Cross-working-condition edge end fault diagnosis method based on domain self-adaption
By using the operating samples and fault type operating samples under a single operating condition obtained by simulation experiments in mechanical equipment fault diagnosis, cloud models are trained and edge models are built, and the problem of low fault identification accuracy in the existing technology under multiple operating conditions and complex conditions is solved, and efficient and accurate fault diagnosis is achieved.
Patent Information
- Application Number
- CN202510233416.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-06-20
AI Technical Summary
The prior art has low accuracy in identifying faults of mechanical equipment under multiple operating conditions and complex conditions, and has a long model training time and high calculation cost.
Through simulation experiments, the operating samples of mechanical equipment under a single operating condition and the operating samples of each fault type are obtained, the cloud model is trained, and the edge model is constructed through weight sharing operation. The edge model can be fault diagnosis under multiple operating conditions.
It significantly reduces the computing cost of edge models while maintaining high-accuracy fault recognition capabilities, enabling higher fault diagnosis accuracy and robustness under multiple operating conditions and complex conditions.
Smart Images
Figure CN120180871A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of large models, and in particular to a cross-condition edge-side fault diagnosis method based on domain adaptation. Background Art
[0002] Modern transmission equipment shows a development trend of complexity, intelligence, and informatization. As an important transmission equipment, each transmission component in mechanical equipment is closely connected, and a fault in one component may affect the normal operation of the entire mechanical equipment. Bearings are key components in mechanical equipment. However, the complex working environment and long-term high-speed operation make bearings extremely prone to faults. If the faults are not detected in time, it is easy to cause a series of equipment damages and even casualties.
[0003] Currently, the method of machine learning based on big data is mostly used to identify faults in each component of mechanical equipment. In this method, the model is usually trained through a large number of operating samples under various conditions, and then the trained model is used to identify faults in each component of mechanical equipment. However, only through this method, due to the need to perform high-complexity calculations on a large amount of data under multiple conditions, the model has extremely high requirements for hardware resources, long training time, and high calculation cost; but if the amount of data used for model training is simply reduced, the fault identification accuracy of the model for mechanical equipment under multiple conditions and complex conditions will be significantly reduced.
[0004] Therefore, there is an urgent need for a cross-condition edge-side fault diagnosis method based on domain adaptation. Summary of the Invention
[0005] This application provides a cross-condition edge-side fault diagnosis method based on domain adaptation, which solves the problem that if the amount of data used for model training is simply reduced, the fault identification accuracy of the model for mechanical equipment under multiple conditions and complex conditions will be significantly reduced.
[0006] In the first aspect of the present application, a cross-condition edge-side fault diagnosis method based on domain adaptation is provided. The method includes: obtaining operation samples under a single condition corresponding to a mechanical equipment and operation samples of various fault types through simulation experiments; training a cloud model based on the operation samples under the single condition and the operation samples of various fault types. The cloud model includes a cloud model pre-feature extractor and a cloud model post-feature extractor; constructing an edge model according to the cloud model and through a weight sharing operation. The edge model includes an edge model pre-feature extractor, an edge model post-feature extractor, and a classifier. The weight sharing operation is used to share the feature extraction performance of the cloud model pre-feature extractor to the edge model pre-feature extractor; inputting the operation samples under the single condition, and outputting cloud model fault diagnosis features according to the cloud model pre-feature extractor; inputting the operation samples under multiple conditions, and outputting edge model fault diagnosis features according to the edge model pre-feature extractor; calculating a target optimization value according to the cloud model fault diagnosis features and the edge model fault diagnosis features, and training the edge model post-feature extractor and the classifier of the edge model according to the target optimization value; outputting a fault diagnosis result of the mechanical equipment according to the trained edge model.
[0007] Optionally, obtaining the operation samples under a single condition corresponding to the mechanical equipment through simulation experiments specifically includes: obtaining multiple operation states in the mechanical equipment, where the operation states include a rotational speed operation state, a load operation state, and a working temperature operation state; obtaining the preset level ranges respectively corresponding to the multiple operation states, with one operation state corresponding to one preset level range; obtaining the condition samples of the mechanical equipment when a first preset condition is satisfied, where the first preset condition is that the multiple operation states are maintained within their respective preset level ranges; and using the condition samples when the first preset condition is satisfied as the operation samples under the single condition.
