each node can be obtained by the following formula:

[0088] E jn =1M∑t=1M((X jn (t)) 2

[0089] Among them, X jn (t) represents the time domain vibration signal, t represents time, M is X jn After decomposition, the number of nodes is ∑iN2 i , i represents the number of neurons, and N is the number of decomposition layers, so as to obtain the energy of the corresponding node.

[0090] In an optional implementation, a fault game model may be established; Figure 3 As shown, the following steps are included:

[0091] S301) generating fault diagnosis questions;

[0092] S302) accepting diagnostic questions and giving result feedback via an agent;

[0093] S303) The game model determines whether the diagnosis result is correct. If it is correct, the total reward is increased by 1 and the game reward value is output; otherwise, the total reward is reduced by 1 and the process returns to step S301) to regenerate the fault diagnosis problem.

[0094] In an optional implementation, a deep neural network model is obtained, such as Figure 4 As shown, the following steps are included:

[0095] S401) adding "damage noise" to the data after denoising preprocessing;

[0096] Incorporating denoising and sparsity constraints into the autoencoder improves the robustness and generalization of the diagnostic method.

[0097] In practice, the noise reduction limit usually uses masknoise as a correction, that is, the input data is randomly set to 0. Let the corrected input vector be x', then the formula is:

[0098] h=f(x′)=S f (Wx′+b)

[0099] Where: x′ is the corrected input vector; S frepresents the nonlinear activation function; W is the encoder weight matrix; b is the encoder bias.

[0100] The sparsity constraint is achieved by controlling the average activation of neurons in the hidden layer. j (x) represents the jth activation unit in the hidden layer, then the average activation of the jth unit in the hidden layer is:

[0101]

[0102] Where: N represents the total number of samples, j represents the hidden layer unit.

[0103] In order to ensure that j In order not to deviate from ρ, a sparse term penalty factor needs to be added to the cost function.

[0104]

[0105] Where: S2 represents the number of neurons in the hidden layer; ρ is the sparsity parameter; KL is the Kullback-Leibler (induced sparsity) divergence; ρ j represents the average activation of the jth unit in the hidden layer.

[0106] After denoising the stacked autoencoder (AE) and adding a sparse penalty factor, a sparse denoising autoencoder (SDAE) is obtained. The original cost function can be expressed as:

[0107] J SDAE =J AE +γPN

[0108] In the formula, γ represents the weight of the sparse penalty factor; J AE is the stacked autoencoder cost function; PN represents the sparse term penalty factor.

[0109] S402) pre-training data samples;

[0110] During the pre-training process, the sparse autoencoder needs to update the connection weights and biases according to the following formula:

[0111]

[0112]

[0113] Where: ε represents the learning rate, J SDAE represents the cost function of the sparse denoising autoencoder, W ij (l) represents the weight between the i-th neuron and the j-th neuron in the l-th layer, b i (l) represents the bias between the lth layer and the ith neuron, Indicates the difference.

[0114] S403) Taking the hidden layer output value of the previous sparse denoising autoencoder as the input value of the next sparse denoising autoencoder, repeat the previous step using a layer-by-layer greedy algorithm until all autoencoders are trained.

[0115] S404) A softmax classifier is added to the end of the sparse denoising autoencoder for classification.

[0116] S405) Using the BP back propagation algorithm to update the weights and biases in each iteration, fine-tuning the parameters of the entire deep network.

[0117] S406) Using the test set to test the classification accuracy of the algorithm.

[0118] S407) Complete the deep neural network model.

[0119] In a specific implementation of step S406), the classification accuracy is tested using a test set to see whether the expected accuracy is achieved. If not, the step S405) is repeated to fine-tune the deep neural network model until the expected accuracy is achieved, thereby obtaining a trained model. The expected accuracy may be 90%.

[0120] The embodiment of the present invention also provides a device for implementing the above-mentioned important plant water pump intelligent diagnosis method, such as Figure 5 As shown, including:

[0121] The acceleration sensor 501 is used to obtain vibration signals of different fault types.

