A bearing intelligent diagnostic method based on multi-source physical information fusion
By using a multi-source physical information fusion method, and employing sparse grouped nonlinear convolution, bidirectional hidden state fusion mechanism, and trilinear attention operation, local and global output feature maps are extracted from multi-source sensor signals. This solves the problem of fault feature extraction for mine hoist bearings under complex working conditions, and enables accurate identification and diagnosis of fault features.
Patent Information
- Application Number
- CN202611580790.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-17
AI Technical Summary
Existing bearing diagnostic methods are difficult to effectively extract fault characteristics of mine hoist bearings under low signal-to-noise ratio and time-varying characteristics, especially under complex working conditions, it is difficult to identify weak fault characteristics and high-frequency impact characteristics.
By employing a multi-source physical information fusion method, local and global output feature maps are extracted from multi-source sensor signals through sparse grouped nonlinear convolution, bidirectional hidden state fusion mechanism and trilinear attention operation, thereby enhancing the high-order interactive expressive ability of the feature maps and obtaining joint feature representation.
It effectively extracts fault features under complex working conditions, enhances the ability to perceive local details such as high-frequency impacts, accurately depicts the long-term and short-term evolution patterns of real-time fault features, alleviates the obstacles of time-varying characteristics to fault identification, and achieves accurate extraction of fault features.
Smart Images

Figure CN123116175A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a bearing intelligent diagnostic method that integrates multi-source physical information, belonging to the field of bearing fault diagnosis technology. Background Technology
[0002] Mine hoists are known as the "throat" of mine operations, responsible for the lifting and transporting of ore, gangue, materials, and personnel. As a crucial supporting component of the mine hoist, the health of the hoist bearings directly affects the overall operational stability and service life of the machine.
[0003] Mine hoists operate under extreme conditions such as strong alternating loads, high humidity, and dust for extended periods, making their main bearings prone to fatigue, wear, and thermal damage. A main bearing failure can lead to anything from abnormal vibration and temperature rise, affecting the overall stability of the machine, to serious accidents like a hoist falling, causing shutdowns, production stoppages, and even casualties. The mine hoist's transmission system uses a motor-driven drum structure with sequentially decreasing speeds and interconnected components. In actual service, mine hoists frequently experience heavy-load starts, acceleration, and deceleration / braking, causing drastic changes in the load and speed of the transmission system. This results in severe fluctuations in the contact stress of the main bearing, leading to fatigue cracks gradually forming on the raceway surface. Simultaneously, under strong alternating loads, these cracks propagate, eventually forming fatigue pitting. The increased surface roughness in the pitted areas further concentrates stress, causing large-scale material peeling off the raceway surface. Furthermore, due to installation errors, axial force fluctuations, and shaft deformation, the bearings experience uneven loading, causing excessive loads in localized areas. This exacerbates fatigue damage in these areas and leads to lubrication failure due to localized overheating, further accelerating wear and spalling. This progressive damage results in the main bearing vibration signal exhibiting multi-source aliasing, weak characteristics, and strong time-varying properties. It is also easily overwhelmed by strong noise from multiple sources, such as structural resonance, making it difficult for existing diagnostic methods to identify subtle fault features. This poses a significant challenge to early fault identification and reliable life prediction. Moreover, the mine hoist's transmission structure, which uses a motor driving the drum via a reducer, results in a long vibration transmission path and aliasing of multiple vibration components. This dilutes and attenuates localized features such as high-frequency impacts, making it difficult for existing models to effectively extract the fault characteristics of individual bearings.
[0004] Therefore, existing bearing diagnostic methods have the problem of not being able to effectively extract fault features from fault signals with time-varying characteristics. Summary of the Invention
[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a bearing intelligent diagnosis method that integrates multi-source physical information to effectively extract fault features from fault signals with low signal-to-noise ratio and time-varying characteristics.
[0006] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, this application provides a bearing intelligent diagnosis method based on multi-source physical information fusion, including: acquiring multi-source sensor signals of a mine hoist bearing.
[0007] Local output feature maps are extracted from the multi-source sensor signals using sparse grouped nonlinear convolution.
[0008] A bidirectional hidden state fusion mechanism is used to extract the global output feature map from the multi-source sensor signals.
[0009] The local output feature map and the global output feature map are concatenated to obtain a comprehensive feature map.
[0010] The high-order interactive representation capability of the integrated feature map is enhanced by using a trilinear attention operation to obtain a joint feature representation.
[0011] The fault diagnosis result is obtained based on the joint feature representation.
[0012] Further, the step of extracting local output feature maps from the multi-source sensor signals using sparse grouped nonlinear convolution includes: organizing the multi-source sensor signals into input feature maps; obtaining seed filters; and dividing the seed filters and the input feature maps along the channel dimension and according to the number of independent branches to obtain multiple input feature maps and multiple branches of seed filters; wherein the number of independent branches is the greatest common divisor between the number of seed filters and the number of input channels.
[0013] Perform standard convolutions individually on each branch of the input feature map and its corresponding seed filter branch to obtain seed feature maps; the formula for obtaining the seed feature maps is as follows: In the formula, For branch index variables, For the first Seed feature maps output from each branch, For the seed filter One branch, For the first input feature map One branch, The kernel size is [size]. for Standard convolution operations of size, For the real number field, and These represent the dimensions of the input feature maps. and These represent the dimensions of the local output feature map. The number of output channels for the local output feature map. Seed feature map, For seed filter, For the input feature map, for Sparse grouping convolution operation of size, The greatest common divisor between the number of seed filters and the number of input channels. For the 1st, 2nd, ..., Sub-seed feature maps output from each branch, For the 1st, 2nd, ..., ...th of the input feature map One branch, For the 1st, 2nd, ..., 3rd seeds of the seed filter One branch, Input the number of channels. This represents the number of seed filters. This is an operation to find the greatest common divisor.
[0014] The sub-seed feature maps of each branch are spliced along the channel dimension to obtain the seed feature map that restores the complete channel dimension.
[0015] A local output feature map is generated based on the seed features that restore the complete channel dimension.
[0016] Furthermore, the step of extracting local output feature maps from the multi-source sensor signals using sparse grouped nonlinear convolution further includes: determining the number of manually grouped groups, wherein the number of manually grouped groups should satisfy the following constraint: In the formula, The number of manually grouped groups. This is a mold taking operation.
[0017] Define the seed trait derivation ratio, which should satisfy the following constraint: In the formula, The pluralization rate is derived from seed characteristics. To generate the number of filters, The reduction ratio is an integer.
[0018] Based on the number of artificially grouped groups and the seed feature derivation ratio, the seed feature map of the recovered complete channel dimension is segmented and copied to obtain a copy feature map.
[0019] A diverse set of nonlinear functions is defined, the size of which is determined based on the number of artificially grouped groups and the seed feature derivation ratio.
[0020] The replica feature maps are subjected to nonlinear transformations using the diverse set of nonlinear functions and then concatenated to obtain the generated feature map; the formula for obtaining the generated feature map is as follows: In the formula, To generate feature maps, For splicing operations, and These are the counting sequence numbers, To generate the 1st, ..., th feature map Each generated feature map To generate the first feature map Each generated feature map For a diverse set of nonlinear functions, For a diverse set of nonlinear functions, the index is nonlinear functions, For the replica feature map, the first The sub-seed feature map output by the group. To index the diverse set of nonlinear functions The nonlinear function of the replica feature map The sub-seed feature maps output by the group undergo a non-linear transformation.
[0021] The local output feature map is obtained using the following formula: In the formula, Output feature maps for the local area. To recover the seed feature map in full channel dimensions.
[0022] Furthermore, the step of extracting the global output feature map from the multi-source sensor signal using the bidirectional hidden state fusion mechanism includes: obtaining a global information representation module composed of multiple bidirectional single-gated weighted time-series units, wherein normalization layers are interspersed between adjacent bidirectional single-gated weighted time-series units in the global information representation module.
[0023] The input feature map is input into the global information representation module for bidirectional propagation and layer-by-layer iterative calculation. Each layer of bidirectional single-gated weighted time series unit outputs a complete time series output sequence. The complete time series output sequence of the last layer of bidirectional single-gated weighted time series unit is used as the global output feature map.
[0024] Each layer of bidirectional single-gated weighted time-series unit outputs a complete time-series output sequence, including: obtaining the hidden state before iteration; obtaining the input features of the current layer of bidirectional single-gated weighted time-series unit; performing a linear transformation and superimposing the input features of the current layer of bidirectional single-gated weighted time-series unit and the hidden state before iteration to obtain the input weighted features of the single-gated structure; the input weighted features of the single-gated structure are obtained by the following formula: In the formula, For the first The hidden state at all times For the input weighting features of a single-gated structure, The historical hidden state weight matrix, The current input feature weight matrix, For the first The input characteristics of the bidirectional single-gated weighted time-series unit at this time step. This is the bias term for the gated linear transformation.
[0025] The input weighted features of the single-gated structure are input into the Sigmoid nonlinear activation function and mapped to the 0~1 interval to obtain the gating weight coefficients; the formula for obtaining the gating weight coefficients is as follows: In the formula, This is the gating weight coefficient. This is a non-linear activation operation for the Sigmoid function.
