Large model driven smart factory equipment predictive maintenance method and system

Abstract

The application relates to the technical field of equipment maintenance, in particular to a large model driven smart factory equipment predictive maintenance method and system, which comprises the following steps: acquiring vibration signals of a factory equipment in each monitoring period; extracting significant points of amplitude out-of-limit in the vibration signals; quantifying overall impact intensity of each monitoring period; calculating impact periodicity of each monitoring period, determining a first characteristic value representing a crack fault of the equipment; determining a second characteristic value representing a wear fault of the equipment; extracting a third characteristic value representing stable operation of the equipment; quantifying a fourth characteristic value representing a multi-fault aliasing and concurrency of the equipment; acquiring a feature vector and inputting the feature vector into a classification model to obtain a fault diagnosis result of the equipment; and analyzing and reasoning by using a large language model according to the fault diagnosis result to generate a maintenance processing scheme. The application significantly improves the accuracy and reliability of the predictive maintenance of the smart factory equipment.

Description

Technical Field

[0001] This application relates to the field of equipment maintenance technology, specifically to a method and system for predictive maintenance of smart factory equipment driven by large models. Background Technology

[0002] With the development of intelligent manufacturing, the continuous and stable operation of smart factory equipment is crucial. Currently, the maintenance of factory equipment mostly relies on reactive repairs or scheduled maintenance, which cannot detect early signs of equipment deterioration. This delayed maintenance approach can easily lead to unplanned downtime, causing huge economic losses, or excessive maintenance, increasing operating costs.

[0003] Therefore, the industry currently uses machine learning classification models to predictively diagnose equipment fault conditions. However, because existing methods fail to deeply integrate the specificity of fault mechanisms during feature extraction, they often only extract conventional features of a single dimension. This leads to the overlapping and masking of vibration features caused by different mechanisms when faced with complex faults such as cracks and severe wear. The original regularity of the signal is destroyed. This limitation in feature extraction makes the front-end classification model prone to misjudgment and omission, resulting in a serious deviation between the fault diagnosis input to the large model and the actual fault situation. The maintenance plan generated by the large model based on the deviation information also deviates from the actual needs, failing to provide accurate and reliable maintenance guidance for operation and maintenance personnel, resulting in poor predictive maintenance effects for factory equipment. Summary of the Invention

[0004] To address the aforementioned technical challenges, a method and system for predictive maintenance of smart factory equipment driven by large models are provided.

[0005] The solution to the technical problem presented in this application is to provide a method and system for predictive maintenance of smart factory equipment driven by a large model, comprising the following steps: In a first aspect, embodiments of this application provide a large-model-driven predictive maintenance method for smart factory equipment, the method comprising the following steps: Acquire vibration signals from factory equipment during each monitoring cycle; Extract significant points in the vibration signal where the amplitude exceeds the limit; quantify the overall impact intensity of each monitoring cycle by the deviation level of the amplitude at the significant points; evaluate the periodic distribution of the vibration signal in the time domain, calculate the impact periodicity of each monitoring cycle, and determine the first characteristic value characterizing the equipment crack failure by combining the overall impact intensity. The difference in energy distribution between high and low frequencies of vibration signals in the frequency domain is analyzed to determine the second characteristic value characterizing equipment wear faults; based on the regularity of energy distribution of vibration signals in the frequency domain, the third characteristic value characterizing stable operation of equipment is extracted; the dispersion of amplitude distribution and impact periodicity at significant points are analyzed to quantify the fourth characteristic value characterizing the concurrent occurrence of multiple faults in the equipment. The first, second, third, and fourth feature values ​​are combined into a feature vector and input into the classification model to obtain the fault diagnosis results of the equipment. Based on the fault diagnosis results, the large language model is used for analysis and reasoning to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures.

[0006] Preferably, the process of obtaining the significant points is as follows: for each monitoring cycle, the absolute value of the amplitude of all sampling points in the vibration signal is taken and then the mean is calculated as the average amplitude; sampling points whose absolute amplitude is greater than the average amplitude are selected and defined as significant points.

