Artificial intelligence-based machine pump device state monitoring and fault diagnosis method

Abstract

The application provides a kind of based on artificial intelligence's machine pump class equipment state monitoring and fault diagnosis method, it is related to data processing technical field, including: based on pre-set vibration sensor obtains the vibration signal of target category's machine pump equipment, and is labeled;Based on pre-set double channel feature extraction model extracts the feature of vibration signal;Based on the feature of vibration signal, preset base fault diagnosis model is iteratively trained until reaching preset training completion condition;The preset number of fault diagnosis models are bundled to build pump class equipment state reasoning model;The running vibration signal of target category's machine pump equipment is acquired, and the running vibration signal is input into pump class equipment state reasoning model, and the fault category is output;Through time-frequency double channel fusion and multi-scale perception, the fault recognition precision is improved;The result rationality and interpretability are enhanced by using physical priori constraint;And through model integration optimization, the classification stability and reliability under complex scene are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a machine pump type equipment state monitoring and fault diagnosis method based on artificial intelligence. BACKGROUND

[0002] With the rapid development of industrial automation and intelligent manufacturing, machine pump type equipment as a key rotating machinery in process industry, its running state is directly related to the safety and stability of the production system. The traditional state monitoring and fault diagnosis method depends on artificial inspection, preset threshold judgment or experience rule construction, which has limitations such as reaction lag, high misjudgment rate and dependence on expert experience, and is difficult to meet the needs of modern industry for real-time perception and intelligent analysis of equipment state. In recent years, fault diagnosis methods based on artificial intelligence have gradually emerged, which use machine learning, deep neural network and other technologies to model and classify vibration signals, significantly improving the automation level of diagnosis. However, the existing methods generally have the following problems: first, most models only use single features in time domain or frequency domain, ignoring the complementarity of the two, resulting in insufficient recognition accuracy of non-stationary fault signals; second, the network structure with fixed receptive field is difficult to adapt to the characteristics of machine pump fault changing across multiple time scales; third, the output layer of the model lacks physical priori constraints, and the classification results are not strongly explainable, which may lead to physically unreasonable judgments; fourth, in complex or similar fault scenarios, the class features are easy to overlap, resulting in decreased model boundary judgment ability. SUMMARY

[0003] The purpose of the present application is to provide a machine pump type equipment state monitoring and fault diagnosis method based on artificial intelligence, which improves the fault recognition accuracy through time-frequency dual-channel fusion and multi-scale perception, enhances the rationality and explainability of the results by using physical priori constraints, and improves the classification stability and reliability in complex scenarios through model integration optimization.

[0004] In the first aspect, the present application provides a machine pump type equipment state monitoring and fault diagnosis method based on artificial intelligence, comprising:

[0005] obtaining the vibration signal of the machine pump equipment of the target category based on the pre-set vibration sensor, and labeling; wherein the label of the labeling represents the running state category of the machine pump equipment of the target category;

[0006] extracting the features of the vibration signal based on the pre-set dual-channel feature extraction model; wherein the dual-channel feature extraction model comprises: a time domain feature extraction channel and a frequency domain feature extraction channel; the features of the vibration signal include: fusion features based on the splicing of the time domain features of the vibration signal and the frequency domain features of the vibration signal;

[0007] The preset basic fault diagnosis model is iteratively trained based on the features of the vibration signal until a preset training completion condition is reached, and the trained basic fault diagnosis model is determined as the fault diagnosis model; wherein the basic fault diagnosis model comprises: a time-frequency dual-channel interactive gated neural network module, a multi-scale residual dilated convolution module and a physical prior fusion output layer structure connected in sequence; the time-frequency dual-channel interactive gated neural network module is used for fusing time domain features and frequency domain features of the vibration signal; the multi-scale residual dilated convolution module is used for extracting multi-period features and enhancing time scale perception; the physical prior fusion output layer structure is used for realizing explanatory classification prediction of the pump equipment state; the final output of the basic fault diagnosis model is constrained based on a preset fault state transition matrix and an output result of the physical prior fusion output layer structure;

[0008] A preset number of fault diagnosis models are bundled to construct a pump equipment state reasoning model; wherein the weight is determined based on the information entropy of the fault category probability distribution of the fault diagnosis model; and the fault category with the largest cumulative weight is determined as the output of the pump equipment state reasoning model;

[0009] The running vibration signal of the target category pump equipment is obtained, and the running vibration signal is input into the pump equipment state reasoning model to output the fault category.

[0010] In some preferred embodiments of the present application, the vibration sensor is arranged at any one of the following positions on the pump equipment: the non-driven end of the motor, the driven end of the motor, the driven end of the pump and the non-driven end of the pump; and the label at least includes one of the following: normal operation, slight imbalance, serious misalignment, bearing damage, abnormal gear engagement and unknown state.

[0011] In some preferred embodiments of the present application, the time domain feature extraction channel is used to perform absolute value taking and impact enhancement coefficient amplification operation based on the vibration signal, and then apply exponential moving average processing to obtain the time domain features of the vibration signal; the frequency domain feature extraction channel is used to perform adaptive frequency band decomposition operation based on the vibration signal, perform fast Fourier transform on the signal in each frequency band and calculate the energy, calculate the weight based on the variance of the energy of each frequency band, and sum the weighted energy of all frequency bands to obtain the frequency domain features of the vibration signal; wherein the adaptive frequency band decomposition divides the signal into weighted frequency bands through a Gaussian frequency band window, the center frequency of the frequency band is calculated based on the spectral envelope curve through clustering, and the frequency band width is calculated and determined based on the frequency axis range and a preset parameter.

[0012] In some preferred embodiments of the present application, the time-frequency dual-channel interactive gated neural network module comprises: a time domain path, a frequency domain path and a cross-channel interaction module.

[0013] The time-domain path is used to input the time-domain features of the vibration signal into the gated loop unit for processing. The gated loop unit calculates the update gate output and reset gate output based on the hidden state of the previous time step and the time-domain features of the current time step. Based on the reset gate output, the hidden state of the previous time step and the time-domain features of the current vibration signal, it calculates the candidate hidden state. Based on the update gate output, the hidden state of the previous time step and the candidate hidden state, it calculates the hidden state of the current time step to capture temporal dependencies.

[0014] The frequency domain path is used to perform one-dimensional convolution operations on the real and imaginary parts of the frequency domain features respectively. The imaginary part convolution result is multiplied by the imaginary unit and added to the real part convolution result to obtain the complex domain convolution result. The complex domain convolution result is processed by the ReLU activation function to obtain the output feature of the frequency domain path, so as to enhance the phase sensing capability of the frequency domain features.

[0015] The cross-channel interaction module is used to concatenate the hidden state of the current time step and the output features of the frequency domain path to obtain concatenated features. The concatenated features are then processed by linear transformation and sigmoid activation function to generate a gating weight vector. Based on the gating weight vector, the hidden state of the current time step and the output features of the frequency domain path are weighted and fused to obtain the final output features of the time-frequency dual-channel interactive gating neural network module, thus realizing the dynamic fusion of time-frequency features.

[0016] In some preferred embodiments of the present invention, the multi-scale residual dilated convolution module includes: a residual block of level 1, a residual block of level 2, and a residual block of level 3 connected in sequence.

[0017] The residual block is constrained by the following formula:

[0018] ;

[0019] in, For the expansion rate The dilated convolution operation; For the first Expansion rate of the layer residual block; For the level The residual block; For residual block level index; Use the LeakyReLU activation function; For skip connection branches, the calculation method is represented as follows: ; This is a 1×1 convolution operation, used to keep the feature dimensions unchanged; This represents the final output feature of the time-frequency dual-channel interactive gated neural network module;

[0020] The output of the multi-scale residual dilated convolution module is constrained by the following formula:

[0021] ;

[0022] in, For the level The residual block; For the level The residual block; For the level The residual block; This represents the output features of the multi-scale residual dilated convolution module.

[0023] In some preferred embodiments of the present invention, an output layer structure that integrates physical priors is used to perform averaging and maximizing operations on the output features of the multi-scale residual dilated convolution module along the time dimension, respectively. The results of the averaging and maximizing operations are concatenated to obtain a compressed feature vector, and the diagnostic results of the model are output based on the compressed feature vector.

[0024] In some preferred embodiments of the present invention, the final output of the basic fault diagnosis model is constrained by a pre-set fault state transition matrix and the output result of the output layer structure that integrates physical priors, including:

[0025] Correction is performed based on a preset transfer matrix and preset harmonic factors; whereby the transfer matrix characterizes whether two fault states can be directly transferred; and the harmonic factors are determined based on the vibration spectrum.