[0026] The input features of the bidirectional single-gated weighted time-series unit in this layer and the gating weight coefficients are subjected to linear transformation and Tanh activation operation to obtain the first... The candidate hidden state at time step; the first The formula for obtaining the candidate hidden state at time step is as follows: In the formula, For the first The candidate hidden state at time step. For Tanh activation operation, Input the feature weight matrix for the candidate branch. Scaling the weight matrix to historical states, For candidate state branch bias terms, This is for the Hadamard product operation.
[0027] The formula for obtaining the gate weight coefficients is used to analyze the hidden state before the iteration and the first iteration. We perform a weighted fusion of the candidate hidden states at each time step to obtain... For the first The hidden state at any given time; the For the first The formula for obtaining the hidden state at a given time is as follows: In the formula, For the first The hidden state at any given moment.
[0028] The bidirectional hidden states are fused using a weight matrix to generate the final hidden state at the current time step; the formula for obtaining the final hidden state at the current time step is as follows: In the formula, Forward feature fusion weight matrix, This is the hidden state for forward propagation. For the reverse feature fusion weight matrix, This is the hidden state for backpropagation. To incorporate the bias term, This is the hidden state of the loop unit. For the first The final hidden state of the moment.
[0029] By concatenating the final hidden states at all times, a complete time-series output sequence is obtained.
[0030] The complete timing output sequence is used to be processed by the normalization layer and then input into the next layer of bidirectional single-gated weighted timing unit.
[0031] Furthermore, the step of using trilinear attention operation to enhance the high-order interactive representation capability of the comprehensive feature map to obtain a joint feature representation includes: mapping the comprehensive feature map into five sets of matrices, the five sets of matrices including a query feature matrix, a primary key feature matrix, a secondary key feature matrix, a primary value feature matrix, and a secondary value feature matrix.
[0032] The feature positions of the synthesized feature map are mapped to the periodic phase space, thereby converting the feature position indices of the synthesized feature map into phase angles; the formula for obtaining the phase angles is as follows: In the formula, For the first feature map Phase angles corresponding to each feature position For the first feature map Each feature location The period length for position encoding.
[0033] A seven-dimensional query vector and two sets of key vectors are constructed by inputting a multilayer perceptron and a linear mapping matrix, respectively; the formulas for obtaining the seven-dimensional query vector and the two sets of key vectors are as follows: In the formula, A seven-dimensional query vector. and These are two sets of key vectors. To query the feature matrix, The primary key feature matrix, For auxiliary key feature matrix, It is a scale constant. For the feature location index of the query, The feature location index of the primary key. For the feature position index of the secondary key. For input to a multilayer perceptron, For the first feature map Phase angles corresponding to each feature position For the first feature map The phase angle corresponding to each feature position.
[0034] The attention score is obtained based on the seven-dimensional query vector and the two sets of key vectors; the formula for obtaining the attention score is as follows: In the formula, This represents the attention score.
[0035] The attention weights are obtained using the following formula: In the formula, For attention weights, and To query location The corresponding two sets of local sliding windows, and For dummy variables in double summation, for and Attention score at that time.
[0036] The output feature at each position is obtained using the following formula: In the formula, For the first Output features at each position, Principal value vector, As an auxiliary value vector, Principal value eigenma matrix This is the auxiliary characteristic matrix.
[0037] The output features from all positions are concatenated to obtain the concatenated output features.
[0038] The joint feature representation is obtained based on the spliced output features.
[0039] Furthermore, a diagnostic model for mine hoist bearings is constructed, which includes a local information representation module, a global information representation module, and a high-order dynamic weight enhancement module.
[0040] The mine hoist bearing diagnostic model is trained using samples to obtain a trained mine hoist bearing diagnostic model.
[0041] The local information representation module in the trained mine hoist bearing diagnostic model is used to perform sparse grouping nonlinear convolution operation on the multi-source sensor signals to obtain the local output feature map.
[0042] The global information representation module in the trained mine hoist bearing diagnostic model is used to operate the multi-source sensor signals using a bidirectional hidden state fusion mechanism to obtain the global output feature map.
[0043] The high-order dynamic weight enhancement module in the trained mine hoist bearing diagnostic model is used to enhance the high-order interactive representation of the comprehensive feature map using trilinear attention operation, thereby obtaining the joint feature representation.
[0044] Furthermore, the higher-order dynamic weight enhancement module includes a cross-attention layer and an output layer. The cross-attention layer is used to enhance the higher-order interactive expressive power of the comprehensive feature map using trilinear attention operations to obtain the concatenated output features.
[0045] The output layer is used to obtain the joint feature representation based on the concatenated output features.
[0046] The process of training the mine hoist bearing diagnostic model using samples to obtain the trained mine hoist bearing diagnostic model includes: obtaining the weight update term through the following formula: In the formula, For weight update items, As a control factor, For activation function, For the output layer The raw input scores of each neuron without normalization.
[0047] The weight update is obtained through the weight update term, and the weight update is obtained by the following formula: In the formula, For the counting sequence number, For weight updates, This refers to the number of samples within a single mini-batch. This represents the total number of training samples in the bearing fault dataset. For sample index within a batch, For the first Weight updates for each sample of training data. This refers to the output error of the mine hoist bearing diagnostic model. For the output layer corresponding to the first The first sample The raw input scores of each neuron without normalization.
[0048] The weights of each parameter in the mine hoist bearing diagnostic model are adjusted based on the updated weights.
[0049] The weights of each parameter in the mine hoist bearing diagnostic model are continuously adjusted until the output error of the mine hoist bearing diagnostic model is less than the threshold, thus obtaining a trained mine hoist bearing diagnostic model.
[0050] Secondly, this application also provides a computer device, including a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it performs the steps of the bearing intelligent diagnosis method of multi-source physical information fusion as described in any embodiment of the first aspect.
[0051] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the bearing intelligent diagnosis method based on multi-source physical information fusion as described in any embodiment of the first aspect.
[0052] Fourthly, this application also provides a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the steps of the bearing intelligent diagnosis method based on multi-source physical information fusion as described in any embodiment of the first aspect.
[0053] Compared with existing technologies, the beneficial effects achieved by this application are as follows: The bearing intelligent diagnosis method based on multi-source physical information fusion provided by this application addresses the characteristics of low signal-to-noise ratio and time-varying characteristics in fault signals under various complex working conditions. Therefore, this application does not rely on a single physical information source, but instead acquires multi-source sensor signals of the mine hoist bearing as the basis for obtaining fault diagnosis results. Sparse grouped nonlinear convolution is used to extract local output feature maps from the multi-source sensor signals. Compared with the feature maps output by traditional schemes, the local output feature maps enhance the perception ability of local details such as high-frequency impacts, thus overcoming the influence of low signal-to-noise ratio in fault signals. A bidirectional hidden state fusion mechanism is used to extract global output feature maps from the multi-source sensor signals. Compared with the feature maps output by traditional schemes, the global output feature maps accurately depict the long-term and short-term evolution laws of real-time fault features, alleviating the obstacle of time-varying fault signal characteristics to fault identification. Trilinear attention operation is used to enhance the high-order interactive expression capability of the comprehensive feature map, focusing on discriminative features, and the resulting joint feature representation can effectively extract fault features. Attached Figure Description
[0054] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0055] Figure 1 This is a flowchart of the steps of the CLGF-Net framework in the embodiments of this application.
[0056] Figure 2 This is a structural diagram of the global information representation module in this application embodiment.
[0057] Figure 3 This is a schematic diagram of the confusion matrix of the CLGF-Net framework in the embodiments of this application.
[0058] Figure 4 This is a comparison curve of the training loss rate of the six models in the embodiments of this application.
[0059] Figure 5 This is a flowchart of the steps of the bearing intelligent diagnosis method based on multi-source physical information fusion provided in the embodiments of this application.
[0060] Figure 6 This is a schematic block diagram of the computer device provided in the embodiments of this application.
[0061] Figure 7 This is a topology diagram of the signal transmission direction of the local information representation module in this application embodiment.
[0062] Figure 8 This is the signal transmission direction topology diagram of the global information representation module in this application embodiment.
[0063] Figure 9 This is a topology diagram of the signal transmission direction of the high-order dynamic weight enhancement module in the embodiments of this application.
[0064] Figure 10 This is a topology diagram showing the signal transmission direction between the local information representation module, the global information representation module, and the high-order dynamic weight enhancement module in this embodiment of the application.
[0065] In the diagram, ⊙ represents the Hadamard product operator, and ⊕ represents the addition operator. Detailed Implementation
[0066] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.
[0067] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0068] Example 1
[0069] Figure 5 This is a flowchart of a bearing intelligent diagnostic method based on multi-source physical information fusion, as described in Embodiment 1 of the present invention. This flowchart merely illustrates the logical sequence of the method described in this embodiment. Provided there are no conflicts, different methods may be used in other possible embodiments of the present invention. Figure 5 Complete the steps shown or described in the order indicated.
[0070] The bearing intelligent diagnostic method based on multi-source physical information fusion provided in this embodiment can be applied to terminals. (See attached image) Figure 5 The method of this implementation specifically includes the following steps: acquiring multi-source sensor signals of the mine hoist bearing.
[0071] Local output feature maps are extracted from the multi-source sensor signals using sparse grouped nonlinear convolution.