[0007] Preferably, the calculation process of the overall impact strength is as follows: calculate the difference between the absolute value of the amplitude of each significant point and the average amplitude, as the relative deviation; the overall impact strength is positively correlated with the relative deviation.

[0008] Preferably, the calculation of the impact periodicity for each monitoring cycle includes: After setting the amplitude of all sampling points except for significant points in each monitoring period to 0, the autocorrelation function is calculated on the amplitude of all sampling points in each monitoring period to output a series of autocorrelation coefficients with different lag orders. For each monitoring period, the peaks of the autocorrelation coefficients of all lag orders are extracted, and the mean of the autocorrelation coefficients at all peaks is calculated as the period intensity. The hysteresis order corresponding to the maximum wave peak is selected and defined as the cycle step size. Based on the cycle step size, all sampling points within the monitoring period are continuously segmented to obtain multiple cycle intervals. The maximum amplitude of all sampling points in each cycle interval is selected from the original vibration signal, and the sampling interval between the maximum amplitudes of two adjacent cycle intervals is calculated. The dispersion of the sampling interval between all two adjacent cycle intervals is calculated. The impact periodicity is positively correlated with the periodic intensity, but negatively correlated with the degree of dispersion.

[0009] Preferably, the first characteristic value is positively correlated with both the overall impact strength and the impact periodicity.

[0010] Preferably, the calculation process of the second characteristic value is as follows: frequency domain analysis is performed on the vibration signal of each monitoring period to obtain a spectrum diagram; the frequency component corresponding to the maximum energy in the spectrum diagram is defined as the fundamental frequency component; the sum of the energies corresponding to all frequency components in the spectrum diagram that are less than the fundamental frequency component is calculated as the low-frequency energy; the sum of the energies corresponding to all frequency components in the spectrum diagram that are greater than the fundamental frequency component is calculated as the high-frequency energy; the total energy of all frequency components in the spectrum diagram is calculated; the difference between the low-frequency energy and the high-frequency energy is calculated, and its proportion in the total energy is used as the second characteristic value of each monitoring period.

[0011] Preferably, the calculation process of the third characteristic value is as follows: calculate the permutation entropy of the energy corresponding to all frequency components in the spectrum, perform a negative mapping on it, and use it as the third characteristic value for each monitoring period.

[0012] Preferably, the calculation process of the fourth characteristic value is as follows: calculate the fluctuation degree of the amplitude of all significant points in each monitoring period, and use the ratio of it to the periodicity of the impact as the fourth characteristic value of each monitoring period.

[0013] Preferably, the step of using a large language model for analysis and reasoning to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures includes: pre-building a local maintenance knowledge base; filling fault diagnosis results into a preset prompt word template and inputting it into a locally deployed large language model; the large language model performing semantic retrieval in the local maintenance knowledge base based on the prompt words, and performing contextual fusion of the retrieval results with the diagnostic information in the prompt words to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures.

[0014] Secondly, embodiments of this application also provide a large model-driven predictive maintenance system for smart factory equipment, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above-described large model-driven predictive maintenance methods for smart factory equipment.

