A device health management method based on vibration signal unified coding
By constructing a device health management method based on unified coding of vibration signals, the problem of existing technologies being difficult to apply under various working conditions is solved, and efficient response and rapid analysis of various device health management tasks are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG INST OF AUTOMATION - CHINESE ACAD OF SCI
- Filing Date
- 2024-12-11
- Publication Date
- 2026-06-12
AI Technical Summary
Existing AI-based equipment health management algorithms struggle to perform various health management tasks under diverse working conditions, limiting their application in practical engineering.
A unified coding method for equipment health management based on vibration signals is constructed. This method involves building a multimodal equipment health management dataset, a fault classification network, and a multimodal equipment health management language model. The large language model is used to perform unified health management task response, and equipment health management is carried out in combination with vibration signals and operating parameters.
It enables reliable response to various device health management tasks under heavy workloads, reduces the professional skill requirements of maintenance personnel, and allows for rapid and efficient device health management on terminals with limited computing power.
Smart Images
Figure CN122196857A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of health management algorithms, specifically a device health management method based on unified encoding of vibration signals. Background Technology
[0002] Mechanical equipment is prone to malfunctions during operation. By installing accelerometers to collect vibration signals, the operating status of the equipment can be analyzed, enabling equipment health management and preventing personal injury and property damage. Existing AI-based health management algorithms often only address one or a few specific operating conditions for particular health management tasks, such as fault diagnosis. In actual engineering projects, the operating conditions of equipment and the required health management tasks are far more diverse, making it difficult to apply existing algorithms to real-world scenarios. Summary of the Invention
[0003] To address the limitations of existing technologies in performing various health management tasks under diverse working conditions, this invention provides a device health management method based on unified vibration signal encoding. This method considers the spectral changes that occur after a device malfunctions and utilizes a large language model to provide a unified health management task response, thus reliably providing a reference for device health management.
[0004] The technical solution adopted by the present invention to achieve the above objectives is as follows:
[0005] A device health management method based on unified vibration signal encoding includes the following steps:
[0006] 1) Construct a multimodal device health management dataset based on unified coding;
[0007] 2) Construct a fault classification network and train it using the equipment health management dataset;
[0008] 3) Construct a multimodal device health management language model, and use the multimodal device health management dataset to perform low-rank adaptive fine-tuning on the weights of the pre-trained language model;
[0009] 4) Using vibration signals and operating parameters, the adjusted language model is used for testing and reasoning to obtain the text response of the equipment health management task.
[0010] Step 1) includes the following steps:
[0011] 1.1) Vibration signals S under various working conditions are collected using an accelerometer. v ∈R 1×∞ ;
[0012] 1.2) Regarding the vibration signal S vUninterrupted sampling is performed, and the sampled data is converted to frequency to obtain the vibration signal X in the frequency domain. v ;
[0013] 1.3) Using natural numbers L v The labels representing vibration signals are used to indicate different operating states of the equipment, where 0 indicates a fault-free state;
[0014] 1.4) For each vibration signal sample, query its working condition. If the same working condition already exists in the database, set the working condition index C of the newly added sample to the working condition index of the queried sample. If not, increment and create a new working condition index value.
[0015] 1.5) Construct health management tasks. For each health management task, set a task text template. Each sample is paired with a task template to obtain a sample-task pair, which is then used with a language model containing a large number of parameters. The modified text is the health management task text X. t ;
[0016] 1.6) Transfer the health management task text X t and vibration signal label L v The text description is input together with the large number of parameters into the language model, and the resulting response is used as the response text reference L. t .
[0017] Step 1.2) specifically refers to:
[0018] Let the sampling rate of the sensor be n. f Then each sample The length of each is n f Using Discrete Cosine Transform to Convert to frequency domain representation The length is still n f ,right The input length and amplitude are normalized to obtain the encoded result X. v .
[0019] The input length normalization specifically involves:
[0020] Let the required input length be n. s When n f <n s Then in Add n to the end s -n f A zero, when for n f ≥n s , then delete The last n f -n sEach value is used to obtain the length-normalized result.
[0021] The normalization of the amplitude specifically refers to:
[0022] calculate The power is calculated, and divided by that power value to make the power 1, then multiplied by a coefficient β to make X... v The amplitude is in the range of [-1, 1].