[0008] Optionally, after using the condition samples when the first preset condition is satisfied as the operation samples under the single condition, the method further includes normalizing the operation samples under the single condition, which specifically includes: obtaining the time-domain condition samples corresponding to the mechanical equipment through an acceleration sensor; converting the time-domain condition samples into frequency-domain condition samples through Fourier transform; and normalizing the time-domain condition samples through the following formula:
[0009] x′ = (x - x min ) / (x max - x min );
[0010] where x′ is the time-domain condition sample after normalization, x is the time-domain condition sample, and x max , x min respectively correspond to the maximum value and the minimum value of the time-domain condition sample.
[0011] Optionally, a weight sharing operation is performed on the edge model pre-feature extractor, which specifically includes: obtaining the weights of the cloud model pre-feature extractor corresponding to the cloud model through iterative training, and migrating the weights of the cloud model pre-feature extractor corresponding to the cloud model to the edge model; under the condition of meeting the second preset condition, locking the edge model pre-feature extractor corresponding to the edge model pre-feature extractor according to the weights of the cloud model pre-feature extractor to complete the weight sharing operation. The second preset condition is to ensure that the module-level structures of the cloud model pre-feature extractor and the edge model pre-feature extractor are the same, and the module-level structure includes a convolutional neural network layer structure, a residual structure, and an activation function structure.
[0012] Optionally, according to the cloud model fault diagnosis features and the edge model fault diagnosis features, a target optimization value is calculated, which specifically includes: calculating a feature loss according to the cloud model fault diagnosis features and the edge model fault diagnosis features and through the dual constraint algorithm of the maximum mean discrepancy; calculating a classification loss according to the cloud model fault diagnosis features and the edge model fault diagnosis features and through a cross-entropy loss function; calculating a feature loss weight corresponding to the feature loss and a classification loss weight corresponding to the classification loss; calculating the target optimization value according to the feature loss, the classification loss, the feature loss weight, and the classification loss weight.
[0013] Optionally, calculating a feature loss weight corresponding to the feature loss and a classification loss weight corresponding to the classification loss specifically includes: calculating a first gradient corresponding to the feature loss and a second gradient corresponding to the classification loss; calculating the feature loss weight and the classification loss weight according to the feature loss, the classification loss, the first gradient, and the second gradient through the following formula:
[0014]
[0015] where α is the feature loss weight, β is the classification loss weight, w a is the Euclidean norm corresponding to the first gradient, w b is the Euclidean norm corresponding to the second gradient, l a is the Euclidean norm corresponding to the feature loss, l b is the Euclidean norm corresponding to the classification loss, and ∈ is a minimum value used to avoid a denominator of 0.
[0016] Optionally, calculating the target optimization value according to the feature loss, the classification loss, the feature loss weight, and the classification loss weight specifically includes: calculating the target optimization value according to the following formula:
[0017]
[0018] Among them, loss is the target optimization value, LOSS_Feature is the feature loss, LOSS_Classify is the classification loss, α is the feature loss weight, and β is the classification loss weight.
[0019] Optionally, the operating conditions include load conditions, speed conditions, lubrication status, and temperature change conditions.
[0020] Optionally, the fault types include wear faults, broken tooth faults, and pitting faults.
[0021] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:
[0022] 1. Obtain the operation samples of the mechanical equipment under a single operating condition and the operation samples of each fault type through simulation experiments, and train a cloud model based on the operation samples under a single operating condition and the operation samples of each fault type. The cloud model includes a cloud model pre-feature extractor and a cloud model post-feature extractor. Then, according to the cloud model, construct an edge model through weight sharing operation. The edge model includes an edge model pre-feature extractor, an edge model post-feature extractor, and a classifier. After that, by inputting the operation samples under a single operating condition, output the cloud model fault diagnosis features according to the cloud model pre-feature extractor, and by inputting the operation samples under multiple operating conditions, output the edge model fault diagnosis features according to the edge model pre-feature extractor; calculate the target optimization value according to the cloud model fault diagnosis features and the edge model fault diagnosis features, and train the edge model post-feature extractor and classifier of the edge model according to the target optimization value; according to the trained edge model, output the fault diagnosis result of the mechanical equipment. Furthermore, by adopting the mechanism of weight sharing, transfer the feature extraction part with complex calculations in the cloud model to the edge model, significantly reducing the calculation cost of the edge model while maintaining a high-accuracy fault recognition ability.