[0072] A bidirectional hidden state fusion mechanism is used to extract the global output feature map from the multi-source sensor signals.
[0073] The local output feature map and the global output feature map are concatenated to obtain a comprehensive feature map.
[0074] The high-order interactive representation capability of the integrated feature map is enhanced by using a trilinear attention operation to obtain a joint feature representation.
[0075] The fault diagnosis result is obtained based on the joint feature representation.
[0076] Fault signals exhibit low signal-to-noise ratio and time-varying characteristics under various complex operating conditions. Therefore, this embodiment does not rely on a single physical information source, but instead acquires multi-source sensor signals from the mine hoist bearing as the basis for obtaining fault diagnosis results. Sparse grouped nonlinear convolution is used to extract local output feature maps from the multi-source sensor signals. Compared to feature maps output by traditional schemes, the local output feature maps enhance the perception of local details such as high-frequency impacts, thus overcoming the impact of low signal-to-noise ratio in fault signals. A bidirectional hidden state fusion mechanism is used to extract global output feature maps from the multi-source sensor signals. Compared to feature maps output by traditional schemes, the global output feature maps accurately depict the long- and short-term evolution patterns of real-time fault features, alleviating the obstacle of time-varying fault signal characteristics to fault identification. Trilinear attention operations are used to enhance the high-order interactive representation capability of the comprehensive feature maps, focusing on discriminative features, and the resulting joint feature representation can effectively extract fault features.
[0077] Example 2
[0078] This embodiment provides a bearing intelligent diagnosis method based on multi-source physical information fusion. This embodiment is an optimization based on Embodiment 1 to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to Embodiment 1.
[0079] As one embodiment, the step of extracting local output feature maps from the multi-source sensor signals using sparse grouped nonlinear convolution includes: organizing the multi-source sensor signals into input feature maps; obtaining seed filters; and dividing the seed filters and the input feature maps along the channel dimension and according to the number of independent branches to obtain multiple input feature maps and multiple branches of seed filters; wherein the number of independent branches is the greatest common divisor between the number of seed filters and the number of input channels.
[0080] Perform standard convolutions individually on each branch of the input feature map and its corresponding seed filter branch to obtain seed feature maps; the formula for obtaining the seed feature maps is as follows: In the formula, For branch index variables, For the first Seed feature maps output from each branch, For the seed filter One branch, For the first input feature map One branch, The kernel size is [size]. for Standard convolution operations of size, For the real number field, and These represent the dimensions of the input feature maps. and These represent the dimensions of the local output feature map. The number of output channels for the local output feature map. Seed feature map, For seed filter, For the input feature map, for Sparse grouping convolution operation of size, The greatest common divisor between the number of seed filters and the number of input channels. For the 1st, 2nd, ..., Sub-seed feature maps output from each branch, For the 1st, 2nd, ..., ...th of the input feature map One branch, For the 1st, 2nd, ..., 3rd seeds of the seed filter One branch, Input the number of channels. This represents the number of seed filters. This is an operation to find the greatest common divisor.
[0081] The sub-seed feature maps of each branch are spliced along the channel dimension to obtain the seed feature map that restores the complete channel dimension.
[0082] A local output feature map is generated based on the seed features that restore the complete channel dimension.
[0083] The method provided in this embodiment enhances CLGF-Net's ability to perceive local details such as high-frequency impacts, effectively addressing the challenges of long vibration transmission paths and the dilution and attenuation of high-frequency impact features due to the mixing of multiple vibration sources in mine hoist transmission systems. In this embodiment, a small number of seed filters replace traditional fully connected convolutional kernels, resulting in a significantly smaller number of seed filters compared to standard convolutional kernels. This substantially reduces the number of trainable parameters in CLGF-Net at the structural level, laying a parameter foundation for source-end deployment. Simultaneously, it refines the basic process of extracting local output feature maps using sparse grouped nonlinear convolutions—dividing the input feature map and seed filters into independent branches along the channel dimension using the greatest common divisor, performing standard convolutions on each branch to obtain seed feature maps, and then concatenating them to restore the complete channel dimension.
[0084] As one embodiment, the step of extracting local output feature maps from the multi-source sensor signals using sparse grouped nonlinear convolution further includes: determining the number of manually grouped groups, wherein the number of manually grouped groups should satisfy the following constraint: In the formula, The number of manually grouped groups. This is a mold taking operation.
[0085] Define the seed trait derivation ratio, which should satisfy the following constraint: In the formula, The pluralization rate is derived from seed characteristics. To generate the number of filters, The reduction ratio is an integer.
[0086] Based on the number of artificially grouped groups and the seed feature derivation ratio, the seed feature map of the recovered complete channel dimension is segmented and copied to obtain a copy feature map.
[0087] A diverse set of nonlinear functions is defined, the size of which is determined based on the number of artificially grouped groups and the seed feature derivation ratio.
[0088] The replica feature maps are subjected to nonlinear transformations using the diverse set of nonlinear functions and then concatenated to obtain the generated feature map; the formula for obtaining the generated feature map is as follows: In the formula, To generate feature maps, For splicing operations, and These are the counting sequence numbers, To generate the 1st, ..., th feature map Each generated feature map To generate the first feature map Each generated feature map For a diverse set of nonlinear functions, For a diverse set of nonlinear functions, the index is nonlinear functions, For the replica feature map, the first The sub-seed feature map output by the group. To index the diverse set of nonlinear functions The nonlinear function of the replica feature map The sub-seed feature maps output by the group undergo a non-linear transformation.
[0089] The local output feature map is obtained using the following formula: In the formula, Output feature maps for the local area. To recover the seed feature map in full channel dimensions.
[0090] This embodiment applies various data-independent nonlinear transformations to the seed feature map using a diverse set of nonlinear functions, enriching the multi-scale and multi-morphological representation capabilities of local spatial features. This enables the model to capture weak local features such as transient impacts and envelope modulation of bearing faults from different nonlinear perspectives, further enhancing its detail perception capabilities. Since the diverse nonlinear functions are fixed and data-independent transformations, the number of feature channels can be expanded several times without increasing any trainable parameters, achieving an extreme compression effect of zero parameter growth and a doubling of feature dimensions.
[0091] As one embodiment, the step of extracting the global output feature map from the multi-source sensor signal using a bidirectional hidden state fusion mechanism includes: obtaining a global information representation module composed of multiple bidirectional single-gated weighted time-series units, wherein normalization layers are interspersed between adjacent bidirectional single-gated weighted time-series units in the global information representation module.
[0092] The input feature map is input into the global information representation module for bidirectional propagation and layer-by-layer iterative calculation. Each layer of bidirectional single-gated weighted time series unit outputs a complete time series output sequence. The complete time series output sequence of the last layer of bidirectional single-gated weighted time series unit is used as the global output feature map.
[0093] Each layer of bidirectional single-gated weighted time-series unit outputs a complete time-series output sequence, including: obtaining the hidden state before iteration; obtaining the input features of the current layer of bidirectional single-gated weighted time-series unit; performing a linear transformation and superimposing the input features of the current layer of bidirectional single-gated weighted time-series unit and the hidden state before iteration to obtain the input weighted features of the single-gated structure; the input weighted features of the single-gated structure are obtained by the following formula: In the formula, For the first The hidden state at all times For the input weighting features of a single-gated structure, The historical hidden state weight matrix, The current input feature weight matrix, For the first The input characteristics of the bidirectional single-gated weighted time-series unit at this time step. This is the bias term for the gated linear transformation.
[0094] The input weighted features of the single-gated structure are input into the Sigmoid nonlinear activation function and mapped to the 0~1 interval to obtain the gating weight coefficients; the formula for obtaining the gating weight coefficients is as follows: In the formula, This is the gating weight coefficient. This is a non-linear activation operation for the Sigmoid function.
[0095] The input features of the bidirectional single-gated weighted time-series unit in this layer and the gating weight coefficients are subjected to linear transformation and Tanh activation operation to obtain the first... The candidate hidden state at time step; the first The formula for obtaining the candidate hidden state at time step is as follows: In the formula, For the first The candidate hidden state at time step. For Tanh activation operation, Input the feature weight matrix for the candidate branch. Scaling the weight matrix to historical states, For candidate state branch bias terms, This is for the Hadamard product operation.
[0096] The formula for obtaining the gate weight coefficients is used to analyze the hidden state before the iteration and the first iteration. We perform a weighted fusion of the candidate hidden states at each time step to obtain... For the first The hidden state at any given time; the For the first The formula for obtaining the hidden state at a given time is as follows: In the formula, For the first The hidden state at any given moment.
[0097] The bidirectional hidden states are fused using a weight matrix to generate the final hidden state at the current time step; the formula for obtaining the final hidden state at the current time step is as follows: In the formula, Forward feature fusion weight matrix, This is the hidden state for forward propagation. For the reverse feature fusion weight matrix, This is the hidden state for backpropagation. To incorporate the bias term, This is the hidden state of the loop unit. For the first The final hidden state of time. and The calculation process and Consistent.
[0098] By concatenating the final hidden states at all times, a complete time-series output sequence is obtained.
[0099] The complete temporal output sequence is processed by the normalization layer and then input into the next layer of bidirectional single-gated weighted temporal unit to finally obtain the global output feature map.