[0023] The health management tasks include:
[0024] Anomaly detection is used to detect whether a fault has occurred.
[0025] Fault diagnosis is used to detect what kind of fault has occurred.
[0026] Maintenance recommendations are provided to suggest how to handle the malfunction;
[0027] Potential risk analysis is used to analyze the potential problems that may result from a failure.
[0028] The fault classification network includes:
[0029] The main feature extraction part consists of three convolutional layers with a kernel size of 16 arranged in parallel, used to extract the frequency domain samples X of the vibration signal. v Trouble-free reference input and residual X res The main features;
[0030] The encoder section consists of three sequentially connected multi-scale channel attention modules, which are used to further extract fault features from the output of the main feature extraction section.
[0031] The decoder consists of two sequentially connected linear layers, L1 and L2, which are used to map fault features to fault labels.
[0032] The multi-scale channel attention module is specifically:
[0033] Use three convolutional layers with different kernel sizes to extract multi-scale features from the input;
[0034] The channel attention mechanism is used to assign weights to features at different scales and then weighted.
[0035] By utilizing the residual structure, the original input is transformed through a convolutional layer with a kernel size of 1, and then added to the weighted multi-scale features to obtain the final feature output.
[0036] The fault classification network was trained using an equipment health management dataset, specifically as follows:
[0037] Extract vibration signal samples X from the datasetv Operating condition index C and label L v ;
[0038] Using the same operating condition index C and fault-free label L v =0 is used as a filtering condition to extract reference input from the dataset.
[0039] Calculate the residual between the input and the reference.
[0040] X v , X res Both are used as network inputs, and the output fault classification results are related to the label L. v We calculate the cross-entropy as the loss and optimize the model weights through gradient backpropagation.
[0041] The construction of the multimodal device health management language model specifically involves:
[0042] Load the pre-trained weights of a language model with a small number of parameters;
[0043] A vibration signal encoder and an alignment layer are added to the original language model, and a low-rank adapter is added to the language model. The vibration signal encoder has the same structure as the encoder in the fault classification network, and the alignment layer is used to convert the encoding result into text embedding.
[0044] User instructions regarding device health management tasks are converted into text embedding vectors through a word segmenter and embedding layer. These vectors are then concatenated with the text embeddings of vibration signals and used as input to the language model.
[0045] The output of the language model is decoded and used as a response to the device health management task.
[0046] The low-rank adaptive fine-tuning of the pre-trained language model weights specifically involves:
[0047] The encoder weights of the fault classification network are directly used as the encoder of the language model. The linear layers L1 and L2 of the fault classification network are transferred to the alignment layers L1 and L2 of the language model. The text descriptions of all device states are converted into text embedding vectors through the word segmenter and embedding layer of the language model, which are used as the weights of L3 in the alignment layer.
[0048] Using the same input as the fault classification network as the input to the language model, the output of the language model is compared with a high-quality response text reference L. t A comparison was made, and the LoRA parameters and alignment layer parameters were fine-tuned based on the comparison results.
[0049] The present invention has the following beneficial effects and advantages:
[0050] 1. It can be applied to a wide range of working conditions and various equipment health management tasks, and can use language models to achieve natural language interaction, reducing the professional skills required of operation and maintenance personnel;
[0051] 2. By fine-tuning the language model with a small number of parameters, its quality in device health management is comparable to that of models with a large number of parameters, and it has a faster running speed, enabling deployment on terminals with limited computing power;
[0052] 3. By fully utilizing the prior knowledge that equipment will vibrate at a certain frequency after a failure, and by comparing the differences between the input vibration frequency domain information and the fault-free reference signal, the accuracy of health management analysis is greatly improved. Attached Figure Description
[0053] Figure 1 This is a structural diagram of a device health management method based on unified coding of vibration signals.
[0054] Figure 2 This diagram illustrates the fault classification network and how to use the pre-trained weights. Detailed Implementation
[0055] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0056] A device health management method based on unified vibration signal encoding includes the following steps:
[0057] 1. Construct a multimodal equipment health management dataset based on unified coding. This dataset includes the operating condition index (IDC) and the vibration signal DCT coding results (X). v Fault Category Code L v Health Management Task Text X t and response text reference V t This is used for subsequent model training.