[0023] 2. Obtain the time-domain operating condition samples corresponding to the mechanical equipment through an acceleration sensor, convert the time-domain operating condition samples into frequency-domain operating condition samples through Fourier transform, and perform normalization processing on the frequency-domain operating condition samples. Thus, use the normalized time-domain operating condition samples as the operating condition samples for feature extraction, which can effectively retain the dynamic feature information of the operating state of the mechanical equipment, and at the same time eliminate the influence caused by the signal amplitude difference, making the data have better comparability and consistency.
[0024] 3. Based on the cloud model fault diagnosis features and edge model fault diagnosis features, calculate the feature loss through the dual-constraint algorithm of maximum mean discrepancy to quantify the distribution difference of relevant sub-domains between the source domain and the target domain; calculate the classification loss through the cross-entropy loss function to evaluate the classification accuracy of the target domain samples; further, calculate the feature loss weight corresponding to the feature loss and the classification loss weight corresponding to the classification loss, and calculate the final target optimization value through weighted synthesis according to the feature loss, classification loss and the corresponding weights, so as to simultaneously optimize the cross-domain feature distribution alignment and classification performance, thereby achieving higher fault diagnosis accuracy and robustness under multi-condition conditions. Brief Description of the Drawings
[0025] Figure 1 is a schematic flowchart of a cross-condition edge-side fault diagnosis method based on domain adaptation provided by an embodiment of the present application;
[0026] Figure 2 is a schematic module diagram of a mechanical equipment fault diagnosis device for cross-conditions provided by an embodiment of the present application;
[0027] Figure 3 is a schematic flowchart of a mechanical equipment fault diagnosis for cross-conditions provided by an embodiment of the present application. Detailed Embodiments
[0028] In order to enable those skilled in the art to better understand the technical solutions in this specification, the following will clearly and completely describe the technical solutions in the embodiments of this specification with reference to the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments.
[0029] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. As used in the specification of the present application, the singular forms "a", "an", "the", "above", "the foregoing", "this" are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term " / and / " used in the present application refers to and includes any or all possible combinations of one or more of the listed items.
[0030] Hereinafter, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, unless otherwise stated, the meaning of "a plurality" is two or more.
[0031] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0032] Please refer to Figure 1 , which shows a schematic flow chart of a cross-condition edge-side fault diagnosis method based on domain adaptation provided by an embodiment of the present application. The flow chart mainly includes the following steps: S101 to S106.
[0033] Step S101, obtain the operation samples under a single condition corresponding to the mechanical equipment and the operation samples of each fault type through simulation experiments.
[0034] Specifically, in the following embodiments of the present application, the mechanical equipment will be taken as an example to illustrate the method steps of fault diagnosis, but the fault diagnosis is not limited to the mechanical equipment. When the user performs a fault diagnosis operation on the mechanical equipment, the operation samples under a single condition corresponding to the mechanical equipment are obtained through simulation experiments. Among them, the condition is the general term for the description of the operation state, including but not limited to: load condition, rotation speed condition, lubrication state, temperature change condition, etc.; the condition samples of each fault type include but not limited to: the condition samples of bearing wear fault, gear meshing fault, gear fracture fault, lubrication poor fault, etc. The condition samples can characterize the vibration characteristics of the mechanical equipment under different operation states and fault conditions, and are used to support the subsequent training and optimization of the fault diagnosis model. For the specific processing method of the target samples, the steps are as follows: obtain the time-domain condition samples corresponding to the mechanical equipment through a three-axis acceleration sensor; convert the time-domain condition samples into frequency-domain condition samples through Fourier transform, and perform fast Fourier transform through the following formula:
[0035]
[0036] Among them, x(F) is the fast Fourier transform function, x is the time-domain condition sample, F is the frequency-domain condition sample, that is, the target frequency component, and can also be understood as the specific frequency point analyzed in the frequency domain. T S is the sampling period, which represents the time interval of signal sampling, defines the time difference between sampling points, t0 represents the initial time of the frequency-domain signal, and is used to correct the initial phase of the signal to ensure that the calculated frequency-domain signal accurately reflects the actual signal. n represents the discrete sampling point number of the time-domain signal and is used to traverse all signal sampling values. F n is the specific frequency value in the spectrum, -j2πFt0 is the complex exponential function, which is used to adjust the phase of the frequency-domain signal and is related to the initial time t0 of the signal. -j2πF n T s is also the complex exponential function, which is used to describe the amplitude and phase relationship of each frequency component of the time-domain signal in the frequency domain. j is the imaginary unit. After that, the time-domain condition samples are normalized through the following formula:
[0037] x′ = (x - x min ) / (x max - x min );
[0038] Wherein, x′ is the time-domain operating condition sample after normalization processing, x is the time-domain operating condition sample, and x max , x min correspond to the maximum and minimum values of the time-domain operating condition sample respectively.
