Mine water pump bearing fault diagnosis method based on multi-dimensional information interaction fusion

Abstract

The invention discloses a multi-dimensional information interactive fusion mine water pump bearing fault diagnosis method, and belongs to the technical field of bearing fault diagnos.The method comprises the steps that firstly, a fault diagnosis experiment platform is built through existing equipment, and experiment data are collected; secondly, filtering and noise reduction processing is carried out on the collected data by utilizing a wavelet threshold function, and signal-picture and signal-point cloud conversion is carried out; and then, building a multi-dimensional information interaction fusion fault diagnosis model, specifically, extracting one-dimensional time domain signal features by using one-dimensional convolution, extracting two-dimensional time frequency signal features by using two-dimensional convolution, mapping the two features into three-dimensional point cloud data through physical model mapping, and then performing interaction fusion with three-dimensional point cloud features extracted by Point CNN to obtain a fault diagnosis result. And fault diagnosis classification is carried out. Finally, the method is applied to an actually acquired data set, and experiments show that the method can describe the feature information more comprehensively, can effectively improve the fault diagnosis accuracy, and has practical engineering application value.

Description

Technical Field

[0001] The present invention belongs to a fault detection method for water supply system equipment, and particularly relates to a fault diagnosis method for mine water pump bearings with multi-dimensional information interaction and fusion. Background Art

[0002] Large water pumps are widely used in mine water supply systems. For example, when supplying water to the ground, large water pumps are often installed in the pump stations at the water source to transport water to various water-using points in the mine. However, with the long-term operation of the water pump, for example, the bearing components will generate high temperatures due to poor lubrication and excessive friction, resulting in increased wear of the bearing materials and ultimately damage. Secondly, when the outlet valve of the water pump suddenly closes, the impact force of the water flow will change instantaneously, causing the axial force and radial force borne by the water pump impeller to increase sharply, resulting in the bearing bearing a load exceeding its design limit and being damaged. Once the above factors occur, it will cause equipment damage at least and endanger personnel safety at worst. Therefore, there is an urgent need to find a method for diagnosing the faults of large water pump bearings to predict the operating state of the water pump in advance and avoid unnecessary losses.

[0003] The current research methods mainly tend to the following three directions: (1) Only using one-dimensional time-domain signal analysis for judgment. For example, Wang et al. collected one-dimensional vibration signals of different faulty motors, then used the short-time Fourier transform (STFT) to preprocess the original signals to obtain the corresponding time-frequency diagrams, and then used a convolutional neural network (CNN) to adaptively extract the features of the time-frequency diagrams to achieve fault classification. Guo et al. used the EMD algorithm for one-dimensional vibration signals to complete signal-noise separation, selected the high-signal-to-noise ratio components for FFT algorithm transformation, and analyzed the bearing fault location through the spectrogram; (2) Using the combination of one-dimensional time-domain signals and two-dimensional time-frequency signals for analysis and judgment. For example, Xu et al. constructed a spatio-temporal feature fusion network, which is composed of a two-dimensional convolutional neural network (CNN) and a recurrent neural network (RNN), and simultaneously extracted the time-frequency information and time-domain information of the motor vibration signals. (3) Only using three-dimensional point cloud signal analysis for judgment. For example, Tan et al. transformed one-dimensional vibration signals into two-dimensional time-frequency diagrams through wavelet transform and then mapped them into three-dimensional point cloud data, and then built a PointNet++ deep learning model to train the point cloud data and perform bearing fault classification.

[0004] However, although the above research methods can achieve a relatively high fault diagnosis accuracy to a certain extent, there are serious phenomena of missed diagnosis and misdiagnosis in practical applications. After analysis, the reasons are as follows. When only using one-dimensional time-domain signal analysis for judgment: (1) It leads to a single information dimension, and only simple features such as the peak value and period of the vibration signal can be seen; (2) The fault feature recognition ability is limited. The inner ring fault and outer ring fault of the bearing may both show periodic pulses in the time-domain signal, but it is difficult to accurately distinguish the subtle differences of these pulses (such as pulse interval, amplitude change law, etc.) solely relying on the time-domain signal. When using the combined analysis of one-dimensional time-domain signal and two-dimensional time-frequency signal for judgment: (1) It leads to the lack of spatial structure information. The two-dimensional time-frequency signal can only provide the frequency distribution and time-frequency joint characteristics of the signal; (2) The complex fault diagnosis ability is limited. When the bearing has both wear and local crack faults at the same time, one-dimensional and two-dimensional signal analysis may only be able to identify that there is a fault, but the spatial-related information such as the specific position relationship and mutual influence degree of the two faults cannot be effectively obtained. When only using three-dimensional point cloud signal analysis for judgment, due to the possible loss of key feature information during the process of the signal changing from one-dimensional to two-dimensional and then to three-dimensional, it affects the accuracy of the final fault diagnosis of the model.

