Deep learning based numerical control tool failure diagnosis method

Abstract

The application relates to the technical field of numerical control machining, and discloses a numerical control tool fault diagnosis method based on deep learning, which comprises the following steps: collecting tool running state data, determining real-time feature offset and a comprehensive fault feature index set; determining the maximum feature analysis capability of a candidate diagnosis model based on equipment compatibility parameters and hidden layer activation parameters; screening the candidate diagnosis model in combination with a benchmark analysis model, selecting a target diagnosis model after calculating a total delay of dynamic processing, and generating a fault diagnosis configuration result; and further selecting an auxiliary monitoring node according to a redundant feature compensation amount of the target model, and generating an auxiliary analysis network configuration result. The method is suitable for tool state monitoring and intelligent maintenance in the field of numerical control machining, and improves the accuracy, real-time performance and system adaptability of tool fault diagnosis through multi-source feature fusion, model adaptation optimization, delay and cost control and redundant monitoring construction.

Description

Technical Field

[0001] This invention relates to the field of CNC machining technology, specifically a CNC tool fault diagnosis method based on deep learning. Background Technology

[0002] In the field of modern CNC machining, cutting tools, as the core components of machine tools performing cutting operations, directly affect machining quality, production efficiency, and equipment safety. With the deepening development of intelligent manufacturing and industrial automation, the complexity and precision requirements of CNC machining systems are constantly increasing, and traditional tool fault diagnosis methods are gradually revealing many limitations.

[0003] Traditional fault diagnosis primarily relies on human experience or simple signal processing techniques, such as vibration monitoring based on threshold judgment and feature extraction based on spectral analysis. These methods have the following significant problems: First, manual diagnosis depends on the experience and subjective judgment of technicians, making it difficult to cope with complex and ever-changing machining conditions, resulting in low diagnostic efficiency and unstable accuracy. Second, traditional signal processing methods require pre-setting feature extraction rules, making it difficult to effectively capture deep-seated fault characteristics for nonlinear and non-stationary tool operation data (such as sudden changes in vibration signals caused by tool wear and breakage during cutting, and complex modulation of acoustic emission signals), leading to insufficient early fault identification capabilities. Furthermore, traditional methods typically analyze data from a single monitoring node, failing to integrate the collaborative information from multi-source heterogeneous data (such as vibration, acoustic emission, and cutting force data from different types of sensors), making it difficult to comprehensively reflect the true operating state of the tool and easily leading to missed or misdiagnosed cases.

[0004] With the rapid development of deep learning technology, it has demonstrated powerful capabilities in areas such as feature extraction and pattern recognition. Introducing deep learning into CNC tool fault diagnosis has become a research hotspot, but current technologies still face the following challenges: Firstly, the sensor layout (e.g., sensor installation location, quantity, and type) varies significantly among different CNC machines, leading to insufficient compatibility between the deployment conditions and analytical capabilities of deep learning models and the equipment, making it difficult to achieve universal diagnosis in different machine tool environments. Secondly, the complexity of deep learning models themselves results in high computational resource consumption and poor real-time performance. For example, insufficient optimization of hidden layer activation parameters may limit feature parsing capabilities, and the balance between data transmission latency and model computation latency is difficult to achieve, affecting the real-time performance and accuracy of fault diagnosis. Furthermore, existing methods lack comprehensive consideration of redundant feature coverage and parsing costs, making it impossible to construct an efficient auxiliary monitoring network, resulting in a difficulty in balancing the reliability and economy of the diagnostic system.

[0005] Therefore, how to construct a deep learning diagnostic model adapted to the sensor layout of equipment based on the multi-source data characteristics of CNC tool operation, and achieve an optimal balance between feature parsing capability, computational latency, and parsing cost, has become a key technical problem that urgently needs to be solved in the current CNC machining field. Developing a tool fault diagnosis method that can adapt to equipment compatibility, integrate multi-source features, and optimize diagnostic efficiency and cost is of great practical significance for improving the intelligence level of CNC machining and reducing downtime losses and maintenance costs. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for diagnosing CNC tool faults, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a CNC tool fault diagnosis method based on deep learning, the method comprising:

[0008] Based on the tool running status data collected by multiple monitoring nodes within a set sampling period, the real-time feature offset corresponding to each monitoring node is determined, and the comprehensive fault feature index set corresponding to multiple monitoring nodes is integrated.

[0009] After parsing the benchmark analysis model required for the comprehensive fault feature index set, the maximum feature parsing capability of each candidate diagnostic model under the corresponding equipment compatibility parameters is determined. The equipment compatibility parameters are used to characterize the adaptability of the sensor layout in the target machine tool to the deployment conditions and parsing capability of the candidate diagnostic model. The maximum feature parsing capability is determined according to the hidden layer activation parameters associated with the candidate diagnostic model.

[0010] Based on the maximum feature parsing capability and the benchmark analysis model, at least one alternative diagnostic model is selected from multiple candidate diagnostic models. After calculating the total dynamic processing delay corresponding to the alternative diagnostic model and multiple associated monitoring nodes, a target diagnostic model is selected from at least one alternative diagnostic model based on at least the total dynamic processing delay, and a fault diagnosis configuration result is generated.

[0011] Preferably, the fault diagnosis configuration result includes all selected target diagnostic models, the maximum feature resolution capability of each target diagnostic model under the target device compatibility parameters, and deployment coordinates.

[0012] Preferably, after generating the fault diagnosis configuration result, the method further includes:

[0013] Obtain second feature parameters of multiple candidate auxiliary monitoring nodes, wherein the second feature parameters include monitoring channel identifier and auxiliary resolution range, and the auxiliary resolution range is used to characterize the redundant feature coverage area of ​​the candidate auxiliary monitoring nodes;

[0014] Based on the maximum feature parsing capability of each target diagnostic model under the target device compatibility parameters, the redundant feature compensation amount information corresponding to each target diagnostic model is determined, wherein the redundant feature compensation amount information is mapped to the corresponding auxiliary parsing coverage information;

[0015] Based on the redundant feature compensation information and the auxiliary parsing range corresponding to each candidate auxiliary monitoring node, at least one alternative auxiliary monitoring node that satisfies the matching of the auxiliary parsing coverage information is selected from multiple candidate auxiliary monitoring nodes.

[0016] Based on the monitoring channel identifier corresponding to each of the candidate auxiliary monitoring nodes and the deployment coordinates of the corresponding target diagnostic model, the total single-transmission delay of the feature compensation between the target diagnostic model and at least one of the candidate auxiliary monitoring nodes is calculated. The comprehensive parsing cost is determined according to the deployment delay corresponding to each of the candidate auxiliary monitoring nodes and the total single-transmission delay. The target auxiliary monitoring node is selected from at least one of the candidate auxiliary monitoring nodes, and the auxiliary parsing network configuration result is generated.

[0017] Preferably, the comprehensive analysis cost is determined based on the deployment delay corresponding to each of the candidate auxiliary monitoring nodes and the total delay for a single operation, including:

[0018] Obtain a preset comprehensive parsing cost threshold, wherein the threshold is used to characterize the maximum allowable latency of the redundant parsing network covering all the monitoring nodes;

[0019] Based on the comprehensive analysis cost, the total network latency corresponding to the configured candidate auxiliary monitoring node combinations is calculated, and the combination with the smallest difference between the total network latency and the comprehensive analysis cost threshold is selected to generate the target auxiliary monitoring node combination.