[0058] 2. Pre-training the fault classification network weights and initializing the alignment layer weights: First, a fault classification network was constructed, utilizing the working condition index (IDC) and the vibration signal DCT encoding result (X). v Fault Category Code L v The fault classification network is pre-trained, and the alignment layer weights are initialized using a pre-defined fault category description code.
[0059] 3. Construct a multimodal equipment health management language model. Using the constructed multimodal equipment health management dataset, fine-tune the weights of the pre-trained language model using low-rank adaptation (LoRA) to further improve the response quality of the language model under various tasks related to equipment health management.
[0060] Example:
[0061] This invention includes the following steps:
[0062] Step 1: Acquire raw signals. Install an accelerometer on the equipment and acquire vibration signals S under various operating conditions. v ∈R 1×∞ (Because of continuous data collection, the length is considered infinite).
[0063] Step 2: Process the signal sample. For the vibration signal S... v Uninterrupted sampling is performed, with each sample lasting 1 second. Let the sensor's sampling rate be n. f Then each sample The length of each is n f Using Discrete Cosine Transform (DCT) to... Convert to frequency domain representation The length is still n f Next, the input length and amplitude are normalized to obtain the encoded result X. v Normalization is calculated using the following method:
[0064]
[0065] Input length normalization is performed using truncation and zero-padding. Let the required input length be n. s For n f <n s In this situation, Add n to the end s -n f A zero, for n f >n s In the case of deletion The last n f -n s There are several values. Let the result after length normalization be... The normalization of the amplitude is calculated first. The power is calculated, and then divided by that power value to make the power 1. Then, it is multiplied by a coefficient β to make X... v The amplitude is in the range of [-1, 1], and the calculation is performed using the enhanced model. Preferably, β is set to 0.01.
[0066] Step 3: Set sample status labels. The label for the vibration signal is L. v The different operating states of the equipment are represented by natural numbers, where 0 represents a fault-free state.
[0067] Step 4: Set the operating condition index. For each vibration signal sample, query its equipment conditions, sensor model and location, working load, and operating speed. If the same operating condition already exists in the database, set the operating condition index C of the newly added sample to the operating condition index of the queried sample. If not, create a new operating condition index value incrementally.
[0068] Step 5: Define the health management task text. Four types of health management tasks were designed: anomaly detection (detecting whether a fault has occurred), fault diagnosis (detecting what type of fault has occurred), maintenance recommendations (suggesting how to handle the fault), and potential risk analysis (analyzing potential problems after a fault occurs). For each health management task, a task text template was first defined. Next, for each sample-task pairing, a language model with a large number of parameters (e.g., ChatGPT) is used to... After polishing and modification, the health management task text X is obtained. t .
[0069] Step 6: Set the response text reference. (This refers to setting the health management task text X.) t and vibration signal label L v The text descriptions are input together into a language model with a large number of parameters, and the resulting response results are recorded as a high-quality response text reference. t .
[0070] Step 7: Construct a fault classification network. For example... Figure 2 First, three convolutional layers with relatively large kernel sizes are used to process the frequency domain samples X of the vibration signal. v Trouble-free reference input and residual X res The main features are extracted, preferably with a kernel size of 16. Next, fault features are further extracted using three consecutive multi-scale channel attention modules. Finally, the fault classification result is output through two linear layers. In the multi-scale channel attention module, three convolutional layers with different kernel sizes are first used to extract the multi-scale features of the input; preferably, the kernel sizes are set to 3, 5, and 7, respectively. Then, a channel attention mechanism is used to assign weights to the features at different scales, and these weights are summed. Finally, using a residual structure, the original input is transformed by a convolutional layer with a kernel size of 1, and then added to the weighted multi-scale features to obtain the final feature output.
[0071] Step 8: Extract samples to train the fault classification network. First, extract vibration signal samples X from the dataset. v Operating condition index C and label L v Next, using the same operating condition index C and the fault-free label L... v =0 is used as a filtering condition to extract reference input from the dataset. Calculate the residual between the input and the reference. X v , X res Both are used as network inputs, and the output fault classification results are related to the label L. v The cross-entropy is calculated as the loss, and the model weights are optimized through gradient backpropagation.