[0005] In summary, the present invention proposes a technical solution. Summary of the Invention

[0006] The purpose of the present invention is to propose a fault diagnosis method for mine water pump bearings with multi-dimensional information interaction and fusion in view of the problems of single information dimension acquisition, loss of key feature information, and incomplete description of features in the process of mine water pump bearing fault diagnosis. It aims to provide an idea for comprehensively describing fault features and enhancing complex fault diagnosis ability, and finally improve the accuracy and reliability of the actual application of fault diagnosis.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A fault diagnosis method for mine water pump bearings with multi-dimensional information interaction and fusion, which specifically includes the following steps:

[0009] Step 1: Collect the axial vibration signal of the centrifugal pump through a vibration test device, and finally export the data in the.mat file format;

[0010] Step 2: Filter the collected vibration signal data;

[0011] Step 3: Convert the filtered vibration signal into signal-picture and signal-point cloud;

[0012] Step 4: Build a multi-dimensional information interaction and fusion fault diagnosis model;

[0013] Step 5: Input the training set into the model for fault diagnosis, set an accuracy threshold, save the model with each training result higher than the threshold, and finally determine whether the model is optimal through the validation set, and select the optimal model as the real-time fault diagnosis model for the water pump bearing.

[0014] As a further solution of the present invention, the description formula for filtering the collected vibration signal data in step 2 is:

[0015]

[0016] In the formula: w j，k represents the wavelet coefficient, represents the quantized wavelet coefficient, T represents the threshold, N represents the signal length, sign(w j，k ) represents the sign function.

[0017] As a further solution of the present invention, the method for converting the filtered vibration signal into signal - image and signal - point cloud in step 3 is:

[0018] Step 301: Perform signal - image conversion on the filtered vibration signal using continuous wavelet transform, and the mathematical description formula is:

[0019]

[0020] In the formula: f(t) represents the signal to be transformed, represents the mother wavelet function, a and τ represent the scale parameter and the translation parameter, and Wt(a, τ) represents the wavelet coefficient of the signal f(t) under the scale parameter a and the translation parameter τ.

[0021] Step 302: Perform signal - point cloud conversion on the filtered vibration signal using continuous wavelet transform, and the mathematical description formula is:

[0022]

[0023] In the formula: x(t) represents the signal to be analyzed, fs represents the sampling frequency, f represents the signal frequency, and cfs represents the wavelet coefficient;

[0024] Step 303: Use the meshgrid function in matlab to generate a two - dimensional grid matrix for time and frequency At the same time, take the absolute value of the wavelet coefficient as the component, and its description formula is:

[0025]

[0026] In the formula: represents the time feature component, Represents the frequency feature components and represents the wavelet coefficient components;

[0027] Step 304: Use the reshape function to and Expand the matrix dimensions of the components, merge and extract them one by one according to the corresponding relationship, and obtain the three-dimensional point cloud data p after the conversion of the one-dimensional vibration signal. Its description formula is:

[0028]

[0029] As a further solution of the present invention, the method for building a multi-dimensional information interaction and fusion fault diagnosis model in step four is:

[0030] Step 401: Use one-dimensional convolution to extract the data features of the one-dimensional vibration signal;

[0031] Step 402: Use two-dimensional convolution to extract the data features of the converted picture;

[0032] Step 403: Normalize the time domain features of the one-dimensional vibration signal and the two-dimensional time-frequency features respectively and map them into three-dimensional features;

[0033] Step 404: Use the three-dimensional PointCNN neural network to extract the data features of the converted point cloud;

[0034] Step 405: Perform mutual interaction and fusion between the mapped three-dimensional features and the data features of the converted point cloud extracted by the three-dimensional PointCNN neural network.

[0035] As a further solution of the present invention, the mathematical description formula of step 401 is:

[0036] F i = f(cov1D(W i , X) + b i ), i = 1,...a

[0037] F A (t) = [F 1 ,..., F a

[0038] a is the length of the data features of the one-dimensional vibration signal extracted by one-dimensional convolution.

