Remote device fault pre-diagnosis method based on AI analysis
Patent Information
- Application Number
- CN202511857821.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-10-27
- Estimated Expiration
- 2045-12-10
AI Technical Summary
[0002]当前,挂壁炉因其所处的运行环境为高温,会存在需要长期进行运维和健康状态监控的需求,目前两种主要模式均存在根本性问题:一是“事后维修”,其在设备故障停机后才进行被动干预,导致非计划性停产、经济损失,并高度依赖维修人员经验,排查过程不确定性强、周期长
[0019]本发明实施例带来了以下有益效果:提出一种基于AI分析的远程设备故障预诊断方法,通过多层级的分析方法实现设备(壁挂炉)精确和高效诊断,首先在停机期间依据多维数据构建对应挂壁炉机械静态部件和动态部件一致性评估模型得到故障的初步动静态故障分析结果,进一步在动静态一致性评估模型评估后,在远程设备运行期间震动数据进行采集,通过调整运行期间的输出参数来完成和模拟以及样机对远程设备的运行故障复现;通过两层确定后,进一步通过深度模型进行深层次故障分析,并基于远程通信直接反馈给用户,这样极大提高和保证了设备故障诊断的准确性和效率。
Smart Images

Figure CN121786365B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of equipment operation and maintenance data analysis technology, and in particular to a remote equipment fault pre-diagnosis method based on AI analysis. Background Technology
[0002] Currently, wall-mounted fireplaces, operating in high-temperature environments, require long-term maintenance and health monitoring. Both existing main maintenance models have fundamental problems: First, "reactive maintenance," which involves passive intervention after equipment failure, leading to unplanned shutdowns and economic losses. It also heavily relies on the experience of maintenance personnel, resulting in highly uncertain and time-consuming troubleshooting processes. For specialized equipment like wall-mounted fireplaces, the lack of early warning systems poses serious safety hazards. Second, traditional fault prediction methods, primarily based on empirical rules and statistical analysis, extract certain rules and fault characteristics. However, fixed maintenance cycles lead to a "one-size-fits-all" approach, inevitably causing "over-maintenance" of healthy equipment and wasting resources. Simultaneously, they fail to effectively identify and intervene in the accelerated deterioration of equipment under harsh conditions, resulting in "under-maintenance" and failing to eliminate potential fault risks. Furthermore, existing remote fault diagnosis methods are not applicable to wall-mounted fireplace maintenance, and their accuracy needs improvement. Summary of the Invention
[0003] The purpose of this invention is to provide a remote equipment fault pre-diagnosis method based on AI analysis to solve the technical problems existing in the prior art.
[0004] In a first aspect, the present invention provides a remote equipment fault pre-diagnosis method based on AI analysis, comprising: Step 1: During the downtime, construct a consistency evaluation model for the corresponding remote equipment's static and dynamic mechanical components based on multidimensional data; The mechanical components of a wall-hung boiler can be divided into static components and dynamic components. Static components are those whose geometric characteristics do not change significantly over time, while dynamic components are those whose geometric characteristics change significantly over time. Considering the operating characteristics of the wall-hung boiler, both dynamic and static components will undergo different degrees of mechanical deformation during operation. Furthermore, in this invention, static components include: a circulating water pump, a fan, a gas valve, a three-way valve, and an automatic exhaust valve; dynamic components include: a heat exchanger, a burner, a shell, pipes, a flue, and an expansion tank. Consistency assessment is performed based on the following formula:
[0005] Where R is the consistency level, ws is the importance weight of static parts; wd is the importance weight of dynamic parts; xs is the measured value of static parts; ys is the actual value of dynamic parts; xd is the measured value of dynamic parts; and 2y is the actual value of the size of dynamic parts.
[0006] Step 2: After the dynamic and static consistency evaluation model in Step 1 is evaluated, vibration data is collected during the operation of the remote equipment. The output parameters during operation are adjusted to complete the simulation and the prototype to reproduce the operation fault of the remote equipment. The vibration signal is converted from the time domain to the frequency domain using Fast Fourier Transform (FFT). The consistency of the frequency domain characteristics is evaluated by calculating the power spectral density and comparing their correlation coefficients. The signal with N sampling points is converted to the frequency domain in O(NlogN) time complexity. For a discrete-time signal x[n] of length N, its discrete Fourier transform X[k] can be calculated using the following formula:
[0007] In the formula, X[k] represents the complex amplitude of signal x[n] at frequency k; x[n] is the signal in discrete time; k is the discrete frequency; and N is the number of sampling points of the signal.