[0015] This application has at least the following beneficial effects: This application abandons the conventional "one-size-fits-all" black-box feature extraction method in data-driven approaches. For different physical phenomena such as high-frequency transient impacts from cracks and low-frequency rumbling caused by wear, it designs a first feature value based on the amplitude deviation and periodic distribution at salient points, and a second feature value based on the energy difference distribution between high and low frequencies. This approach deeply integrates the physical mechanisms and specific characterizations of different types of mechanical degradation, effectively avoiding the limitations of conventional single-dimensional feature extraction and significantly improving the classification model's accuracy in identifying the operating states of various individual devices. Addressing the problem of feature masking and disruption of the original signal period due to multiple concurrent faults in actual working conditions, a fourth feature characterizing the concurrent occurrence of multiple faults is extracted by deeply analyzing the correlation between the amplitude distribution discreteness and impact periodicity at salient points. The value effectively quantifies the nonlinear disordered oscillation state caused by compound degradation, solving the problem of missed diagnosis and misdiagnosis caused by feature masking in multi-fault concurrent scenarios. The diagnostic results based on the multi-dimensional specific feature vector output in the time-frequency domain have extremely high accuracy. Using this as the basis for input into the large language model, the hidden danger of "misleading reasoning" in subsequent decision-making caused by the fuzziness of front-end feature extraction is completely eliminated. Combined with the maintenance processing scheme generated by the analysis and reasoning of the large language model, the transformation from discrete state labels to specific cause analysis and executable maintenance actions is realized. This effectively avoids over-maintenance and unplanned downtime, reduces the overall operation and maintenance cost of factory equipment, significantly improves the accuracy and reliability of predictive maintenance of smart factory equipment, and improves the intelligence and automation level of smart factory equipment management. Attached Figure Description

[0016] The following section provides a more detailed explanation of the large-model-driven predictive maintenance method for smart factory equipment in this application, with reference to the accompanying drawings.

[0017] Figure 1 A flowchart illustrating the steps of a large-model-driven predictive maintenance method for smart factory equipment provided in an embodiment of this application; Figure 2 A flowchart illustrating the steps of the method for obtaining the first feature value provided in the embodiments of this application. Detailed Implementation

[0018] The following, in conjunction with the accompanying drawings and embodiments, provides a more detailed description of the large-model-driven predictive maintenance method and system for smart factory equipment proposed in this application.

[0019] Please see Figure 1 The diagram illustrates a flowchart of a large-model-driven predictive maintenance method for smart factory equipment according to an embodiment of this application. The method includes the following steps: Step 1: Acquire vibration signals of factory equipment in each monitoring cycle; Step 2: Extract significant points in the vibration signal where the amplitude exceeds the limit; quantify the overall impact intensity of each monitoring cycle by the deviation level of the amplitude at the significant points; evaluate the periodic distribution of the vibration signal in the time domain, calculate the impact periodicity of each monitoring cycle, and determine the first characteristic value characterizing the equipment crack failure by combining the overall impact intensity. Step 3: Analyze the difference in high and low frequency energy distribution of the vibration signal in the frequency domain to determine the second characteristic value characterizing the wear fault of the equipment; based on the regularity of the energy distribution of the vibration signal in the frequency domain, extract the third characteristic value characterizing the stable operation of the equipment; analyze the dispersion of amplitude distribution and the periodicity of impact at significant points to quantify the fourth characteristic value characterizing the concurrent occurrence of multiple faults in the equipment. Step 4: Combine the first feature value, the second feature value, the third feature value, and the fourth feature value into a feature vector and input it into the classification model to obtain the fault diagnosis results of the equipment. Based on the fault diagnosis results, use the large language model to perform analysis and reasoning to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures.

[0020] Smart factory equipment is susceptible to wear and tear and cracking during operation, and traditional reactive maintenance can lead to production interruptions. Predictive maintenance addresses this by collecting physical signals from equipment operation and extracting their features using algorithms. This approach identifies potential faults and provides maintenance solutions before equipment downtime occurs, preventing unexpected production line shutdowns.

[0021] In step 1, the specific process is as follows: An acceleration vibration sensor is installed on a rigid structure near the vibration source of the equipment, such as a bearing housing or gearbox housing, to collect vibration signals of the factory equipment in real time during each monitoring cycle. The signal acquisition frequency is 20 kHz and the monitoring cycle duration is 1 second. As other implementation methods, the implementer can set the frequency according to the actual situation.

[0022] The collected vibration signal is filtered using a Gaussian filtering algorithm to reduce noise interference during the data acquisition process. The Gaussian filtering algorithm is a well-known technique and will not be described in detail here.