[0026] In some preferred embodiments of the present invention, the loss function of the fault diagnosis model is based on the joint constraint of standard cross-entropy loss and feature decoupling regularization loss; wherein, the feature decoupling regularization term includes: intra-class aggregation and inter-class separation; the fault diagnosis model updates parameters based on error backpropagation of gradient projection.

[0027] In some preferred embodiments of the present invention, each fault diagnosis model in the pump equipment state reasoning model is trained based on a fixed proportion of random training data.

[0028] The final classification result of the pump equipment state reasoning model is constrained by the following formula:

[0029] ;

[0030] ;

[0031] ;

[0032] in, For the first The weights of each member model; is a positive integer, indicating the index of the member model; is the entropy of the prediction probability distribution of the first is the prediction probability distribution of the first is a small constant, preventing division by zero errors; is the total number of classes; is the prediction probability of the first is the final classification result class; is an indicator function; is the class index c that maximizes the expression; c is the class index; is the number of effective member models in the ensemble model; is the class index k that maximizes the expression; κ is the slope factor; is the prediction probability of the first is the class index that maximizes the prediction probability of the first that is, characterizes when the condition of the indicator function is true.

[0033] In some preferred embodiments of the present application, the method further comprises:

[0034] The labeled target class vibration signal of the pump equipment is input into the conventional time domain feature combined with the support vector machine method, the wavelet transform combined with the neural network method, the one-dimensional convolutional neural network method and the pump equipment state reasoning model respectively, the fault class is output respectively, and chart analysis is performed.

[0035] The present application brings the following beneficial effects:

[0036] ​​​​​​​The application provides a kind of based on artificial intelligence's machine pump class equipment state monitoring and fault diagnosis method, comprising: based on the vibration signal of the vibration sensor of pre-set target category machine pump equipment is obtained, and is labeled;Wherein, the label of labeling represents the running state category of target category machine pump equipment;Based on the feature of pre-set double-channel feature extraction model vibration signal is extracted;Wherein, double-channel feature extraction model includes: time domain feature extraction channel and frequency domain feature extraction channel;The feature of vibration signal includes: based on vibration signal time domain feature and vibration signal frequency domain feature splicing fusion feature;Based on the feature of vibration signal, preset basic fault diagnosis model is iteratively trained until the preset training completion condition is reached, and the basic fault diagnosis model of complete training is determined as fault diagnosis model;Wherein, basic fault diagnosis model includes: time-frequency double-channel interactive gate neural network module, multi-scale residual dilated convolution module and fusion physical prior output layer structure connected in turn;Time-frequency double-channel interactive gate neural network module is used to fuse vibration signal time domain feature and vibration signal frequency domain feature;Multi-scale residual dilated convolution module is used to extract multi-period feature and enhance time scale perception;Interpretive classification prediction of machine pump equipment state is realized;The final output of basic fault diagnosis model is based on the output result constraint of pre-set fault state transition matrix and fusion physical prior output layer structure;The preset number of fault diagnosis models is bundled to build pump class equipment state reasoning model;Wherein, the weight is determined based on the information entropy of the fault category probability distribution of fault diagnosis model;The fault category of maximum cumulative weight is determined as the output of pump class equipment state reasoning model;The running vibration signal of target category machine pump equipment is obtained, and the running vibration signal is input into pump class equipment state reasoning model, and the fault category is output;Through time-frequency double-channel fusion and multi-scale perception, the fault recognition precision is improved;The result rationality and explainability are enhanced by using physical prior constraint;And through model integration optimization, the classification stability and reliability under complex scene are improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the following specific embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor.

[0038] Figure 1 A flow chart of a kind of based on artificial intelligence's machine pump class equipment state monitoring and fault diagnosis method provided by the embodiment of the present application;

[0039] Figure 2 A vibration sensor position schematic diagram provided by the embodiment of the present application;

[0040] Figure 3 A comparison chart of the accuracy of different models provided by the embodiment of the present application on various faults;

[0041] Figure 4 A comparison chart of the robustness of various methods provided by the embodiment of the present application under different signal-to-noise ratio conditions;

[0042] Figure 5 A comparison chart of the characteristics of the method provided by the embodiment of the present application and the conventional method;

[0043] Figure 6 A comparison chart of the convergence speed provided by the embodiment of the present application;

[0044] Figure 7 A structural schematic diagram of an electronic device provided by the embodiment of the present application.

[0045] Icon: 400 - memory; 401 - processor; 402 - bus; 403 - communication interface. DETAILED DESCRIPTION

[0046] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the accompanying drawings herein can be arranged and designed in various different configurations.

[0047] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0048] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0049] In the description of the present application, it should be pointed out that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product of the application is usually placed, which is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, the terms "first", "second", "third" and the like are only used for differentiation and cannot be understood as indicating or implying relative importance.

[0050] In addition, the terms "horizontal", "vertical", "overhanging" and the like do not mean that the components must be absolutely horizontal or overhanging, but can be slightly inclined. For example, "horizontal" only means that it is more horizontal relative to "vertical", and does not mean that the structure must be completely horizontal, but can be slightly inclined.

[0051] In the description of the present application, it should be further pointed out that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0052] The artificial intelligence-based machine pump device state monitoring and fault diagnosis method of the present application has the following technical effects:

[0053] Improve fault recognition accuracy and robustness: the vibration signal features are extracted in parallel by the dual-channel feature extraction model (time domain and frequency domain), and the time-frequency dual-channel interactive gating neural network module is used for feature fusion, which overcomes the limitation of using only single domain features in existing methods, fully excavates the complementarity of time domain and frequency domain features, significantly enhances the recognition ability of non-stationary fault signals, and reduces the misjudgment rate.

[0054] Enhance multi-scale time adaptability: a multi-scale residual dilated convolution module is used to capture fault features in vibration signals across multiple time scales through convolution layers with different dilation rates, avoiding the shortcomings of fixed receptive field network structure, making the model flexible to adapt to the complex change law of machine pump faults, and improving the perception ability of long-term dependence and short-term fluctuation characteristics.

[0055] To improve the interpretability and physical rationality of diagnostic results: A constraint mechanism that integrates physical priors is introduced into the model output layer. The classification results are corrected based on the fault state transition matrix, ensuring that the output conforms to the physical laws of equipment state changes, reducing physically unreasonable judgments, enhancing the interpretability of diagnostic results, and making it easier for engineers to understand and trust.

[0056] Improving classification boundary judgment in complex scenarios: By bundling multiple fault diagnosis models to construct a state inference model for pump-type equipment, and dynamically allocating weights based on the information entropy of the fault category probability distribution, adaptive integration of model outputs is achieved. This method effectively alleviates the problem of overlapping category features in complex or similar fault scenarios, strengthens the model's ability to distinguish decision boundaries, and improves the overall stability and accuracy of diagnosis.

[0057] In summary, this method not only enables real-time intelligent monitoring and fault diagnosis of pump equipment status, but also comprehensively solves the limitations mentioned in the background technology, such as slow response, high misjudgment rate, and reliance on expert experience, through multi-feature fusion, multi-scale perception, physical constraints, and model integration. It provides a more reliable and efficient status perception solution for industrial automation and intelligent manufacturing.

[0058] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0059] Example 1

[0060] This invention provides an artificial intelligence-based method for condition monitoring and fault diagnosis of pumps and other equipment. (See also...) Figure 1 The flowchart shown in this embodiment of the invention provides a method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence. The method includes:

[0061] Step S102: Obtain vibration signals of pumps and machinery of the target category based on pre-set vibration sensors and label them; wherein, the labeled tags represent the operating status category of the pumps and machinery of the target category.

[0062] Specifically, by arranging high-sensitivity vibration sensors at key positions of the machine pump equipment, continuous collection of vibration signals generated during operation of the machine pump equipment is realized; the sensor type used is an acceleration type charge output sensor, which can realize high-precision response to low-frequency and high-frequency vibration components; the sensor is connected to an industrial computer through a data acquisition card, and the acquisition frequency is set to 5 kHz to 20 kHz to cover the main frequency range of fault induction; the vibration signal maintains the original amplitude and time sequence during data acquisition, and the acquisition content includes time domain waveforms of the machine pump equipment under normal, slight abnormal and typical fault and other operating states, ensuring the universality and representativeness of sample distribution; in order to ensure data synchronization and calibration accuracy, the acquisition system uses a high-precision clock and records timestamp information, providing accurate timing reference for subsequent modeling.