[0039] In a possible implementation manner, step S101 further includes: obtaining preset level ranges corresponding to multiple operating states respectively, where one operating state corresponds to one preset level range; obtaining the operating condition sample of the mechanical equipment when the first preset condition is satisfied, and the first preset condition is that multiple operating states are maintained within their respective corresponding preset level ranges; using the operating condition sample when the first preset condition is satisfied as the operating condition sample under a single operating condition.
[0040] Specifically, obtaining multiple operating states of the mechanical equipment, including but not limited to operating parameters such as rotational speed, load, and working temperature, defining corresponding preset level ranges for each operating state. For example, the range of rotational speed is the upper and lower deviations of a specific rotational speed value, the range of load is the percentage interval of the rated load, and the range of working temperature is the fluctuation interval of the normal working temperature of the equipment; when the operating states are maintained within their respective corresponding preset level ranges, record the operating data of the mechanical equipment and generate the operating condition sample; define the set of all operating condition samples that satisfy the first preset condition as the operating condition sample under a single operating condition, which is used to characterize the performance characteristics of the mechanical equipment in the normal operating state and provide reference data for fault diagnosis.
[0041] Step S102, training a cloud model based on the operating samples under a single operating condition and the operating samples of each fault type.
[0042] Specifically, construct a cloud model according to the target operating condition sample. The composition of the cloud model mainly includes: a cloud model pre-feature extractor, a post-feature extractor, and a classifier. Among them, the cloud model pre-feature extractor is the pre-feature extractor of the cloud model; a large number of residual connection modules are adopted in the feature extractor. The residual module can be implemented in the form of skip-layer connection, adding the input of the unit directly to the output of the unit and then activating. This method can improve the performance of the cloud model and reduce the problems of gradient disappearance and training difficulties during the training process. Train the cloud model based on the obtained large number of target operating condition samples, and obtain the optimal weight of the cloud model through iterative training.
[0043] Step S103, construct an edge model according to the cloud model and through weight sharing operation.
[0044] Specifically, according to the cloud model and through weight sharing operations, an edge model is constructed. The edge model includes: an edge model pre-feature extractor, a post-feature extractor, and a classifier. Among them, the edge model pre-feature extractor is the pre-feature extractor of the edge model. A weight sharing operation is performed on the edge model pre-feature extractor of the edge model, enabling the edge model to fully utilize the feature extraction capabilities of the cloud model part. This process means that after the weight sharing operation, the edge model can effectively inherit and share the feature representations learned by the cloud model, thereby reducing the computational cost of training this part of the feature extraction capabilities on the edge device and effectively reducing the computational burden of the edge model. After completing the above weight sharing operation, that is, after locking the weights of the pre-feature extractor of the edge model, the edge model to be trained is constructed by training its post-extractor and classifier.
[0045] In a possible implementation manner, step S104 further includes: obtaining the weights of the cloud model pre-feature extractor corresponding to the cloud model through iterative training, and migrating the weights of the cloud model pre-feature extractor corresponding to the cloud model to the edge model; under the condition of meeting the second preset condition, locking the edge model pre-feature extractor corresponding to the edge model pre-feature extractor according to the weights of the cloud model pre-feature extractor to complete the weight sharing operation. The second preset condition is to ensure that the module-level structures of the cloud model pre-feature extractor and the edge model pre-feature extractor are the same. The module-level structure includes a convolutional neural network layer structure, a residual structure, and an activation function structure.