[0023] During long-term operation, high-load operating parts such as bearings and gears inside factory equipment may suffer severe wear or crack damage. Cracks manifest as discontinuities inside or on the surface of the material, causing periodic abrupt changes in stiffness and nonlinear vibration characteristics. Severe wear, on the other hand, occurs when the surface material of the parts is gradually lost during equipment operation, which manifests as increased gear clearance, a slow decrease in contact surface stiffness, and a slow increase in vibration amplitude.

[0024] In step 2, the first feature value is calculated. The flowchart of the method for obtaining the first feature value provided in this embodiment is as follows: Figure 2As shown, the specific process is as follows: Step 201: For each monitoring cycle, take the absolute value of the amplitude of all sampling points in the vibration signal and then calculate the mean value as the average amplitude; select sampling points whose absolute amplitude is greater than the average amplitude and define them as significant points. Step 202: Calculate the difference between the absolute value of the amplitude and the average amplitude at each significant point as the relative deviation; the overall impact intensity of each monitoring cycle is positively correlated with the relative deviation. It should be noted that a positive correlation means that the dependent variable increases as the independent variable increases and decreases as the independent variable decreases.

[0025] Specifically, the calculation process for relative deviation is as follows: calculate the difference between the absolute value of the amplitude of each significant point and the average amplitude, and take the ratio of the absolute value of the amplitude to the average amplitude as the relative deviation; secondly, take the average of the relative deviations of all significant points in each monitoring period as the overall impact intensity. It should be noted that the greater the relative deviation, the greater the amplitude of the significant point is compared to the overall average level, reflecting that the equipment is subjected to a high-intensity transient physical impact at this time; the greater the overall impact intensity, the higher the overall intensity of the abnormal impact within the monitoring period, reflecting the presence of deep cracks on the surface of the equipment components, leading to severe local stiffness abrupt changes during operation.

[0026] Step 203: After setting the amplitude of all sampling points except for significant points in each monitoring period to 0, perform autocorrelation function calculation on the amplitude of all sampling points in each monitoring period to output a series of autocorrelation coefficients with different lag orders; for each monitoring period, extract the peaks of the autocorrelation coefficients of all lag orders and calculate the mean of the autocorrelation coefficients at all peaks as the period intensity. Specifically, the range of values ​​for the lag order is: Where N is the total number of sampling points in each monitoring period, and the autocorrelation function is a well-known technique, which will not be elaborated here; It should be noted that the greater the periodic intensity, the higher the peak value at different peaks, reflecting the high regularity and continuity of the periodic impact of the vibration signal within the monitoring period.

[0027] Step 204: Select the hysteresis order corresponding to the maximum wave peak and define it as the cycle step size; based on the cycle step size, divide all sampling points in the monitoring period into continuous segments to obtain multiple cycle intervals; select the maximum amplitude of all sampling points in each cycle interval from the original vibration signal, and count the sampling interval between the maximum amplitudes of two adjacent cycle intervals; calculate the dispersion of the sampling interval between all two adjacent cycle intervals. Specifically, the process of obtaining the loop interval is as follows: The loop step size is denoted as... Starting from the first sampling point within this monitoring period, with Each sampling point serves as the length of an interval. The monitoring period is sequentially divided into several continuous and non-overlapping intervals. Each of these intervals is defined as a cyclic interval. When the number of remaining sampling points at the end of the monitoring period is less than one cyclic step, the monitoring period continues. In the first step, the remaining sampling points are treated as a separate cyclic interval. The second step involves processing the dispersion as follows: extracting the sampling intervals of all two adjacent cyclic intervals and calculating their sum; dividing each sampling interval by this sum to obtain the probability value corresponding to each sampling interval, thus forming a sampling interval probability distribution sequence; and calculating the information entropy of this sampling interval probability distribution sequence to measure the dispersion of the sampling interval. The calculation of information entropy is a well-known technique and will not be elaborated upon here.