[0020] Preferably, the calculation of the total dynamic processing latency corresponding to the candidate diagnostic model and the associated multiple monitoring nodes includes:

[0021] Based on the data transmission path parameters of each monitoring node and the corresponding candidate diagnostic model, the single-point parsing delay of the candidate diagnostic model in processing the feature data of the monitoring node is determined, and the total parsing delay corresponding to the candidate diagnostic model is obtained by summarizing.

[0022] Determine the activation response parameter associated with the maximum feature resolution capability corresponding to each of the candidate diagnostic models, and determine the periodic computation cost of the hidden layer activation based on the response parameter;

[0023] The total dynamic processing latency is generated based on the cycle calculation cost, model update cost, and total parsing latency.

[0024] Preferably, selecting a target diagnostic model from at least one of the candidate diagnostic models, based at least on the total dynamic processing delay, includes:

[0025] Based on the benchmark analysis model corresponding to each candidate diagnostic model, the required amount of training data and storage adaptation amount are determined, and the node with the shortest transmission path and meeting the required amount of training data is selected from multiple candidate data source nodes to generate the target data source node. The scheme with the shortest adaptation path and meeting the required storage adaptation amount is selected from multiple candidate storage configuration schemes to generate the target storage configuration.

[0026] Based on the transmission path, adaptation path, and corresponding unit processing coefficient, the data base cost of the candidate diagnostic model is calculated, and the data base cost and the total dynamic processing delay are summed to generate the comprehensive analysis cost of the candidate diagnostic model.

[0027] The target diagnostic model is selected from at least one of the candidate diagnostic models in order of increasing comprehensive analysis cost.

[0028] Preferably, based on the maximum feature parsing capability and the benchmark analysis model, at least one candidate diagnostic model is selected from the plurality of candidate diagnostic models, including:

[0029] Determine the feature coverage of the candidate diagnostic model under the maximum feature parsing capability corresponding to each of the device compatibility parameters;

[0030] Based on the matching degree between the coverage and the benchmark analysis model, a set of configuration parameters containing at least one candidate diagnostic model is generated.

[0031] Based on the adaptability score corresponding to the device compatibility parameters of each model in the set, the set with the highest total score is selected as the candidate diagnostic model combination.

[0032] Preferably, determining the maximum feature parsing capability of each candidate diagnostic model under the corresponding device compatibility parameters includes:

[0033] Obtain the resolution capability parameter table, wherein the table includes equipment compatibility parameters, hidden layer activation threshold, reference distance to the core components of the target machine tool, and the relationship between the three;

[0034] Based on the deployment coordinates of the candidate diagnostic model, calculate its deviation value from the baseline distance, and query the hidden layer activation threshold corresponding to the deviation value and device compatibility parameters in the table;

[0035] The maximum feature parsing capability is generated based on the activation threshold.

[0036] Preferably, obtaining the monitoring channel identifier includes:

[0037] The operating parameters of the monitoring node are collected in real time through a preset sensor network. The operating parameters include vibration amplitude, acoustic emission frequency and cutting force.

[0038] Preferably, the acquisition frequency of the operating parameters is adaptively adjusted according to the fluctuation range of the real-time feature offset.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] In terms of feature processing and model adaptation, by collecting operational status data from multiple monitoring nodes, calculating real-time feature offsets, and integrating a comprehensive fault feature index set, the system can comprehensively capture multi-source heterogeneous features (such as vibration amplitude, acoustic emission frequency, and cutting force) during tool operation, avoiding diagnostic blind spots caused by single data dimensions. Simultaneously, equipment compatibility parameters are introduced to characterize the adaptability of the machine tool sensor layout to the candidate diagnostic model. Combined with hidden layer activation parameters, the system determines the model's maximum feature resolution capability, enabling dynamic optimization based on the sensor deployment conditions of different machine tools. This significantly improves the model's compatibility with diverse equipment and the specificity of feature resolution, solving the problem of poor model versatility in traditional methods.

[0041] In terms of model selection and latency optimization, candidate diagnostic models are selected based on maximum feature parsing capability and benchmark analysis models. Factors such as data transmission path parameters, the periodic computation cost of hidden layer activation, and model update costs are comprehensively considered to accurately calculate the total dynamic processing latency. By combining the basic data cost (including training data transmission paths and storage adaptation paths) with the total dynamic processing latency to generate a comprehensive parsing cost, a quantitative evaluation of model computational efficiency and resource consumption is achieved. This ensures that the selected target diagnostic model achieves an optimal balance between feature parsing capability and real-time performance, effectively avoiding the diagnostic latency problem caused by excessive computational complexity in deep learning models, and meeting the requirements of real-time monitoring in CNC machining.

[0042] In terms of constructing the auxiliary monitoring network, by introducing the redundant feature coverage area (auxiliary parsing range) of candidate auxiliary monitoring nodes and combining it with the redundant feature compensation information of the target diagnostic model, suitable auxiliary monitoring nodes are dynamically selected. The overall parsing cost is calculated by comprehensively considering deployment delay and the total single-transmission delay of feature compensation, thus constructing an efficient auxiliary parsing network. This design not only enhances the system's coverage of tool fault features and improves diagnostic reliability through redundant feature compensation, but also reduces parsing costs by optimizing node deployment and transmission paths. This achieves a dual improvement in the diagnostic system's reliability and economy, resolving the contradiction between redundant monitoring and cost control in traditional methods.

[0043] In terms of intelligent data acquisition and processing, the acquisition frequency of the monitoring node's operating parameters can be adaptively adjusted according to the fluctuation range of real-time feature offset. When the tool's operating state tends to be stable, the acquisition frequency is reduced to decrease data redundancy and computational load; when abnormal fluctuations in feature offset are detected, the acquisition frequency is automatically increased to capture subtle fault characteristics, realizing dynamic optimization of data acquisition and improving system operating efficiency and resource utilization while ensuring diagnostic accuracy. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the working principle of the deep learning-based CNC tool fault diagnosis method described in this invention.

[0045] Figure 2 Design diagram for configuring the network for auxiliary monitoring nodes;

[0046] Figure 3 Design diagram for dynamic total delay calculation logic;

[0047] Figure 4 Design diagram for selecting the logic of the target diagnostic model;

[0048] Figure 5 This is a design diagram for the logic of determining the maximum feature parsing capability. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see Figures 1-5 The present invention relates to a deep learning-based CNC tool fault diagnosis method, the specific implementation steps of which are as follows:

[0051] Based on multiple monitoring nodes (such as vibration sensors, acoustic emission sensors, and cutting force sensors), tool operating status data, including parameters such as vibration amplitude, acoustic emission frequency, and cutting force, is collected in real time within a set sampling period. For each monitoring node, real-time feature offset is calculated by comparing the currently collected data with historical baseline data. This offset reflects the degree of difference between the current state and the normal operating state. Subsequently, the real-time feature offsets from multiple monitoring nodes are fused and combined with a preset fault feature weight matrix to form a comprehensive fault feature index set. This index set contains information such as the offset degree and correlation of features in each dimension, which is used for subsequent analysis of the fault diagnosis model.

[0052] The pre-built benchmark analysis model is analyzed, which defines the distribution range, association rules, and typical patterns of fault characteristics under normal operating conditions. For each candidate diagnostic model (such as deep learning models like convolutional neural networks and recurrent neural networks), its corresponding equipment compatibility parameters are obtained. These parameters characterize the adaptability of the sensor layout within the target machine tool to the model's deployment conditions and analytical capabilities, specifically including sensor type, installation location, and data transmission protocol. Based on the hidden layer activation parameters (such as activation function type and threshold) associated with the candidate diagnostic model, its maximum feature resolution capability under the corresponding equipment compatibility parameters is calculated. This capability reflects the model's ability to extract and identify complex fault features.