[0100] This embodiment implements the core temporal modeling function of the global information representation module. By simultaneously capturing forward and backward temporal dependencies through a bidirectional propagation mechanism, it accurately characterizes the long- and short-term evolution patterns of real-time bearing fault features, effectively addressing the challenge of highly time-varying characteristics where features continuously drift with changing operating conditions such as frequent start-stop cycles. Compared to the traditional GRU's dual-gated update gate + reset gate structure, SWTS-Net simplifies the gate structure to a single gate, significantly reducing the number of parameters and computational overhead while maintaining the ability to learn temporal dependencies, making the global information representation module adaptable to limited computing power at the source end.
[0101] As one embodiment, the step of using trilinear attention operation to enhance the high-order interactive representation capability of the comprehensive feature map to obtain a joint feature representation includes: mapping the comprehensive feature map into five sets of matrices, the five sets of matrices including a query feature matrix, a primary key feature matrix, a secondary key feature matrix, a primary value feature matrix, and a secondary value feature matrix.
[0102] The feature positions of the synthesized feature map are mapped to the periodic phase space, thereby converting the feature position indices of the synthesized feature map into phase angles; the formula for obtaining the phase angles is as follows: In the formula, For the first feature map Phase angles corresponding to each feature position For the first feature map Each feature location The period length for position encoding.
[0103] A seven-dimensional query vector and two sets of key vectors are constructed by inputting a multilayer perceptron and a linear mapping matrix, respectively; the formulas for obtaining the seven-dimensional query vector and the two sets of key vectors are as follows: In the formula, A seven-dimensional query vector. and These are two sets of key vectors. To query the feature matrix, The primary key feature matrix, For auxiliary key feature matrix, It is a scale constant. For the feature location index of the query, The feature location index of the primary key. For the feature position index of the secondary key. For input to a multilayer perceptron, For the first feature map Phase angles corresponding to each feature position For the first feature map The phase angle corresponding to each feature position.
[0104] The attention score is obtained based on the seven-dimensional query vector and the two sets of key vectors; the formula for obtaining the attention score is as follows: In the formula, This represents the attention score.
[0105] The attention weights are obtained using the following formula: In the formula, For attention weights, and To query location The corresponding two sets of local sliding windows, and For dummy variables in double summation, for and Attention score at that time.
[0106] The output feature at each position is obtained using the following formula: In the formula, For the first Output features at each position, Principal value vector, As an auxiliary value vector, Principal value eigenma matrix This is the auxiliary characteristic matrix.
[0107] The output features from all positions are concatenated to obtain the concatenated output features.
[0108] The joint feature representation is obtained based on the spliced output features.
[0109] This embodiment implements the core attention mechanism of the high-order dynamic weight enhancement module. Specifically, it uses a third-order determinant attention model to focus on discriminative features, adaptively enhancing bearing fault-sensitive information and suppressing various vibration source interferences such as structural resonance, thereby improving the feature discriminativeness of CLGF-Net. Simultaneously, periodic phase position encoding enables the model to perceive the temporal positional relationship of features, enhancing its ability to model fault evolution patterns. By limiting the candidate range of primary and secondary keys within two sets of local sliding windows, the computational complexity of the high-order attention model is reduced from the third-order order for the entire sequence to the first-order order, effectively reducing computational complexity and memory usage, allowing the high-order attention mechanism to run in real-time under limited computing power at the source end.
[0110] As one embodiment, a diagnostic model for mine hoist bearings is constructed, which includes a local information representation module, a global information representation module, and a high-order dynamic weight enhancement module.
[0111] The mine hoist bearing diagnostic model is trained using samples to obtain a trained mine hoist bearing diagnostic model.
[0112] The local information representation module in the trained mine hoist bearing diagnostic model is used to perform sparse grouping nonlinear convolution operation on the multi-source sensor signals to obtain the local output feature map.
[0113] The global information representation module in the trained mine hoist bearing diagnostic model is used to operate the multi-source sensor signals using a bidirectional hidden state fusion mechanism to obtain the global output feature map.
[0114] The high-order dynamic weight enhancement module in the trained mine hoist bearing diagnostic model is used to enhance the high-order interactive representation of the comprehensive feature map using trilinear attention operation, thereby obtaining the joint feature representation.
[0115] This embodiment establishes a hierarchical feature extraction mechanism at the system level, consisting of detail perception, evolution tracking, and interference suppression. The local information representation module handles detail perception, the global information representation module handles evolution tracking, and the high-order dynamic weight enhancement module handles interference suppression. These three modules work together to form a complete feature extraction chain, effectively solving the problem that existing models struggle to accurately capture the degradation and evolution patterns of mine hoist bearings under complex operating conditions. This embodiment also embodies the collaborative design concept of parameter compression and computational optimization, enabling CLGF-Net to overcome the bottleneck of existing multi-source information fusion models' difficulty in source-end deployment, and achieving rapid online diagnosis at the edge of the mine hoist monitoring system.
[0116] As one embodiment, the high-order dynamic weight enhancement module includes a cross-attention layer and an output layer. The cross-attention layer is used to enhance the high-order interactive expressive power of the comprehensive feature map using trilinear attention operations to obtain the concatenated output features.
[0117] The output layer is used to obtain the joint feature representation based on the concatenated output features.
[0118] The process of training the mine hoist bearing diagnostic model using samples to obtain the trained mine hoist bearing diagnostic model includes: obtaining the weight update term through the following formula: In the formula, For weight update items, As a control factor, For activation function, For the output layer The raw input scores of each neuron without normalization.
[0119] The weight update is obtained through the weight update term, and the weight update is obtained by the following formula: In the formula, For the counting sequence number, For weight updates, This refers to the number of samples within a single mini-batch. This represents the total number of training samples in the bearing fault dataset. For sample index within a batch, For the first Weight updates for each sample of training data. This refers to the output error of the mine hoist bearing diagnostic model. For the output layer corresponding to the first The first sample The raw input scores of each neuron without normalization.
[0120] The weights of each parameter in the mine hoist bearing diagnostic model are adjusted based on the updated weights.
[0121] The weights of each parameter in the mine hoist bearing diagnostic model are continuously adjusted until the output error of the mine hoist bearing diagnostic model is less than the threshold, thus obtaining a trained mine hoist bearing diagnostic model.
[0122] This embodiment addresses the problem that loss functions exhibit high-dimensional non-convexity, multi-peaked surfaces, and susceptibility to saddle points caused by multi-source heterogeneous data fusion. LDCD improves the weight update direction through second-order derivative information, enabling it to more effectively traverse saddle points and local minima. Simultaneously, it introduces only a single control factor. This significantly reduces the burden of parameter tuning. LDCD enables the model to quickly lock into the global optimum, ensuring that the model breaks through the limitations of local representation at a single scale and forms a continuous multi-scale representation that takes into account local transient impacts, envelope modulation features, and global temporal evolution. LDCD's accelerated convergence characteristics fundamentally ensure that the model maintains excellent generalization ability when facing complex and variable working conditions in mines, allowing the hierarchical feature extraction mechanism of detail perception-evolution tracking-interference suppression to fully learn the optimal parameter configurations at each level during the training phase. Notably, LDCD's rapid convergence characteristics significantly shorten the debugging cycle before model deployment, providing a guarantee of training efficiency for the real-time nature of online diagnostics.
[0123] As one embodiment, the full name of the mine hoist bearing diagnostic model is the Curvature Optimizer-Enabled Local-to-Global Fusion Mine Hoist Bearing Diagnostic Model.
[0124] Example 3
[0125] refer to Figure 1 , Figure 7 , Figure 8 , Figure 9 and Figure 10 This embodiment provides a bearing intelligent diagnosis method based on multi-source physical information fusion. This embodiment is an optimization based on embodiment two to improve the technical effect and refine the technical solution. For details not described in this embodiment, please refer to embodiment two.
[0126] First, let's introduce the application background of this embodiment. Mine hoists are considered the "throat" of the mine operation system, responsible for the lifting and transporting of ore, gangue, materials, and personnel. The bearings of the hoist, as crucial supporting components, directly affect the overall operational stability and service life of the machine. In practical applications, mine hoists operate under extreme conditions such as strong alternating loads, high humidity, and dust, leading to fatigue, wear, and thermal damage in the main bearings. Once a main bearing fails, it can cause abnormal vibration and temperature rise, affecting the overall operational stability, or even trigger major accidents such as hoist collapses, resulting in shutdowns, production stoppages, and even casualties. Therefore, conducting research on fault diagnosis for mine main bearings is of significant engineering importance for improving maintenance efficiency and ensuring mine safety.
[0127] The mine hoist transmission system employs a structure where a motor drives a drum via a reducer, with each component's speed decreasing sequentially and coupled to the others. In actual service, the mine hoist frequently experiences varying operating conditions such as heavy-load starts, acceleration, and deceleration / braking, causing drastic changes in the load and speed of the transmission system. This leads to severe fluctuations in the contact stress of the main bearings, causing fatigue cracks to gradually form on the raceway surface. Simultaneously, under strong alternating loads, the cracks propagate, eventually forming fatigue pitting. Furthermore, the increased surface roughness in the pitted areas further concentrates stress, resulting in large-scale material spalling from the raceway surface. In addition, installation errors, axial force fluctuations, and shaft deformation cause uneven bearing loads, leading to excessively high loads in localized areas. This accelerates fatigue damage in these areas and, due to localized overheating, causes lubrication failure, further accelerating wear and spalling. The aforementioned progressive damage causes the main bearing vibration signal to exhibit characteristics of multi-source aliasing, weak features, and strong time-varying nature. It is also easily submerged by strong noise from multiple sources, such as structural resonance, making it difficult for existing diagnostic methods and identification models to effectively extract the weak fault features. This poses a severe challenge to early fault identification and reliable life prediction.
