A deep learning-based intelligent diagnosis and state monitoring method for equipment failure

Abstract

The application discloses a kind of based on deep learning equipment fault intelligent diagnosis and state monitoring method, specifically related to deep learning technical field, edge node gathers time-stamped equipment operation energy consumption data transmission to cloud, generates associated data set and constructs normal evolution trajectory, confidence interval and LSTM-BP prediction model, calculates trajectory deviation and energy consumption deviation, when abnormal, start causal tracing and locate fault, adopt dynamic weight particle swarm optimization to generate operation strategy, linkage industrial energy scheduling system executes and generates adjustment scheme and inspection work order.The application adopts trajectory deviation and energy consumption deviation double abnormal detection mechanism, captures early performance degradation signal by calculating the cumulative deviation of real-time data and normal evolution trajectory, simultaneously constructs theoretical energy consumption benchmark in combination with energy consumption dynamic correlation coefficient, reduces the false positive rate and false negative rate of traditional single threshold detection method.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for intelligent diagnosis and status monitoring of equipment faults based on deep learning. Background Technology

[0002] The field of deep learning technology encompasses research on neural network algorithms and their practical application in industrial scenarios. In intelligent equipment fault diagnosis and condition monitoring, it utilizes a data-driven approach to analyze and process the massive amounts of data generated during the operation of industrial equipment. The core content of this technology covers industrial data acquisition and transmission, equipment operation feature extraction, abnormal state identification, fault type determination, and operation trend prediction. It is widely used in equipment operation and maintenance management across multiple industries, including industrial energy, heavy machinery, and power grids. It serves as the core technological support for achieving intelligent equipment inspection and smart operation and maintenance, providing a technological foundation for the full lifecycle management of industrial systems.

[0003] One such method is a deep learning-based intelligent equipment fault diagnosis and condition monitoring approach, which addresses the needs for condition monitoring and fault diagnosis during the operation of industrial equipment. This method encompasses the acquisition and transmission of industrial equipment operating data, the extraction and analysis of equipment operating characteristics, the identification and judgment of abnormal equipment states, the classification and location of equipment fault types, and the prediction of equipment operating trends. It employs a model integrating Long Short-Term Memory (LSTM) networks and backpropagation neural networks to learn and analyze equipment operating data. A dynamic weighted particle swarm optimization algorithm is introduced to adjust model parameters. The method relies on an edge-cloud collaborative architecture for edge preprocessing and cloud-based deep analysis of data, supporting data access via multiple industrial protocols such as Modbus and OPCUA.

[0004] Existing technologies fail to perform in-depth correlation analysis between energy consumption data and other operating status parameters during equipment operation data processing. This results in insufficient data alignment accuracy, making it difficult to accurately reflect the intrinsic relationship between equipment operating status and energy consumption changes. Anomaly identification relies solely on single feature judgments, failing to capture potential equipment anomalies in advance. Fault location lacks a causal tracing mechanism, making it difficult to quickly pinpoint the root cause of the fault. Operational strategy generation does not fully integrate energy consumption and status prediction information, leading to low energy utilization efficiency. Maintenance and inspection lack specificity, easily resulting in over-inspection or missed inspections, increasing equipment maintenance costs and affecting the long-term stable operation of industrial systems. Summary of the Invention

[0005] The main objective of this invention is to provide a method for intelligent diagnosis and status monitoring of equipment faults based on deep learning, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based intelligent equipment fault diagnosis and status monitoring method includes the following steps: Edge acquisition nodes on the equipment side collect equipment operating parameters, generate raw equipment operating energy consumption data with unique timestamps, and transmit them to the cloud; the cloud sorts and aligns the received data according to equipment number and timestamp, generating an equipment operating energy consumption associated dataset; based on the associated dataset, it constructs the normal evolution trajectory and corresponding confidence interval of each operating status parameter, and simultaneously constructs an LSTM-BP fusion operating status prediction model; it extracts real-time data points of operating status parameters from the real-time received data, calculates the vertical distance between these points and the normal evolution trajectory as the trajectory deviation, sums the trajectory deviations over consecutive time moments to generate a cumulative deviation, and generates comprehensive potential anomaly features; it calculates the dynamic correlation coefficient between each operating status parameter and energy consumption, generating theoretical energy consumption benchmark values ​​and energy consumption deviation values; when the energy consumption deviation value exceeds the initial tolerance threshold, it initiates a causal tracing process to locate the faulty component; it uses a dynamic weighted particle swarm optimization algorithm combined with the prediction model output to generate an energy system operating strategy; it combines comprehensive potential anomaly features, fault information, and prediction results to generate operating parameter adjustment schemes and inspection work orders, and links the industrial energy dispatch system to execute the operating strategy.

[0007] Preferably, the edge acquisition nodes on the device side collect all measurable operating status parameters of the device and the instantaneous energy consumption parameters at the corresponding time intervals to form the raw energy consumption data of the device operation. The integrity of each set of raw energy consumption data of the device operation is verified, the data that fails the verification is discarded, and the data that passes the verification is classified and stored according to the parameter type and then synchronously transmitted to the cloud.