[0063] In order to realize supervised training of the model, the collected vibration monitoring data is labeled. The labeling method combines artificial inspection records, machine pump equipment maintenance logs and expert experience evaluation, and determines the operating state category of the vibration signal by synchronously referring to the operation and maintenance information in the corresponding time period.

[0064] Further, in some preferred embodiments of the present application, the vibration sensor is arranged at any one of the following positions on the machine pump equipment: motor non-driving end, motor driving end, pump driving end and pump non-driving end; and the label at least includes one of the following: normal operation, slight imbalance, serious misalignment, bearing damage, gear meshing abnormality and unknown state.

[0065] Specifically, referring to the vibration sensor position diagram provided by the embodiment of the present application shown in Figure 2 The vibration sensor position diagram provided by the embodiment of the present application shown in

[0066] The labeled categories include typical working conditions such as "normal operation", "slight imbalance", "serious misalignment", "bearing damage" and "gear meshing abnormality", and "unknown state" is provided as a buffer category for uncategorized samples.

[0067] In step S104, the features of the vibration signal are extracted based on a pre-set double-channel feature extraction model; the double-channel feature extraction model includes a time domain feature extraction channel and a frequency domain feature extraction channel; and the features of the vibration signal include fusion features based on vibration signal time domain features and vibration signal frequency domain features.

[0068] Specifically, vibration signals from pumps and other equipment exhibit strong non-stationarity, with frequency components changing significantly over time. This makes conventional Fourier transform ineffective in processing local transient impact signals. While wavelet transform possesses some time-frequency locality, it relies on basis function selection and has poor adaptability. Furthermore, directly inputting the original signal or single-domain features into the model can cause bias in the model's identification of impact-type and continuous faults, resulting in a lack of comprehensive perception capabilities. This invention addresses the problem of insufficient extraction of time and frequency domain information from non-stationary vibration signals by constructing an adaptive frequency band decomposition and dual-channel feature fusion mechanism.

[0069] Furthermore, in some preferred embodiments of the present invention, the time-domain feature extraction channel is used to perform absolute value taking and impact enhancement coefficient amplification operations based on the vibration signal, and then apply exponential moving average processing to obtain the time-domain features of the vibration signal; the frequency-domain feature extraction channel is used to perform adaptive frequency band decomposition based on the vibration signal, perform fast Fourier transform on the signal in each frequency band and calculate the energy, calculate the weight based on the variance of the energy of each frequency band, and sum the weighted energy of all frequency bands to obtain the frequency-domain features of the vibration signal; wherein, the adaptive frequency band decomposition divides the signal into weighted frequency bands through a Gaussian frequency band window, the center frequency of the frequency band is obtained by clustering calculation based on the spectral envelope curve, and the bandwidth is determined by calculation based on the frequency axis range and preset parameters.

[0070] Specifically, for the time-domain characteristics of the vibration signal, based on the original vibration signal, absolute value taking and impact enhancement coefficient amplification operations are performed, and then an exponential moving average is applied to the result to obtain the time-domain characteristics of the vibration signal. This achieves the effect of enhancing the impact component while preserving the signal trend characteristics, expressed as:

[0071] ;

[0072] in, Represents the time-domain characteristics of vibration signals; Indicates the first The original vibration signal of each sample; A positive integer, representing the sample index; This indicates the absolute value operation; This represents the impact enhancement factor, used to amplify the impact component, such as... ; The exponential moving average operator is used to smooth signals while preserving trend characteristics. The recursive implementation of the exponential moving average operator is represented as follows: ; For the current number Temporal eigenvalues ​​of a time step; A positive integer representing the time step index; For the first Temporal eigenvalues ​​of a time step; For the first The value of the i-th original vibration signal sample at time step; For example, a smoothing factor. .

[0073] It should be noted that, The term representation employs an exponential smoothing shock enhancement operator in the time-domain channel, through... The nonlinear amplification highlights transient impact components, such as pulses generated by bearing cracks, while utilizing... control The exponential moving average of the term retains the overall trend information of equipment operation, and the synergistic extraction of impact components and trend features enables the model to both detect sudden failures and track the slow-developing degradation process.

[0074] For the frequency domain characteristics of vibration signals, an adaptive frequency band decomposition operation is performed based on the original vibration signal. This adaptive frequency band decomposition divides the signal into weighted frequency bands using a Gaussian frequency band window. The center frequency of each band is obtained through clustering calculation based on the spectral envelope curve, and the bandwidth is determined based on the frequency axis range and preset parameters. Then, a Fast Fourier Transform is performed on the signal within each band, and its energy is calculated. Finally, weights are calculated based on the variance of the energy in each band, and the weighted energies of all bands are summed to obtain the frequency domain characteristics of the vibration signal. This achieves adaptive extraction of key frequency band energy distribution information, represented as:

[0075] ;

[0076] in, Represents the frequency domain characteristics of vibration signals; This represents the Fast Fourier Transform, which converts a time-domain signal to the frequency domain. The frequency axis represents the signal's frequency spectrum and is calculated using the frequency points obtained through Fast Fourier Transform. ; The sampling frequency can be set from 5kHz to 20kHz. The signal length; Indicates the first A Gaussian frequency band window is used for adaptive frequency band decomposition, and its calculation method is expressed as follows: ; A positive integer, representing the frequency band index; Indicates the total number of adaptive frequency bands, such as, ; Indicates the first Each frequency band center frequency, through -means clustering is calculated on the spectral envelope curve; Indicates the first The standard deviation of each frequency band, controlling the bandwidth, is calculated as follows: ; Indicates the bandwidth scaling factor, such as, ; Indicates the maximum value of the frequency axis; Indicates the minimum value of the frequency axis; It is a natural exponential function; Represents the L2 norm; Indicates the first The energy variance weights for each frequency band are calculated as follows: ; This represents the variance calculation function; It is a positive integer, representing a frequency band index that is distinct from k; Indicates the first A Gaussian frequency band window.

[0077] It should be noted that, unlike fixed frequency band division, the center frequency is dynamically determined through spectral envelope clustering. , making the Gaussian frequency window Automatic focusing on energy concentration areas, such as the characteristic frequency band of bearing failures, and energy variance weighting. Determined by the energy variance of each frequency band, the fault-sensitive frequency band is given a higher weight, which can enhance the model's ability to separate complex faults. When multiple faults occur simultaneously, such as the coexistence of imbalance and bearing damage, the adaptive frequency band can independently capture the frequency domain characteristics of different faults and avoid feature confusion. Conventional methods are difficult to achieve this kind of dynamic frequency domain focusing for specific equipment.

[0078] The time-domain and frequency-domain features of the vibration signal are concatenated using a matrix concatenation operation, and the concatenation result is transposed to form a dual-channel feature matrix. This achieves the fusion of complementary information from the time and frequency domains, as shown below:

[0079] ;

[0080] in, This represents a synthesized dual-channel feature matrix, used in pump equipment condition monitoring models to process dual-channel features separately. This represents a matrix concatenation operation; This represents the matrix transpose operation.

[0081] In step S106, the preset basic fault diagnosis model is iteratively trained based on the features of the vibration signal until a preset training completion condition is reached, and the trained basic fault diagnosis model is determined as the fault diagnosis model; wherein the basic fault diagnosis model comprises: a time-frequency dual-channel interactive gate neural network module, a multi-scale residual dilated convolution module and a physical prior fusion output layer structure connected in sequence; the time-frequency dual-channel interactive gate neural network module is used for fusing the time domain features of the vibration signal and the frequency domain features of the vibration signal; the multi-scale residual dilated convolution module is used for extracting multi-period features and enhancing time scale perception; the physical prior fusion output layer structure is used for realizing explanatory classification prediction of the state of the pump equipment; and the final output of the basic fault diagnosis model is constrained based on a pre-set fault state transition matrix and an output result of the physical prior fusion output layer structure.

[0082] Specifically, the structure of the basic fault diagnosis model adopts an end-to-end modeling manner, takes the input dual-channel features as a starting point, gradually extracts deep semantic representations through multiple layers of non-linear mapping, and finally outputs a state classification result. The model training process adopts a decoupling regular term and gradient projection mechanism to improve the model discrimination ability and training stability.

[0083] Further, in some preferred embodiments of the present application, the time-frequency dual-channel interactive gate neural network module comprises: a time domain path, a frequency domain path and a cross-channel interaction module.