[0053] Based on the matching degree between the maximum feature resolution capability of each candidate diagnostic model and the benchmark analysis model, at least one candidate diagnostic model is selected. For each candidate diagnostic model, the total dynamic processing latency corresponding to its multiple associated monitoring nodes is calculated. This latency is determined through the following steps: First, based on the data transmission path parameters (such as transmission distance, bandwidth, protocol overhead, etc.) between the monitoring nodes and the candidate diagnostic models, the single-point resolution latency is calculated and summarized into the total resolution latency; second, the activation response parameters (such as the computational complexity of the activation function, number of iterations, etc.) associated with the maximum feature resolution capability are determined, and the periodic computational cost of hidden layer activation is calculated; finally, combined with the model update cost (such as parameter update frequency, computational resource consumption, etc.), the total dynamic processing latency is generated. Based on at least the total dynamic processing latency, a target diagnostic model is selected from the candidate diagnostic models, and a fault diagnosis configuration result containing the target diagnostic model, its maximum feature resolution capability under the target device compatibility parameters, and deployment coordinates is generated.

[0054] Example 1:

[0055] In the deep learning-based CNC tool fault diagnosis method, the fault diagnosis configuration results include all selected target diagnostic models, the maximum feature resolution capability of each target diagnostic model under the target equipment compatibility parameters, and deployment coordinates. The determination of the deployment coordinates is closely related to the physical structure of the target machine tool, sensor layout, and data transmission characteristics. The specific implementation method is as follows:

[0056] A digital model of the sensor layout for the target machine tool needs to be established. This model must accurately map the physical installation positions of each monitoring node (such as vibration sensors, acoustic emission sensors, cutting force sensors, etc.), including their three-dimensional coordinates (X, Y, Z axis coordinates) in the machine tool coordinate system, their spatial orientation relative to the tool machining area (e.g., near the spindle, tool holder, worktable, etc.), and the relative distances between the sensors and the core moving components of the machine tool (e.g., spindle motor, guide rails, ball screws, etc.). For example, a vibration sensor may be installed at a specific location on the spindle housing, with coordinates (X1, Y1, Z1) and a radial distance of D1 from the spindle axis; an acoustic emission sensor may be embedded near the tool holder, with coordinates (X2, Y2, Z2) and a linear distance of D2 from the cutting edge of the tool. This spatial location information is obtained through machine tool design drawings, sensor installation process documents, or on-site measurements and stored in the system's sensor layout database.

[0057] For each candidate diagnostic model, the deployment form of its computing nodes needs to be determined. Computing nodes can be physical hardware devices (such as edge computing units or industrial computers) or virtual computing resources (such as containers or virtual machines in a cloud computing platform). For physical hardware devices, the deployment coordinates correspond to their installation location in a machine tool control cabinet, equipment rack, or field cabinet, taking into account heat dissipation, electromagnetic compatibility, and cabling distance from sensors. For virtual computing resources, the deployment coordinates correspond to their logical address in the network topology (such as IP address or port number), taking into account network latency, bandwidth usage, and routing paths for data transmission. For example, if a candidate diagnostic model adopts an edge computing architecture, its physical computing nodes may be deployed in a cabinet near the machine tool to shorten the data transmission distance with local sensors; if a cloud-edge collaborative architecture is adopted, some computing tasks may be allocated to cloud servers, and their logical deployment coordinates are the network address of the cloud data center.

[0058] When determining deployment coordinates, the following core factors need to be considered:

[0059] Data transmission latency: The length of the data transmission path between the monitoring node and the diagnostic model computing node directly affects the latency. For physical deployment scenarios, the transmission path length is determined by the length of the cable or fiber optic cable between the sensor and the computing node; for network deployment scenarios, it is determined by the network hop count, link bandwidth, and protocol overhead. For example, if the vibration sensor is connected to the local edge computing node via a USB cable, the transmission latency mainly depends on the cable length and the processing time of the USB protocol; if it is transmitted to the cloud via a Wi-Fi network, the latency is determined by the wireless signal transmission distance, the access point load, and the internet routing efficiency. To minimize transmission latency, computing nodes should be deployed closer to the sensors, or high-speed data transmission links (such as fiber optics or 5G networks) should be selected.

[0060] Balanced Processing Efficiency: When multiple monitoring nodes are associated with the same diagnostic model, it is necessary to ensure that the computing nodes can process multi-source data in parallel. The selection of deployment coordinates should consider whether the physical location of the computing nodes is convenient for accessing the data interfaces of multiple sensors, or whether the logical location has efficient data aggregation capabilities. For example, if a diagnostic model needs to simultaneously analyze spindle vibration data, tool cutting force data, and feed axis current data, its computing nodes can be deployed in the machine tool electrical control cabinet, and simultaneously access the signal acquisition modules of each sensor via fieldbus (such as PROFINET, EtherCAT) to reduce intermediate steps in data aggregation.

[0061] Device compatibility parameter adaptation: Device compatibility parameters include sensor type, data format, sampling frequency, etc., and the deployment coordinates must match these parameters. For example, if the sensor uses analog signal output (such as a 4-20mA current signal), the computing node needs to be configured with the corresponding analog input module and deployed close to the sensor to avoid signal attenuation; if the sensor outputs digital signal (such as an RS485 signal based on the Modbus protocol), the computing node needs to have a compatible communication interface (such as an RS485 interface), and its deployment coordinates can be flexibly selected according to the network topology, but the consistency of the communication protocol must be ensured.

[0062] System scalability and maintenance convenience: The planning of deployment coordinates should reserve space for future additions of sensors or diagnostic models. For example, spare installation locations or network ports should be reserved in the machine tool control cabinet so that when auxiliary monitoring nodes are added later, there is no need to make large-scale adjustments to the deployment of existing computing nodes. At the same time, the physical deployment coordinates should be easily accessible to maintenance personnel, such as by installing them on easily removable racks, or by having clear network naming rules for logical deployment coordinates to facilitate remote debugging and troubleshooting.

[0063] The specific implementation steps are as follows:

[0064] The first step is to obtain the device compatibility parameters of the candidate diagnostic models, including a list of sensor types, data transmission protocol requirements, and computing resource requirements (such as CPU model, memory capacity, and storage media type). For example, a convolutional neural network model needs to process vibration data with a high sampling frequency, requiring the data transmission protocol to support the real-time EtherCAT bus, and the computing nodes to have GPU acceleration capabilities.

[0065] The second step is to select a set of monitoring nodes associated with the candidate diagnostic model based on the digital model of the sensor layout. For example, if the model is used to diagnose tool wear faults, the associated monitoring nodes may include vibration sensors, acoustic emission sensors, and current sensors of the spindle motor installed near the tool.

[0066] The third step involves calculating the transmission path parameters from the physical location of the sensor to the deployment location of the candidate computing node for each associated monitoring node. For physical connections, the path parameters include cable length and cabling direction (e.g., whether it crosses areas with electromagnetic interference); for network connections, the path parameters include the number of hops in the IP network, link bandwidth, and historical latency statistics. For example, the cable length from the vibration sensor to the edge computing node is 5 meters, and the cabling needs to bypass the strong electromagnetic field near the spindle motor, with an estimated transmission delay of 0.5ms; the network hop count to the cloud server is 10 hops, with an average latency of 20ms.