[0046] Specifically, the weights of the cloud model pre-feature extractor of the cloud model are obtained through iterative training. The target weights are optimized by the cloud model pre-feature extractor of the cloud model after feature learning on multiple operating state samples; the obtained weights of the cloud model pre-feature extractor are migrated to the edge model pre-feature extractor of the edge model. Under the condition of meeting the second preset condition, the edge model pre-feature extractor of the edge model pre-feature extractor is further locked to complete the weight sharing operation. Among them, the second preset condition is to ensure that the module-level structures of the cloud model pre-feature extractor and the edge model pre-feature extractor are the same. The module-level structure includes a convolutional neural network layer structure for extracting local features of signals; a residual structure to alleviate the problem of gradient disappearance in deep networks through skip connections; and an activation function structure for introducing non-linear characteristics to enhance the feature expression ability, thereby ensuring the consistency of the feature extraction capabilities between the cloud model and the edge model.
[0047] In step S104, by inputting the operation samples under a single working condition, the cloud model fault diagnosis features are output according to the cloud model pre-feature extractor; by inputting the operation samples under multiple working conditions, the edge model fault diagnosis features are output according to the edge model pre-feature extractor.
[0048] Specifically, by inputting the operation samples under a single working condition, the cloud model fault diagnosis features are output according to the cloud model pre-feature extractor. By inputting the operation samples under multiple working conditions, the edge model fault diagnosis features are output according to the edge model pre-feature extractor. This step includes inputting the working condition samples under a single working condition into the cloud model pre-feature extractor of the cloud model to obtain the cloud model fault diagnosis features, inputting the working condition samples of each fault type into the edge model pre-feature extractor of the edge model to obtain the edge model fault diagnosis features, and denoting the cloud model fault diagnosis features as Denoting the edge model fault diagnosis features as
[0049] Step S105: Calculate the target optimization value according to the cloud model fault diagnosis features and the edge model fault diagnosis features, and train the edge model posterior feature extractor and classifier of the edge model according to the target optimization value.
[0050] Specifically, calculate the target optimization value according to the cloud model fault diagnosis features, the second fault diagnosis features, and the labels corresponding to each target working condition sample, and train the edge model according to the target optimization value.
[0051] In a possible implementation manner, step S105 further includes: calculating the feature loss according to the cloud model fault diagnosis features and the edge model fault diagnosis features through the dual constraint algorithm of the maximum mean discrepancy; calculating the classification loss according to the cloud model fault diagnosis features and the edge model fault diagnosis features through the cross-entropy loss function; calculating the feature loss weight corresponding to the feature loss and the classification loss weight corresponding to the classification loss; calculating the target optimization value according to the feature loss, the classification loss, the feature loss weight, and the classification loss weight.
[0052] Specifically, based on the cloud model fault diagnosis features and the edge model fault diagnosis features, the feature loss is calculated by the dual constraint algorithm of the maximum mean discrepancy. The feature loss is the distribution difference of relevant subdomains. The specific calculation steps of the distribution difference of relevant subdomains are as follows: Map the feature representations of the source domain and the target domain to a high-dimensional kernel space, calculate the means of the features of different subdomains respectively, and then calculate the mean discrepancy between the source domain and the target domain based on the kernel method; By establishing a dual constraint relationship between subdomains, enhance the sensitivity of the algorithm to the local feature distribution difference, and perform a weighted average of these differences to finally obtain the measurement result of the distribution difference of relevant subdomains, which is used to optimize the cross-domain alignment ability of the model and ensure the effectiveness of the target domain features in the fault diagnosis task. According to the cloud model fault diagnosis features and the edge model fault diagnosis features, the classification loss is calculated by the cross-entropy loss function. The specific calculation steps of the classification loss are as follows: First, use the classifier of the edge model to generate the predicted class probability distribution for the target domain samples, and perform label smoothing on the predicted classes to reduce the model overfitting that may be caused by label noise; Then, compare the smoothed predicted class distribution with the true class labels, and calculate the classification loss through the cross-entropy loss function. This loss value reflects the optimization degree of the classification accuracy of the model on the target domain; Finally, take the classification loss as one of the optimization objectives, and adjust the classifier weights through backpropagation to improve the classification performance of the edge model in the target domain fault diagnosis task. Denote the feature loss as LOSS_Feature and the classification loss as LOSS_Classify. Assign weights A and B to LOSS_Feature and LOSS_Feature respectively, and calculate LOSS through the formula: LOSS = A * LOSS Classify + B * LOSS_Feature, and take it as the minimization optimization objective. Then, calculate the first gradient corresponding to the feature loss and the second gradient corresponding to the classification loss according to the following formula:
[0053]
[0054] where Grad_Classify is the first gradient, and the first gradient is the gradient of LOSS_Classify with respect to the output value of the edge model ; Grad_Feature is the second gradient, and the second gradient is the gradient of LOSS_Feature with respect to the output value of the edge model ; Gradient() represents the gradient calculation of the output value of the edge model .