[0008] In order to quickly process running vibration signals, this invention proposes a Fast Fourier Transform (FFT) algorithm that decomposes the discrete Fourier transform of N sampling points into multiple smaller transforms, thereby achieving lower time complexity.
[0009] The correlation coefficient is a statistic used to measure the strength and direction of the linear relationship between two variables. For two variables X and Y, the correlation coefficient can be calculated using the following formula:
[0010] In the formula, and It is the i-th observation of variables X and Y; and is the i-th mean of variables X and Y; N is the number of observations.
[0011] Further frequency domain feature analysis was performed on the obtained vibration data. The power spectral density was calculated from the frequency domain data to obtain the energy distribution of the signal in the frequency domain. Then, the correlation coefficient was used to describe the correlation between the two vibration signals in the power spectral density. Calculating the power spectral density function of two vibration data points and their correlation coefficient essentially compares whether their distributions in the frequency domain are similar. If the energy distributions of the two vibration data points are similar in the frequency domain, they are considered consistent with the vibration characteristics of a certain type of fault, thus preliminarily determining the corresponding fault category.
[0012] Step 3: Remote equipment operation fault reproduction. After determining the initial fault category in Step 2, the operation data is uploaded to the remote server in real time via a 4G / 5G module. The remote server records and analyzes the corresponding data; further analysis of the fault is conducted based on multi-dimensional operation parameters and a deep network classification model. First, multi-dimensional operation data is collected by sensors, specifically including outlet water temperature, return water temperature, combustion chamber temperature, flue gas temperature, gas flow rate, water flow rate, gas pressure, water pressure, ignition current, valve opening, flue gas oxygen content, and carbon monoxide concentration.
[0013] Furthermore, the collected data is aligned, cleaned, and sliced for labeling. By aligning timestamps and handling missing values caused by communication interruptions, this invention employs forward padding or linear interpolation to perform Z-score normalization on each sensor channel. For steady-state faults (such as reduced fan efficiency), a fixed-length sequence (optional 300 points, i.e., 5 minutes) is extracted from the stable operation phase as a sample. For transient faults (such as ignition failure), a specific-length sequence (optional 60 points, i.e., 1 minute) is extracted starting from a critical event (such as the issuance of an ignition signal) as a sample, with each sample corresponding to at least one fault label.
[0014] Based on labeled data, a six-layer improved deep network model is proposed for further analysis, specifically including: Layer 1: The input layer receives preprocessed multi-channel sensor time-series data; its input data dimension is [batch size, number of channels C, sequence length L], preferably, the number of channels C is 7 and the sequence length L is 300; this layer acts as a data distribution node, synchronously feeding the input data to the convolutional neural network branch and the divide-and-conquer Transformer branch.
[0015] Layer 2: Convolutional Neural Network Branch The convolutional neural network branch, acting as a local feature extractor, is used to capture local temporal morphological features within each sensor channel; this branch includes, in sequence: Initial convolutional module: Consists of a one-dimensional convolutional kernel, activation function, and batch normalization layer connected sequentially, used to achieve initial feature fusion and dimensionality enhancement; Multiple cascaded improved depthwise separable convolutional modules: Each module includes sequentially connected up-dimensional point convolution, depthwise convolution, squeezed excitation attention module and down-dimensional point convolution, and introduces residual connection structure; the feature map size is compressed stepwise by setting a stride greater than 1 for depthwise convolution; Global Pooling and Projection Module: Composed of a global average pooling layer and a fully connected layer, it is used to aggregate the compressed feature map into a fixed-length local feature vector. Layer 3: Divide and Conquer Transformer Branches The divide-and-conquer Transformer branch, acting as a global relation extractor, is used to capture long-term dependencies across sensor channels; this branch includes, in sequence: Embedding and Encoding Module: Consists of a fully connected layer and a learnable position encoder, used to map the input sequence into a high-dimensional embedding vector and inject position information; The divide-and-conquer Transformer encoder is characterized by first dividing the sensor channels into K non-overlapping groups according to physical semantics; then, performing multi-head self-attention computation in parallel within each group to capture long-term dependencies within the group; and finally, fusing the output features of the K groups through an inter-group interaction module. Sequence Pooling and Projection Module: Composed of a sequence average pooling layer and a fully connected layer, it is used to aggregate the fused sequence features into a global feature vector of fixed length; the final output of the divide-and-conquer Transformer branch is the global feature vector.