[0128] Deep learning, with its ability to autonomously learn complex nonlinear mapping relationships, has significant application potential in the field of intelligent fault diagnosis of rotating machinery. However, most existing models rely on single signals for fault identification, making it difficult to comprehensively characterize the multi-dimensional fault information of mine hoist bearings under complex operating conditions, thus limiting diagnostic accuracy and stability. In recent years, multi-source data fusion models have significantly improved diagnostic accuracy, generalization ability, and noise resistance by automatically extracting and fusing features from various sources, and have been widely studied.
[0129] However, existing models still face several unresolved issues when processing multi-source data and information fusion: First, the mine hoist's transmission structure, where a motor drives the drum via a reducer, results in a long vibration transmission path and overlapping of multiple vibration components. This leads to the dilution and attenuation of local features such as high-frequency impacts, making it difficult for existing models to effectively extract fault features. Simultaneously, the main bearing experiences rapid changes in speed and load under frequent start-stop and other variable operating conditions, causing features to drift continuously with the operating conditions. The fault signal exhibits strong time-varying characteristics, making it difficult for existing models to stably represent it. Furthermore, the superposition of interference from multiple vibration sources, such as structural resonance, creates strong background noise, resulting in a low signal-to-noise ratio for the main bearing signal. Weak fault information is submerged, making it difficult for existing models to achieve global and robust feature extraction.
[0130] Secondly, because status signals need to be collected from locations such as bearing housings, the data acquisition and processing link of the mine hoist bearing monitoring system needs to be deployed at the source end. Simultaneously, limitations in onboard installation space, power supply, and heat dissipation make it difficult to equip the source end with high-performance computing equipment. Furthermore, the main bearing's operating conditions are complex and variable, and faults evolve rapidly, requiring diagnostic models to possess fast real-time inference capabilities. However, while existing multi-source fusion models can improve fault characterization capabilities by leveraging multi-source information, they also significantly increase the number of model parameters and computational complexity, further exacerbating the pressure on computing power, power consumption, and real-time performance at the source end.
[0131] Thirdly, the fusion of multi-source heterogeneous mine hoist bearing data significantly increases the dimensionality of the input features. Cross-modal interactions and nonlinear mapping introduce numerous parameter coupling terms, resulting in a loss function exhibiting high-dimensional non-convex characteristics. This leads to a multi-peaked surface with numerous local minima and saddle points, making the solution path extremely complex. Existing optimization algorithms struggle to achieve stable convergence in the gradient update direction on such high-dimensional non-convex loss surfaces, often requiring the introduction of excessive hyperparameters and additional computation for parameter tuning, increasing the difficulty of weight parameter optimization and the training burden.
[0132] To solve the above-mentioned technical problems, this embodiment builds a fault diagnosis framework based on embodiment two to support the various steps mentioned in embodiment two.
[0133] The first step involves using various sensors, including vibration, acoustic emission, current, strain, temperature, and rotational speed sensors, to collect multi-source signals from the mine hoist bearings under different health conditions. The collected raw signals undergo preprocessing operations such as denoising and normalization. Then, a sliding window method is used to partition the samples, constructing a bearing fault dataset, which is randomly divided into training and testing sets at a 7:3 ratio. This process achieves the acquisition of multi-source sensor signals from the mine hoist bearings.
[0134] The second step involves designing a curvature optimizer-empowered local-global fusion bearing diagnosis method and system. Specifically, based on the size and number of categories of multi-source bearing signals, a CLGF-Net framework for bearing fault diagnosis is designed. This framework mainly consists of three parts: a local information representation module, a global information representation module, and a high-order dynamic weight enhancement module. The weights and bias parameters of the proposed framework are initialized. CLGF-Net stands for "A Curvature Optimizer-Empowered Local-Global Fusion Method and System for BearingDiagnosis," also known as a curvature optimizer-empowered local-global fusion bearing diagnosis model for mine hoists.
[0135] The local information representation module includes a sparse grouping seed convolutional layer, a channel splicing layer, a channel grouping layer, a diversified nonlinear transformation layer, a derived feature splicing layer, and an output splicing layer. First, the multi-source bearing signal is input into the sparse grouping seed convolutional layer, and then... Convolution calculations are performed on seed filters to obtain... First, a seed feature map is generated. Then, the seed feature maps are concatenated through a channel concatenation layer to obtain a seed feature map. Next, the seed feature map is input into a channel grouping layer to group the seed feature maps. Then, the grouping result is input into a diversified nonlinear transformation layer to obtain a sub-generated feature map through nonlinear transformation. Finally, the sub-generated feature map is input into a derived feature concatenation layer to obtain a generated feature map. Finally, in the output concatenation layer, the seed feature map and the generated feature map are concatenated and output to obtain the local output feature map.
[0136] Convolutional neural networks extract local features by sliding convolutional kernels across an image and expand the receptive field by stacking layers. However, when using extremely large convolutional kernels to capture long-range dependencies, the effective receptive field distribution is disrupted, causing the distribution of pixel influences in the network to deviate from the ideal asymptotic Gaussian distribution. This makes it difficult for the model to focus on key regions and introduces a large amount of redundant computation, thereby reducing the model's efficiency and expressive power.
[0137] To address the aforementioned issues, a local information representation module is proposed. This module uses SparseGroupNon-LinearConvolution (SGNLConv) as the basic operational unit. Through operations such as group sparsification, seed feature reuse, and diverse nonlinear transformations, it achieves lightweight and efficient local feature extraction. Specifically, the input feature map and seed filter are equally divided along the channel dimension. One independent branch: , In the formula, the input feature map satisfies The seed filter satisfies The seed filter is a small number of convolution kernels, also known as seed convolution. The branch index variable satisfies... The first input feature map Branches satisfy ,in This represents the number of input channels for a single branch. This represents the number of seed filters for a single branch.
[0138] Perform the operation separately for each branch sub-input feature map and its corresponding sub-seed filter. Standard convolution yields independent sub-seed feature maps for each branch: In the formula, the first Seed feature maps output by each branch satisfy The seed feature map is the feature map generated by the seed filter, and the sub-seed feature map is the feature map generated by the sub-seed filter.
[0139] The first The sub-seed features calculated independently from each branch are concatenated and rearranged along the channel dimension to restore the complete channel-dimensional seed feature map: , .
[0140] The seed feature map is in Divided into dimensions Group: In the formula, For the first Group output sub-seed feature map; number of manually grouped groups The modulo result satisfies , For mold taking operation, For the first The sub-seed feature map output by the group and Count sequence number In the above formula, is the group index variable and satisfies . .
[0141] Obtaining multiple sets of diverse nonlinear functions: Defining a generator filter The quantity is Define integer reduction ratio for ; The seed feature derivation ratio satisfies: .
[0142] Therefore, through segmentation and replication, it is necessary to obtain Zhang and Zi Seed Characteristics Figure 1 A replica feature map. Introducing a replica feature map... Different nonlinear transformations independent of data, namely diversified nonlinear functions (DNLFs), can be represented as: .
[0143] For the duplicate feature map, DNLFs with different random hyperparameter settings are used to obtain the corresponding sub-generated feature maps: , In the formula, the counting sequence number In the above formula, each replica branch index is defined and satisfies the following conditions: , satisfy , satisfy .
[0144] The globally generated feature map is concatenated with the seed feature map to produce the local output feature map of the local information representation module, as shown below: In the formula, the local output feature map .
[0145] In practical implementation, the proposed global information representation module mainly consists of batch normalization layers and multiple stacked SWTS-Net basic units. First, the measured multi-source bearing signals are input into the global information representation module, with an input feature sequence size of 1×1000×6. To fully capture the temporal dependencies of bearing fault features, this feature sequence is directly fed into the temporal modeling backbone network composed of six bidirectional SWTS-Net basic units. Leveraging the unique shared gating and lightweight design of SWTS-Net units, this module achieves accurate capture of long- and short-term temporal dependencies while maintaining parameter efficiency. Finally, the bidirectional hidden states are integrated through a state fusion mechanism to obtain the global output feature map. Each unit is trained and stabilized through batch normalization layers, accelerating convergence and significantly improving the robustness of the temporal features.
[0146] GRU can preserve the temporal order of data when processing large batches of long-sequence data. However, the GRU structure faces computational complexity challenges when dealing with high-dimensional, multi-dimensional temporal features. Therefore, inspired by lightweight network design principles, a global information representation module is proposed. This module includes a novel gated recurrent temporal network structure, named Single-gatedweighttimeseriesnetwork (SWTS-Net). SWTS-Net reduces the number of gates to one, effectively reducing weight parameters and computational cost while maintaining the ability to learn temporal dependencies.
[0147] The specific calculation process is as follows: The input features at the current time step are superimposed with the hidden state at the previous time step through a linear transformation to obtain the weighted input features at the current time step: In the formula, The input weighted features are for a single-gated structure, namely SWTS-Net.