[0084] Specifically, conventional neural network structures mostly adopt a single-channel modeling manner, only utilize one kind of information in the time domain or the frequency domain, and the channel splicing manner lacks an information interaction mechanism, cannot fully fuse the information flow of the two types of features, and is difficult to realize unified modeling of dynamic dependence modeling and spectral pattern perception. The embodiments of the present application aim at the dynamic interactive fusion of time-frequency dual-channel information, adopt a dual-channel interactive gate network structure, capture time dependence by adopting a gated recurrent unit in the time domain path, enhance the frequency spectrum phase perception ability by adopting a complex convolution in the frequency domain path, and realize adaptive weighted fusion of the time domain and frequency domain features through a cross-channel gating mechanism, thereby improving the perception ability of the model to multi-class fault features.

[0085] The time domain path is used for inputting the time domain features of the vibration signal into the gated recurrent unit for processing, wherein the gated recurrent unit calculates update gate output and reset gate output based on the hidden state of the previous time step and the time domain features of the current time step, calculates the candidate hidden state based on the reset gate output, the hidden state of the previous time step and the current time domain features of the vibration signal, calculates the hidden state of the current time step based on the update gate output, the hidden state of the previous time step and the candidate hidden state, so as to capture the time sequence dependence.

[0086] Specifically, the time domain path is constrained by the following formula:

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] in, For the first The update gate output of the time step controls the degree of update of the hidden state; For the current number Temporal eigenvalues ​​of a time step; The weight matrix for updating the gate is a trainable parameter; The bias vector for updating the gate is a trainable parameter; For the first The hidden state of the time step; For the first The hidden state of the time step; For the first The time step reset gate output controls the degree of forgetting of historical information; The weight matrix for the reset gate is a trainable parameter; The bias vector for resetting the gate is a trainable parameter; For the first Candidate hidden states at a time step; Let be the weight matrix of the candidate hidden states, and be trainable parameters; The bias vector is the candidate hidden state, and the parameters are trainable parameters; It is the hyperbolic tangent function; Use the Sigmoid activation function; For Hadama accumulation.

[0092] The frequency domain path is used to perform one-dimensional convolution operations on the real and imaginary parts of the frequency domain features respectively. The imaginary part convolution result is multiplied by the imaginary unit and added to the real part convolution result to obtain the complex domain convolution result. The complex domain convolution result is processed by the ReLU activation function to obtain the output feature of the frequency domain path, so as to enhance the phase sensing capability of the frequency domain features.

[0093] Specifically, the frequency domain path is constrained by the following formula:

[0094] ;

[0095] ;

[0096] in, The output features of the frequency domain path are the result after complex convolution and activation. is a ReLU activation function; represents a complex domain 1D convolution operation; is a one-dimensional convolution operation, which performs local feature extraction on input features using a convolution kernel, for example, the size of the convolution kernel can be set to 3; is an imaginary unit; is an operation for extracting a real part; is an operation for extracting an imaginary part.

[0097] The cross-channel interaction module is configured to splice the hidden state of the current time step and the output feature of the frequency domain path to obtain spliced features, perform linear transformation and Sigmoid activation function processing on the spliced features in sequence to generate a gating weight vector, and perform weighted fusion on the hidden state of the current time step and the output feature of the frequency domain path based on the gating weight vector to obtain the final output feature of the time-frequency dual-channel interaction gating neural network module, thereby realizing dynamic fusion of time-frequency features.

[0098] Specifically, the cross-channel interaction module is constrained by the following formula:

[0099] ;

[0100] wherein, is a gating weight vector used for dynamically fusing time domain and frequency domain features, and the calculation manner is represented as ; is a linear transformation operation for mapping the spliced features to a gating weight space; is an output feature of the dual-channel interaction gating network.

[0101] Further, the fixed receptive field structure cannot adapt to fault features of different time scales, which is prone to cause insufficient extraction of low-frequency components of long-period changes or fuzzy phenomenon of short-period impact components, thereby affecting the comprehensiveness and accuracy of fault recognition. In some preferred embodiments of the present application, a multi-scale residual dilated convolution module with a cascade structure is adopted, different scale features are superimposed through residual connection, and the receptive field is dynamically changed by using an adaptive dilated rate to enhance the extraction ability of the model for fault features of different time scales. While retaining high-level abstract semantic information, the bottom-layer fine-grained perception ability is strengthened, and effective fusion of multi-period signal features is realized.

[0102] Specifically, the multi-scale residual dilated convolution module comprises: a residual block with a level of 1, a residual block with a level of 2 and a residual block with a level of 3 connected in sequence. Based on the level of the residual block, a cyclically changing dilated rate is calculated through a modulo operation to realize adjustment of the receptive field size of the convolution kernel according to a preset period, and the dilated rate is constrained by the following formula:

[0103] ;

[0104] in, This is the residual block level index, with values ​​ranging from 1 to 3, corresponding to the first residual block, the second residual block, and the third residual block, respectively. This represents the modulo operation; For the first The dilation rate of the residual blocks controls the size of the receptive field of the convolution kernel.

[0105] It should be noted that, The result is cyclical: 1→1, 2→2, 3→0, 4→1, 5→2, 6→0, ..., and so on. Therefore... This indicates that the expansion rate is achieved periodically. cycle.

[0106] It should also be noted that bearing damage generates microsecond-level impacts, gear abnormalities exhibit millisecond-level modulations, and imbalance faults manifest as second-level vibrations. Convolutional networks with fixed receptive fields struggle to accommodate features at different scales, leading to the smoothing of short-term impacts or the neglect of long-period variations. Make the receptive field of the convolution kernel periodically Cyclic changes, when When the temporal dilatation rate is 1, it represents the basic receptive field, which can precisely capture instantaneous impacts. The time expansion rate 2 covers the medium cycle. The time dilation rate of 4 senses long-term trends, and three layers of cascaded residual blocks can cover the entire time scale, with computational efficiency higher than traditional dilated convolution.

[0107] Based on the residual block hierarchy and decay control parameters, the skip connection weights are calculated using the Sigmoid activation function, achieving weight decay as the hierarchy increases. The skip connection weights are constrained by the following formula:

[0108] ;

[0109] in, The decay control parameter adjusts the decay rate of the weight as the level increases, for example, ; For the first Layer skip connection weights.

[0110] It should be noted that skip connection weights To achieve weight decay with network depth, In smaller areas, shallow residual blocks retain more original feature details, making them suitable for capturing high-frequency impacts. Large time, deep residual block depends on the abstract features extracted by the main branch, which is suitable for analyzing low-frequency trends, and solves the multi-scale feature conflict problem in rotating machinery fault diagnosis. In the gearbox compound fault diagnosis, this structure can clearly extract the high-frequency impact of gear tooth breakage and the low-frequency vibration of shaft bending at the same time.

[0111] Based on the input features, the main branch performs dilated convolution operation with a specified dilated rate and LeakyReLU activation function processing, the skip connection branch performs one-dimensional convolution operation and multiplies the skip connection weight of this level, and finally the main branch result and the skip connection branch result are added to obtain the residual block output, realizing feature extraction and information reservation. The residual block is constrained by the following formula:

[0112] ;

[0113] Wherein, is the dilated convolution operation with a dilated rate of ; is the dilated rate of the th residual block; is the residual block with level ; is the residual block level index; is the LeakyReLU activation function; is the skip connection branch, and the calculation method is represented as ; is the 1x1 convolution operation, which is used to keep the feature dimension unchanged; is the final output feature of the time-frequency dual-channel interaction gate neural network module.

[0114] Based on the final output feature of the input dual-channel interaction gate network, it is processed through three residual blocks of different levels in turn to obtain the output feature of the multi-scale residual dilated convolution module, realizing the fusion of multi-scale time features. The output of the multi-scale residual dilated convolution module is constrained by the following formula:

[0115] ;

[0116] Wherein, is the residual block with level ; is the residual block with level ; is the residual block with level ; is the output feature of the multi-scale residual dilated convolution module.

[0117] Further, in some preferred embodiments of the present application, a fusion physical prior output layer structure is used to respectively perform average value and maximum value operations on the output features of the multi-scale residual dilated convolution module along the time dimension, the results of the average value operation and the maximum value operation are spliced to obtain a compressed feature vector, and the diagnostic result of the model is output based on the compressed feature vector.