[0055] Based on the feature loss, classification loss, first gradient and second gradient, calculate the feature loss weight and classification loss weight according to the following formula:
[0056]
[0057] where α is the feature loss weight, β is the classification loss weight, w a is the Euclidean norm corresponding to the first gradient, w b is the Euclidean norm corresponding to the second gradient, l a is the Euclidean norm corresponding to the feature loss, l b is the Euclidean norm corresponding to the classification loss, and ∈ is a minimum value used to avoid a zero denominator.
[0058] Finally, the target optimization value is calculated by the following formula:
[0059]
[0060] Step S106, according to the trained edge model, output the fault diagnosis result of the mechanical equipment.
[0061] Specifically, at the edge side, based on a small number of samples, the excellent feature extraction ability of the cloud model is transferred to the lightweight edge model. At the same time, the diagnostic ability in the source domain is generalized to the corresponding target domain. After the above knowledge transfer process, the edge model is deployed and run. By inputting real-time multi-condition operation data, the fault diagnosis result can be obtained quickly and accurately near the mechanical equipment.
[0062] Please refer to Figure 2 , which shows a schematic diagram of the distributed architecture of an edge computing solution provided by an embodiment of the present application. In this figure, the cloud server includes a cloud model, the edge device includes an edge model. By locking the optimal weights of the cloud model to the edge model and giving the factual diagnosis result of the mechanical equipment according to the edge model, Equipment01 to EquipmentN are all mechanical equipment for real-time diagnosis. Please refer to Figure 3 , which shows a schematic diagram of the process of mechanical equipment fault diagnosis under cross-conditions provided by an embodiment of the present application.
[0063] By adopting the above method, this application obtains the operation samples under a single working condition corresponding to the mechanical equipment and the operation samples of various fault types through simulation experiments, and trains a cloud model based on the operation samples under a single working condition and the operation samples of various fault types. The cloud model includes a cloud model pre-feature extractor and a cloud model post-feature extractor. Then, according to the cloud model, an edge model is constructed through weight sharing operation. The edge model includes an edge model pre-feature extractor, an edge model post-feature extractor, and a classifier. After that, by inputting the operation samples under a single working condition, the cloud model fault diagnosis features are output according to the cloud model pre-feature extractor, and by inputting the operation samples under multiple working conditions, the edge model fault diagnosis features are output according to the edge model pre-feature extractor; according to the cloud model fault diagnosis features and the edge model fault diagnosis features, the target optimization value is calculated, and the edge model post-feature extractor and classifier of the edge model are trained according to the target optimization value; according to the trained edge model, the fault diagnosis result of the mechanical equipment is output. Furthermore, by adopting the mechanism of weight sharing, the feature extraction part with complex calculations in the cloud model is migrated to the edge model, significantly reducing the calculation cost of the edge model while maintaining a high-accuracy fault recognition ability.
Claims
1. A cross-operating-condition edge fault diagnosis method based on domain adaptation, characterized in that: The method comprises: Through simulation experiments, operation samples under single working conditions and operation samples of various fault types corresponding to mechanical equipment are obtained; Training a cloud model based on the operation samples under the single working condition and the operation samples of each fault type, wherein the cloud model includes a cloud model pre-feature extractor and a cloud model posterior feature extractor; According to the cloud model, an edge model is constructed through a weight sharing operation, wherein the edge model includes an edge model pre-feature extractor, an edge model posterior feature extractor, and a classifier, and the weight sharing operation is used to share the feature extraction performance of the cloud model pre-feature extractor to the edge model pre-feature extractor; By inputting the operation samples under the single working condition, the cloud model fault diagnosis features are output according to the cloud model pre-feature extractor; by inputting the operation samples under multiple working conditions, the edge model fault diagnosis features are output according to the edge model pre-feature extractor; Calculating a target optimization value according to the cloud model fault diagnosis feature and the edge model fault diagnosis feature, and training the edge model posterior feature extractor and the classifier of the edge model according to the target optimization value; According to the trained edge model, the fault diagnosis result of the mechanical equipment is output.