[0148] The input weighted features are fed into the Sigmoid activation function, mapped to the 0-1 interval, and gating weight coefficients are generated to control the retention and forgetting of information. In the formula, The Sigmoid non-linear activation function is expressed as follows: ,in For calculating algebra.
[0149] Combining the input features with the gated hidden states, a linear transformation and Tanh activation are applied to generate the candidate hidden states for the current time step: .
[0150] The hidden state at the current time step is obtained by adaptively weighting and fusing the historical hidden state with the current candidate state using gating weight coefficients. .
[0151] The forward and backward hidden states are fused using a weight matrix to generate the final hidden state at the current time step: .
[0152] By performing calculations on all time steps and concatenating them, the complete temporal output sequence of the current layer is obtained. After batch normalization, it is fed into the next layer Bi-SWTS-Net unit. After six layers of iterative calculation, the output of the sixth layer is the global output feature map.
[0153] After obtaining the model's predicted output vector at the current time, the backpropagation algorithm is used to iteratively update the model parameters and solve for the gradient of the loss function with respect to each weight and bias. The specific derivation process is as follows: , , , , , , , , , , , , , In the formula, To determine the sign of the derivative, For the first The time-matter model predicts the output vector. For the actual label vector, For linear intermediate variables of the output layer, For the first The loss function at time step 1. This is the output layer weight matrix. For output layer bias terms, The total step size of the input time series sequence. This is the last moment of the time series, with no subsequent gradient backpropagation. For the intermediate time step, the reverse gradient of the next time step needs to be superimposed. For the first The time-limited unique gated input weighted sum of SWTS-Net, For the first The time-limited unique gated input weighted sum of SWTS-Net, A unified representation of the gradient of the loss function with respect to the corresponding variable. , , These represent the gradients of the loss function with respect to the linear pre-value of the output layer, the weights of the output layer, and the bias of the output layer, respectively. , These are the gradients of the loss function with respect to the final hidden state and the weights of the historical states in the gated branch, respectively. , These are the gradients of the loss function with respect to the candidate hidden state and the single-gated state, respectively. The gradient of the loss function with respect to the gated weighted sum. , and These are the gradients of the loss function with respect to the historical state weights of the gated branch, the input weights of the candidate hidden state branch, and the scaling weights of the historical state, respectively. , , These are the derivatives of the Sigmoid functions of the output layer, the gated layer, and the Tanh activation function of the candidate layer, respectively. During the backpropagation iteration, all learnable parameters of the network are adjusted based on these gradient update amounts to minimize the temporal loss at a single time step. The network iterative optimization is completed; the meanings of the other symbols in the above formula have appeared elsewhere in this article and will not be repeated here.
[0154] In specific implementation, the high-order dynamic weight enhancement module consists of a cross-attention layer and an output layer. This module takes as input a comprehensive feature map the result of concatenating the local feature output map from the local information representation module and the global feature output map from the global information representation module. First, the comprehensive feature map is input to the cross-attention layer based on a 2-simplicial attention mechanism. Through five sets of projection vectors and trilinear interaction operations, it adaptively captures the complex long-range dependencies within the comprehensive features and dynamically assigns attention weights based on their discriminative importance for fault diagnosis, thereby focusing on the most discriminative key patterns and achieving feature enhancement and refinement. Subsequently, the enhanced features are input to the output layer, where they are transformed and compressed through a fully connected layer and a nonlinear activation function to obtain a highly discriminative joint feature representation. This module, through this feature self-enhancement strategy in a serial architecture, fully mines the deep fault information contained in multi-source bearing signals, providing highly discriminative feature inputs for fault classification tasks.
[0155] As one embodiment, the higher-order dynamic weight enhancement module enhances the higher-order interaction expression capability between multiple feature units through trilinear interaction operations, thereby improving the model's ability to perceive complex contextual information. Furthermore, by introducing sliding window constraints and an efficient computational kernel, this module reduces the complexity of long program sequence modeling while maintaining CLGF-Net's ability to focus on key information in mining equipment bearing fault information.
[0156] Its core computational process specifically includes the following: mapping the comprehensive feature map into five sets of feature matrices: In the formula, , , , , These are the query feature matrix, primary key feature matrix, secondary key feature matrix, primary value feature matrix, and secondary value feature matrix, respectively. The input is the comprehensive feature map. , , , and These are the query mapping matrix, primary key mapping matrix, secondary key mapping matrix, primary value mapping matrix, and secondary value mapping matrix, respectively.
[0157] The feature locations of the synthesized feature map are mapped to the periodic phase space, thereby converting the feature location index into a phase angle: .
[0158] A seven-dimensional query vector and two sets of key vectors are constructed by inputting a multilayer perceptron and a feature matrix, respectively: .
[0159] To illustrate the calculation process of attention scores for feature triples, based on trigonometric identities, and with the help of given... and Derivation of the calculation process: , In the formula, and These are the matrices used to derive the formulas. , and The angle used to derive the formula.
[0160] The results were: , , In the formula, Determinant operations: , , , , , , , , In the formula, , , For the first group of three-dimensional sub-vectors, , , This is the second group of three-dimensional sub-vectors. This is the first group of third-order determinants. This is the second group of third-order determinants.
[0161] key vector and The slicing method is the same. For regular feature triples, the seventh dimension of both key vectors is 0, therefore the seventh dimension does not participate in the determinant calculation.
[0162] For empty feature triples, their attention scores are constructed using the seventh dimension of the query vector: , In the formula, For a fixed blank key vector, Attention score when there are blank features.
[0163] Within the two sliding window ranges, the attention weights are obtained through normalization: In the formula, For attention weights, and To query location The corresponding two sets of local sliding windows, and For dummy variables in double summation, for and Attention score at time Limited primary key index The optional range, Limited secondary key index The selectable range and window sizes are respectively and For pre-set hyperparameters, the window uses... Centered on the sequence, only temporal indices within the local neighborhood are selected for ternary attention calculation, and blank key vectors are used to fill when the window exceeds the sequence boundary; Traversing the sliding window All primary key candidate positions within the range. Traversing the sliding window All candidate positions of auxiliary keys; summation This represents the sum of the attention index scores for all triples within the window, used as the denominator for Softmax normalization. It also represents a characteristic triple consisting of the query, primary key, and secondary key. Attention score.
[0164] By weighting the element-wise products of the two sets of value vectors using normalized weights, we obtain: In the formula, For the first The output features at each position are concatenated to obtain the complete output of the cross-attention layer. and These are the principal value vector and the auxiliary value vector, respectively, obtained through the principal value feature matrix. and auxiliary characteristic matrix The proposed computational method can simultaneously characterize the ternary higher-order association between the query feature and two sets of key-value features within a local sliding window. This allows the module to not only focus on the importance of a single time location but also model the collaborative change relationships between multiple bearing fault feature units. Furthermore, the sliding window constraint avoids calculating the three-dimensional attention tensor across the entire sequence, effectively reducing the computational complexity and memory usage of the higher-order attention mechanism.
[0165] The training set data is input into the proposed framework for forward propagation. The prediction error is calculated using the cross-entropy loss function, and the network parameters are updated using a low-rank diagonal curvature descent optimizer. Through multiple iterations of training, the loss function value is continuously reduced, improving the model's diagnostic accuracy until the optimal diagnostic model for mine hoist bearings is obtained.
[0166] Traditional SGD optimizers may suffer from local convergence and slow convergence speed, while algorithms such as Adam, although capable of adaptively adjusting the learning rate, are computationally complex and have limited generalization capabilities. To address these limitations, this paper proposes a novel multi-source information fusion model optimization algorithm, named Low-rank Diagonal Curvature Descent (LDCD). The core idea of this algorithm is to leverage the second derivative information of the objective function to achieve fast convergence and possess strong global optimization capabilities.
[0167] The core calculation process is as follows: Softmax is used as the activation function of the output layer, and Newton's method is improved to replace the traditional gradient term in the weight update formula. The original theoretical value of the weight update term based on Newton's method is shown below: In the formula, This represents the original theoretical value of the weight update term for Newton's method. for The first derivative, for The second derivative; due to The range of values is (0,1), therefore the denominator The value range of is (-1, 1). At this time, It may take a negative value, and when Sometimes the denominator will be zero. This is due to iterative descent in the optimization algorithm, i.e. Considering that the value should be greater than 0 and that the denominator cannot contain a value of 0, the LDCD method introduces a control factor into the denominator of the above equation. Make corrections to ensure that its value range is always greater than 0.
[0168] The revised formula is shown below: In the formula, As a control factor, compared to the learning rate in SGD With the same function and a value range of (0,1], it is easy to directly compare its performance with that of SGD. The value is set to 0.5, which is commonly used in SGD. Compared to the commonly used Adam method, which requires two control factors, this design simplifies the process by reducing the number of control factors required and making the algorithm more concise.
[0169] Based on the principles of generalized incremental rules and mini-batch learning, the weight update is calculated as follows: In the formula, The output error of the mine hoist bearing diagnostic model is calculated as the target output. With model output The error between them.
[0170] Adjusting weights: In the formula, For the adjusted network connection weights, This represents the network connection weights before adjustment.
[0171] Repeat the above steps until the error is less than or equal to the set threshold.