[0067] The fourth step is to establish a mapping relationship between transmission delay and deployment coordinates. For physical deployment scenarios, this can be achieved using formula T. phy =L×k cab +t proc Calculate the delay, where L is the cable length and k cab t represents the delay factor per unit length of cable (e.g., 0.1 μs / m for a USB 3.0 cable). proc This refers to the processing time of the interface module; for network deployment scenarios, the latency value T is obtained through historical data statistics or network simulation tools. net .

[0068] The fifth step is to assess whether the processing capabilities of the computing nodes meet the requirements for parallel processing of multi-source data. For example, if the CPU of the edge computing node has a clock speed of 2.5 GHz and memory of 8 GB, it is necessary to calculate the computing resource utilization required to simultaneously process three channels of vibration data with a sampling frequency of 20 kHz, one channel of acoustic emission data with a sampling frequency of 10 kHz, and two channels of cutting force data with a sampling frequency of 5 kHz, ensuring that it does not exceed the maximum load of the node (such as 80% CPU utilization).

[0069] Step 6: Score each candidate deployment coordinate based on a comprehensive evaluation of transmission latency, processing capacity, and device compatibility parameters. The scoring metrics include: latency score (40%), processing load score (30%), compatibility score (20%), and scalability score (10%). The latency score is calculated as the ratio of the actual latency to the system's maximum allowable latency threshold (e.g., 5ms); the processing load score is calculated as the reciprocal of the computing resource utilization rate; the compatibility score is the sum of binary indicators such as sensor type and protocol matching degree (1 point for matching, 0 points for non-matching); and the scalability score is quantified based on the number of reserved installation space or network ports.

[0070] Step 7: Select the deployment coordinates with the highest score as the deployment location of the target diagnostic model. For example, if an edge computing node deployment scheme scores 92 points, which is better than the cloud deployment scheme's 78 points, then select that edge node as the deployment coordinates and record its physical coordinates (e.g., the left slot on the third layer of the machine tool control cabinet) or logical coordinates (e.g., IP address 192.168.1.100, port number 8080).

[0071] During deployment, if multiple candidate diagnostic models exist, deployment coordinates must be determined independently for each model to avoid computational resource conflicts or data transmission congestion. For example, for tool wear diagnostic models and spindle fault diagnostic models running simultaneously, they can be deployed in different edge computing nodes or cloud containers to ensure that their deployment coordinates are independent and meet performance requirements.

[0072] Furthermore, the determination of deployment coordinates must be coordinated with the machine tool's electrical design and network planning. In terms of physical layout, industrial equipment installation standards must be followed to ensure the power supply stability, heat dissipation, and vibration resistance of computing nodes. In terms of network architecture, independent industrial control network segments must be defined to ensure the security and real-time performance of data transmission. For future equipment upgrades or process adjustments, the deployment coordinates of the diagnostic model can be dynamically adjusted by reassessing the sensor layout and computing resource requirements to adapt to new fault diagnosis needs.

[0073] Example 2:

[0074] In deep learning-based CNC tool fault diagnosis methods, after generating the fault diagnosis configuration results containing the target diagnostic model and its deployment coordinates, an auxiliary parsing network needs to be further constructed to improve the parsing capability of fault features. The construction of the auxiliary parsing network is based on the principle of redundant feature compensation, introducing additional monitoring nodes to cover feature regions that the target diagnostic model cannot fully parse. The specific implementation is as follows:

[0075] The system needs to acquire the second characteristic parameters of multiple candidate auxiliary monitoring nodes. These parameters include the monitoring channel identifier and the auxiliary resolution range. The monitoring channel identifier is a unique identifier for each candidate auxiliary monitoring node, which can be encoded using a combined encoding method, such as "SENSOR-<type>-<location>-<serial number>", where the type can be divided into vibration (VIB), temperature (TEMP), noise (NOISE), etc., the location is represented by a coordinate range in the machine tool coordinate system, and the serial number is the number of the sensor of that type at a specific location. For example, the monitoring channel identifier of a temperature sensor installed near the spindle bearing housing can be "SENSOR-TEMP-(X100,Y200,Z300)-01". The auxiliary resolution range defines the redundant feature area that each candidate auxiliary monitoring node can effectively cover. This area is represented by a set of feature vectors in the fault feature space, such as vibration characteristics in a specific frequency range (e.g., high-frequency vibration components of 5-10kHz), temperature change characteristics under specific cutting parameters (e.g., the tool temperature rise trend when the feed rate is greater than 0.2mm / r), etc.

[0076] The system needs to calculate the degree of redundant feature requirements for each target diagnostic model based on its maximum feature resolution capability under the target equipment compatibility parameters. The maximum feature resolution capability of a target diagnostic model is determined by its hidden layer structure, activation function, and training parameters, and can be quantified by the model's feature response curve. For example, the feature response curve of a certain convolutional neural network model for tool chipping faults shows that its feature extraction accuracy is 90% in the 1-5kHz frequency range, but drops to 70% in the 5-10kHz range, indicating that the model has a feature resolution blind spot in the high-frequency band. By analyzing such feature response curves, the system determines the degree of feature requirements of the model for different frequency bands and different time domains, and generates corresponding redundant feature compensation information. This information is stored in the form of a feature vector matrix, where each element represents the compensation requirement value in a specific feature dimension. For example, the matrix element [3,5] corresponds to a vibration feature compensation requirement value of 0.3 (range 0-1) in the 5-10kHz frequency band.

[0077] Based on the redundant feature compensation information and the auxiliary resolution range of candidate auxiliary monitoring nodes, the system selects candidate auxiliary monitoring nodes with matching coverage ranges. The matching process uses a feature space overlap calculation method, with the following steps: The redundant feature compensation information is converted into a demand region in the feature space, where each demand region is defined by a feature dimension (e.g., frequency, time-domain statistics) and demand intensity (compensation demand value); the auxiliary resolution range of each candidate auxiliary monitoring node is converted into a supply region in the feature space, also defined by a feature dimension and coverage intensity (sensor sensitivity, measurement accuracy, etc.); the overlap volume between each supply region and the demand region is calculated, with a larger overlap volume indicating a higher matching degree. For example, if the auxiliary resolution range of a temperature sensor covers the temperature change characteristics of the tool cutting area at a specific feed rate, and this characteristic happens to have a high demand intensity in the redundant feature compensation information of the target diagnostic model, then the sensor has a high matching degree. The system sets an overlap threshold (e.g., 0.6) and selects candidate auxiliary monitoring nodes with overlap exceeding the threshold as candidate nodes.

[0078] For each candidate auxiliary monitoring node, the system needs to calculate its total single-transmission delay for feature compensation. This delay is determined by the transmission path of data from the auxiliary node to the deployment coordinates of the target diagnostic model. The specific calculation steps are as follows: First, determine the physical or network path of data transmission. The physical path includes cable length, the number of signal conditioning modules, etc., while the network path includes network hop count, switch forwarding delay, etc. Second, calculate the transmission delay based on the path parameters. For example, for analog signal transmission, the delay is determined by the cable propagation speed and signal conditioning time; for digital signal transmission, the delay is determined by network bandwidth, data packet size, and protocol overhead. Finally, sum the delays of each path segment to obtain the total single-transmission delay for feature compensation. For example, data from a candidate temperature sensor needs to pass through a signal amplifier, an A / D converter, and an industrial Ethernet switch before finally being transmitted to the target diagnostic model deployed on the edge computing node. Its total delay is the sum of the delays of each component (e.g., amplifier delay 0.2ms + A / D conversion delay 0.3ms + network transmission delay 1.5ms = 2ms).