2. The method according to claim 1, characterized in that The obtaining of operation samples under a single working condition corresponding to the mechanical equipment through simulation experiments specifically includes: Acquire multiple operating states of the mechanical equipment, wherein the operating states include a speed operating state, a load operating state, and an operating temperature operating state; Acquire preset horizontal ranges corresponding to a plurality of the operating states respectively, wherein one operating state corresponds to one preset horizontal range; Acquire a working condition sample of the mechanical equipment when a first preset condition is met, wherein the first preset condition is that a plurality of the operating states are maintained within the respective corresponding preset level ranges; The operating condition sample when the first preset condition is met is used as the operation sample under the single operating condition.
3. The method according to claim 2, characterized in that After taking the operating condition samples satisfying the first preset condition as the operation samples under the single operating condition, the method further includes normalizing the operation samples under the single operating condition, specifically including: Acquiring a time domain working condition sample corresponding to the mechanical equipment through an acceleration sensor; Converting the time domain working condition samples into frequency domain working condition samples by Fourier transform; The time domain working condition samples are normalized by the following formula: x′=(x-x min ) / (x max -x min ); Wherein, x′ is the normalized time domain working condition sample, x is the time domain working condition sample, and x max 、x min They correspond to the maximum and minimum values of the time domain operating condition samples respectively.
4. The method according to claim 1, characterized in that: The weight sharing operation is performed on the edge model pre-feature extractor, specifically including: Obtaining a cloud model pre-feature extractor weight corresponding to the cloud model through iterative training, and migrating the cloud model pre-feature extractor weight corresponding to the cloud model to the edge model; Under the second preset condition, the edge model pre-feature extractor corresponding to the edge model pre-feature extractor is locked according to the weight of the cloud model pre-feature extractor to complete the weight sharing operation, and the second preset condition is to ensure that the module-level structure of the cloud model pre-feature extractor is the same as that of the edge model pre-feature extractor, and the module-level structure includes a convolutional neural network layer structure, a residual structure, and an activation function structure.
5. The method according to claim 1, characterized in that The calculating the target optimization value according to the cloud model fault diagnosis feature and the edge model fault diagnosis feature specifically includes: According to the cloud model fault diagnosis feature and the edge model fault diagnosis feature, the feature loss is calculated by a dual constraint algorithm of maximum mean difference; Calculate the classification loss according to the cloud model fault diagnosis feature and the edge model fault diagnosis feature and through a cross entropy loss function; Calculate the feature loss weight corresponding to the feature loss and the classification loss weight corresponding to the classification loss; The target optimization value is calculated according to the feature loss, the classification loss, the feature loss weight and the classification loss weight.
6. The method according to claim 5, characterized in that The calculating the feature loss weight corresponding to the feature loss and the classification loss weight corresponding to the classification loss specifically includes: Calculating a first gradient corresponding to the feature loss and a second gradient corresponding to the classification loss; The feature loss weight and the classification loss weight are calculated according to the following formula using the feature loss, the classification loss, the first gradient, and the second gradient: Among them, α is the feature loss weight, β is the classification loss weight, and w a is the Euclidean norm corresponding to the first gradient, w b is the Euclidean norm corresponding to the second gradient, l a is the Euclidean norm corresponding to the feature loss, l b is the Euclidean norm corresponding to the classification loss, ∈ is a minimum value used to avoid the denominator being 0.
7. The method according to claim 5, characterized in that The calculating the target optimization value according to the feature loss, the classification loss, the feature loss weight and the classification loss weight specifically includes: The target optimization value is calculated according to the following formula: Among them, loss is the target optimization value, LOSS_Feature is the feature loss, LOSS_Classify is the classification loss, α is the feature loss weight, and β is the classification loss weight.
8. The method according to claim 1, characterized in that The operating conditions include load conditions, speed conditions, lubrication conditions and temperature change conditions.
9. The method according to claim 1, characterized in that: The fault types include wear fault, tooth breakage fault and pitting fault.