[0172] The preprocessed test set data is input into the fully trained CLGF-Net, and the probability distribution of fault types is obtained through forward propagation, thereby realizing end-to-end intelligent identification and classification of bearing faults.
[0173] Comparative experiments were conducted on the same test set using the proposed framework and five mainstream diagnostic methods to obtain the corresponding confusion matrix, training loss rate curve, and various evaluation indicators. The advancement, reliability, and engineering applicability of the proposed framework were verified through comprehensive analysis.
[0174] It is worth noting that, for reference Figure 2 Let's learn about the global information representation module, in which... and The first The hidden states of forward and backward propagation at any given moment. and The first The hidden states of forward and backward propagation at any given moment. and The first The hidden states of forward and backward propagation at any given moment. and The first The hidden states of forward and backward propagation at any given moment. For the first The final hidden state of time. For the first The final hidden state of time. For the first The final hidden state of time. For the first The final hidden state of time. For the first The current input feature vector at time step [time]. For the first The current input feature vector at time step [time]. For the first The current input feature vector at time step [time]. For the first The current input feature vector at time step [time]. This can be achieved through [method / method]. Figure 3 To understand the CLGF-Net framework, you can... Figure 4 A horizontal evaluation of the performance of the method provided in this embodiment is conducted. Figure 4 In the diagram, (a) represents the CLGF-Net framework, and (b), (c), (d), (e), and (f) represent existing technologies such as UniConv-SWTS-Net, UniConv-Siattention-Net, SWTS-Net-Siattention-Net, 1DCNN-SWTS-Net, and UniConv-Bi-GRU-Net, respectively.
[0175] Under the complex operating conditions of frequent start-stop, the fault signals of mine hoist bearings exhibit characteristics such as weak, highly time-varying, and low signal-to-noise ratio. Furthermore, the deployment of existing models is constrained by space and cannot meet the real-time requirements of diagnosis. In addition, existing optimizers are difficult to converge stably when faced with the non-convex characteristics of loss functions caused by multi-source data.
[0176] In comparison, the multi-source physical information fusion bearing intelligent diagnosis method provided in this embodiment has the following advantages over existing technologies: First, addressing the problem that the bearing fault signals of mine hoists under complex operating conditions with frequent start-stop cycles exhibit weak features, strong time-varying characteristics, and low signal-to-noise ratio, making feature extraction difficult, this embodiment proposes a curvature optimizer-enhanced local-global fusion mine hoist bearing diagnosis model, CLGF-Net. Specifically, the local information representation module extracts local spatial features through sparse grouping nonlinear convolution, enhancing CLGF-Net's ability to perceive local details such as high-frequency impacts; the global information representation module models global evolution features through a single-gated bidirectional temporal network, accurately depicting the long- and short-term evolution patterns of real-time bearing fault features; and the high-order dynamic weight enhancement module focuses discriminative features through third-order determinant attention, adaptively enhancing bearing fault-sensitive information and suppressing various vibration source interferences such as structural resonance, thereby improving the feature discriminativeness of CLGF-Net. This hierarchical feature extraction mechanism, which combines detail perception, evolution tracking, and interference suppression, effectively solves the problem of existing models failing to accurately capture the degradation and evolution patterns of mine hoist bearings under complex operating conditions. Furthermore, it fundamentally ensures that CLGF-Net maintains its keen perception of bearing fault characteristics even when facing frequent changes in mine operating conditions. Compared to existing technologies, this embodiment can accurately identify and diagnose fault features from multi-source sensor signals.
[0177] Secondly, addressing the spatial constraints of the source-end deployment of the mine hoist bearing monitoring system and the high computational complexity of existing multi-source information fusion models that prevent them from meeting real-time requirements, this embodiment employs a multi-layered lightweight design for CLGF-Net. The local information representation module expands feature channels using a small number of seed convolutional kernels combined with fixed, diverse nonlinear transformations, significantly compressing the number of trainable parameters in CLGF-Net. The global information representation module simplifies recurrent units through a single-gated structure, significantly reducing CLGF-Net parameters and computational overhead. The high-order dynamic weight enhancement module avoids constructing a three-dimensional attention tensor through double sliding window constraints, reducing the computational complexity of high-order attention from the third order to the first order. Thanks to the collaborative design of parameter compression and computational optimization, CLGF-Net overcomes the bottleneck of existing multi-source information fusion models' difficulty in source-end deployment, enabling millisecond-level online diagnosis at the edge of the mine hoist monitoring system. This moves bearing fault identification capabilities from the backend analysis center to the front line of equipment operation, providing direct support for rapid response by on-site maintenance personnel.
[0178] Thirdly, this embodiment addresses the problems of high-dimensional non-convex loss functions, multi-peaked surfaces, and susceptibility to saddle points caused by the fusion of multi-source heterogeneous mine hoist bearing data, as well as the difficulties in convergence and numerous hyperparameters in existing optimization algorithms. It proposes a lightweight training algorithm, LDCD, based on Newton's method. This algorithm improves weight updates through second-order derivatives and stabilizes the update direction with a single control factor, thereby accelerating convergence of CLGF-Net and reducing the burden of parameter tuning. The LDCD algorithm enables the proposed mine hoist bearing diagnostic model to quickly lock into the global optimum, ensuring that the model overcomes the limitations of single-scale local representation and forms a continuous multi-scale representation that considers local transient impacts, envelope modulation features, and global temporal evolution. This data-driven automated optimization mechanism not only significantly shortens the debugging cycle before model deployment but also fundamentally ensures that the model maintains excellent generalization ability and real-time online diagnostics when facing complex and variable mine conditions.
[0179] Example 4
[0180] refer to Figure 5 This embodiment provides a computer device, including a processor and a memory connected to the processor. The memory stores a computer program. When the computer program is executed by the processor, it performs the steps of the bearing intelligent diagnosis method based on multi-source physical information fusion as provided in Embodiments 1, 2 or 3.
[0181] The computer device may be a server or an electronic terminal, as one embodiment, see reference. Figure 6The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data acquired and generated in the bearing intelligent diagnostic method based on multi-source physical information fusion. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the bearing intelligent diagnostic method based on multi-source physical information fusion provided in Embodiments 1, 2, or 3.
[0182] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0183] The computer device provided in this embodiment has the same technical effects as those in Embodiments 1, 2, or 3, and will not be described again here.
[0184] Example 5
[0185] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the bearing intelligent diagnosis method based on multi-source physical information fusion provided in Embodiment 1 or Embodiment 2.
[0186] The computer-readable storage medium provided in this embodiment has the same technical effects as those in Embodiments 1, 2, or 3, and will not be described again here.
[0187] Example 6
[0188] This embodiment provides a computer program product storing a computer program that, when executed by a processor, implements the steps of the bearing intelligent diagnostic method based on multi-source physical information fusion provided in Embodiment 1 or Embodiment 2. The computer program product provided in this embodiment can be transmitted, distributed, and downloaded via the Internet in the form of signals.
[0189] The computer program product provided in this embodiment has the same technical effects as those in Embodiments 1, 2, or 3, and will not be described again here.
[0190] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0191] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0192] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0193] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0194] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A bearing intelligent diagnostic method based on multi-source physical information fusion, characterized in that, include: Acquire multi-source sensor signals from the bearings of mine hoists; Local output feature maps are extracted from the multi-source sensor signals using sparse grouped nonlinear convolution. A bidirectional hidden state fusion mechanism is used to extract the global output feature map from the multi-source sensor signals; By concatenating the local output feature map and the global output feature map, a comprehensive feature map is obtained. The high-order interactive representation capability of the comprehensive feature map is enhanced by using trilinear attention operations to obtain a joint feature representation; The fault diagnosis result is obtained based on the joint feature representation.
2. The bearing intelligent diagnosis method based on multi-source physical information fusion according to claim 1, characterized in that, The extraction of local output feature maps from the multi-source sensor signals using sparse grouped nonlinear convolution includes: The multi-source sensor signals are organized into an input feature map; a seed filter is obtained, and the seed filter and the input feature map are divided along the channel dimension and according to the number of independent branches to obtain multiple input feature maps and multiple branches of the seed filter; the number of independent branches is the greatest common divisor between the number of seed filters and the number of input channels; Perform standard convolutions individually on each branch of the input feature map and its corresponding seed filter branch to obtain seed feature maps; the formula for obtaining the seed feature maps is as follows: , In the formula, For branch index variables, For the first Seed feature maps output from each branch, For the seed filter One branch, For the first input feature map One branch, The kernel size is [size]. for Standard convolution operations of size, For the real number field, and These represent the dimensions of the input feature maps. and These represent the dimensions of the local output feature map. The number of output channels for the local output feature map. Seed feature map, For seed filter, For the input feature map, for Sparse grouping convolution operation of size, The greatest common divisor between the number of seed filters and the number of input channels. For the 1st, 2nd, ..., Sub-seed feature maps output from each branch, For the 1st, 2nd, ..., ...th of the input feature map One branch, For the 1st, 2nd, ..., 3rd seeds of the seed filter One branch, Input the number of channels. This represents the number of seed filters. This is an operation to find the greatest common divisor; The sub-seed feature maps of each branch are spliced along the channel dimension to obtain the seed feature map that restores the complete channel dimension. A local output feature map is generated based on the seed features that restore the complete channel dimension.