[0016] Layer 4: Adaptive Fusion Gate Module The adaptive fusion gate module is used to dynamically and weightedly fuse features from the convolutional neural network branch and the divide-and-conquer Transformer branch; this module includes, in sequence: Feature concatenation unit: used to concatenate the input local feature vector and global feature vector along the feature dimension; Weight generation unit: composed of one or more fully connected layers and a Softmax normalization function, used to learn and generate a pair of normalized fusion weights based on the concatenated features; Feature fusion unit: used to perform weighted summation operation, the calculation formula is: fused feature = α × local feature vector + β × global feature vector; the final output of the adaptive fusion gate module is the fused feature vector; Layer 5: Classifier Layer The classifier layer is used to determine the fault type based on the fused feature vector; this layer consists of a fully connected layer with the number of neurons equal to the number of predefined fault categories; the output of this layer is the unnormalized discrimination score for each fault category.
[0017] Layer 6: Output Layer The output layer, consisting of a Softmax function, is used to convert the discrimination scores output by the classifier layer into a probability distribution. The output of this layer is a probability vector, in which the category with the highest probability value is determined as the final fault diagnosis result.
[0018] Step 4: Through data analysis, users can receive feedback on the analysis results via mobile phones and tablets through 4G / 5G networks, allowing them to intuitively understand the equipment's operation and fault status.
[0019] The embodiments of this invention bring the following beneficial effects: A remote equipment fault pre-diagnosis method based on AI analysis is proposed. This method achieves accurate and efficient equipment (wall-hung boiler) diagnosis through multi-level analysis. First, during downtime, a consistency evaluation model of the corresponding static and dynamic mechanical components of the wall-hung boiler is constructed based on multi-dimensional data to obtain preliminary dynamic and static fault analysis results. Further, after evaluation by the dynamic and static consistency evaluation model, vibration data is collected during remote equipment operation. By adjusting the output parameters during operation, simulations and prototypes are used to reproduce the operational faults of the remote equipment. After these two layers of determination, a deep fault analysis is further performed using a deep model, and feedback is directly given to the user based on remote communication. This greatly improves and ensures the accuracy and efficiency of equipment fault diagnosis.
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of the present invention, the drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0022] Figure 1 Flowchart of the diagnostic method in this application; Figure 2 An example of an improved deep network model structure from one embodiment of this application. Detailed Implementation
[0023] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used only to describe differences in name and should not be construed as indicating or implying relative importance. Physical quantities in formulas, unless otherwise specified, should be understood as basic quantities in the International System of Units (SI), or derived quantities derived from basic quantities through mathematical operations such as multiplication, division, differentiation, or integration.
[0025] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0026] like Figure 1 As shown in the first aspect of the embodiments of the present invention, the present invention provides a remote equipment fault pre-diagnosis method based on AI analysis, comprising: Step 1: During the downtime, construct a consistency evaluation model for the corresponding remote equipment's static and dynamic mechanical components based on multidimensional data; The mechanical components of a wall-hung boiler can be divided into static components and dynamic components. Static components are those whose geometric characteristics do not change significantly over time, while dynamic components are those whose geometric characteristics change significantly over time. Considering the operating characteristics of the wall-hung boiler, both dynamic and static components will undergo different degrees of mechanical deformation during operation. Furthermore, in this invention, static components include: a circulating water pump, a fan, a gas valve, a three-way valve, and an automatic exhaust valve; dynamic components include: a heat exchanger, a burner, a shell, pipes, a flue, and an expansion tank. Consistency assessment is performed based on the following formula:
[0027] R represents the consistency level; ws represents the importance weight of static parts; wd represents the importance weight of dynamic parts; xs represents the measured value of static parts; ys represents the actual value of dynamic parts; xd represents the measured value of dynamic parts; 2y represents the actual size of dynamic parts. In this embodiment, R being greater than a certain threshold indicates good consistency.
[0028] Step 2: After the dynamic and static consistency evaluation model in Step 1 is evaluated, vibration data is collected during the operation of the remote equipment. The output parameters during operation are adjusted to complete the simulation and the prototype to reproduce the operation fault of the remote equipment. The vibration signal is converted from the time domain to the frequency domain using Fast Fourier Transform (FFT). The consistency of the frequency domain characteristics is evaluated by calculating the power spectral density and comparing their correlation coefficients. The signal with N sampling points is converted to the frequency domain in O(NlogN) time complexity. For a discrete-time signal x[n] of length N, its discrete Fourier transform X[k] can be calculated using the following formula:
[0029] In the formula, X[k] represents the complex amplitude of signal x[n] at frequency k; x[n] is the signal in discrete time; k is the discrete frequency; and N is the number of sampling points of the signal.