[0008] Preferably, all data within the continuous fault-free operation period of the equipment are extracted from the equipment operation energy consumption association dataset, and a normal operation time series is generated by splitting each operation status parameter separately. The time series is divided into continuous data segments of fixed length, and trend fitting is performed on each data segment to generate a local evolution trajectory. All local evolution trajectories are connected to generate a complete normal evolution trajectory, and the dispersion of each data point on the trajectory is calculated to generate the initial confidence interval of the corresponding position.

[0009] Preferably, the energy consumption-related dataset of equipment operation is divided into a training set and a validation set according to a preset ratio. The training set is used to train an LSTM network to extract long-term temporal features of the equipment operation status. The extracted long-term temporal features are input into a BP network for regression prediction. The validation set is used to adjust the model parameters and generate the final LSTM-BP fusion operation status prediction model. The model outputs the operation status parameters and energy consumption trend data of the equipment within a fixed period of time in the future.

[0010] Preferably, the real-time data points and corresponding timestamps of each operating status parameter are extracted from the raw energy consumption data of the device operation received in real time. Each real-time data point is projected onto the corresponding complete normal evolution trajectory to determine the position of the projection point. The vertical distance between the real-time data point and the projection point is calculated as the trajectory deviation at that moment. The trajectory deviations of the most recent fixed number of consecutive moments are stored and summed to generate the cumulative deviation. The cumulative deviation is compared with the upper limit of the initial confidence interval of the corresponding position.

[0011] Preferably, when the cumulative deviation is greater than the upper limit of the initial confidence interval, the corresponding parameter is marked as having a potential anomaly. The trajectory deviation data of the parameter at the most recent fixed number of consecutive moments are extracted, the duration, frequency of occurrence and maximum deviation value of the deviation are calculated, and the comprehensive potential anomaly features of the parameter are generated and stored in the potential anomaly feature set.

[0012] Preferably, all operating status parameters and corresponding instantaneous energy consumption data within the same time window are extracted. The correlation between each operating status parameter and instantaneous energy consumption within the time window is calculated to generate an initial correlation coefficient. The initial correlation coefficient is then weighted and averaged with the initial correlation coefficients of a previous fixed number of time windows to generate a dynamic correlation coefficient. The current values ​​of all operating status parameters are multiplied by their corresponding dynamic correlation coefficients and then summed to generate the theoretical energy consumption benchmark value for the time window. The difference between the actual instantaneous energy consumption and the theoretical energy consumption benchmark value within the current time window is calculated to obtain the energy consumption deviation value. The energy consumption deviation value is then compared with a preset initial tolerance threshold.

[0013] Preferably, when initiating the causal tracing process, all operational status parameter data within a fixed time window before and after the energy consumption deviation occurs are extracted. The difference between the average values ​​of each parameter within the two time windows is calculated to obtain the change amplitude. The change amplitude is multiplied by the corresponding dynamic correlation coefficient to obtain the contribution value. All contribution values ​​are sorted from largest to smallest, and the top three parameters are selected. The current values ​​of the three parameters are compared with the confidence intervals of the corresponding positions on their respective normal evolution trajectories.

[0014] Preferably, with the dual objectives of minimizing total system energy consumption and optimizing equipment health status, the future operating status parameters and energy consumption trend data output by the LSTM-BP fusion operating status prediction model are used as constraints to initialize the position and velocity of the particle swarm, dynamically adjust the inertial weight of each particle, iteratively calculate to obtain the optimal solution, and generate an energy system operation strategy that includes the allocation of operating loads and parameter settings for each device.

[0015] Preferably, a comprehensive potential anomaly feature is extracted from the potential anomaly feature set, all unrelated fault information is extracted, and a list of equipment performance degradation priorities is generated by sorting the cumulative deviation from large to small. The future state trend output by the prediction model is combined to generate an equipment operating parameter adjustment plan and send it to the corresponding equipment edge acquisition node for execution. The cumulative deviation and energy consumption deviation value before and after adjustment are compared in the cloud. When the two values ​​are still higher than the corresponding threshold after adjustment, the equipment performance degradation priority is updated, and an inspection work order containing the equipment location, abnormal parameters and suggested inspection content is generated and sent to the operation and maintenance personnel.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention improves the accuracy and timeliness of equipment fault diagnosis, realizing a shift from passive maintenance to proactive early warning. It adopts a dual anomaly detection mechanism of trajectory deviation and energy consumption deviation, capturing early performance degradation signals of equipment by calculating the cumulative deviation between real-time data and the normal evolution trajectory. At the same time, it combines the dynamic correlation coefficient of energy consumption to construct a theoretical energy consumption benchmark, reducing the false alarm rate and false negative rate of traditional single threshold detection methods. Based on this, the causal tracing process designed can quantify the contribution of each parameter to energy consumption deviation by multiplying the parameter change amplitude and the correlation coefficient, reducing the risk of unplanned equipment downtime.