[0072] Step 9: Construct a language model that can be used for equipment health management tasks. For example... Figure 1 First, pre-trained weights (e.g., Qwen2.5) of a language model with a small number of parameters are loaded. Then, a vibration signal encoder and alignment layer are added to the original model, along with a low-rank adapter (LoRA) for the language model. The vibration signal encoder has the same structure as the encoder in the fault classification network described above, and the alignment layer converts the encoded result into a text embedding. The user's instructions regarding the equipment health management task are converted into text embedding vectors through a tokenizer and embedding layer, concatenated with the text embedding of the vibration signal, and used as input to the language model. The output of the language model is decoded as the response to the equipment health management task.
[0073] The multimodal equipment health management language model is based on an existing small-parameter general language model, but adds a vibration signal encoder, enabling the language model to understand the health status of equipment through vibration signals. This vibration signal encoder has the same structure as the encoder in the aforementioned fault classification network, and uses an alignment layer to convert the encoding result into a text embedding. User instructions regarding equipment health management tasks are converted into text embedding vectors through a tokenizer and embedding layer, and concatenated with the text embedding of the vibration signal as input to the language model. The output of the language model is decoded as the response to the equipment health management task. Using text pairs from the constructed multimodal equipment health management dataset, the small-parameter language model is trained using LoRA. By adjusting the LoRA parameters, the multimodal equipment health management language model can achieve equipment health management task response quality comparable to large-parameter models under conditions of relatively few parameters.
[0074] Step 10: Pre-trained weight transfer. Transfer the weights of the fault classification network from Step 8 to the language model built in Step 9. For example... Figure 2 The encoder weights of the fault classification network will be directly used as the encoder of the language model. The linear layers L1 and L2 of the fault classification network are transferred to the L1 and L2 layers of the language model alignment layer. The L3 weights in the language model alignment layer are obtained using the text descriptions of all device states from step 3. These text descriptions are converted into text embedding vectors by the word segmenter and embedding layer of the language model in step 9, and used as the weights of L3 in the alignment layer.
[0075] Step 11: Fine-tune the language model. Using the same input as in Step 8 as the input to the language model, compare the output with the high-quality response text reference L obtained in Step 6. t By comparing the results, the LoRA parameters and alignment layer parameters were fine-tuned to optimize the output of the device health management language model.
[0076] Step 12: Perform testing and reasoning. In practical applications, the vibration signal and operating parameters are known. First, the vibration signal is processed using the method described in Step 2. Next, the operating parameters are used to obtain the operating condition index, and a fault-free reference signal under the same operating condition is queried in the database. Finally, the processed vibration signal and the reference signal are used as encoder inputs, and the required health management task is input to obtain the equipment health management task text response.
Claims
1. A method for equipment health management based on unified encoding of vibration signals, characterized in that, Includes the following steps: 1) Construct a multimodal device health management dataset based on unified coding; 2) Construct a fault classification network and train it using the equipment health management dataset; 3) Construct a multimodal device health management language model, and use the multimodal device health management dataset to perform low-rank adaptive fine-tuning on the weights of the pre-trained language model; 4) Using vibration signals and operating parameters, the adjusted language model is used for testing and reasoning to obtain the text response of the equipment health management task.
2. The equipment health management method based on unified encoding of vibration signals according to claim 1, characterized in that, Step 1) includes the following steps: 1.1) Vibration signals S under various working conditions are collected using an accelerometer. v ∈R 1×∞ ; 1.2) Regarding the vibration signal S v Uninterrupted sampling is performed, and the sampled data is converted to frequency to obtain the vibration signal X in the frequency domain. v ; 1.3) Using natural numbers L v The labels representing vibration signals are used to indicate different operating states of the equipment, where 0 indicates a fault-free state; 1.4) For each vibration signal sample, query its working condition. If the same working condition already exists in the database, set the working condition index C of the newly added sample to the working condition index of the queried sample. If not, increment and create a new working condition index value. 1.5) Construct health management tasks. For each health management task, set a task text template. Each sample is paired with a task template to obtain a sample-task pair, which is then used with a language model containing a large number of parameters. The modified text is the health management task text X. t ; 1.6) Transfer the health management task text X t and vibration signal label L v The text description is input together with the large number of parameters into the language model, and the resulting response is used as the response text reference L. t .