[0118] Specifically, the conventional output layer mostly uses a simple full connection structure, which ignores the physical causal relationship between the vibration fault and the frequency response, and is easy to cause the model classification result to fail to reflect the physical interpretability of the fault feature, especially when the accuracy decreases obviously in the presence of class confusion. The embodiment of the present application adopts a weighted output layer structure based on physical priori, utilizes the theoretical relationship between different fault categories and frequency band energy distribution, adjusts the attention intensity of the output layer to different frequency domain features through a physical priori matrix, combines multi-scale statistical features to form a final prediction result, thereby realizing the collaborative optimization between physical interpretability and data-driven model, and the specific steps include: multi-scale feature compression and physical priori matrix injection.

[0119] The multi-scale feature compression includes: based on the output features of the multi-scale residual dilated convolution module, respectively performing average value and maximum value operations along the time dimension, then splicing the two results to obtain a compressed feature vector, realizing fusion of time series statistical information, and the compressed feature vector is constrained by the following formula:

[0120] ;

[0121] Wherein, represents a splicing operation; is a compressed feature vector, which fuses time series average and maximum value information; is the value of the output feature of the multi-scale residual dilated convolution module at the i-th time step; represents the operation of taking the maximum value of the i-th time step; is the total number of time steps of the output feature of the multi-scale residual dilated convolution module. The physical priori matrix injection includes: based on the fault category and the frequency band index, a physical priori matrix is constructed, the elements of which represent the theoretical energy weight of a specific fault in a specific frequency band, then, based on the compressed feature vector, a linear transformation is performed, at the same time, combining the average vector of the frequency domain feature and the physical priori matrix, the predicted fault category probability distribution is obtained through Softmax processing, realizing the fusion of the fault physical mechanism constraint classification result, and the physical priori matrix is constrained by the following formula:

[0122]

[0123] ; ​​

[0124] where, is the energy weight of the c-th failure mode in the k-th frequency band. The c-th failure mode is in the k-th frequency band. The energy weight of the c-th failure mode in the k-th frequency band.

[0125] In some preferred embodiments of the present application, the theoretical energy proportion of the c-th failure mode in the k-th frequency band is obtained based on failure mechanism knowledge, such as bearing damage energy concentrating in high frequency band, gear abnormality in meshing frequency, and the energy distribution proportion of each frequency band is calculated through historical failure data or simulation, and the theoretical energy proportion of the k-th frequency band is calculated based on the average energy of the c-th failure mode in the k-th frequency band, which is represented as , which is obtained by statistical analysis of the annotated failure sample spectrum.

[0126] Further, the output layer calculation is:

[0127] ;

[0128] where, is a Softmax function that converts the output into a probability distribution; is the average vector of the frequency domain features, which is calculated by averaging the frequency domain features of each sample along the time step; is the prior injection intensity coefficient, which adjusts the weight of the physical constraint, such as ; is the predicted failure class probability distribution; is the weight matrix of the output layer, which is a trainable parameter; is the failure-frequency band energy correlation matrix, and the element in the i-th row and the k-th column of the matrix represents the energy weight of the c-th failure mode in the k-th frequency band.

[0129] It should be noted that the present application encodes prior knowledge such as bearing damage needing to focus on high frequency and gear abnormality needing to focus on meshing frequency into matrix elements. The physical constraint is injected into the output layer through the Softmax function calculation, so that the model learns both data rules and physical rules, and the controls the strength of the physical constraint to achieve a balanced adjustment between knowledge-driven and data-driven.

[0130] Further, the probability distribution output by the conventional neural network may not satisfy the actual failure evolution logic, such as some high-order failures necessarily containing pre-failure states, but logical conflicts may occur in model prediction, affecting the reliability of the results.

[0131] Therefore, in some preferred embodiments of the present invention, the present invention constructs a finite state machine based on known fault evolution relationships, combines the spectral harmonic factor calibration prediction results, uses the state transition matrix to filter physically unreasonable state transition paths, and dynamically corrects the probability distribution of classification results by linking with the harmonic factor weights calculated from the spectral features, thereby improving the rationality and consistency of prediction.

[0132] The final output of the basic fault diagnosis model is based on the output result constraints of the pre-set fault state transition matrix and the output layer structure that integrates physical priors, including: correction based on the preset transition matrix and preset harmonic factors; wherein, the transition matrix represents whether two fault states can be directly transferred; the harmonic factors are determined based on the vibration spectrum.

[0133] Specifically, the element values ​​in the fault state transition matrix indicate whether a fault state can directly transition to another fault state. For example, imbalance can transition to misalignment, which is set to 1; bearing cracks cannot directly transition to shaft bending, which is set to 0. This is to encode the fault evolution rules. The fault state transition matrix is ​​represented as follows:

[0134] ;

[0135] in, This is the fault state transition matrix; For matrix The element represents the first element. The first fault to the first The possibility of fault transfer; This serves as the fault category index for the fault state transition matrix calculation process; To distinguish from The fault category index in the fault state transition matrix calculation process.

[0136] Based on the vibration spectrum, harmonic factors are calculated. Based on these harmonic factors, fault category-specific harmonic factor thresholds, and slope factors, the harmonic factor weights for each fault category are calculated using the Sigmoid activation function. Finally, the initial predicted probability distribution is multiplied by the harmonic factor weight vector, and then multiplied by the fault state transition matrix to obtain the corrected fault probability distribution. This achieves the correction of the predicted probability based on spectral harmonic characteristics and fault transition rules. The corrected predicted probability is constrained by the following formula:

[0137] ;

[0138] in, To correct the fault probability distribution; The harmonic factor weight vector is calculated as follows: ; For the first Harmonic factor weights for fault types; Indexed by category; This is the slope factor, for example, set to 1.0; The harmonic factor characterizes the distribution of harmonic energy in the spectrum, and its calculation method is expressed as follows: ; For the first The harmonic factor threshold for fault types is set according to the physical characteristics of the fault type. For example, bearing damage is significantly caused by high-frequency harmonics, so the threshold is set accordingly. ; For the first in the spectrum The amplitude of the second harmonic; A positive integer, representing the harmonic order; This is the amplitude of the fundamental wave in the spectrum, which is also the amplitude of the first harmonic.

[0139] It should be noted that, and The fundamental and harmonic amplitude values ​​are obtained by extracting the spectrum of the vibration signal after fast Fourier transform. The amplitude of the first harmonic. For the first Second harmonic amplitude.

[0140] It should also be noted that neural networks may output results that violate fault evolution logic, such as skipping "slight imbalance" and directly diagnosing "bearing fracture". This invention addresses this by using a state transition matrix. Block physically impossible transfer paths, such as Using harmonic factor weight vector The category probability is dynamically adjusted based on the actual spectral characteristics, including the harmonic factor. Quantifying harmonic richness: When equipment simultaneously exhibits imbalances with high fundamental frequency amplitude and bearing damage with abundant harmonics... The value triggers an increase in the harmonic factor weight of the bearing damage category, while the transfer matrix allows two types of faults to coexist, improving the identification rate of composite faults.

[0141] Because there may be feature overlap between different fault categories, the model is prone to confusion when distinguishing boundary samples. Conventional regularization terms such as L2 or dropout may fail to explicitly distinguish inter-class characteristics at the feature level.

[0142] Furthermore, in some preferred embodiments of the present invention, the loss function of the fault diagnosis model is based on the joint constraint of standard cross-entropy loss and feature decoupling regularization loss; wherein, the feature decoupling regularization term includes: intra-class aggregation and inter-class separation; the fault diagnosis model updates parameters based on error backpropagation of gradient projection.

[0143] Specifically, in this embodiment of the invention, a feature decoupling regularization term is used in the loss function, including intra-class aggregation and inter-class separation. By maximizing the mean distance of features between different categories and minimizing feature fluctuations between samples of the same class, the model learns more discriminative feature representations, enhancing the clarity of boundaries between fault categories. Constructing the loss function includes the following steps A1 to A4:

[0144] Step A1, calculate the intra-class aggregate items:

[0145] Based on the samples belonging to the same fault category in each training batch and their corresponding multi-scale residual dilated convolutional module output features, the feature mean vector of that category is calculated. Then, the Euclidean distance between the batch mean vector and the preset target mean vector of that category is calculated. Finally, the distances of all categories are summed to obtain the intra-class aggregation loss term, thereby constraining the clustering of features of similar samples, as expressed as:

[0146] ;

[0147] in, This is the intra-class aggregation loss term; Total number of categories; For the current training batch belonging to the first The number of samples in each category; For the first The output features of the multi-scale residual dilated convolution module corresponding to each sample; For the first The target vector of the feature mean vector of each category is calculated based on the global training data.