[0122] The data acquisition unit 502 is used to isolate and regulate the input voltage, and convert the analog input quantity into the digital quantity required by the microcomputer system so as to interface with the CPU.

[0123] The signal processing unit 503 is used to perform wavelet threshold denoising preprocessing on the acquired signal.

[0124] The intelligent diagnosis unit 504 is used to establish a fault game model and a deep neural network model, intelligently identify fault signals, and use a deep neural network to automatically determine the fault type.

[0125] The display unit 505 is used to label the classified fault type on the human-computer interaction interface, and to promptly warn of the corresponding fault type when a fault occurs.

[0126] The storage unit 506 is used to store data in real time and read and write data.

[0127] The reset unit 507 is used to reset the current CPU and operating data to zero and then start up without power failure.

[0128] In the embodiment of the present invention, the data acquisition unit 502 is connected to the LabVIEW measurement system via an acceleration sensor, and the intelligent diagnosis unit 504 is completed by reading and processing data via LabVIEW.

[0129] The device and method embodiments in the embodiments of the present invention are based on the same inventive concept.

[0130] In the embodiments of the present invention, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiment.

[0131] It will be apparent to those skilled in the art that the structure of the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0132] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for intelligent diagnosis of important plant water pumps, characterized in that: The steps include: (S1) collecting vibration signals of important plant water pumps under different working conditions to obtain sample points of vibration signals containing different fault states; (S2) performing wavelet threshold denoising preprocessing on the sample points; (S3) Establish a fault game model to provide an interactive environment for the fault diagnosis agent to observe, act and obtain rewards; The specific process includes: (S301) generating a fault diagnosis question; (S302) accepting diagnostic questions and providing result feedback via an agent; (S303) judging whether the diagnosis result is correct for the game model, if it is correct, the total reward is increased by 1, and the game reward value is output; otherwise, the total reward is decreased by 1, and the fault diagnosis problem is regenerated by returning to (S301); (S4) reducing the dimensions and extracting features of multiple hidden layers by stacking autoencoder neural networks, and optimizing initial parameters by using BP neural networks to obtain a deep neural network model with feature extraction and pattern recognition functions; (S5) Output the diagnosis result.

2. The method for intelligent diagnosis of important plant water pumps as claimed in claim 1, characterized in that: In step (S1), the vibration signals under different fault states are collected by a measurement system composed of an acceleration sensor and LabVIEW software.

3. The method for intelligent diagnosis of important plant water pumps as claimed in claim 1, characterized in that: In step (S1), the causes of the fault include: severe wear and cavitation at the mouth ring, the rotor center and the volute center do not coincide, the blades and the diffuser are not properly aligned, the inlet and outlet pipelines are not designed reasonably, the inlet and outlet pipelines have no straight pipe section or the straight pipe section is too short, causing turbulence before the fluid enters the impeller, and the blades are not operating under the design conditions, so the fluid is prone to turbulence, flow separation, impact, and cavitation, which causes random and uneven impact on the blades.

4. The method for intelligent diagnosis of important plant water pumps as claimed in claim 1, characterized in that: The method for performing denoising preprocessing on the sample points in step (S2) comprises the following steps: (S201) De-noising the noise data by improving the threshold function, and selecting the improved threshold th as follows: Among them, σ 2 is the noise variance, σ = median, median is the median of the absolute value of the high-frequency wavelet coefficients, n is the signal length, j is the number of wavelet decomposition layers, and β is the control factor; (S202) extracting wavelet packet energy of the denoised data using wavelet packet transform; (S203) performing wavelet packet decomposition on the vibration signal of the fault state; (S204) After decomposition, the wavelet packet energy of each node is obtained by the following formula: Yes jn =1M∑t=1M((X jn (t)) 2 Among them, X jn (t) represents the time domain vibration signal, t represents time, and M represents the node X jn The number of samples in .

5. The method for intelligent diagnosis of important plant water pumps as claimed in claim 4, characterized in that: The wavelet basis selected in step (S203) is db3, and the number of decomposition levels is 4.