[0039] As a further solution of the present invention, the mathematical description formula of step 402 is:

[0040] F i = f(cov2D(W i , X) + b i ), i = 1,...bc ​

[0041] F BC (t)=[F 1 ,...,F bc

[0042] Where b×c is the size of the features of the converted image data extracted by two-dimensional convolution.

[0043] As a further solution of the present invention, in step 403, first, standard deviation normalization is performed on the time-domain features of the obtained one-dimensional vibration signal and the features of the converted image data extracted by two-dimensional convolution. Then, the time-domain feature vector is used as one dimension of the newly constructed three-dimensional feature. When it is used as the z-axis direction, for the two-dimensional time-frequency feature matrix, its elements are expanded into a vector with a length of bc in the order of rows or columns, and are used as the x and y-axis directions respectively. A three-dimensional feature vector is constructed through the mapping method of the physical model, and the description formula is:

[0044]

[0045] In the formula: μ represents the mean of this feature in the original dataset, δ represents the standard deviation normalization of this feature in the original dataset, p represents the mapped three-dimensional point cloud data, and f represents a physical mapping function.

[0046] As a further solution of the present invention, step 405 includes:

[0047] First, multiply the mapped point cloud data by the feature matrix of the point cloud data extracted by PointCNN;

[0048]

[0049] Secondly, calculate the point cloud feature attention distribution generated by the two methods row by row;

[0050]

[0051] Then, multiply the point cloud feature attention distribution generated by the two methods by the feature matrix to obtain the attention representation matrix;

[0052]

[0053] Next, use the multiplicative gating mechanism to obtain the point cloud feature interaction attention information matrix generated by the two methods;

[0054]

[0055] Finally, splice and fuse the two;

[0056]

[0057] ​In the formula: "·" represents matrix multiplication, and "⊙" represents element-wise matrix multiplication; represents the combined three-dimensional feature vector.

[0058] Advantages of the present invention:

[0059] 1. Compared with only using one-dimensional time-domain signal analysis for judgment, using a combination of one-dimensional time-domain signal and two-dimensional time-frequency signal analysis for judgment, or only using three-dimensional point cloud signal analysis, the present invention uses a mapping method of a physical model to map one-dimensional time-domain feature information and two-dimensional time-frequency feature information into three-dimensional point cloud feature information. Not only does the data presentation method change from abstract time series and frequency distribution to point clouds with spatial structure, but also the position of fault features in space can be clearly shown.

[0060] 2. Compared with other methods that only use PointNet to extract three-dimensional point cloud feature information, the PointCNN adopted by the present invention has stronger local feature extraction ability, better adaptability to irregular point clouds, and simultaneously extracts feature information in multiple dimensions using one-dimensional and two-dimensional convolutional neural networks, which can effectively avoid the loss of key information and ultimately improve the accuracy of water pump bearing fault diagnosis.

[0061] 3. Compared with other fault diagnoses that only utilize the feature information interaction at one-dimensional or two-dimensional levels, the present invention adopts the information interaction among three-dimensional levels. It can not only describe fault features more comprehensively, but also locate faults more accurately, and ultimately improve the accuracy and reliability of fault diagnosis. Description of the Drawings

[0062] The following further illustrates the present invention in conjunction with the drawings.

[0063] Figure 1 is the flowchart of the fault diagnosis method described in the present invention;

[0064] Figure 2 is the schematic diagram of the multi-dimensional information interaction and fusion fault diagnosis model described in the present invention;

[0065] Figure 3 is the schematic diagram of the data acquisition process described in the present invention;

[0066] Figure 4 is the schematic diagram of the one-dimensional vibration signal described in the present invention;

[0067] Figure 5 is the schematic diagram of the signal-picture conversion of the one-dimensional vibration signal described in the present invention;

[0068] Figure 6 is the schematic diagram of the signal-point cloud conversion of the one-dimensional vibration signal described in the present invention;

[0069] Figure 7Schematic diagram of the accuracy and loss change curves of model training described in the present invention. Detailed implementation manners

[0070] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0071] A fault diagnosis method for the bearing of a mine water pump with multi-dimensional information interaction and fusion, as Figure 1 shown, the method specifically includes the following steps:

[0072] Step 1: Install a magnetic vibration sensor at the axial position of the centrifugal pump and motor bearing housing, set the parameters of the data acquisition instrument to collect the axial vibration signal of the centrifugal pump, and finally store the data in the.mat file format;

[0073] Step 1 specifically includes:

[0074] Step 101: Use the LMS vibration test equipment to collect vibration signals and export them in the.mat file format;

[0075] Step 102: Import the vibration signals into Matlab for encoding different fault categories, and divide the data set according to the ratio of 7:2:1.