[0030] In order to quickly process running vibration signals, this invention proposes a Fast Fourier Transform (FFT) algorithm that decomposes the discrete Fourier transform of N sampling points into multiple smaller transforms, thereby achieving lower time complexity.
[0031] The correlation coefficient is a statistic used to measure the strength and direction of the linear relationship between two variables. For two variables X and Y, the correlation coefficient can be calculated using the following formula:
[0032] In the formula, and It is the i-th observation of variables X and Y; and is the i-th mean of variables X and Y; N is the number of observations.
[0033] Further frequency domain feature analysis was performed on the obtained vibration data. The power spectral density was calculated from the frequency domain data to obtain the energy distribution of the signal in the frequency domain. Then, the correlation coefficient was used to describe the correlation between the two vibration signals in the power spectral density. Calculating the power spectral density function of two vibration data points and their correlation coefficient essentially compares whether their distributions in the frequency domain are similar. If the energy distributions of the two vibration data points are similar in the frequency domain, they are considered consistent with the vibration characteristics of a certain type of fault, thus preliminarily determining the corresponding fault category.
[0034] Step 3: Remote equipment operation fault reproduction. After determining the initial fault category in Step 2, the operation data is uploaded to the remote server in real time via a 4G / 5G module. The remote server records and analyzes the corresponding data; further analysis of the fault is conducted based on multi-dimensional operation parameters and a deep network classification model. First, multi-dimensional operation data is collected by sensors, specifically including outlet water temperature, return water temperature, combustion chamber temperature, flue gas temperature, gas flow rate, water flow rate, gas pressure, water pressure, ignition current, valve opening, flue gas oxygen content, and carbon monoxide concentration.
[0035] Furthermore, the collected data is aligned, cleaned, and sliced for labeling. By aligning timestamps and handling missing values caused by communication interruptions, this invention employs forward padding or linear interpolation to perform Z-score normalization on each sensor channel. For steady-state faults (such as reduced fan efficiency), a fixed-length sequence (optional 300 points, i.e., 5 minutes) is extracted from the stable operation phase as a sample. For transient faults (such as ignition failure), a specific-length sequence (optional 60 points, i.e., 1 minute) is extracted starting from a critical event (such as the issuance of an ignition signal) as a sample, with each sample corresponding to at least one fault label.
[0036] See appendix Figure 2 Based on labeled data samples, this invention proposes a six-layer improved deep network model for further analysis, specifically including: Layer 1: The input layer receives preprocessed multi-channel sensor time-series data; its input data dimension is [batch size, number of channels C, sequence length L], preferably, the number of channels C is 7 and the sequence length L is 300; this layer acts as a data distribution node, synchronously feeding the input data to the convolutional neural network branch and the divide-and-conquer Transformer branch.
[0037] Layer 2: Convolutional Neural Network Branch The convolutional neural network branch, acting as a local feature extractor, is used to capture local temporal morphological features within each sensor channel; this branch includes, in sequence: Initial convolutional module: Consists of a one-dimensional convolutional kernel, activation function, and batch normalization layer connected sequentially, used to achieve initial feature fusion and dimensionality enhancement; Multiple cascaded improved depthwise separable convolutional modules: Each module includes sequentially connected up-dimensional point convolution, depthwise convolution, squeezed excitation attention module and down-dimensional point convolution, and introduces residual connection structure; the feature map size is compressed stepwise by setting a stride greater than 1 for depthwise convolution; Global Pooling and Projection Module: Composed of a global average pooling layer and a fully connected layer, it is used to aggregate the compressed feature map into a fixed-length local feature vector. Layer 3: Divide and Conquer Transformer Branches The divide-and-conquer Transformer branch, acting as a global relation extractor, is used to capture long-term dependencies across sensor channels; this branch includes, in sequence: Embedding and Encoding Module: Consists of a fully connected layer and a learnable position encoder, used to map the input sequence into a high-dimensional embedding vector and inject position information; The divide-and-conquer Transformer encoder is characterized by first dividing the sensor channels into K non-overlapping groups according to physical semantics; then, performing multi-head self-attention computation in parallel within each group to capture long-term dependencies within the group; and finally, fusing the output features of the K groups through an inter-group interaction module. Sequence Pooling and Projection Module: Composed of a sequence average pooling layer and a fully connected layer, it is used to aggregate the fused sequence features into a global feature vector of fixed length; the final output of the divide-and-conquer Transformer branch is the global feature vector.