[0079] In addition to transmission latency, the system also needs to consider the deployment latency of alternative auxiliary monitoring nodes. Deployment latency includes physical installation time (such as the time required for sensor mounting, wiring, and calibration) and logical configuration time (such as the time required for communication parameter settings and data acquisition program deployment). Physical installation time is related to the sensor type and the complexity of the installation location; for example, embedded sensors typically take longer to install than surface-mount sensors. Logical configuration time is related to the difficulty of system integration; for example, sensors that need to be integrated with existing PLC systems have longer configuration times. The system determines the deployment latency value for each alternative auxiliary monitoring node based on historical project data or engineering estimates.

[0080] By combining the total latency of a single feature compensation transmission with the deployment latency, the system determines the comprehensive analysis cost of each candidate auxiliary monitoring node. The comprehensive analysis cost is calculated using a weighted summation method, with weighting factors set according to the system's priority on real-time performance and deployment efficiency. For example, if the system prioritizes real-time performance, the transmission latency weight can be set to 0.7, and the deployment latency weight to 0.3; conversely, if deployment efficiency is prioritized, the weights can be adjusted to 0.5 and 0.5 respectively. The calculation formula is: Comprehensive Analysis Cost = Transmission Latency × Transmission Weight + Deployment Latency × Deployment Weight.

[0081] To select the optimal combination from the candidate auxiliary monitoring nodes, the system obtains a preset comprehensive analysis cost threshold. This threshold is set based on the overall system performance requirements and resource constraints. For example, if the system requires the total latency of the auxiliary analysis network to not exceed 10ms and the deployment time to not exceed 8 hours, then the comprehensive analysis cost threshold can be set to the corresponding value. The system calculates the total network latency of all possible combinations of candidate auxiliary monitoring nodes. The total network latency is obtained by summing the total latency of each node's characteristic-compensated transmission, taking into account the data synchronization overhead between nodes. For example, if three candidate auxiliary monitoring nodes are activated simultaneously, with transmission latencies of 2ms, 3ms, and 4ms respectively, and a data synchronization overhead of 1ms, then the total network latency is 2 + 3 + 4 + 1 = 10ms.

[0082] The system selects the combination with the smallest difference between total delay and threshold as the target auxiliary monitoring node combination. If multiple combinations have the same difference, the deployment costs (such as sensor procurement costs and installation fees) are further compared, and the combination with the lowest cost is selected. Finally, an auxiliary parsing network configuration result containing the target auxiliary monitoring node combination and its configuration parameters is generated. The configuration parameters include the installation location of each node, communication protocol, sampling frequency, data preprocessing algorithm, etc., which are determined according to the requirements of the target diagnostic model and the characteristics of the candidate auxiliary monitoring nodes.

[0083] During deployment, if the target diagnostic model requires multiple auxiliary monitoring nodes to cover different redundant feature areas, the system must ensure that data acquisition and transmission between nodes do not conflict. For example, time slicing technology can be used to allocate data acquisition time periods for each node to avoid network congestion caused by nodes with the same sampling frequency transmitting data simultaneously; a data priority mechanism can be adopted to assign higher transmission priority to key feature data (such as features corresponding to high compensation demand values) to ensure that they arrive at the target diagnostic model in a timely manner.

[0084] Furthermore, the auxiliary analysis network configuration results need to work in conjunction with the original fault diagnosis configuration results. The data collected by the target auxiliary monitoring nodes needs to be fused with the data from the original monitoring nodes. The fusion algorithm can employ weighted averaging, Kalman filtering, or multimodal fusion methods from deep learning. The fused data is then input into the target diagnostic model to improve the model's ability to analyze complex fault characteristics.

[0085] The auxiliary analysis network is scalable for later system upgrades or process adjustments. The system can reassess the redundancy feature compensation requirements of the target diagnostic model and dynamically adjust the combination of auxiliary monitoring nodes. For example, when the machining process changes from milling to turning, the distribution of tool fault characteristics changes. The system can enable or disable some auxiliary monitoring nodes, or adjust the sampling parameters of the nodes, according to the new compensation requirements, to adapt to the new fault diagnosis needs.

[0086] The auxiliary analytical network constructed in the above manner can effectively cover the feature analysis blind spots of the target diagnostic model, improving the comprehensiveness and accuracy of CNC tool fault diagnosis. This process is based on specific feature space analysis, transmission path calculation, and cost assessment, and does not rely on hypothetical experimental data, ensuring the engineering practicality and operability of the method.

[0087] Example 3:

[0088] In deep learning-based CNC tool fault diagnosis methods, the calculation of dynamic processing total latency is a crucial step in selecting the target diagnostic model. This latency is determined by multiple factors, including data transmission latency between the monitoring node and the candidate diagnostic model, the computational cost of model hidden layer activation, and the resource overhead required for model updates. The specific implementation method is as follows:

[0089] First, for each monitoring node and candidate diagnostic model's data transmission path, the system needs to extract several key parameters to calculate the single-point resolution latency. These parameters include transmission path length, network topology, and data compression ratio. For physical connections, transmission path length refers to the actual length of the cable or fiber optic cable between the sensor and the computing node; for network connections, it refers to the total physical distance traveled by data packets within the network. Network topology describes the number and connection method of network devices (such as switches and routers) through which data transmission occurs; different topologies introduce different forwarding delays. Data compression ratio reflects the degree to which data is compressed before transmission; a higher compression ratio results in a smaller amount of data transmitted, but may increase the computation time required for decompression.

[0090] Based on the above parameters, the system determines the single-point resolution delay through the following steps: For the physical transmission path, the delay mainly consists of the signal propagation time in the medium and the processing time of the interface device. The signal propagation time can be expressed by the formula t. prop =Calculated as L / v, where L is the path length and v is the speed of signal propagation in the medium (e.g., approximately 2 × 10⁻⁶ in copper cable). 8 m / s, approximately 2.2 × 10⁻⁶ m / s in optical fiber. 8The processing time of the interface device includes the time required for signal conversion, protocol encapsulation, and other operations, which can be obtained from the device specifications or actual measurements. For the network transmission path, the delay consists of four parts: propagation delay, queuing delay, processing delay, and transmission delay. Propagation delay is related to the physical path length, queuing delay depends on the network device load, processing delay refers to the time it takes for the device to process data packets (such as checking headers and looking up routing tables), and transmission delay refers to the time required to send data packets to the link, which is related to the link bandwidth and data packet size. The system sums the delays of each path segment to obtain the single-point resolution delay t from the monitoring node to the alternative diagnostic model. single .

[0091] The total resolution delay T can be obtained by summing the individual resolution delays of all monitoring nodes. parse For example, if the system contains three monitoring nodes, their single-point resolution latency is t. single1 t single2 t single3 The total resolution delay is T. parse =t single1 +t single2 +t single3 .