[0076] Step 2: Filter the vibration signal data after encoding and division, and the specific method is:

[0077]

[0078] where: w j，k represents the wavelet coefficient, represents the quantized wavelet coefficient, T represents the threshold, N represents the signal length, sign(w j，k ) represents the sign function.

[0079] Step 3: Convert the vibration signal after filtering into signal-image and signal-point cloud conversions, and the specific method is;

[0080] Step 301: Perform signal-image conversion on the filtered vibration signal using continuous wavelet transform, and the mathematical description formula is:

[0081]

[0082] where: f(t) represents the signal to be transformed, Let \(\psi(t)\) represent the mother wavelet function, \(a\) and \(\tau\) represent the scale parameter and the translation parameter respectively, and \(W_t(a,\tau)\) represents the wavelet coefficient of the signal \(f(t)\) under the scale parameter \(a\) and the translation parameter \(\tau\).

[0083] Step 302: Perform signal-point cloud conversion on the filtered vibration signal using continuous wavelet transform. The mathematical description formula is:

[0084]

[0085] In the formula: \(x(t)\) represents the signal to be analyzed, \(f_s\) represents the sampling frequency, \(f\) represents the signal frequency, and \(c_f\) represents the wavelet coefficient;

[0086] Step 303: Use the meshgrid function in matlab to generate a two-dimensional grid matrix for time and frequency At the same time, take the absolute value of the wavelet coefficient as component, and its description formula is:

[0087]

[0088] In the formula: represents the time feature component, represents the frequency feature component, and represents the wavelet coefficient component;

[0089] Step 304: Use the reshape function to expand the matrix dimensions of and components, and extract and merge them one by one according to the corresponding relationship to obtain the three-dimensional point cloud data \(P\) after the conversion of the one-dimensional vibration signal;

[0090]

[0091] Step Four: The multi-dimensional information interaction and fusion fault diagnosis model built by the present invention is as Figure 2 shown;

[0092] Step 401: Use one-dimensional convolution to extract the one-dimensional vibration signal data features with a length of \(a\). The mathematical description formula is:

[0093] F i = f(covlD(W i , X)+b i ), i = 1,...a

[0094] F A (t)=[F 1 ,..., F a

[0095] ​Step 402: Use two-dimensional convolution to extract the feature of the picture data with the size of b×c after continuous wavelet transform. The description formula is:

[0096] F i = f(conv2D(W i , X)+b i ), i = 1,..bc

[0097] F BC (t) = [F 1 ,..., F bc

[0098] Step 403: First, perform standard deviation normalization on the time-domain feature of the obtained one-dimensional vibration signal and the feature of the picture data after extraction and conversion by two-dimensional convolution. Then, use the time-domain feature vector as one dimension of the newly constructed three-dimensional feature. For example, use it as the z axis direction. For the two-dimensional time-frequency feature matrix, its elements can be expanded into a vector with a length of bc in the order of rows or columns, and used as the x and y axis directions respectively. A three-dimensional feature vector is constructed through the mapping method of the physical model. The mathematical description formula is:

[0099]

[0100] In the formula: μ represents the mean of this feature in the original dataset, δ represents the standard deviation normalization of this feature in the original dataset, p represents the mapped three-dimensional point cloud data, and f represents a mapping function;

[0101] Step 404: Interact and fuse the point cloud data features extracted by the three-dimensional PointCNN neural network and the newly mapped point cloud data features through the interactive attention mechanism. The mathematical description formula is:

[0102] First, multiply the feature matrices of the mapped point cloud data and the point cloud data extracted by PointCNN;

[0103]

[0104] Secondly, calculate the point cloud feature attention distribution generated by the two methods row by row;

[0105]

[0106] Then, multiply the point cloud feature attention distributions generated by the two methods with the feature matrix to obtain the attention representation matrix;

[0107]

[0108] Next, use the multiplicative gating mechanism to obtain the point cloud feature interactive attention information matrix generated by the two methods;​

[0109]

[0110] Finally, splice and fuse the two;

[0111]

[0112] In the formula: "·" represents matrix multiplication, and "⊙" represents element-wise matrix multiplication; represents the combined three-dimensional feature vector; represents the combined three-dimensional feature vector.