[0017] This invention achieves synergistic optimization of equipment health status and energy consumption, reducing total system energy consumption while ensuring safe equipment operation. The constructed LSTM-BP fusion operation status prediction model can predict the equipment's operation status and energy consumption trend for the next 24 hours. Combined with the dynamic weighted particle swarm optimization algorithm, an energy system operation strategy is generated with the dual objectives of minimizing total system energy consumption and optimizing equipment health status. This achieves dynamic and reasonable allocation of equipment load, extends equipment lifespan, and reduces the operating cost of the equipment throughout its entire life cycle.

[0018] This invention constructs a complete intelligent closed-loop management system for operation and maintenance, improving the efficiency of operation and maintenance management of industrial equipment. It can automatically generate equipment operating parameter adjustment plans and distribute them to edge nodes for execution. For equipment that still has abnormalities after adjustment, it automatically generates inspection work orders containing location, abnormal parameters, and suggested inspection contents and pushes them to operation and maintenance personnel. It realizes full-process automation of anomaly detection, parameter adjustment, fault location, work order dispatch, and result feedback. This system reduces the workload of manual inspection, avoids subjective errors caused by human experience judgment, realizes data-driven operation and maintenance, and provides support for the intelligent management of industrial equipment. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a flowchart of the data acquisition and preprocessing process of the present invention; Figure 3 This is a flowchart of the anomaly diagnosis and fault tracing process of the present invention; Figure 4 This is a flowchart illustrating the strategy generation and operation / maintenance execution process of this invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Please see Figures 1 to 4 This invention discloses a method for intelligent diagnosis and status monitoring of equipment faults based on deep learning. The following section further discloses the specific implementation process of this invention using a group of 10 200kW stamping motors in the stamping workshop of an automotive parts manufacturing plant as the application object.

[0022] Example 1: This example mainly aims to align edge acquisition, preprocessing, and cloud transmission of equipment operation data, providing high-quality basic data support for subsequent data analysis and fault diagnosis. Specifically, it includes: edge acquisition nodes on the device side collecting equipment operation parameters, generating raw equipment operation energy consumption data with unique timestamps, and transmitting it to the cloud; the cloud sorts and aligns the received data according to device number and timestamp, generating a related dataset of equipment operation energy consumption; edge acquisition nodes on the device side collecting all measurable operating status parameters of the device and the corresponding instantaneous energy consumption parameters at fixed time intervals to form raw equipment operation energy consumption data; performing integrity verification on each set of raw equipment operation energy consumption data, discarding data that fails verification, and storing the data that passes verification according to parameter type before synchronously transmitting it to the cloud.

[0023] In the specific implementation process, a total of 10 edge acquisition nodes were deployed in the stamping workshop, with each node corresponding to one stamping motor, numbered sequentially from CM001 to CM010. Each edge acquisition node has a built-in high-precision current transformer, voltage transformer, temperature sensor, vibration sensor, and speed sensor, which can simultaneously collect eight operating status parameters of the motor: three-phase current, three-phase voltage, stator temperature, bearing temperature, radial vibration, axial vibration, and output speed, as well as the corresponding instantaneous energy consumption parameters. The acquisition time interval of the edge acquisition nodes is set to 100ms, that is, all operating status parameters and instantaneous energy consumption parameters are collected once every 100ms. After each acquisition, the edge acquisition node immediately performs an integrity check on the set of data. The check rule is that all eight operating status parameters and instantaneous energy consumption parameters are not empty values ​​and are within the preset reasonable range. The reasonable ranges for three-phase current are 0A to 500A, three-phase voltage is 380V±10% (342V to 418V), stator temperature is -40℃ to 150℃, bearing temperature is -40℃ to 120℃, radial and axial vibration are 0mm / s to 20mm / s, output speed is 0r / min to 1500r / min, and instantaneous energy consumption is 0kW to 250kW. Data that fails verification is discarded by the edge acquisition node without storage or transmission. Data that passes verification is categorized into six types based on current, voltage, temperature, vibration, speed, and energy consumption parameters, and stored in six local circular buffers, each with a capacity of 10,000 data records. Meanwhile, the edge acquisition nodes add a unique timestamp to the verified data, accurate to the millisecond, in the format "YYYY-MM-DDHH:MM:SS.sss". This timestamp is then synchronously transmitted to the cloud server via a 5G industrial private network, with a communication latency of no more than 20ms, ensuring real-time and reliable data transmission. The cloud server adopts a distributed architecture, configured with 12 compute nodes with 8 cores and 16GB of memory each, and 6 storage nodes with 2TB of capacity, capable of processing operational data from over 100 devices simultaneously. After receiving the raw data from all devices, the cloud server first groups the data by device number, and then sorts each group by timestamp in ascending order. For multiple data entries from the same device at the same timestamp, the average value is taken as the valid data for that moment. Finally, a device operational energy consumption correlation dataset is generated, containing the device number, timestamp, 8 operational status parameters, and instantaneous energy consumption.In the specific implementation process, the amount of data collected by a single edge acquisition node is about 864,000 per day. After integrity verification, the effective data rate remains above 99.5%. The discarded data is mainly null data caused by instantaneous sensor failure. The cloud server generates a complete device operation energy consumption related dataset every 10 minutes. Each dataset contains all the operation data of 10 devices in the past 10 minutes, totaling 6,000 data records.