6. The method for intelligent diagnosis of important plant water pumps as claimed in claim 1, characterized in that: The specific process of step (S4) includes: (S401) adding damage noise to the preprocessed data, adding noise reduction restriction and sparsity restriction to the autoencoder to obtain a sparse noise reduction autoencoder; (S402) pre-training a sparse denoising autoencoder; (S403) taking the hidden layer output value of the previous sparse denoising autoencoder as the input value of the next sparse denoising autoencoder, and repeating step (S402) using a layer-by-layer greedy algorithm until all sparse denoising autoencoders are trained; (S404) adding a softmax classifier to the end of the sparse denoising autoencoder for classification; (S405) using the BP back propagation algorithm to update the weights and biases in each iteration and fine-tune the parameters of the entire deep network; (S406) testing the classification accuracy of the algorithm using a test set; (S407) Complete the deep neural network model.

7. The method for intelligent diagnosis of important plant water pumps as claimed in claim 6, characterized in that: In step (S401), the noise reduction restriction uses masknoise as a correction, that is, the input data is randomly set to 0, and the corrected input vector is denoted as x′. The activation expression is: h=f(x′)=S f (Wx′+b) Where: x′ is the corrected input vector; S f represents the nonlinear activation function; W is the encoder weight matrix; b is the encoder bias; The sparsity restriction is achieved by controlling the average activation of neurons in the hidden layer. Assuming a j (x) represents the jth activation unit in the hidden layer, then the average activation of the jth unit in the hidden layer is: Where: N represents the total number of samples, j represents the hidden layer unit; In order to ensure that j In order not to deviate from ρ, a sparse term penalty factor needs to be added to the cost function. Where: S2 represents the number of neurons in the hidden layer; ρ is the sparsity parameter; KL is the Kullback-Leibler divergence; ρ j represents the average activation of the jth unit in the hidden layer; After denoising the stacked autoencoder AE and adding a sparse penalty factor, a sparse denoising autoencoder SDAE is obtained. The original cost function is expressed as: I SDAE =J AE +γPN In the formula, γ represents the weight of the sparse penalty factor; J AE is the stacked autoencoder cost function; PN represents the sparse term penalty factor.

8. The method for intelligent diagnosis of important plant water pumps as claimed in claim 6, characterized in that: In step (S402), the sparse denoising autoencoder updates the connection weights and biases according to the following formula during pre-training: Where: ε represents the learning rate, J SDAE represents the cost function of the sparse denoising autoencoder, W ij (l) represents the weight between the i-th neuron and the j-th neuron in the l-th layer, b i (l) represents the bias between the lth layer and the ith neuron, Indicates the difference.

9. The method for intelligent diagnosis of important plant water pumps as claimed in claim 1, characterized in that: The specific process of step (S5) includes: the display unit in the computer device warns the fault type corresponding to the vibration signal, and stores it in the storage unit in time.

10. A device for implementing the intelligent diagnosis method for important plant water pumps as claimed in any one of claims 1 to 9, characterized in that: include: Acceleration sensor, used to obtain vibration signals of different fault types; Data acquisition unit, used to isolate and regulate input voltage, and convert analog input quantity into digital quantity required by microcomputer system so as to interface with CPU; A signal processing unit, used for performing wavelet threshold denoising preprocessing on the acquired signal; The intelligent diagnosis unit is used to establish a fault game model and a deep neural network model, intelligently identify fault signals, and automatically determine the fault type using a deep neural network; the fault game model provides an interactive environment for the fault diagnosis agent to observe, act, and obtain rewards. The specific process includes: (S301) generating a fault diagnosis question; (S302) accepting diagnostic questions and providing result feedback via an agent; (S303) judging whether the diagnosis result is correct for the game model, if it is correct, the total reward is increased by 1, and the game reward value is output; otherwise, the total reward is decreased by 1, and the fault diagnosis problem is regenerated in (S301); The display unit is used to label the classified fault type on the human-computer interaction interface, and when a fault occurs, timely warn the corresponding fault type; Storage unit, used to store data and read and write data in real time; The reset unit is used to reset the current CPU and running data to zero and then start up without powering off.

Images