[0148] Step A2, calculate the inter-class separation term:

[0149] Based on the feature mean vectors of all different fault category pairs, the Euclidean distance between them is calculated; when this distance is less than a preset minimum distance threshold, a penalty term is generated; finally, the penalty terms of all category pairs that meet the conditions are summed and multiplied by the decoupling strength coefficient to obtain the inter-class separation loss term, thereby forcing the features of different classes to separate from each other, expressed as:

[0150] ;

[0151] in, This is the inter-class separation loss term; For the first The feature mean vector of each category; For the first The feature mean vector of each category; The minimum inter-class spacing threshold, such as, =2.0; Characterizing the first The category and the Euclidean distance between class mean vectors; is the maximum value function; characterizes the positive penalty when the inter-class distance is less than , otherwise outputs 0; is the decoupling strength coefficient, which controls the weight of the inter-class separation term, such as .

[0152] Step A3, synthesis of decoupling regularization loss:

[0153] Add the intra-class aggregation loss term and the inter-class separation loss term to obtain the feature decoupling regularization loss, denoted as:

[0154] ;

[0155] wherein, is the feature decoupling regularization loss.

[0156] It should be noted that for similar faults such as "slight imbalance" and "early bearing damage", the feature overlap of is the main cause of misjudgment, and the intra-class aggregation loss term forces samples of the same class to be close to the class center, and the inter-class separation term pushes the distance between different class centers to at least , so as to realize the identification of the transition state of the fault, such as the development from imbalance to misalignment, and the feature space forms a continuous trajectory rather than a discrete jump, which is more in line with the device degradation law, and secondly, the decoupling constraint enables the model to focus on the relative pattern rather than the absolute amplitude.

[0157] Step A4, define the total loss function:

[0158] Add the standard cross-entropy loss and the feature decoupling regularization loss to obtain the total loss function for model training, denoted as:

[0159] ;

[0160] wherein, is the total loss function of the model; is the cross-entropy loss, which calculates the standard cross-entropy loss between the model prediction result and the true label, and is calculated based on the oneHot encoding of the true label and the corrected fault probability distribution .

[0161] In the prior art, although the feature decoupling loss can improve the discrimination ability, it will cause a gradient direction conflict with the original training target of the model, affecting the overall convergence, and the conventional optimizer does not have a control mechanism for the conflicting gradient in this scenario.

[0162] The embodiment of the application adopts a gradient direction orthogonal projection strategy, projects a decoupling regularization term gradient component to a direction orthogonal to a main task gradient, and then synthesizes the final update vector with a cross-entropy gradient, so as to relieve an optimization conflict between loss functions, improve training stability and accelerate convergence speed, and specifically comprises the following steps B1 to B2.

[0163] Step B1, gradient projection:

[0164] The gradient vectors of the cross-entropy loss part and the feature decoupling regularization loss part in the total loss function on the model parameters are calculated respectively, the feature decoupling regularization loss gradient vector is projected onto a direction orthogonal to the cross-entropy loss gradient vector, the projected decoupling gradient vector is obtained, the conflict with the main task gradient is eliminated, and is expressed as:

[0165] ;

[0166] wherein, is the gradient vector of the cross-entropy loss, and the calculation mode is expressed as ; is the gradient vector of the feature decoupling regularization loss, and the calculation mode is expressed as ; is the gradient of the cross-entropy loss on the model parameters ; is all the trainable parameters of the model, which is a parameter set; is the gradient of the parameter ; is the projected decoupling gradient, which is orthogonal to the cross-entropy gradient. It should be noted that,

[0167] the term adopts a conflict gradient orthogonal decomposition mode, projects the gradient vector of the feature decoupling regularization loss onto a subspace orthogonal to the gradient vector of the cross-entropy loss , eliminates the optimization direction conflict, and the orthogonal projection automatically suppresses abnormal gradient directions caused by noise. Step B2, adaptive update:

[0168] The cross-entropy loss gradient vector and the projected decoupling gradient vector are added according to the projection weight to obtain the final gradient vector, a proportion factor of the norm of the final gradient vector based on the final gradient vector, a basic learning rate and the norm of the cross-entropy loss gradient vector, the model parameter is updated, the stable convergence is realized, and is expressed as:

[0169]

[0170] ;

[0171] ;​

[0172] wherein, is the gradient vector for final parameter update; is the projection weight, such as, ; is the parameter update operation; is the base learning rate, such as, .

[0173] Further, in some preferred embodiments of the present application, the model is prevented from overfitting and the training efficiency is improved by setting a training termination condition.

[0174] First, the maximum number of iterations, such as 300 rounds, is set as the upper limit of training during the model training process; second, 10% of the samples in the training set can be divided as a validation set before training, and the validation set cross-validation mechanism is adopted, the loss function value and accuracy are calculated on the independent validation set after each training, and when the validation loss does not decrease significantly or the accuracy does not improve for 10 consecutive rounds, the early stopping mechanism is started to terminate the training in advance.

[0175] In addition, the optimal performance indicators of the model during the training process are recorded, and the parameters are rolled back to the optimal state before termination, ensuring that the final deployed model achieves the best balance between performance and generalization ability.

[0176] Step S108, a preset number of fault diagnosis models are bundled to construct a pump equipment state reasoning model; wherein, the weight is determined based on the information entropy of the fault category probability distribution of the fault diagnosis model; and the fault category with the largest cumulative weight is determined as the output of the pump equipment state reasoning model.

[0177] Specifically, a single model shows overfitting or misjudgment tendency when facing complex working conditions or non-steady state signals, and conventional integration methods often average the results of each sub-model, ignoring the difference in prediction uncertainty. The embodiment of the present application adopts a deterministic weight integration strategy, trains multiple sub-models through the Bootstrap method, and calculates the weight according to the information entropy of each model prediction result. The model with low uncertainty has a higher weight, thereby improving the robustness and judgment stability of the integrated model under non-steady state conditions, and realizing high-credibility final fault recognition output.

[0178] Further, in some preferred embodiments of the present application, each fault diagnosis model in the pump equipment state reasoning model is trained based on a random number of fixed proportion of training data; a plurality of sub-models are trained, and the sub-model fault diagnosis model is, for example, 100 is constructed, each sub-model is trained based on part of the training data sampled with replacement, and a random discard operation is applied to the gating recurrent unit in the double-channel interactive gating network during the training process. Specifically, 80% of the training data is randomly sampled with replacement.

[0179] Based on the probability distribution of fault categories predicted by each sub-model, its information entropy is calculated. Then, the weights of each sub-model are calculated based on the information entropy, so that higher weights are assigned to more certain predictions. For each candidate fault category, the weights of all sub-models whose prediction results are for that category are accumulated. The final classification result is the fault category with the largest accumulated weight, thus achieving the effect of robust decision-making by the ensemble model.

[0180] The final classification result of the pump equipment state reasoning model is constrained by the following formula:

[0181] ;

[0182] ;

[0183] ;

[0184] in, For the first The weights of each member model; A positive integer representing the index of the member model; For the first The entropy of the predicted probability distribution of a model measures uncertainty; For the first The predicted probability distribution of each model; It should be a small constant to prevent division by zero errors; Total number of categories; For the first The model for the first The predicted probabilities of each category; The final classification result category; For indicator functions; To find the category index c that maximizes the expression; c is the category index; This represents the number of valid member models in the ensemble model. To obtain the category index k that maximizes the expression; κ is the slope factor; For the first The model for the first The predicted probabilities of each category; For the first The category index with the highest predicted probability for each model is equal to... ,Right now, Characterization when When the condition for the indicator function is met.

[0185] Step S110: Obtain the operating vibration signal of the pump equipment of the target category, input the operating vibration signal into the pump equipment state reasoning model, and output the fault category.

[0186] Specifically, after the pump equipment state inference model is trained, the inference stage is entered to realize automatic classification and judgment of the pump equipment state.

[0187] Specifically, the real-time vibration signal of the to-be-tested pump equipment is input into the model process, and the pre-processing, feature construction, time-frequency fusion, state inference and other operations are sequentially performed, and finally the running state category of the pump equipment is output.

[0188] The model realizes high-precision discrimination of a plurality of states such as "normal running", "slight imbalance", "serious misalignment", "bearing damage" and "gear meshing anomaly" according to the time sequence and frequency domain feature rules learned in the training stage, in combination with physical priori and multi-scale expression.

[0189] The inference result is output in a structured format, including fault category, classification confidence and the like, and can be used for real-time monitoring and early warning decision support of the pump equipment operation and maintenance system.