[0092] Next, the system needs to calculate the cyclic computation cost of the hidden layer activation of the candidate diagnostic model. This cost is closely related to the model's maximum feature resolution capability. A stronger maximum feature resolution capability usually means a more complex model structure, more hidden layer neurons, and a greater computational cost. The system first obtains the hidden layer activation parameters associated with the maximum feature resolution capability of the candidate diagnostic model. These parameters include the activation function type (such as ReLU, Sigmoid, Tanh, etc.), threshold adjustment frequency, and neuron connection weight update strategy. Different activation functions have different computational complexities. For example, the Sigmoid function requires exponential operations, making its computational cost higher than that of the ReLU function. A higher threshold adjustment frequency means more comparison and judgment operations are required for each calculation. The connection weight update strategy affects the computational cost of parameter updates.

[0093] Based on these hidden layer activation parameters, the system determines the activation response parameter R. The activation response parameter is a comprehensive indicator reflecting the computational resources and time required for hidden layer activation when processing a unit amount of data. For example, for a hidden layer containing 1000 neurons, if the Sigmoid activation function is used, each activation requires 1000 exponentiation operations and 1000 division operations, resulting in a relatively high activation response parameter R; if the ReLU activation function is used, each activation only requires 1000 comparison operations, resulting in a relatively low R. By analyzing the model structure and activation parameters, and considering the computational performance of the hardware platform (such as CPU clock speed and GPU floating-point operation capability), the system calculates the cycle computational cost C for hidden layer activation. act The calculation formula is:

[0094] C act =R×B×T cycle

[0095] Where R is the activation response parameter, reflecting the computational complexity per unit of data; B is the batch size of data processed each time (i.e., the number of data samples input into the model at one time); T cycle The time required for a hardware platform to complete a basic computational operation (such as the time it takes for the CPU to perform a floating-point operation).

[0096] Model update cost is the third component of the total latency in dynamic processing. During real-time fault diagnosis, alternative diagnostic models need to continuously update their parameters based on new data to adapt to changes in tool condition. Model update cost depends on the frequency of parameter updates and the amount of data. The parameter update frequency is related to the model's learning rate and the rate of data change; a higher learning rate and faster data changes result in a higher parameter update frequency. The amount of data is related to the model's size (e.g., the number of parameters) and the update method (e.g., full update or incremental update).

[0097] The system first determines the amount of data D required for each parameter update. update This data volume includes parameter values ​​that need to be updated, gradient information, etc. Then, based on the data transmission bandwidth B... bandwidth and data processing rate P rate Calculate the model update cost C update Data transmission time can be determined via D. update / B bandwidth Calculation and data processing time can be reduced by D update / P rate The sum of the two is the model update cost C. update .

[0098] Total resolution delay T parse Cycle calculation cost C act and model update cost C updateBy performing a weighted summation, the total dynamic processing delay T can be generated. total The weighting factors are set according to the system's priorities regarding real-time performance, computational resource consumption, and model adaptability. For example, if the system has high real-time requirements, the weight of total parsing latency can be set to 0.5, the weight of periodic computation cost to 0.3, and the weight of model update cost to 0.2; if more emphasis is placed on model adaptability, the weights can be adjusted to 0.3, 0.3, and 0.4. The formula for dynamically processing total latency is:

[0099] T total =w1×T parse +w2×C act +w3×C update

[0100] Where w1, w2, and w3 are the weighting factors for total parsing delay, periodic computation cost, and model update cost, respectively, and w1+w2+w3=1.

[0101] In actual calculations, the system needs to consider the dynamic characteristics of each component. For example, the total parsing latency may fluctuate with changes in network load, the periodic computation cost may vary depending on the complexity of the data processed by the model, and the model update cost may be affected by the parameter update strategy. To accurately reflect these dynamic characteristics, the system uses a sliding window statistical method to collect historical data for each parameter within a certain time window and calculate its average, maximum, and standard deviation, etc., as the basis for calculating the total latency for dynamic processing.

[0102] For complex systems containing multiple monitoring nodes and candidate diagnostic models, the calculation of the total dynamic processing latency needs to be performed in parallel. The system creates an independent computation thread for each candidate diagnostic model, simultaneously calculating its total dynamic processing latency in combination with each monitoring node. During the computation process, the system monitors the execution status of each thread in real time to ensure the rational allocation and utilization of computing resources.

[0103] The calculated total dynamic processing latency will serve as a crucial basis for selecting the target diagnostic model. The system compares the total dynamic processing latency of each candidate diagnostic model and selects the model with the lowest latency as the target diagnostic model. Furthermore, the total dynamic processing latency can also be used to evaluate the real-time performance of the entire fault diagnosis system, providing a reference for system optimization. For example, if the total dynamic processing latency of a candidate diagnostic model is mainly contributed by the total resolution latency, then optimizing the sensor layout or data transmission path can be considered; if it is mainly contributed by the cycle calculation cost or model update cost, then adjusting the model structure or update strategy can be considered.

[0104] During system operation, the total dynamic processing latency varies with changing operating conditions. For example, when tool wear intensifies, the complexity of monitoring data increases, potentially leading to higher cycle calculation costs; when network congestion occurs, the total parsing latency may increase. The system continuously monitors the total dynamic processing latency and adjusts its diagnostic strategies in real time to ensure the real-time nature and accuracy of fault diagnosis.

[0105] The total dynamic processing latency calculated using the above method comprehensively considers multiple factors such as data transmission, model calculation, and parameter updates, accurately reflecting the performance of the alternative diagnostic models in a real-world operating environment. This calculation process is based on specific system parameters and operating conditions, and does not rely on hypothetical experimental data, ensuring the objectivity and practicality of the method.

[0106] Example 4:

[0107] In deep learning-based CNC tool fault diagnosis methods, the selection of the target diagnostic model needs to comprehensively consider both the total dynamic processing latency and the basic data cost, the latter involving the efficiency optimization of training data transmission and storage configuration. The following describes the specific implementation method in detail, using a specific application scenario:

[0108] Assume the target machine tool is a five-axis machining center of a certain model, equipped with three types of monitoring nodes: vibration, acoustic emission, and cutting force, deployed near the spindle, tool holder, and feed axis motor, respectively. A target model needs to be selected from three candidate diagnostic models (Model A, Model B, and Model C). The training data requirements for each model's baseline analysis model are as follows: Model A requires 50GB of historical vibration data and 10GB of acoustic emission data, with a storage capacity of 200GB; Model B requires 30GB of vibration data and 20GB of cutting force data, with a storage capacity of 150GB; Model C requires 40GB of vibration data, 15GB of acoustic emission data, and 5GB of cutting force data, with a storage capacity of 180GB. Candidate data source nodes include local edge storage (located in the machine tool control cabinet, storing the most recent 3 months of data), a workshop-level server (storing 1 year of data, 50 meters from the machine tool), and a cloud database (storing full lifecycle data, with a network latency of approximately 20ms). The candidate storage configuration options include local SSD storage (read / write speed of 500MB / s), distributed file system (deployed on the workshop server, read / write speed of 300MB / s), and cloud storage (read / write speed fluctuates with the network, averaging 200MB / s).

[0109] For each candidate diagnostic model, the system needs to determine its required training data volume and storage capacity. Taking model A as an example, its training data volume consists of 50GB of vibration data and 10GB of acoustic emission data, and the storage capacity requirement is that the model parameters and intermediate results occupy no more than 200GB of storage space. The system selects nodes from the candidate data source nodes that meet the data volume requirements: local edge storage only stores data from the most recent 3 months, which may not be able to cover the historical data volume required by model A; workshop-level servers store 1 year of data, which can meet the complete acquisition of vibration and acoustic emission data; although cloud databases store data throughout the entire life cycle, the impact of network transmission latency on training efficiency needs to be considered.