[0024] Example 2, building upon the equipment operation energy consumption correlation dataset generated in Example 1, further constructs a normal operation benchmark model, including the normal evolution trajectory of each operating state parameter, the corresponding confidence interval, and an LSTM-BP fusion operating state prediction model. This provides a benchmark reference and model support for subsequent anomaly detection and state prediction. Specifically, it includes constructing the normal evolution trajectory and corresponding confidence interval of each operating state parameter based on the correlation dataset, and simultaneously constructing an LSTM-BP fusion operating state prediction model; extracting all data from the equipment operation energy consumption correlation dataset during the continuous fault-free operation period, and separately generating a normal operation time sequence for each operating state parameter. The system divides the time series into continuous data segments of fixed length. For each data segment, trend fitting is performed to generate a local evolution trajectory. All local evolution trajectories are connected to generate a complete normal evolution trajectory. The dispersion of each data point on the trajectory is calculated to generate the initial confidence interval for the corresponding position. The energy consumption-related dataset of the equipment operation is divided into training and validation sets according to a preset ratio. The training set is used to train an LSTM network to extract long-term time series features of the equipment operation status. The extracted long-term time series features are input into a BP network for regression prediction. The validation set is used to adjust the model parameters and generate the final LSTM-BP fusion operation status prediction model. The model outputs the operation status parameters and energy consumption trend data of the equipment within a fixed future time period.

[0025] In the specific implementation process, firstly, all data of stamping motor CM001 from January 1, 2026 to January 31, 2026 were extracted from the equipment operation energy consumption correlation dataset. During this period, motor CM001 operated continuously without faults, and a total of 2,678,400 valid data records were collected. Then, these data were separately split according to eight operating status parameters, generating eight independent normal operation time series, each containing 2,678,400 data points. Next, each time series was divided into continuous data segments of length 3600, meaning each data segment corresponds to one hour of operating data, resulting in 744 data segments for each time series. For each data segment, a cubic polynomial trend fitting was performed using the least squares method to generate the local evolution trajectory of that data segment. The general form of the cubic polynomial is y = ax³ + bx² + cx + d, where x is the time point and y is the value of the operating status parameter. The local evolution trajectories of all data segments were connected in chronological order to generate the complete normal evolution trajectory of that operating status parameter. Finally, the standard deviation of each data point in the original time series and its corresponding value on the normal evolution trajectory is calculated. Three times the standard deviation is used as the upper limit of the initial confidence interval for that position, i.e., the confidence interval is [trajectory value - 3σ, trajectory value + 3σ]. In the specific implementation, the normal evolution trajectory value of the stator temperature parameter of motor CM001 at 10:00 on January 15, 2026 is 75.2℃, and the standard deviation at this position is 1.2℃. Therefore, the initial confidence interval is [71.6℃, 78.8℃]. Simultaneously, the energy consumption correlation dataset of the CM001 stamping motor from January 1, 2026 to January 28, 2026 is divided into training and validation sets in an 8:2 ratio. The training set contains 2,142,720 data records, and the validation set contains 535,680 data records. The LSTM network is set to a 2-layer structure, with each layer containing 64 hidden units. The ReLU activation function is used, and the dropout rate is set to 0.2 to prevent overfitting. The BP network is configured with a 3-layer structure. The input layer contains 64 neurons, corresponding to the long-term temporal features output by the LSTM network; the hidden layer contains 32 neurons, using the ReLU activation function; and the output layer contains 9 neurons, corresponding to 8 operational status parameters and instantaneous energy consumption parameters. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 128, and 100 training epochs. After each training epoch, the model is validated using a validation set, and the mean squared error (MSE) is calculated. Training is terminated early when the MSE of the validation set no longer decreases for 10 consecutive epochs. The final LSTM-BP fusion operational status prediction model can output the operational status parameters and energy consumption trend data of the device for the next 24 hours, with a prediction interval of 10 minutes.In practical implementation, the model's prediction mean square error for stator temperature is 0.85℃², and its prediction mean square error for instantaneous energy consumption is 2.3kW², with prediction accuracy meeting the requirements of industrial applications.

[0026] Example 3, building upon the normal evolution trajectory and confidence interval established in Example 2, further implements potential anomaly detection based on trajectory deviation, generating comprehensive potential anomaly features to provide quantitative basis for subsequent fault warning and performance degradation assessment. Specifically, this includes extracting real-time data points of operating status parameters from real-time received data, calculating their vertical distance from the normal evolution trajectory as the trajectory deviation, summing the trajectory deviations over consecutive time points to generate a cumulative deviation, and generating comprehensive potential anomaly features; extracting real-time data points and corresponding timestamps for each operating status parameter from the raw energy consumption data of the equipment received in real-time, and then... The location of the projection point is determined by projecting the data point onto the corresponding complete normal evolution trajectory. The vertical distance between the real-time data point and the projection point is calculated as the trajectory deviation at that moment. The trajectory deviations of the most recent fixed number of consecutive moments are stored and summed to generate the cumulative deviation. The cumulative deviation is compared with the upper limit of the initial confidence interval of the corresponding position. When the cumulative deviation is greater than the upper limit of the initial confidence interval, the corresponding parameter is marked as having a potential anomaly. The trajectory deviation data of the most recent fixed number of consecutive moments of the parameter are extracted, and the duration, frequency of occurrence, and maximum deviation value of the deviation are calculated. The comprehensive potential anomaly features of the parameter are generated and stored in the potential anomaly feature set.