[0028] It should be noted that the smaller the degree of dispersion, the more stable the time interval between the occurrence of the maximum impact in adjacent cycle intervals.

[0029] Step 205: The impact periodicity of each monitoring cycle is positively correlated with the cycle intensity, but negatively correlated with the degree of dispersion. It should be noted that a negative correlation means that the dependent variable decreases as the independent variable increases, and increases as the independent variable decreases.

[0030] Specifically, the ratio of periodic intensity to dispersion is taken as the periodicity of the impact. When calculating the ratio, in order to avoid the denominator being 0, a parameter adjustment factor is added to the denominator. In this embodiment, the parameter adjustment factor is set to 0.01. As for other implementation methods, the implementer can set it according to the actual situation.

[0031] It should be noted that the greater the periodicity of the impact, the stronger the overall periodicity of the signal, and the more fixed the physical intervals at which the micro-impacts occur.

[0032] Step 206: The first characteristic value of each monitoring cycle is positively correlated with the overall impact intensity and the impact periodicity; Specifically, the product of the overall impact intensity and the impact periodicity is used as the first characteristic value; It should be noted that the larger the first characteristic value, the more likely the vibration signal of the monitoring period has both a violent impact amplitude that far exceeds the normal level and maintains a very high time regularity, reflecting a higher possibility of serious cracks or faults inside the equipment, resulting in a highly periodic and violently abrupt change in the vibration signal.

[0033] Furthermore, when the equipment is in normal operation, the unit has good rigidity and the rotor is under balanced force. The vibration signal is mainly dominated by a single and stable main operating frequency of the equipment. When equipment parts experience severe wear, the gap between contact surfaces increases and the rigidity decreases. Under the action of gravity and inertial forces, loose parts will produce non-periodic collisions. This irregular movement breaks the original mechanical balance, mainly exciting the low-frequency resonant mode of the equipment structure, resulting in a severe distortion of the signal energy distribution, causing the signal energy to tend towards the low-frequency part. Therefore, in step 3, the specific process is as follows: Step 301: Perform frequency domain analysis on the vibration signals of each monitoring period to obtain a spectrum diagram; define the frequency component corresponding to the maximum energy in the spectrum diagram as the fundamental frequency component; calculate the sum of the energies corresponding to all frequency components in the spectrum diagram that are less than the fundamental frequency component as low-frequency energy; calculate the sum of the energies corresponding to all frequency components in the spectrum diagram that are greater than the fundamental frequency component as high-frequency energy; calculate the total energy of all frequency components in the spectrum diagram; calculate the difference between low-frequency energy and high-frequency energy, and use its proportion in the total energy as the second characteristic value of each monitoring period. Specifically, frequency domain analysis is performed using Fast Fourier Transform (FFT) to obtain the spectrum. FFT is a well-known technique and will not be elaborated here. Secondly, the calculation process of the second eigenvalue is as follows: the difference between low-frequency energy and high-frequency energy is calculated, and the ratio of this difference to the total energy is used as the second eigenvalue.

[0034] It should be noted that the larger the second eigenvalue, the more abnormal low-frequency vibration energy the equipment exhibits, reflecting that there may be severe excessive wear inside the equipment, causing non-periodic collisions due to loose parts, which in turn excites low-frequency resonant modes of the structure in the vibration signal.

[0035] Secondly, when the equipment is in normal operation, the unit's stiffness and rotor mechanical balance are well maintained. The energy of the vibration signal is highly concentrated on the equipment's main fundamental frequency, while other frequency bands remain at extremely low levels. At this time, the spectral energy distribution sequence is highly predictable and has very low disorder. When the equipment suffers early damage or deterioration due to faults, the accompanying impact excitation, high-frequency friction, or low-frequency loosening will excite a large number of broadband stray components. This will cause the originally highly concentrated energy to spread to other frequency bands, producing irregular fluctuations in the spectrum.