[0110] Calculate the transmission path length between the candidate data source node and the deployment coordinates of the alternative diagnostic model. Assuming model A is deployed at an edge computing node near the machine tool (physical distance 0.5 meters from local edge storage, 50 meters from the workshop server, and connected to the cloud via the internet), the transmission path parameters are as follows: Local edge storage is connected via a local bus (e.g., PCIe), with negligible transmission latency; the workshop server is connected via Gigabit Ethernet, with a transmission rate of approximately 125 MB / s, and the data transfer time for 50 GB is approximately 50 × 1024 / 125 ≈ 409.6 seconds; cloud transmission is affected by network latency, with an average rate of approximately 50 MB / s and a transmission time of approximately 60 × 1024 / 50 ≈ 1228.8 seconds. Based on the principle of "shortest transmission path while meeting training data volume requirements," model A preferentially selects the workshop-level server as the target data source node, as it meets the data volume requirements and has a significantly shorter transmission time than the cloud.

[0111] Regarding the selection of storage configuration, Model A requires 200GB of storage. Local SSD storage (256GB capacity) meets the space requirements and offers the fastest read / write speed, making it suitable for storing frequently accessed model parameters. While distributed file systems and cloud storage have sufficient capacity, their lower read / write speeds may affect model inference efficiency. Therefore, the target storage configuration for Model A is local SSD storage, with the adaptation path being a PCIe link between the edge computing node and the local SSD, ensuring extremely low latency.

[0112] Similarly, the 30GB vibration data and 20GB cutting force data required for Model B can be obtained from local edge storage (assuming sufficient storage capacity). The transmission path length is 0.5 meters, connected via a local bus, resulting in the shortest transmission time. A storage capacity of 150GB can be met by a local SSD or a distributed file system, but a local SSD offers better speed; therefore, a local SSD is chosen as the target storage configuration. Model C requires mixed-type data (vibration, acoustic emission, and cutting force). Local edge storage may only store vibration data; acoustic emission and cutting force data must be obtained from the workshop server. The transmission path length is 50 meters, and the transmission time is approximately (40+15+5)×1024 / 125≈491.52 seconds. A storage capacity of 180GB can be met by a local SSD or a distributed file system; a local SSD is preferred.

[0113] After determining the target data source node and storage configuration, the system calculates the basic data cost. The basic data cost includes transmission cost and storage cost. The former is related to the transmission path length and data volume, while the latter is related to the storage adaptation path and storage capacity. Taking Model A as an example, the transmission cost mainly comes from the transmission of 50GB of vibration data and 10GB of acoustic emission data from the workshop server. The path length is 50 meters. The unit processing coefficient (such as energy consumption and time cost) can be estimated through hardware specifications: the energy consumption for transmitting 1GB of data via Gigabit Ethernet is approximately 0.1Wh, and the time cost is approximately 8 seconds (125MB / s transmission rate). Therefore, the transmission cost is (50+10)×(0.1Wh+8 seconds)=60×0.1Wh+60×8 seconds=6Wh+480 seconds. The storage cost is the space occupied cost of the local SSD. Assuming a unit storage coefficient of 0.01Wh / GB·day, the cost of 200GB of storage per day is 200×0.01=2Wh. The adaptation path is a local link, and latency costs are negligible.

[0114] Model B has lower data infrastructure costs because data is obtained from local edge storage, resulting in shorter transmission paths and lower energy consumption. Model C requires some data to be obtained from the workshop server, leading to slightly higher transmission costs than Model B, but lower than Model A. Regarding total dynamic processing latency, assuming Model A has a complex hidden layer structure and higher periodic computation costs, Model B has a simple structure but needs to process multiple data types, resulting in longer total parsing latency, while Model C falls somewhere in between. Summing the data infrastructure costs and total dynamic processing latency yields the comprehensive parsing cost: Model A: Transmission cost (6Wh + 480 seconds) + computation latency (assuming 300 seconds) = 780 seconds equivalent cost; Model B: Transmission cost (close to 0) + computation latency (400 seconds) = 400 seconds equivalent cost; Model C: Transmission cost (approximately 5Wh + 492 seconds) + computation latency (350 seconds) = 842 seconds equivalent cost.

[0115] Based on the overall analysis cost ranked from lowest to highest, Model B (400 seconds) < Model A (780 seconds) < Model C (842 seconds). Therefore, Model B was selected as the target diagnostic model. In practical applications, the "shortest" transmission path refers not only to physical distance but also to the optimization of logical paths. For example, prioritizing data source nodes within the same local area network (LAN) rather than cloud nodes across a wide area network (WAN) can reduce network latency and packet loss. The "shortest" storage adaptation path is reflected in the physical proximity of the storage medium and the computing node (e.g., integration on the same hardware platform) or high-speed interface connection (e.g., NVMe SSD).

[0116] If multiple candidate data source nodes meet the data volume requirements, the system needs to further compare the reliability of the transmission path (e.g., historical failure rate of network links) and the freshness of the data (e.g., higher data update frequency of edge storage). For example, when the model requires real-time data for online training, even if the transmission path to the workshop server is long, if the data update delay of edge storage exceeds the model training cycle, the workshop server or the cloud should still be selected as the data source. Storage configuration schemes also need to consider data security. For example, local SSD storage may face the risk of physical damage, while distributed file systems have data redundancy mechanisms; a trade-off must be made based on the system's reliability requirements.

[0117] Furthermore, the calculation of data infrastructure costs needs to consider the reusability of hardware resources. For example, if the workshop server already provides data services to other systems, its additional processing of data requests from Model A may lead to increased load, and the actual transmission latency may be higher than the theoretically calculated value. In this case, it is necessary to obtain the dynamic transmission rate through real-time monitoring. If cloud storage is chosen as the storage configuration solution, the additional bandwidth usage for encrypted data transmission must also be considered, which may increase transmission costs.

[0118] Through the above steps, the system quantifies the basic data cost and total dynamic processing latency into comparable comprehensive indicators, avoiding the limitations of subjective judgment. This process closely revolves around the physical layout of specific equipment, data characteristics, and resource constraints. Through layer-by-layer screening and cost accumulation, it achieves the scientific selection of the target diagnostic model, ensuring optimal overall efficiency of the diagnostic system in data acquisition, storage, and computation.

[0119] Example 5:

[0120] In deep learning-based CNC tool fault diagnosis methods, when the system faces feature confusion caused by complex working conditions, it is necessary to construct a dynamic feature space mapping matrix to improve the accuracy of fault diagnosis. This matrix can adjust the feature dimension weights according to real-time monitoring data, effectively distinguishing different types of fault features. The implementation of Example 5 is described in detail below with reference to a specific application scenario:

[0121] Suppose a CNC milling machine, while machining high-strength alloy steel, simultaneously faces three types of faults: tool wear, chipping, and chatter. The initial feature space includes 10 dimensions, such as time-domain statistics of the vibration signal (e.g., mean, variance, skewness), frequency-domain features (e.g., dominant frequency amplitude, energy distribution), cutting force fluctuation characteristics, and spindle current change rate. Traditional fixed-weight feature extraction methods confuse wear (slowly changing features) with chatter (high-frequency periodic features), leading to decreased diagnostic accuracy.