[0027] In the specific implementation process, real-time data points and corresponding timestamps for each operating status parameter are extracted from the raw energy consumption data of the equipment received in real time. Each real-time data point is projected onto the corresponding complete normal evolution trajectory to determine the position of the projection point. The vertical distance between the real-time data point and the projection point is calculated as the trajectory deviation at that moment. For the stator temperature parameter of stamping motor CM001, the real-time data point collected at 10:00 on February 1, 2026 is 82.5℃. Projecting this data point onto the complete normal evolution trajectory of the stator temperature, the value of the projection point is 75.6℃. Therefore, the trajectory deviation at that moment is |82.5-75.6|=6.9℃. The edge acquisition node stores trajectory deviation data for the most recent 10 consecutive moments, i.e., trajectory deviation data within the most recent second. From 10:00:00.000 on February 1, 2026 to 10:00:00.900 on February 1, 2026, the trajectory deviations at the 10 moments were 6.2℃, 6.5℃, 6.7℃, 6.9℃, 7.1℃, 7.3℃, 7.5℃, 7.7℃, 7.9℃, and 8.1℃, respectively. Summing these trajectory deviations at 10 moments yields a cumulative deviation of 6.2 + 6.5 + 6.7 + 6.9 + 7.1 + 7.3 + 7.5 + 7.7 + 7.9 + 8.1 = 71.9℃. Simultaneously, the average upper limit of the initial confidence interval for the corresponding positions at these 10 time points was calculated to be 3.6℃. Comparing the cumulative deviation of 71.9℃ with the average upper limit of the initial confidence interval of 3.6℃, ​​it is clear that the cumulative deviation is much greater than the upper limit of the initial confidence interval. At this point, a potential anomaly in the stator temperature parameter is immediately identified, and then the trajectory deviation data for the parameter over the most recent 600 consecutive time points (i.e., the trajectory deviation data within the most recent minute) is extracted. The calculated duration of the deviation is 60 seconds, from 09:59:01.000 on February 1, 2026 to 10:00:00.900 on February 1, 2026. The trajectory deviation at all times during this period is greater than the upper limit of the initial confidence interval for the corresponding position; the frequency of deviation occurrence is 100%, meaning that the trajectory deviation at all times out of the 600 time points exceeds the threshold; the maximum deviation is 8.1℃, occurring at 10:00:00.900 on February 1, 2026. A comprehensive potential anomaly characteristic for the stator temperature parameter is formed by combining a duration of 60 seconds, an occurrence frequency of 100%, and a maximum deviation of 8.1℃. This characteristic is then stored in a potential anomaly characteristic set. The potential anomaly characteristic set is stored in key-value pair format, where the key is a combination of the device number and parameter name, and the value is the corresponding comprehensive potential anomaly characteristic and its occurrence time. In practice, the potential anomaly characteristic set is updated every 5 minutes, and historical anomaly characteristic data older than 7 days is automatically deleted.

[0028] Example 4, building upon the normal operation benchmark model constructed in Example 2 and the comprehensive potential anomaly features generated in Example 3, further realizes energy consumption deviation calculation and fault causal tracing and location, providing precise guidance for subsequent fault handling and maintenance. Specifically, it includes calculating the dynamic correlation coefficient between each operating state parameter and energy consumption, generating theoretical energy consumption benchmark values ​​and energy consumption deviation values; when the energy consumption deviation value exceeds the initial tolerance threshold, initiating the causal tracing process to locate the faulty component; extracting all operating state parameters and corresponding instantaneous energy consumption data within the same time window, calculating the correlation between each operating state parameter and instantaneous energy consumption within that time window to generate an initial correlation coefficient, and performing a weighted average calculation with the initial correlation coefficients of previous consecutive fixed-number time windows to generate a dynamic correlation coefficient. The correlation coefficient is calculated by multiplying the current values ​​of all operating status parameters by their corresponding dynamic correlation coefficients and summing the results to generate the theoretical energy consumption baseline value for the current time window. The difference between the actual instantaneous energy consumption and the theoretical energy consumption baseline value within the current time window is calculated to obtain the energy consumption deviation value. The energy consumption deviation value is then compared with a preset initial tolerance threshold. When the causal tracing process is initiated, all operating status parameter data within the fixed time window before and after the energy consumption deviation occurs are extracted. The difference between the average values ​​of each parameter in the two time windows is calculated to obtain the change amplitude. The change amplitude is multiplied by the corresponding dynamic correlation coefficient to obtain the contribution value. All contribution values ​​are sorted from largest to smallest, and the top three parameters are selected. The current values ​​of the three parameters are then compared with the confidence intervals of their corresponding positions on their respective normal evolution trajectories.