[0113] Embodiment

[0114] S1: Data acquisition

[0115] To explore the practical value of the large water pump bearing fault diagnosis model with multi-dimensional information interaction and fusion proposed by the present invention in the fault diagnosis of water pump bearings in the actual industrial environment, the present invention collected a fault data set of a water pump unit in a copper mine pump house in Tongling City for experiments. The rotational speeds of the main shaft of the water pump drive motor and the water pump rotating shaft in this pump house are both 1485 r / min, and the sampling frequency during the signal acquisition process is set to 12 kHz. Vibration acceleration data sampling is carried out on the drive end of the water pump drive motor under three different fault operating states and the normal operating state. The acquisition process is as Figure 3 shown, and the sampling duration for each time is set to 30 s. The one-dimensional vibration signals after acquisition are as Figure 4 shown.

[0116] S2: Data preprocessing

[0117] Import the vibration signals under different fault operating states and the normal operating state into matlab for encoding different fault categories, and divide the data set according to the ratio of 7:2:1, as shown in Table 1.

[0118] Table 1 Division of the sample data set

[0119]

[0120] Then, perform signal-picture and signal-point cloud conversions on the one-dimensional vibration signals after filtering, as shown in Figure 5 , Figure 6 shown.

[0121] S3: Experimental analysis

[0122] The deep learning framework used in the experiments of this invention is the Pytorch framework developed and maintained by the Facebook AI research team. At the same time, Python is used as the front-end language, which is easy to integrate and combine with other Python libraries and tools. Other configurations are shown in Table 2.

[0123] Table 2 Training Environment Configuration

[0124]

[0125] When evaluating the performance of the network model, there are two extremely crucial core indicators, namely accuracy and loss rate, which play a decisive role in comprehensively and deeply measuring the advantages and disadvantages of the model. Therefore, this invention uses these two indicators to judge the model performance, as Figure 7 shown in the accuracy and loss change curves of the model during the iterative training process. It can be seen from this that the classification accuracy of the model for the vibration data of the water pump motor reaches about 95.23% after only 125 rounds and begins to tend to converge. After 200 rounds of iterative training process, these two indicators are basically in a stable state. Among them, the average accuracy reaches 95.31%, and the highest reaches 98.83%. This data fully reflects the applicability and reliability of the fault diagnosis model of this invention under the real fault data set.

[0126] The prior art includes one-dimensional vibration signal analysis, the combination analysis of one-dimensional time-domain signal and two-dimensional time-frequency signal, and the separate analysis of vibration signals using PointNet, etc. As shown in Table 3, under the condition of the same data set, it can be clearly compared that the method proposed in this invention has a higher fault diagnosis accuracy than using one-dimensional vibration signal analysis alone; higher than using the combination analysis of one-dimensional time-domain signal and two-dimensional time-frequency signal; higher than using PointNet to analyze vibration signals alone. Thus, it can be proved that multi-dimensional information fusion has advantages and can improve the accuracy of fault diagnosis.

[0127] Table 3 Fault Accuracy of Prior Art Methods

[0128]

[0129] In summary, a fault diagnosis method for mine water pump bearings with multi-dimensional information interaction and fusion proposed in this invention can break through the limitation of single-dimensional analysis compared with the prior art, integrate multi-dimensional feature information, provide a more comprehensive and intuitive method for bearing fault diagnosis, can dig out more fault features hidden in the signals, and ultimately improve the accuracy of fault diagnosis, opening up a new research direction for the field of fault diagnosis.

[0130] The above content is only an example and illustration of the present invention. Those skilled in the art to which the present technology pertains can make various modifications, supplements, or use similar methods of substitution to the specific embodiments described. As long as they do not deviate from the invention or exceed the scope defined by the claims of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A mine water pump bearing fault diagnosis method based on multi-dimensional information interaction and fusion, characterized in that: The method specifically comprises the following steps: Step 1: Collect the axial vibration signal of the centrifugal pump through the vibration test equipment, and export the final data in .mat file format; Step 2: Filter the collected vibration signal data; Step 3: Convert the filtered vibration signal into signal-image and signal-point cloud; Step 4: Build a multi-dimensional information interaction fusion fault diagnosis model; Step 5: Input the training set into the model for fault diagnosis, set an accuracy threshold, save the model that is above the threshold for each training, and finally determine whether the model is optimal through the validation set, and select the optimal model as the real-time fault diagnosis model for the water pump bearing.

2. According to the multi-dimensional information interactive fusion method for mine water pump bearing fault diagnosis according to claim 1, it is characterized in that: The description formula for filtering the collected vibration signal data in step 2 is: Where: w j，k represents the wavelet coefficients, represents the quantized wavelet coefficient, T represents the threshold, N represents the signal length, sign(w j，k ) represents a sign function.