[0190] Further, in some preferred embodiments of the present application, the method further comprises: inputting the vibration signal of the pump equipment of the target category after labeling into a conventional time domain feature combined with a support vector machine method, a wavelet transform combined with a neural network method, a one-dimensional convolutional neural network method and a pump equipment state inference model respectively, outputting fault categories respectively, and performing chart analysis.

[0191] Specifically, the recognition ability of the method on a plurality of pump equipment fault types is evaluated, and comparison is made with three mainstream technologies (conventional time domain feature combined with support vector machine method, wavelet transform combined with neural network method and one-dimensional convolutional neural network method), as shown in Figure 3 The embodiment of the present application provides a comparison diagram of accuracy rates of different models on various faults, wherein the horizontal axis shows five typical pump equipment states, and the vertical axis represents the classification accuracy rate, the method of the present application is presented in a prominent red color in the column chart and is filled with a star pattern, and the experimental results show that all methods perform well for the normal running state, but the accuracy rate of the conventional method significantly decreases for complex faults such as bearing damage and gear meshing anomaly, and the method of the present application still maintains a high position, especially for the gear meshing anomaly fault with obvious high-frequency characteristics, the method of the present application effectively captures the fault characteristics through adaptive frequency band decomposition and dual-channel feature fusion, and the leading advantage in accuracy rate is most prominent.

[0192] Further, in the present embodiment, a common noise interference scene in an industrial field is simulated, the signal-to-noise ratio (horizontal axis) of the system is changed to test the noise resistance of each method, and reference is made to Figure 4A different signal-to-noise ratio condition under which the embodiment of the application provides a comparison diagram of the robustness of each method, the left low signal-to-noise ratio area in the figure simulates a strong noise environment, and the right high signal-to-noise ratio area is close to an ideal working condition. Experimental data show that when the signal-to-noise ratio is lower than 5dB, the accuracy of the conventional method decreases sharply, and the wavelet transform method also fluctuates obviously. However, the method of the application (red pentagonal marker line) still maintains a gentle downward trend under strong noise, indicating that the time-domain impact enhancement and frequency band energy weighting mechanism in the method of the application can effectively suppress noise interference. The scatter points around each signal-to-noise ratio point in the figure represent the results of multiple repeated experiments. The scatter points of the method of the application are more concentrated, verifying the stability advantage thereof.

[0193] Further, in the present embodiment, the effect of the feature decoupling regularization mechanism is intuitively displayed by feature space distribution. Referring to Figure 5 The feature comparison diagram of the conventional method and the method provided by the present embodiment is shown. In the feature distribution of the conventional method and the feature distribution, two most representative feature dimensions are selected. In the conventional method, the bearing damage (blue point), gear meshing abnormality (orange point), and serious misalignment (green point) three types of samples are seriously overlapped in the feature space, the boundary is blurred, leading to classification confusion, while the three types of samples of the method of the application form a closely aggregated cluster group, and the distance between classes is significantly increased. The red pentagonal star in the figure represents the center point of each class. The distance between the class centers of the method of the application is larger, and the distribution of the same type of samples around the center is more compact. It is shown that the intra-class aggregation and inter-class separation constraint in the loss function of the application enables the model to better distinguish similar faults and significantly reduces the misjudgment rate.

[0194] Further, in the present embodiment, the convergence characteristics of each method in the training process are also compared, referring to Figure 6 A convergence speed comparison diagram provided by the embodiment of the application is shown. The horizontal axis represents the training round, and the vertical axis is the accuracy of the verification set. The method of the application (red curve) rapidly rises in the early training stage, and reaches a high-precision platform at about 30 rounds, while the conventional method needs more than 60 rounds to reach the maximum. The golden pentagonal star marks the convergence point (round of reaching stable accuracy) of each method. The method of the application converges the earliest. Experimental results show that the gradient projection mechanism alleviates the optimization conflict, the multi-scale residual structure accelerates feature extraction, and provides effective guidance for the physical prior matrix. It is worth noting that the method of the application not only converges faster, but also has higher final accuracy. The end of the curve still maintains an upward trend, indicating that the model has stronger continuous learning ability.

[0195] The embodiment of the application realizes joint perception of transient impact and spectral mode through time domain impact enhancement + adaptive frequency band energy extraction, constructs time-frequency dual-channel features, and realizes dynamic weighted fusion through interactive gating neural network. A multi-scale residual network with cascaded inflation rate is proposed, which adjusts the receptive field periodically to cover different time scales from microseconds to seconds, and accurately extracts the vibration features of different faults. The fault-frequency band energy matrix and the state transition rule are used to perform prior injection and harmonic correction in the output layer, so that the model has physical interpretability and classification consistency, and the composite fault recognition ability is improved. In the loss function, the regularization term of intra-class aggregation and inter-class separation is used, and the conflict between the main task and the regularization term is solved through the gradient projection mechanism, so that the boundary sample distinguishability and the training convergence are simultaneously improved.

[0196] The application provides a kind of based on artificial intelligence's machine pump class equipment state monitoring and fault diagnosis method, comprising: based on pre-set vibration sensor obtains the vibration signal of target category's machine pump equipment, and is labeled;Wherein, the label of labeling represents the running state category of target category's machine pump equipment;Based on pre-set double-channel feature extraction model extracts the feature of vibration signal;Wherein, double-channel feature extraction model includes: time domain feature extraction channel and frequency domain feature extraction channel;The feature of vibration signal includes: based on vibration signal time domain feature and vibration signal frequency domain feature splicing fusion feature;Based on the feature of vibration signal, preset basic fault diagnosis model is iteratively trained until reaching preset training completion condition, and the basic fault diagnosis model of completion training is determined as fault diagnosis model;Wherein, basic fault diagnosis model includes: time-frequency dual-channel interactive gating neural network module, multi-scale residual dilated convolution module and fusion physical prior output layer structure connected in turn;Time-frequency dual-channel interactive gating neural network module is used to fuse vibration signal time domain feature and vibration signal frequency domain feature;Multi-scale residual dilated convolution module is used to extract multi-period feature and enhance time scale perception;Realize the explanatory classification prediction of machine pump equipment state;The final output of basic fault diagnosis model is based on pre-set fault state transition matrix and the output result constraint of fusion physical prior output layer structure;A preset number of fault diagnosis models are bundled to build pump class equipment state reasoning model;Wherein, the weight is determined based on the information entropy of the fault category probability distribution of fault diagnosis model;The fault category with the largest cumulative weight is determined as the output of pump class equipment state reasoning model;Obtain the running vibration signal of target category's machine pump equipment, input running vibration signal into pump class equipment state reasoning model, and output fault category;Through time-frequency dual-channel fusion and multi-scale perception, the fault recognition precision is improved;The rationality and interpretability of the result are enhanced by using physical prior constraint;And through model integration optimization, the classification stability and reliability in complex scene are improved.

[0197] Embodiment two

[0198] The embodiment of the application further provides an electronic device for running the machine pump type equipment state monitoring and fault diagnosis method based on artificial intelligence; see Figure 7 The electronic device provided by the embodiment of the application has the structure as shown in the figure, and the electronic device comprises a memory 400 and a processor 401, wherein the memory 400 is used for storing one or more computer instructions, and the one or more computer instructions are executed by the processor 401 to realize the machine pump type equipment state monitoring and fault diagnosis method based on artificial intelligence.

[0199] Further, Figure 7 The electronic device further comprises a bus 402 and a communication interface 403, and the processor 401, the communication interface 403 and the memory 400 are connected through the bus 402.

[0200] The memory 400 can contain a high-speed random access memory (RAM), and can also include a non-volatile memory, for example, at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 403 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network and the like can be used. The bus 402 can be an ISA bus, a PCI bus or an EISA bus and the like. The bus can be divided into an address bus, a data bus, a control bus and the like. For the convenience of representation, Figure 7 Only one bidirectional arrow is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0201] The processor 401 can be an integrated circuit chip having a signal processing capability. In the implementation process, each step of the above method can be completed by the integrated logic circuit or the instruction in the form of software in the processor 401. The processor 401 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium in the art. The storage medium is located in the memory 400, and the processor 401 reads the information in the memory 400, and combines the hardware to complete the steps of the method of the above-mentioned embodiments.