[0029] In the specific implementation process, the time window length was set to 1 minute, meaning that the theoretical energy consumption baseline value and the energy consumption deviation value were calculated every 1 minute. At 10:00 on February 1, 2026, all operating status parameters and instantaneous energy consumption data of stamping motor CM001 within the time window from 09:59:00 to 10:00:00 on February 1, 2026 were extracted, and the Pearson correlation coefficient between each operating status parameter and instantaneous energy consumption was calculated as the initial correlation coefficient. The calculated initial correlation coefficients were 0.92 for three-phase current, 0.15 for three-phase voltage, 0.78 for stator temperature, 0.65 for bearing temperature, 0.52 for radial vibration, 0.48 for axial vibration, and 0.89 for output speed. Then, a weighted average is calculated between the initial correlation coefficient of the current time window and the initial correlation coefficients of the previous nine consecutive time windows. The weighting coefficients, from most recent to oldest, are 0.2, 0.15, 0.15, 0.1, 0.1, 0.1, 0.08, 0.06, 0.04, and 0.02, generating dynamic correlation coefficients. The calculated dynamic correlation coefficients are 0.76 for stator temperature, 0.87 for output speed, 0.63 for bearing temperature, and 0.51 for radial vibration. The theoretical energy consumption baseline for this time window is 185.2 kW, calculated by multiplying the current average values ​​of all operating parameters by their corresponding dynamic correlation coefficients and summing the results. The actual instantaneous average energy consumption within the current time window is 212.7 kW, therefore the energy consumption deviation is 212.7 - 185.2 = 27.5 kW. The preset initial tolerance threshold is 15kW. Comparing the energy consumption deviation value of 27.5kW with the initial tolerance threshold of 15kW, it is clear that the energy consumption deviation value exceeds the initial tolerance threshold, and the causal tracing process is immediately initiated. All operating status parameter data are extracted within the time window from 09:54:00 to 09:59:00 on February 1, 2026 (5 minutes before the energy consumption deviation occurred), and within the time window from 10:00:00 to 10:05:00 on February 1, 2026 (5 minutes after the energy consumption deviation occurred). The difference between the average values ​​of each parameter in the two time windows is calculated to obtain the change amplitude. The change amplitude for stator temperature is 8.2℃, for bearing temperature is 5.7℃, for radial vibration is 3.2mm / s, and for output speed is -12r / min. The contribution value is obtained by multiplying the change range of each parameter by the corresponding dynamic correlation coefficient. The contribution value of stator temperature is 8.2×0.76=6.23, the contribution value of bearing temperature is 5.7×0.63=3.59, the contribution value of radial vibration is 3.2×0.51=1.63, and the contribution value of output speed is |-12|×0.87=10.44.All contribution values ​​were sorted from largest to smallest, with the top three parameters being output speed, stator temperature, and bearing temperature. The current average values ​​of these three parameters were compared to their respective confidence intervals on their normal evolution trajectories. The current average value of output speed was 1488 r / min, the normal evolution trajectory value was 1500 r / min, and the confidence interval was [1490 r / min, 1510 r / min], meaning the current value was below the lower limit of the confidence interval. The current average value of stator temperature was 82.3℃, the normal evolution trajectory value was 75.6℃, and the confidence interval was [72.0℃, 79.2℃], meaning the current value was above the upper limit of the confidence interval. The current average value of bearing temperature was 68.5℃, the normal evolution trajectory value was 62.1℃, and the confidence interval was [59.3℃, 64.9℃], meaning the current value was above the upper limit of the confidence interval. Based on the above analysis, the faulty component was identified as the motor bearing, and the cause of the fault was increased friction due to bearing wear, leading to a decrease in speed, an increase in temperature, and an increase in energy consumption.

[0030] Example 5: Building upon the LSTM-BP fusion operation status prediction model constructed in Example 2, this example further optimizes the generation of a dual-objective energy system operation strategy. This reduces total system energy consumption while ensuring equipment health. Specifically, it employs a dynamic weighted particle swarm optimization algorithm combined with the prediction model output to generate an energy system operation strategy. With the dual objectives of minimizing total system energy consumption and optimizing equipment health, the future operation status parameters and energy consumption trend data output by the LSTM-BP fusion operation status prediction model are used as constraints. The position and velocity of the particle swarm are initialized, the inertial weight of each particle is dynamically adjusted, and the optimal solution is obtained through iterative calculation. This generates an energy system operation strategy that includes the load allocation and parameter settings for each device.

[0031] In the specific implementation, the particle swarm size is set to 50 particles, each with a dimension of 10, corresponding to the operating load distribution ratio of 10 stamping motors. The position of each particle is initialized as a random number between 0 and 1, and the velocity as a random number between -0.1 and 0.1. The inertia weight adopts a linear decreasing strategy, with an initial inertia weight of 0.9 and a final inertia weight of 0.4, decreasing linearly with the number of iterations. The number of iterations is set to 100, and the individual optimal solution and the global optimal solution for each particle are updated after each iteration. The operating status parameters and energy consumption trend data of each device output by the LSTM-BP fusion operating status prediction model for the next 24 hours are used as constraints. These constraints include that the operating load of each device does not exceed 90% of its rated load, the stator temperature of each device does not exceed 120℃, and the bearing temperature of each device does not exceed 100℃. The dual objectives are to minimize total system energy consumption and optimize device health. Device health is represented by the sum of the cumulative deviations of all devices; the smaller the sum of cumulative deviations, the better the device health. After 100 iterations of calculation, the global optimal solution was obtained, generating the energy system operation strategy. In the actual implementation, the optimized system's total energy consumption was 42,860 kWh / day, a 9.8% reduction compared to the unoptimized 47,520 kWh / day; the sum of the cumulative deviations of all equipment was 125.6℃, a significant improvement in equipment health compared to the unoptimized 218.3℃.