3. The equipment health management method based on unified vibration signal coding according to claim 2, characterized in that, Step 1.2) specifically refers to: Let the sampling rate of the sensor be n. f Then each sample The length of each is n f Using Discrete Cosine Transform to Convert to frequency domain representation The length is still n f ,right The input length and amplitude are normalized to obtain the encoded result X. v .
4. The equipment health management method based on unified encoding of vibration signals according to claim 3, characterized in that, The input length normalization specifically involves: Let the required input length be n. s When n f <n s Then in Add n to the end s -n f A zero, when for n f ≥n s , then delete The last n f -n s Each value is used to obtain the length-normalized result. The normalization of the amplitude specifically refers to: calculate The power is calculated, and divided by that power value to make the power 1, then multiplied by a coefficient β to make X... v The amplitude is in the range of [-1, 1].
5. The equipment health management method based on unified encoding of vibration signals according to claim 1, characterized in that, The health management tasks include: Anomaly detection is used to detect whether a fault has occurred. Fault diagnosis is used to detect what kind of fault has occurred. Maintenance recommendations are provided to suggest how to handle the malfunction; Potential risk analysis is used to analyze the potential problems that may result from a failure.
6. The equipment health management method based on unified encoding of vibration signals according to claim 1, characterized in that, The fault classification network includes: The main feature extraction part consists of three convolutional layers with a kernel size of 16 arranged in parallel, used to extract the frequency domain samples X of the vibration signal. v Trouble-free reference input and residual X res The main features; The encoder section consists of three sequentially connected multi-scale channel attention modules, which are used to further extract fault features from the output of the main feature extraction section. The decoder consists of two sequentially connected linear layers, L1 and L2, which are used to map fault features to fault labels.
7. The equipment health management method based on unified vibration signal encoding according to claim 6, characterized in that, The multi-scale channel attention module is specifically: Use three convolutional layers with different kernel sizes to extract multi-scale features from the input; We use a channel attention mechanism to assign weights to features at different scales and then perform weighted summation. By utilizing the residual structure, the original input is deformed through a convolutional layer with a kernel size of 1, and then added to the weighted multi-scale features to obtain the final feature output.
8. The equipment health management method based on unified vibration signal coding according to claim 1, characterized in that, The fault classification network was trained using an equipment health management dataset, specifically as follows: Extract vibration signal samples X from the dataset v Operating condition index C and label L v ; Using the same operating condition index C and fault-free label L v =0 is used as a filtering condition to extract reference input from the dataset. Calculate the residual between the input and the reference. X v , X res Both are used as network inputs, and the output fault classification results are related to the label L. v We calculate the cross-entropy as the loss and optimize the model weights through gradient backpropagation.
9. The equipment health management method based on unified vibration signal coding according to claim 1, characterized in that, The construction of the multimodal device health management language model specifically involves: Load the pre-trained weights of a language model with a small number of parameters; A vibration signal encoder and an alignment layer are added to the original language model, and a low-rank adapter is added to the language model. The vibration signal encoder has the same structure as the encoder in the fault classification network, and the alignment layer is used to convert the encoding result into text embedding. User instructions regarding device health management tasks are converted into text embedding vectors through a word segmenter and embedding layer. These vectors are then concatenated with the text embeddings of vibration signals and used as input to the language model. The output of the language model is decoded and used as a response to the device health management task.
10. A method for equipment health management based on unified encoding of vibration signals according to claim 1, characterized in that, The low-rank adaptive fine-tuning of the pre-trained language model weights specifically involves: The encoder weights of the fault classification network are directly used as the encoder of the language model. The linear layers L1 and L2 of the fault classification network are transferred to the alignment layers L1 and L2 of the language model. The text descriptions of all device states are converted into text embedding vectors through the word segmenter and embedding layer of the language model, which are used as the weights of L3 in the alignment layer. Using the same input as the fault classification network as the input to the language model, the output of the language model is compared with a high-quality response text reference L. t A comparison was made, and the LoRA parameters and alignment layer parameters were fine-tuned based on the comparison results.