[0036] Step 302: Calculate the permutation entropy of the energy corresponding to all frequency components in the spectrum, perform a negative mapping on it, and use it as the third characteristic value of each monitoring period; The negative mapping process is as follows: the reciprocal of the permutation entropy is used as the third feature value. During the calculation of permutation entropy, since the number of effective data points in the spectrum is relatively limited, an excessively high embedding dimension will lead to a low theoretical probability of each permutation pattern, resulting in a sparse statistical distribution and a significant decrease in the reliability of entropy estimation. Simultaneously, there is a strong local correlation between adjacent frequency components in the spectrum. Setting the embedding dimension to 3 is sufficient to effectively capture the relative magnitude ordering pattern of adjacent frequency point amplitudes, balancing pattern representation ability and statistical reliability. Secondly, the order relationship between adjacent frequency components in the spectrum directly reflects the local fluctuation characteristics of the spectrum. Setting the time delay parameter to 1 can retain the finest local ordering information without causing pattern distortion due to frequency hopping. As other implementation methods, implementers can set this according to their actual situation. Permutation entropy is a well-known technique and will not be elaborated upon here. When calculating the reciprocal, to avoid the denominator being 0, a parameter tuning factor is added to the denominator. In this embodiment, the parameter tuning factor is set to 0.01. As other implementation methods, implementers can set this according to their actual situation.

[0037] It should be noted that the smaller the permutation entropy, the larger the third characteristic value, indicating that the energy distribution in the spectrum is more regular, and reflecting that the equipment is more likely to be in a healthy operating phase with excellent condition.

[0038] Furthermore, when parts such as bearings and gears simultaneously exhibit cracks and severe wear, the originally regular mechanical impact environment is completely disrupted. The random collisions and frictions caused by wear will superimpose high-intensity irregular oscillations throughout the entire time domain. This not only leads to a sharp increase in the fluctuation of amplitude at significant points, but also interferes with and disrupts the original temporal regularity of cracks, resulting in a significant decrease in impact periodicity.

[0039] Step 303: Calculate the fluctuation degree of the amplitude of all significant points in each monitoring period, and use the ratio of it to the periodicity of the impact as the fourth characteristic value of each monitoring period. Specifically, the process for handling the degree of fluctuation is as follows: the amplitudes of all significant points within each monitoring period are statistically analyzed, and their sum is calculated; the amplitude of each significant point is divided by the sum to obtain the probability value corresponding to each amplitude, and an amplitude probability distribution sequence is constructed; the degree of fluctuation is measured by calculating the Shannon entropy of the amplitude probability distribution sequence, where Shannon entropy is a well-known technique and will not be elaborated here.

[0040] It should be noted that the larger the fourth characteristic value, the more unstable the amplitude of the significant point is, and the impact time is completely irregular. This reflects that there may be serious crack defects and wear and loosening of the contact surface in the equipment at the same time. As a result, the regular crack impact is superimposed and submerged by a large number of random impacts. The mechanical performance of the equipment has been severely degraded and faces the risk of collapse at any time.

[0041] Furthermore, based on the first feature value, the second feature value, the third feature value, and the fourth feature value, the fault state of the equipment is identified. In step 4, the specific process is as follows: Step 401: For each monitoring period, the first feature value, the second feature value, the third feature value, and the fourth feature value are combined to form a feature vector; It should be noted that the feature vector jointly characterizes the equipment's operating status from multiple dimensions. By integrating multi-dimensional features representing single faults (cracks, wear), normal health, and severe concurrent damage into the same vector, it can avoid the risk of single features failing under complex operating conditions. This provides a complete and complementary fault identification basis for subsequent models, ensuring an accurate characterization of the equipment's current real operating status.