[0032] Example 6, building upon Example 3's generation of comprehensive potential anomaly features, Example 4's location of faulty components, and Example 5's generation of energy system operation strategies, further realizes intelligent operation and maintenance execution and closed-loop management. It links with the industrial energy dispatch system to execute operation strategies, forming a complete closed loop for intelligent equipment fault diagnosis and status monitoring. Specifically, it includes generating operation parameter adjustment schemes and inspection work orders by combining comprehensive potential anomaly features, fault information, and prediction results, and linking with the industrial energy dispatch system to execute operation strategies; extracting comprehensive potential anomaly features from the potential anomaly feature set that are not associated with fault information, sorting them by cumulative deviation from largest to smallest to generate a list of equipment performance degradation priorities, generating equipment operation parameter adjustment schemes based on the future state trends output by the prediction model, and issuing them to the corresponding equipment-side edge acquisition nodes for execution; comparing the cumulative deviation and energy consumption deviation values ​​before and after adjustment in the cloud; when both values ​​are still higher than the corresponding thresholds after adjustment, updating the equipment performance degradation priority, generating inspection work orders containing equipment location, abnormal parameters, and suggested inspection content, and issuing them to operation and maintenance personnel.

[0033] In the specific implementation process, comprehensive potential anomaly features of all unrelated fault information are extracted from the potential anomaly feature set, involving 5 parameters of 3 devices. These parameters are sorted by cumulative deviation from largest to smallest to generate a priority list of device performance degradation. The priorities, from highest to lowest, are: radial vibration of motor CM003, stator temperature of motor CM005, axial vibration of motor CM007, bearing temperature of motor CM002, and output speed of motor CM009. Combining the changing trends of these parameters output by the LSTM-BP fusion operating status prediction model for the next 24 hours, an equipment operating parameter adjustment plan is generated. This plan mainly includes appropriately reducing the operating load of motors CM003, CM005, and CM007 by 10%, 8%, and 5%, respectively. The operating parameter adjustment plan is then distributed to the corresponding device-side edge acquisition nodes for execution. The edge acquisition nodes communicate with the motor controllers via the Modbus protocol to adjust the motor operating parameters in real time. Simultaneously, the energy system operation strategy generated in Example 5 is linked to the industrial energy dispatch system for execution, achieving optimized scheduling of the entire stamping workshop's energy system. One hour after the adjustment plan was implemented, the cloud platform compared the cumulative deviation and energy consumption deviation before and after the adjustment. For motor CM003, the cumulative deviation decreased from 45.2 mm / s to 28.7 mm / s, and the energy consumption deviation decreased from 12.3 kW to 8.5 kW, both below the corresponding thresholds. For motor CM005, the cumulative deviation decreased from 52.6℃ to 31.4℃, and the energy consumption deviation decreased from 14.7 kW to 9.2 kW, both below the corresponding thresholds. For motor CM007, the cumulative deviation decreased from 38.9 mm / s to 35.2 mm / s, still above the corresponding threshold of 25 mm / s, while the energy consumption deviation decreased from 11.8 kW to 10.5 kW, below the corresponding threshold. Therefore, the equipment performance degradation priority list was updated, and the axial vibration of motor CM007 was prioritized as the first priority. Then, an inspection work order is generated. The work order details the equipment location as workstation 3 in the stamping workshop, equipment number CM007, abnormal parameter as axial vibration, and suggested checks for motor bearing lubrication, loose motor base mounting bolts, and bending deformation of the motor shaft. The work order is then sent to Engineer Zhang, responsible for the stamping workshop's operations and maintenance, via WeChat. The work order requires completion and handling within 2 hours. After completing the inspection, Engineer Zhang uploads the results to the cloud server. The cloud server updates the equipment's operating status information, closes the inspection work order, and completes the entire operations and maintenance loop.

[0034] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A method for intelligent diagnosis and condition monitoring of equipment faults based on deep learning, characterized in that, Includes the following steps: The device-side edge acquisition node collects device operating parameters, generates raw device operating energy consumption data with a unique timestamp, and transmits it to the cloud. The cloud sorts and aligns the received data by device number and timestamp, and generates a dataset related to device operating energy consumption. Based on the associated dataset, we construct the normal evolution trajectory and corresponding confidence interval of each operating state parameter, and at the same time construct an LSTM-BP fusion operating state prediction model. The real-time data points of the running status parameters are extracted from the real-time received data. The vertical distance between the data points and the normal evolution trajectory is calculated as the trajectory deviation. The trajectory deviations at consecutive time points are summed to generate the cumulative deviation, and a comprehensive potential anomaly feature is generated. Calculate the dynamic correlation coefficients between various operating state parameters and energy consumption, and generate theoretical energy consumption benchmark values ​​and energy consumption deviation values; When the energy consumption deviation exceeds the initial tolerance threshold, the cause-and-effect tracing process is initiated to locate the faulty component. The dynamic weighted particle swarm optimization algorithm is combined with the output of the prediction model to generate an energy system operation strategy; By combining comprehensive potential anomaly characteristics, fault information and prediction results, an operational parameter adjustment plan and inspection work order are generated, and the industrial energy dispatch system is linked to execute the operational strategy.

2. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, The edge acquisition nodes on the device side collect all measurable operating status parameters of the device and the instantaneous energy consumption parameters at the corresponding time intervals to form the raw energy consumption data of the device operation. The integrity of each set of raw energy consumption data of the device operation is verified. Data that fails the verification is discarded, and the data that passes the verification is classified and stored according to parameter type and then synchronously transmitted to the cloud.

3. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, Extract all data from the equipment's continuous fault-free operation period from the equipment operation energy consumption association dataset, split each operating state parameter separately to generate a normal operation time series, divide the time series into fixed-length continuous data segments, perform trend fitting on each data segment to generate a local evolution trajectory, connect all local evolution trajectories to generate a complete normal evolution trajectory, calculate the dispersion of each data point on the trajectory to generate the initial confidence interval for the corresponding position.

4. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, The energy consumption data set associated with equipment operation is divided into training and validation sets according to a preset ratio. The training set is used to train an LSTM network to extract long-term time-series features of equipment operation status. The extracted long-term time-series features are input into a BP network for regression prediction. The validation set is used to adjust the model parameters and generate the final LSTM-BP fusion operation status prediction model. The model outputs the operation status parameters and energy consumption trend data of the equipment within a fixed period of time in the future.

5. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, Extract real-time data points and corresponding timestamps for each operating status parameter from the raw energy consumption data of the device received in real time. Project each real-time data point onto the corresponding complete normal evolution trajectory to determine the position of the projection point. Calculate the vertical distance between the real-time data point and the projection point as the trajectory deviation at that moment. Store the trajectory deviations of the most recent fixed number of consecutive moments and perform a summation operation to generate the cumulative deviation. Compare the cumulative deviation with the upper limit of the initial confidence interval of the corresponding position.

6. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 5, characterized in that, When the cumulative deviation exceeds the upper limit of the initial confidence interval, the corresponding parameter is marked as potentially abnormal. The trajectory deviation data of the parameter at the most recent fixed number of consecutive moments are extracted, the duration, frequency of occurrence and maximum deviation value of the deviation are calculated, and the comprehensive potential anomaly features of the parameter are generated and stored in the potential anomaly feature set.

7. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, Extract all operating status parameters and corresponding instantaneous energy consumption data within the same time window, calculate the correlation between each operating status parameter and instantaneous energy consumption within the time window to generate an initial correlation coefficient, perform a weighted average operation on the initial correlation coefficient with the initial correlation coefficients of the previous fixed number of time windows to generate a dynamic correlation coefficient, multiply the current values ​​of all operating status parameters by their corresponding dynamic correlation coefficients and sum them to generate the theoretical energy consumption benchmark value for the time window, calculate the difference between the actual instantaneous energy consumption and the theoretical energy consumption benchmark value within the current time window to obtain the energy consumption deviation value, and compare the energy consumption deviation value with a preset initial tolerance threshold.

8. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, When the causal tracing process is initiated, all operating status parameter data within a fixed time window before and after the energy consumption deviation occurs are extracted. The difference between the average values ​​of each parameter in the two time windows is calculated to obtain the change amplitude. The change amplitude is multiplied by the corresponding dynamic correlation coefficient to obtain the contribution value. All contribution values ​​are sorted from largest to smallest, and the top three parameters are selected. The current values ​​of the three parameters are compared with the confidence intervals of the corresponding positions on their respective normal evolution trajectories.

9. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, With the dual objectives of minimizing total system energy consumption and optimizing equipment health, the future operating state parameters and energy consumption trend data output by the LSTM-BP fusion operating state prediction model are used as constraints to initialize the position and velocity of the particle swarm, dynamically adjust the inertial weight of each particle, and iteratively calculate to obtain the optimal solution, generating an energy system operation strategy that includes the allocation of operating loads and parameter settings for each device.

10. The method for intelligent equipment fault diagnosis and status monitoring based on deep learning according to claim 1, characterized in that, Extract comprehensive potential anomaly features from the potential anomaly feature set, which contain all unrelated fault information. Sort the features by cumulative deviation from largest to smallest to generate a list of equipment performance degradation priorities. Combine the future state trend output by the prediction model to generate equipment operating parameter adjustment plans and send them to the corresponding equipment edge acquisition nodes for execution. Compare the cumulative deviation and energy consumption deviation values ​​before and after adjustment in the cloud. If the two values ​​are still higher than the corresponding threshold after adjustment, update the equipment performance degradation priority, generate an inspection work order containing equipment location, abnormal parameters and suggested inspection contents, and send it to maintenance personnel.

Images