3. The bearing intelligent diagnosis method based on multi-source physical information fusion according to claim 2, characterized in that, The step of extracting local output feature maps from the multi-source sensor signals using sparse grouped nonlinear convolution further includes: The number of manually grouped groups shall be determined, and the number of manually grouped groups shall satisfy the following constraint: , In the formula, The number of manually grouped groups. For mold taking operation; Define the seed trait derivation ratio, which should satisfy the following constraint: , In the formula, The pluralization rate is derived from seed characteristics. To generate the number of filters, The reduction ratio is an integer. The seed feature map for restoring the complete channel dimension is segmented and copied according to the number of artificially grouped groups and the seed feature derivation ratio to obtain a copy feature map; A diverse set of nonlinear functions is defined, the size of which is determined based on the number of artificially grouped groups and the seed feature derivation ratio. The replica feature maps are subjected to nonlinear transformations using the diverse set of nonlinear functions and then concatenated to obtain the generated feature map; the formula for obtaining the generated feature map is as follows: , In the formula, To generate feature maps, For splicing operations, and These are the counting sequence numbers, To generate the 1st, ..., th feature map Each generated feature map To generate the first feature map Each generated feature map For a diverse set of nonlinear functions, For a diverse set of nonlinear functions, the index is nonlinear functions, For the replica feature map, the first The sub-seed feature map output by the group. To index the diverse set of nonlinear functions The nonlinear function of the replica feature map The sub-seed feature maps output by the group undergo a non-linear transformation. The local output feature map is obtained using the following formula: ; In the formula, Output feature maps for the local area. To recover the seed feature map in full channel dimensions.
4. The bearing intelligent diagnosis method based on multi-source physical information fusion according to claim 3, characterized in that, The extraction of the global output feature map from the multi-source sensor signals using a bidirectional hidden state fusion mechanism includes: A global information representation module consisting of multiple bidirectional single-gated weighted time-series units is obtained, wherein normalization layers are interspersed between adjacent bidirectional single-gated weighted time-series units in the global information representation module. The input feature map is input into the global information representation module for bidirectional propagation and layer-by-layer iterative calculation. Each layer of bidirectional single-gated weighted time series unit outputs a complete time series output sequence. The complete time series output sequence of the last layer of bidirectional single-gated weighted time series unit is used as the global output feature map. Each layer of bidirectional single-gated weighted timing unit outputs a complete timing output sequence, including: Get the hidden state before iteration: Obtain the input features of the current layer's bidirectional single-gated weighted time-series unit, and then perform a linear transformation and superposition of the input features of the current layer's bidirectional single-gated weighted time-series unit and the hidden state before iteration to obtain the input weighted features of the single-gated structure; the input weighted features of the single-gated structure are obtained by the following formula: , In the formula, For the first The hidden state at all times For the input weighting features of a single-gated structure, The historical hidden state weight matrix, The current input feature weight matrix, For the first The input characteristics of the bidirectional single-gated weighted time-series unit at this time step. This is the bias term for the gated linear transformation; The input weighted features of the single-gated structure are input into the Sigmoid nonlinear activation function and mapped to the 0~1 interval to obtain the gating weight coefficients; the formula for obtaining the gating weight coefficients is as follows: , In the formula, This is the gating weight coefficient. This is a Sigmoid non-linear activation operation; The input features of the bidirectional single-gated weighted time-series unit in this layer and the gating weight coefficients are subjected to linear transformation and Tanh activation operation to obtain the first... The candidate hidden state at time step; the first The formula for obtaining the candidate hidden state at time step is as follows: , In the formula, For the first The candidate hidden state at time step. For Tanh activation operation, Input the feature weight matrix for the candidate branch. Scaling the weight matrix to historical states, For candidate state branch bias terms, For Hadamard product operations; The formula for obtaining the gate weight coefficients is used to analyze the hidden state before the iteration and the first iteration. We perform a weighted fusion of the candidate hidden states at each time step to obtain... For the first The hidden state at any given time; the For the first The formula for obtaining the hidden state at a given time is as follows: , In the formula, For the first The hidden state at any given moment; The bidirectional hidden states are fused using a weight matrix to generate the final hidden state at the current time step; the formula for obtaining the final hidden state at the current time step is as follows: , In the formula, Forward feature fusion weight matrix, This is the hidden state for forward propagation. For the reverse feature fusion weight matrix, This is the hidden state for backpropagation. To incorporate the bias term, This is the hidden state of the loop unit. For the first The final hidden state of the moment; By concatenating the final hidden states at all times, a complete time-series output sequence is obtained; The complete timing output sequence is used to be processed by the normalization layer and then input into the next layer of bidirectional single-gated weighted timing unit.
5. The bearing intelligent diagnosis method based on multi-source physical information fusion according to claim 4, characterized in that, The method of using trilinear attention operations to enhance the high-order interactive representation capability of the comprehensive feature map to obtain a joint feature representation includes: The comprehensive feature map is mapped into five sets of matrices, including a query feature matrix, a primary key feature matrix, an auxiliary key feature matrix, a primary value feature matrix, and an auxiliary value feature matrix. The feature positions of the synthesized feature map are mapped to the periodic phase space, thereby converting the feature position indices of the synthesized feature map into phase angles; the formula for obtaining the phase angles is as follows: , In the formula, For the first feature map Phase angles corresponding to each feature position For the first feature map Each feature location The period length for position encoding; A seven-dimensional query vector and two sets of key vectors are constructed by inputting a multilayer perceptron and a linear mapping matrix, respectively; the formulas for obtaining the seven-dimensional query vector and the two sets of key vectors are as follows: , In the formula, A seven-dimensional query vector. and These are two sets of key vectors. To query the feature matrix, The primary key feature matrix, For auxiliary key feature matrix, It is a scale constant. For the feature location index of the query, The feature location index of the primary key. For the feature position index of the secondary key. For input to a multilayer perceptron, For the first feature map Phase angles corresponding to each feature position For the first feature map Phase angles corresponding to each feature position; The attention score is obtained based on the seven-dimensional query vector and the two sets of key vectors; the formula for obtaining the attention score is as follows: , In the formula, Attention score; The attention weights are obtained using the following formula: , In the formula, For attention weights, and To query location The corresponding two sets of local sliding windows, and For dummy variables in double summation, for and Attention score at time; The output feature at each position is obtained using the following formula: , In the formula, For the first Output features at each position, Principal value vector, As an auxiliary value vector, Principal value eigenma matrix For auxiliary characteristic matrices; The output features from all positions are concatenated to obtain the concatenated output features; The joint feature representation is obtained based on the spliced output features.
6. The bearing intelligent diagnosis method based on multi-source physical information fusion according to claim 5, characterized in that, A diagnostic model for mine hoist bearings is constructed, comprising a local information representation module, a global information representation module, and a high-order dynamic weight enhancement module. The mine hoist bearing diagnostic model is trained using samples to obtain a well-trained mine hoist bearing diagnostic model. The local information representation module in the trained mine hoist bearing diagnostic model is used to perform a sparse grouping nonlinear convolution operation on the multi-source sensor signals to obtain the local output feature map. The global information representation module in the trained mine hoist bearing diagnostic model is used to operate the multi-source sensor signals using a bidirectional hidden state fusion mechanism to obtain the global output feature map. The high-order dynamic weight enhancement module in the trained mine hoist bearing diagnostic model is used to enhance the high-order interactive representation of the comprehensive feature map using trilinear attention operation, thereby obtaining the joint feature representation.
7. The bearing intelligent diagnosis method based on multi-source physical information fusion according to claim 6, characterized in that, The higher-order dynamic weight enhancement module includes a cross-attention layer and an output layer. The cross-attention layer is used to enhance the higher-order interactive expressive power of the comprehensive feature map using trilinear attention operations to obtain the concatenated output features. The output layer is used to obtain the joint feature representation based on the concatenated output features; The process of training the mine hoist bearing diagnostic model using samples to obtain a trained mine hoist bearing diagnostic model includes: The weight update term is obtained using the following formula: , In the formula, For weight update items, As a control factor, For activation function, For the output layer The raw input scores of each neuron without normalization; The weight update is obtained through the weight update term, and the weight update is obtained by the following formula: , In the formula, For the counting sequence number, For weight updates, This refers to the number of samples within a single mini-batch. This represents the total number of training samples in the bearing fault dataset. For sample index within a batch, For the first Weight updates for each sample of training data. This refers to the output error of the mine hoist bearing diagnostic model. For the output layer corresponding to the first The first sample The raw input scores of each neuron without normalization; The weights of each parameter in the mine hoist bearing diagnostic model are adjusted based on the updated weights. The weights of each parameter in the mine hoist bearing diagnostic model are continuously adjusted until the output error of the mine hoist bearing diagnostic model is less than the threshold, thus obtaining a trained mine hoist bearing diagnostic model.
8. A computer device, characterized in that, The method includes a processor and a memory connected to the processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the bearing intelligent diagnosis method based on multi-source physical information fusion as described in any one of claims 1 to 7 are performed.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the bearing intelligent diagnostic method based on the fusion of multi-source physical information as described in any one of claims 1 to 7.
10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the bearing intelligent diagnostic method based on the fusion of multi-source physical information as described in any one of claims 1 to 7.