[0042] Step 402: Input the feature vector into the classification model and output the fault diagnosis result of the device; The classification model uses a support vector machine (SVM) model, and the training process for this model is as follows: Obtain a priori fault dataset of the same model of equipment from the historical database. The priori fault dataset contains multiple historical vibration signals and fault labels known from historical maintenance records or expert experience and manually annotated. The fault labels include normal operating status, crack fault, wear fault, and fault with both wear and crack. For each historical vibration signal in the prior fault dataset, extract its corresponding feature vector according to the above process; The maximum and minimum values ​​of all feature vectors in the prior fault dataset in each dimension are selected. The maximum and minimum value normalization method is used to normalize the same dimension of all feature vectors in the prior fault dataset to obtain the normalized feature vectors. Combined with their fault labels, the support vector machine model is trained. The support vector machine model uses the RBF kernel function and the hinge loss function. The number of iterations is set to 100. The prior state dataset includes four fault categories: normal operation, crack fault, wear fault, and wear and crack coexistence fault. Each fault category has no less than 200 samples. To address the problem of class imbalance caused by the relative scarcity of composite fault samples in actual industrial scenarios, a synthetic minority class oversampling technique is used to amplify the samples of composite fault categories in the feature space. After the number of the four fault categories is balanced, model training is carried out. Then, by using the maximum and minimum values ​​of all feature vectors in each dimension in the prior fault dataset, the current feature vector is normalized. The normalized feature vector is then input into the classification model to output the fault diagnosis result of the device. The training of the support vector machine model and the maximum / minimum normalization method are well-known techniques and will not be elaborated here.

[0043] Furthermore, based on the output fault diagnosis results, and combined with the large model, specific maintenance solutions are output, as follows: Step 403: Construct a local maintenance knowledge base; fill the output fault diagnosis results into the preset prompt word template, input the filled prompt word template into the locally deployed large language model, the large language model performs semantic retrieval in the local maintenance knowledge base according to the prompt words, extracts historical maintenance data related to the fault diagnosis results, the large language model performs contextual fusion of the retrieved historical maintenance data and the diagnostic information in the prompt words, generates a maintenance processing plan that includes fault cause analysis and suggested maintenance measures, and outputs feedback; The process of building the local maintenance knowledge base is as follows: 1) Obtain the maintenance logs and equipment manuals of the factory equipment; 2) Clean and deduplicate the maintenance logs to extract text records containing fault phenomena, root causes, and solutions; 3) Decompose the equipment manuals into structured text containing titles, content, and applicable equipment models by chapter; 4) Convert the extracted text records and structured text into text vectors using the Embedding model and store them in a local vector database as the local maintenance knowledge base; 5) The Embedding model is a well-known technology and will not be elaborated upon here.

[0044] Secondly, the preset prompt template includes a defined text frame and variable slots. For example, the template is "The current device has been diagnosed with [fault diagnosis result]. Please provide the corresponding anomaly analysis and handling solution." It should be noted that the large language model is a well-known technology and will not be elaborated here. In this embodiment, the large language model adopts the open-source model Qwen2.5-7B-Instruct or LLaMA-3-8B-Instruct.

[0045] Based on the same inventive concept as the above methods, embodiments of this application also provide a large model-driven predictive maintenance system for smart factory equipment, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described large model-driven predictive maintenance methods for smart factory equipment.

Claims

1. A large-model-driven predictive maintenance method for smart factory equipment, characterized in that, The method includes the following steps: Acquire vibration signals from factory equipment during each monitoring cycle; Extract significant points in the vibration signal where the amplitude exceeds the limit; quantify the overall impact intensity of each monitoring cycle by the deviation level of the amplitude at the significant points; evaluate the periodic distribution of the vibration signal in the time domain, calculate the impact periodicity of each monitoring cycle, and determine the first characteristic value characterizing the equipment crack failure by combining the overall impact intensity. The difference in energy distribution between high and low frequencies of vibration signals in the frequency domain is analyzed to determine the second characteristic value characterizing equipment wear faults; based on the regularity of energy distribution of vibration signals in the frequency domain, the third characteristic value characterizing stable operation of equipment is extracted; the dispersion of amplitude distribution and impact periodicity at significant points are analyzed to quantify the fourth characteristic value characterizing the concurrent occurrence of multiple faults in the equipment. The first, second, third, and fourth feature values ​​are combined into a feature vector and input into the classification model to obtain the fault diagnosis results of the equipment. Based on the fault diagnosis results, the large language model is used for analysis and reasoning to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures.

2. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 1, characterized in that, The process of obtaining the significant points is as follows: for each monitoring period, the absolute value of the amplitude of all sampling points in the vibration signal is taken and then the mean is calculated as the average amplitude; sampling points whose absolute amplitude is greater than the average amplitude are selected and defined as significant points.

3. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 2, characterized in that, The calculation process for the overall impact strength is as follows: calculate the difference between the absolute value of the amplitude at each significant point and the average amplitude, which is taken as the relative deviation; the overall impact strength is positively correlated with the relative deviation.

4. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 1, characterized in that, The calculation of the impact periodicity for each monitoring cycle includes: After setting the amplitude of all sampling points except for significant points in each monitoring period to 0, the autocorrelation function is calculated on the amplitude of all sampling points in each monitoring period to output a series of autocorrelation coefficients with different lag orders. For each monitoring period, the peaks of the autocorrelation coefficients of all lag orders are extracted, and the mean of the autocorrelation coefficients at all peaks is calculated as the period intensity. The hysteresis order corresponding to the maximum wave peak is selected and defined as the cycle step size. Based on the cycle step size, all sampling points within the monitoring period are continuously segmented to obtain multiple cycle intervals. The maximum amplitude of all sampling points in each cycle interval is selected from the original vibration signal, and the sampling interval between the maximum amplitudes of two adjacent cycle intervals is calculated. The dispersion of the sampling interval between all two adjacent cycle intervals is calculated. The impact periodicity is positively correlated with the periodic intensity, but negatively correlated with the degree of dispersion.

5. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 1, characterized in that, The first eigenvalue is positively correlated with both the overall impact intensity and the impact periodicity.

6. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 1, characterized in that, The calculation process for the second characteristic value is as follows: perform frequency domain analysis on the vibration signal of each monitoring period to obtain the spectrum; define the frequency component corresponding to the maximum energy in the spectrum as the fundamental frequency component; and calculate the sum of the energies corresponding to all frequency components in the spectrum that are less than the fundamental frequency component as the low-frequency energy. The sum of the energies of all frequency components greater than the fundamental frequency in the statistical spectrum is taken as the high-frequency energy; the total energy of all frequency components in the statistical spectrum is also taken as the high-frequency energy. The difference between low-frequency energy and high-frequency energy is calculated, and its proportion in the total energy is used as the second characteristic value for each monitoring cycle.

7. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 6, characterized in that, The calculation process of the third characteristic value is as follows: calculate the permutation entropy of the energy corresponding to all frequency components in the spectrum, perform a negative mapping on it, and use it as the third characteristic value for each monitoring period.

8. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 1, characterized in that, The calculation process for the fourth characteristic value is as follows: calculate the fluctuation degree of the amplitude of all significant points in each monitoring period, and use the ratio of this fluctuation to the periodicity of the impact as the fourth characteristic value for each monitoring period.

9. The large-model-driven predictive maintenance method for smart factory equipment as described in claim 1, characterized in that, The process of using a large language model for analysis and reasoning to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures includes: pre-building a local maintenance knowledge base; filling fault diagnosis results into a preset prompt word template and inputting it into a locally deployed large language model; the large language model performing semantic retrieval in the local maintenance knowledge base based on the prompt words, and contextually fusing the retrieval results with the diagnostic information in the prompt words to generate a maintenance solution that includes fault cause analysis and suggested maintenance measures.

10. A large-model-driven predictive maintenance system for smart factory equipment, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the large model-driven predictive maintenance method for smart factory equipment as described in any one of claims 1-9.

Images