[0038] Layer 4: Adaptive Fusion Gate Module The adaptive fusion gate module is used to dynamically and weightedly fuse features from the convolutional neural network branch and the divide-and-conquer Transformer branch; this module includes, in sequence: Feature concatenation unit: used to concatenate the input local feature vector and global feature vector along the feature dimension; Weight generation unit: composed of one or more fully connected layers and a Softmax normalization function, used to learn and generate a pair of normalized fusion weights based on the concatenated features; Feature fusion unit: used to perform weighted summation operation, the calculation formula is: fused feature = α × local feature vector + β × global feature vector; The final output of the adaptive fusion gate module is the fused feature vector; Layer 5: Classifier Layer The classifier layer is used to determine the fault type based on the fused feature vector; this layer consists of a fully connected layer with the number of neurons equal to the number of predefined fault categories; the output of this layer is the unnormalized discrimination score for each fault category.
[0039] Layer 6: Output Layer The output layer, consisting of a Softmax function, is used to convert the discrimination scores output by the classifier layer into a probability distribution. The output of this layer is a probability vector, in which the category with the highest probability value is determined as the final fault diagnosis result.
[0040] Step 4: Through data analysis, users can receive feedback on the analysis results via mobile phones and tablets through 4G / 5G networks, allowing them to intuitively understand the equipment's operation and fault status.
[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A remote equipment fault pre-diagnosis method based on AI analysis, characterized in that... Include: Step 1: During the downtime, construct a consistency evaluation model for the corresponding remote equipment's static and dynamic mechanical components based on multidimensional data; Step 2: After the dynamic and static consistency evaluation model in Step 1 is evaluated, vibration data is collected during the operation of the remote equipment. The output parameters during operation are adjusted to complete and simulate the operation failure of the remote equipment. The vibration signal is converted from the time domain to the frequency domain by fast Fourier transform, and the consistency of the frequency domain characteristics is evaluated by calculating the power spectral density and comparing their correlation coefficients. Convert the signal with N sampling points to the frequency domain in O(NlogN) time complexity. For a discrete-time signal x[n] of length N, its discrete Fourier transform X[k] is calculated using the following formula: In the formula, X[k] represents the complex amplitude of signal x(n) at frequency k, x(n) is the signal in discrete time, k is the discrete frequency, and N is the number of sampling points of the signal; The discrete Fourier transform of N sampling points is decomposed into multiple smaller transforms, thereby achieving lower time complexity; Further frequency domain feature analysis is performed on the obtained vibration data; by calculating the power spectral density of the frequency domain data, the energy distribution of the signal in the frequency domain is obtained, and then the correlation coefficient is used to describe the correlation between the two vibration signals in the power spectral density; when calculating the power spectral density function of two vibration data and calculating the correlation coefficient between them, it is actually comparing whether their distribution in the frequency domain is similar; if the energy distribution of the two vibration data in the frequency domain is similar, it is considered to be consistent with the vibration characteristics of a certain type of fault, and the corresponding fault category is preliminarily determined. Step 3: Remote device operation fault reproduction. After determining the initial fault category in Step 2, the operation data is uploaded to the remote server in real time via the 4G / 5G module. The remote server records and analyzes the corresponding data; further analysis of the fault is conducted based on multivariate operation parameters and a deep network classification model. Step 4: Through data analysis, users can receive feedback on the analysis results via mobile phones and tablets through 4G / 5G networks, allowing them to intuitively understand the equipment's operation and fault status.
2. The method according to claim 1, characterized in that... In step 1, the specific remote device is a wall-hung boiler, which includes mechanical components divided into static and dynamic components. Static components are those whose geometric characteristics do not change significantly over time, while dynamic components are those whose geometric characteristics change significantly over time. Considering the operating characteristics of the wall-hung boiler, both dynamic and static components will undergo different degrees of mechanical deformation during operation. Static components include: circulating water pump, fan, gas valve, three-way valve, and automatic exhaust valve. Dynamic components include: heat exchanger, burner, shell, pipes, flue pipe, and expansion tank. Consistency assessment is performed based on the following formula: Where R represents the consistency level. The importance weight of static parts; Weights for the importance of dynamic parts; These are measurements taken at static locations; This refers to the actual value of the dynamic part; These are measurements taken at dynamic locations. This represents the actual dimensions of the dynamic part.