[0202] The embodiment of the present application also provides a computer readable storage medium, the computer readable storage medium stores computer executable instructions, when the computer executable instructions are called and executed by the processor, the computer executable instructions cause the processor to implement the above-mentioned artificial intelligence-based machine pump equipment state monitoring and fault diagnosis method, and the specific implementation can be referred to the method embodiment, and will not be repeated here.

[0203] The computer program product of the artificial intelligence-based machine pump equipment state monitoring and fault diagnosis method, device and electronic equipment provided by the embodiment of the present application includes a computer readable storage medium storing program codes, the instructions included in the program codes can be used to execute the method in the foregoing method embodiment, and the specific implementation can be referred to the method embodiment, and will not be repeated here.

[0204] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and / or device described above can refer to the corresponding process in the foregoing method embodiment, and will not be repeated here.

[0205] In addition, in the description of the embodiments of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood in a broad sense, for example, can be fixedly connected, or can be detachably connected, or integrally connected; can be mechanically connected, or can be electrically connected; can be directly connected, or indirectly connected through an intermediate medium; can be internal communication of two elements. For those skilled in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0206] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the present application which essentially contributes to the prior art or the part of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0207] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An artificial intelligence-based machine pump device state monitoring and fault diagnosis method, characterized by, include: Vibration signals of pumps and machinery of the target category are acquired based on pre-set vibration sensors and labeled; wherein, the labeled label represents the operating status category of the pumps and machinery of the target category; The vibration signal features are extracted based on a pre-set dual-channel feature extraction model; wherein, the dual-channel feature extraction model includes: a time-domain feature extraction channel and a frequency-domain feature extraction channel; the vibration signal features include: fused features based on the splicing of the vibration signal's time-domain features and frequency-domain features. Based on the features of the vibration signal, a preset basic fault diagnosis model is iteratively trained until a preset training completion condition is met. The trained basic fault diagnosis model is then determined as the fault diagnosis model. The basic fault diagnosis model includes: a time-frequency dual-channel interactive gated neural network module, a multi-scale residual dilated convolution module, and an output layer structure fused with physical priors, connected sequentially. The time-frequency dual-channel interactive gated neural network module is used to fuse the time-domain features and frequency-domain features of the vibration signal. The multi-scale residual dilated convolution module is used to extract multi-period features and enhance time-scale perception. The output layer structure fused with physical priors is used to achieve interpretative classification and prediction of the pump equipment status. The final output of the basic fault diagnosis model is constrained by a pre-set fault state transition matrix and the output results of the output layer structure fused with physical priors. A predetermined number of fault diagnosis models are bundled together to construct a state reasoning model for pump equipment; wherein, the weights are determined based on the information entropy of the fault category probability distribution of the fault diagnosis models; and the fault category with the largest accumulated weight is determined as the output of the state reasoning model for pump equipment. The operating vibration signal of the pump equipment of the target category is obtained, the operating vibration signal is input into the state reasoning model of the pump equipment, and the fault category is output. The multi-scale residual dilated convolution module comprises: a residual block with a level of 1, a residual block with a level of 2 and a residual block with a level of 3 connected in sequence; the residual block is constrained by the following formula: ; wherein, is a dilated convolution operation with a dilated rate of ; is a dilated rate of the residual block of the th layer; is a residual block with a level of ; is a residual block level index; is a LeakyReLU activation function; is a skip connection branch, and the calculation mode is represented as ; is a 1*1 convolution operation for keeping the feature dimension unchanged; is the final output feature of the time-frequency dual-channel interaction gate neural network module; the output of the multi-scale residual dilated convolution module is constrained by the following formula: ; wherein, is the residual block with a level of ; is the residual block with a level of ; is the residual block with a level of ; is the output feature of the multi-scale residual dilated convolution module; The final output of the basic fault diagnosis model is based on the pre-set fault state transition matrix and the output result constraints of the output layer structure that integrates physical priors, including: Correction is performed based on a preset transfer matrix and a preset harmonic factor; wherein, the transfer matrix represents whether two fault states can be directly transferred; and the harmonic factor is determined based on the vibration spectrum. 2.The AI-based machine-pump device condition monitoring and fault diagnosis method of claim 1, wherein, The vibration sensor is installed at any of the following locations on the pump: non-drive end of the motor, drive end of the motor, drive end of the pump, and non-drive end of the pump; the label includes at least one of the following: normal operation, slight imbalance, severe misalignment, bearing damage, abnormal gear meshing, and unknown state.

3. The method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence according to claim 1, characterized in that, The time-domain feature extraction channel is used to perform absolute value and impact enhancement coefficient amplification operations on the vibration signal, and then apply exponential moving average processing to obtain the time-domain features of the vibration signal. The frequency-domain feature extraction channel is used to perform adaptive frequency band decomposition based on the vibration signal, perform fast Fourier transform on the signal in each frequency band and calculate the energy, calculate the weight based on the variance of the energy of each frequency band, and sum the weighted energy of all frequency bands to obtain the frequency-domain features of the vibration signal. The adaptive frequency band decomposition divides the signal into weighted frequency bands through a Gaussian frequency band window, the center frequency of the frequency band is obtained by clustering calculation based on the spectral envelope curve, and the bandwidth is determined by calculation based on the frequency axis range and preset parameters.

4. The method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence according to claim 1, characterized in that, The time-frequency dual-channel interactive gated neural network module includes: a time-domain path, a frequency-domain path, and a cross-channel interaction module; The time-domain path is used to input the time-domain features of the vibration signal into a gated loop unit for processing. The gated loop unit calculates the update gate output and the reset gate output based on the hidden state of the previous time step and the time-domain features of the current time step. Based on the reset gate output, the hidden state of the previous time step and the current time-domain features of the vibration signal, it calculates the candidate hidden state. Based on the update gate output, the hidden state of the previous time step and the candidate hidden state, it calculates the hidden state of the current time step to capture temporal dependencies. The frequency domain path is used to perform one-dimensional convolution operations on the real and imaginary parts of the frequency domain features respectively, and multiply the imaginary part convolution result by an imaginary unit and add it to the real part convolution result to obtain the complex domain convolution result. The complex domain convolution result is processed by the ReLU activation function to obtain the output feature of the frequency domain path, so as to enhance the phase sensing capability of the frequency domain features. The cross-channel interaction module is used to concatenate the hidden state and the output features of the frequency domain path at the current time step to obtain concatenated features. The concatenated features are then subjected to linear transformation and Sigmoid activation function processing to generate a gating weight vector. Based on the gating weight vector, the hidden state and the output features of the frequency domain path at the current time step are weighted and fused to obtain the final output features of the time-frequency dual-channel interactive gating neural network module, thereby realizing dynamic fusion of time-frequency features.

5. The method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence according to claim 1, characterized in that, The output layer structure that integrates physical priors is used to perform average and maximum value operations on the output features of the multi-scale residual dilated convolution module along the time dimension, respectively. The results of the average operation and the maximum value operation are concatenated to obtain a compressed feature vector, and the diagnostic results of the model are output based on the compressed feature vector.

6. The method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence according to claim 1, characterized in that, The loss function of the fault diagnosis model is based on the joint constraints of standard cross-entropy loss and feature decoupling regularization loss; wherein, the feature decoupling regularization term includes: intra-class aggregation and inter-class separation; the fault diagnosis model updates parameters based on backpropagation of gradient projection error.

7. The method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence according to claim 1, characterized in that, Each of the fault diagnosis models in the pump equipment state reasoning model is trained based on a fixed proportion of random training data. The final classification result of the pump equipment state reasoning model is constrained by the following formula: ； ； ； in, For the first The weights of each member model; A positive integer representing the index of the member model; For the first The entropy of the predicted probability distribution of a model measures uncertainty; For the first The predicted probability distribution of each model; It should be a small constant to prevent division by zero errors; Total number of categories; For the first The model for the first The predicted probabilities of each category; The final classification result category; For indicator functions; To find the category index c that maximizes the expression; c is the category index; This represents the number of valid member models in the ensemble model. To obtain the category index k that maximizes the expression; For the first The model for the first The predicted probabilities of each category; For the first The class index with the highest predicted probability for each model is equal to... ,Right now, Characterization when When the condition for the indicator function is met.

8. The method for condition monitoring and fault diagnosis of pumps and other equipment based on artificial intelligence according to claim 1, characterized in that, The method further includes: The vibration signals of pumps and motors of the target category after being labeled are input into the conventional time-domain feature combined with support vector machine method, wavelet transform combined with neural network method, one-dimensional convolutional neural network method and the state reasoning model of the pump equipment respectively, and the fault category is output and analyzed by chart.

Images