3. The method for diagnosing bearing faults of a mine water pump based on multi-dimensional information interaction and fusion according to claim 1 is characterized in that: In step 3, the method for converting the filtered vibration signal into signal-image and signal-point cloud is as follows: Step 301: Continuous wavelet transform is used to convert the filtered vibration signal into a signal-to-image conversion. The mathematical description formula is: Where: f(t) represents the signal to be transformed, represents the mother wavelet function, a and τ represent the scale parameter and translation parameter, and Wt(a, τ) represents the wavelet coefficient of the signal f(t) under the scale parameter a and translation parameter τ. Step 302: Continuous wavelet transform is used to convert the filtered vibration signal into a point cloud. The mathematical description formula is: Where: x(t) represents the signal to be analyzed, fs represents the sampling frequency, f represents the signal frequency, and cfs represents the wavelet coefficient; Step 303: Use the meshgrid function in MATLAB to generate a two-dimensional grid matrix for time and frequency At the same time, the absolute value of the wavelet coefficient is used as The component is described by the formula: Where: Represents the time characteristic component, represents the frequency characteristic component and Represents the wavelet coefficient component; Step 304: Use the reshape function to and The matrix dimensions of the components are expanded, and they are merged and extracted one by one according to the corresponding relationship to obtain the three-dimensional point cloud data p after the one-dimensional vibration signal is converted. The description formula is:

4. The method for diagnosing bearing faults of a mine water pump based on multi-dimensional information interaction and fusion according to claim 1 is characterized in that: The method for building a multi-dimensional information interaction fusion fault diagnosis model in step 4 is: Step 401: extracting one-dimensional vibration signal data features using one-dimensional convolution; Step 402: extracting features of the converted image data using two-dimensional convolution; Step 403: normalize the time domain features and the two-dimensional time-frequency features of the one-dimensional vibration signal respectively and map them into three-dimensional features; Step 404: extracting features of the converted point cloud data using a three-dimensional PointCNN neural network; Step 405: interactively fuse the mapped three-dimensional features with the point cloud data features extracted and converted using the three-dimensional PointCNN neural network.

5. The method for diagnosing bearing faults of a mine water pump based on multi-dimensional information interaction and fusion according to claim 4 is characterized in that: The mathematical description formula of step 401 is: F i =f(cov1D(W i ,X)+b i ),i=1,...a F A (t)=[F1,...,F a ] a is the characteristic length of the one-dimensional vibration signal data extracted using one-dimensional convolution.

6. A mine water pump bearing fault diagnosis method with multi-dimensional information interactive fusion according to claim 5, characterized in that: The mathematical description formula of step 402 is: F i =f(cov2D(W i ,X)+b i ),i=1,...bc F BC (t)=[F1,...,F bc ] Where b×c is the size of the converted image data features extracted using two-dimensional convolution.

7. A mine water pump bearing fault diagnosis method with multi-dimensional information interactive fusion according to claim 6, characterized in that: In step 403, the standard deviation of the acquired one-dimensional vibration signal time domain features and the image data features after the two-dimensional convolution extraction conversion is first normalized, and then the time domain feature vector is used as a dimension of the newly constructed three-dimensional feature. When it is used as the z-axis direction, for the two-dimensional time-frequency feature matrix, its elements are expanded into a vector of length bc in the order of rows or columns, respectively as the x- and y-axis directions, and a three-dimensional feature vector is constructed by the mapping method of the physical model. The description formula is: Where: μ represents the mean value of the feature in the original data set, δ represents the normalized standard deviation of the feature in the original data set, p represents the mapped three-dimensional point cloud data, and f represents a physical mapping function.

8. The method for diagnosing bearing faults of a mine water pump based on multi-dimensional information interaction and fusion according to claim 7 is characterized in that: Step 405 includes: First, multiply the mapped point cloud data and the feature matrix of the point cloud data extracted by PointCNN; Secondly, the attention distribution of point cloud features generated by the two methods is calculated row by row; Then, the point cloud feature attention distribution generated by the two methods is multiplied by the feature matrix to obtain the attention representation matrix; Next, the multiplication gating mechanism is used to obtain the interactive attention information matrix of the point cloud features generated in two ways; Finally, the two are spliced ​​and merged; Where: "·" represents matrix multiplication, "⊙" represents element-by-element matrix multiplication; Represents the combined three-dimensional feature vector.

Images