3. The method according to claim 1, characterized in that... Step 3 specifically includes: collecting data from multiple operating sensors, including data such as outlet water temperature, return water temperature, combustion chamber temperature, flue gas temperature, gas flow rate, water flow rate, gas pressure, water pressure, ignition current, valve opening, flue gas oxygen content, and carbon monoxide concentration.
4. The method according to claim 3, characterized in that... Step 3 also includes: aligning, cleaning, and labeling the collected data, and handling missing values caused by communication interruptions by aligning timestamps.
5. The method according to claim 4, characterized in that... Step 3 also includes: performing Z-Score normalization on each sensor channel using forward padding or linear interpolation; for steady-state faults: extracting a 5-minute sequence from the stable operation phase as a sample; for transient faults: extracting a 1-minute sequence starting from the critical event as a sample, with each sample corresponding to at least one fault label.
6. The method according to claim 5, characterized in that... Step 3 also includes: proposing a six-layer improved deep network model based on labeled data for further analysis, specifically including: Layer 1: The input layer receives preprocessed multi-channel sensor time-series data; its input data dimensions are [batch size, number of channels C, sequence length L]; this layer acts as a data distribution node, synchronously feeding the input data to the convolutional neural network branch and the divide-and-conquer Transformer branch; Layer 2: Convolutional Neural Network Branch The convolutional neural network branch, as a local feature extractor, is used to capture the local temporal morphological features within each sensor channel; This branch road includes, in sequence: Initial convolutional module: Consists of a one-dimensional convolutional kernel, activation function, and batch normalization layer connected sequentially, used to achieve initial feature fusion and dimensionality enhancement; Multiple cascaded improved depthwise separable convolutional modules: Each module includes sequentially connected up-dimensional point convolution, depthwise convolution, squeezed excitation attention module and down-dimensional point convolution, and introduces residual connection structure; the feature map size is compressed stepwise by setting a stride greater than 1 for depthwise convolution; Global Pooling and Projection Module: Composed of a global average pooling layer and a fully connected layer, it is used to aggregate the compressed feature map into a fixed-length local feature vector. Layer 3: Divide and Conquer Transformer Branches The divide-and-conquer Transformer branch, acting as a global relation extractor, is used to capture long-term dependencies across sensor channels; this branch includes, in sequence: Embedding and Encoding Module: Consists of a fully connected layer and a learnable position encoder, used to map the input sequence into a high-dimensional embedding vector and inject position information; The divide-and-conquer Transformer encoder is characterized by first dividing the sensor channels into K non-overlapping groups according to physical semantics; then, performing multi-head self-attention computation in parallel within each group to capture long-term dependencies within the group; and finally, fusing the output features of the K groups through an inter-group interaction module. Sequence Pooling and Projection Module: Composed of a sequence average pooling layer and a fully connected layer, used to aggregate the fused sequence features into a fixed-length global feature vector; the final output of the divide-and-conquer Transformer branch is the global feature vector; Layer 4: Adaptive Fusion Gate Module The adaptive fusion gate module is used to dynamically and weightedly fuse features from the convolutional neural network branch and the divide-and-conquer Transformer branch; this module includes, in sequence: Feature concatenation unit: used to concatenate the input local feature vector and global feature vector along the feature dimension; Weight generation unit: composed of one or more fully connected layers and a Softmax normalization function, used to learn and generate a pair of normalized fusion weights based on the concatenated features; Feature fusion unit: used to perform weighted summation operation, the calculation formula is: fused feature = α × local feature vector + β × global feature vector; the final output of the adaptive fusion gate module is the fused feature vector; Layer 5: Classifier Layer The classifier layer is used to determine the fault type based on the fused feature vector; this layer consists of a fully connected layer with the number of neurons equal to the number of predefined fault categories; the output of this layer is the unnormalized discrimination score for each fault category. Layer 6: Output Layer The output layer, consisting of a Softmax function, is used to convert the discrimination scores output by the classifier layer into a probability distribution. The output of this layer is a probability vector, in which the category with the highest probability value is determined as the final fault diagnosis result.
7. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, can implement the method described in any one of claims 1-6.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, characterized in that, when executed by a processor, the program can implement the method described in any one of claims 1-6.
Citation Information
Patent Citations
Rolling bearing rolling body weak fault feature extraction method under transmission path
CN113158896A
Transformer state analysis method and system based on digital twinning technology
CN117786515A