[0122] The system first performs cluster analysis on historical fault samples to identify the typical characteristic distribution of each fault type. For example, tool wear faults show a gradual increase in energy in the low-frequency range (0-1kHz) of the vibration signal, with the cutting force fluctuation amplitude slowly increasing over time; chipping faults produce sudden energy spikes in the high-frequency range (5-10kHz), with instantaneous changes in cutting force; chatter faults manifest as periodic vibrations at specific frequencies (e.g., 2-3kHz), with synchronous fluctuations in spindle current. Through principal component analysis (PCA), the system found that among the original 10 feature dimensions, 3 principal components can explain approximately 85% of the fault variability: Principal component 1 mainly contains low-frequency vibration energy and the mean cutting force, principal component 2 contains high-frequency vibration energy and the rate of change of current, and principal component 3 contains the vibration amplitude at specific frequencies.

[0123] Based on the clustering results, the system constructs an initial feature space mapping matrix. Each row of this matrix corresponds to a fault type (e.g., wear, chipping, chatter), and each column corresponds to a feature dimension (e.g., mean vibration, dominant frequency amplitude, etc.). The values ​​of the matrix elements represent the importance weight of that feature dimension to a specific fault type, determined by calculating the mutual information or correlation coefficient between the feature and the fault category. For example, for wear faults, the weight of low-frequency vibration energy might be 0.8, while the weight of high-frequency vibration energy might be 0.2; for chipping faults, the weight of high-frequency vibration energy might be 0.9, and the weight of the mean cutting force might be 0.1. The initial matrix reflects prior knowledge of the fault characteristics but does not consider the influence of changes in operating conditions.

[0124] During real-time diagnostics, the system collects sensor data every 50ms, standardizes the current feature vector, and inputs it into a dynamic feature space mapping matrix. The matrix dynamically adjusts its weights based on the statistical characteristics of the current data. For example, if a sudden increase in the overall energy of the vibration signal is detected, the system determines that a chipping or chattering fault may have occurred, and automatically increases the weights of high-frequency vibration energy and specific frequency amplitudes in the mapping matrix. If the cutting force shows a slow upward trend, the weights of dimensions related to the cutting force are increased. This dynamic adjustment is achieved by introducing an adaptive factor, which is related to indicators such as the rate of change of the current data and the outlierness of the feature distribution.

[0125] To evaluate the effectiveness of dynamic adjustments, the system calculates the projection distance of the feature vector onto each fault type subspace before and after adjustment. For example, if the projection distance of the current feature vector in the chipping fault subspace is significantly smaller than that in other fault subspaces after adjustment, the system initially identifies it as a chipping fault. Simultaneously, the system calculates the confidence score of the feature vector, which is determined based on the relative difference in projection distance and the density distribution of the feature space. If the difference in projection distance is small and the feature space density is low, it indicates that the current feature may be at the boundary of a fault category, resulting in a low confidence score. In this case, the system triggers a more detailed feature analysis process.

[0126] In regions of feature confusion, the system initiates a multi-dimensional feature joint analysis mechanism. For example, when vibration features simultaneously indicate wear and chatter, the system incorporates cutting force and spindle current features for joint judgment. By calculating the cross-correlation between different feature dimensions, it is found that wear faults typically exhibit a positive correlation between vibration and cutting force, while chatter faults show synchronous fluctuations in vibration and spindle current. The system utilizes these correlation patterns to construct a decision tree to further distinguish between confused fault types.

[0127] When the confidence level of the diagnostic result falls below a preset threshold (e.g., 0.7), the system initiates a feature space reconstruction process. This process first identifies the feature dimensions that cause confusion. For example, it may be found that the dominant frequency amplitude of a vibration signal responds to both chipping and flutter under certain operating conditions, making differentiation difficult. The system expands the feature space by introducing new feature dimensions (e.g., the phase change rate of the dominant frequency) or transforming existing dimensions (e.g., converting time-domain features to wavelet packet decomposition coefficients). Then, the correlation between features and fault categories is recalculated, the weights of the mapping matrix are updated, and a more discriminative feature subspace is formed.

[0128] The system also records the feature data of each diagnosis and the final confirmed fault type, forming a closed-loop feedback mechanism. Once a sufficient number of new samples have been accumulated, the system uses an incremental learning algorithm to update the initial feature space mapping matrix. For example, for newly discovered mixed fault modes (such as the simultaneous presence of wear and minor chipping), the system identifies their feature distribution through cluster analysis, adds a new fault type row to the mapping matrix, and adjusts the weights of each feature dimension. This incremental update does not affect the ability to identify existing fault types while enhancing the system's adaptability to new fault modes.

[0129] In actual operation, the system may face feature distortion problems caused by sensor drift or noise interference. To address this, a robust weight adjustment mechanism is introduced into the dynamic feature space mapping matrix. When a fluctuation in data of a certain feature dimension is detected to exceed the normal range, the system reduces the weight of that dimension in the mapping matrix while increasing the weight of other stable dimensions. For example, if a vibration sensor generates abnormal high-frequency noise due to loose installation, the system will automatically reduce the weight of the high-frequency vibration energy dimension to avoid misjudgment.

[0130] For scenarios involving multiple tools working collaboratively, the system establishes an independent dynamic feature space mapping matrix for each tool. When a tool malfunctions, the distribution of its feature space changes significantly, while the feature spaces of other tools remain relatively stable. By comparing the changing trends of the feature spaces of each tool, the system can accurately identify the faulty tool and avoid misjudgments caused by interference features generated by the normal operation of other tools.

[0131] The computational load of the dynamic feature space mapping matrix is ​​optimized through an edge-cloud collaborative architecture. The construction of the basic matrix and the analysis of large-scale historical data are completed on the cloud server, while feature extraction of real-time data, dynamic matrix adjustment, and fault diagnosis inference are performed on edge computing nodes. This division of labor reduces the computational burden on edge nodes while ensuring the real-time nature of the diagnosis. After data preprocessing and feature extraction at the edge nodes, only key feature parameters and diagnostic results are uploaded to the cloud, reducing network transmission volume.

[0132] During long-term operation, the system periodically evaluates the performance of the dynamic feature space mapping matrix. Evaluation metrics include fault identification rate, false alarm rate, and false negative rate. When performance degradation is detected, the system automatically triggers a matrix update process, retraining the mapping matrix using the latest historical data to ensure the system maintains high diagnostic accuracy over the long term.

[0133] Through the above implementation methods, the dynamic feature space mapping matrix can effectively address the feature confusion problem under complex working conditions, improving the accuracy and robustness of CNC tool fault diagnosis. The entire process is based on data characteristics and fault modes in the actual production environment, without relying on hypothetical experimental results data, demonstrating the engineering practicality of the industrial intelligent diagnostic system.

[0134] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A CNC tool fault diagnosis method based on deep learning, characterized in that, include: Based on the tool running status data collected by multiple monitoring nodes within a set sampling period, the real-time feature offset corresponding to each monitoring node is determined, and the comprehensive fault feature index set corresponding to multiple monitoring nodes is integrated. After parsing the benchmark analysis model required for the comprehensive fault feature index set, the maximum feature parsing capability of each candidate diagnostic model under the corresponding equipment compatibility parameters is determined. The equipment compatibility parameters are used to characterize the adaptability of the sensor layout in the target machine tool to the deployment conditions and parsing capability of the candidate diagnostic model. The maximum feature parsing capability is determined according to the hidden layer activation parameters associated with the candidate diagnostic model. Based on the maximum feature parsing capability and the benchmark analysis model, at least one candidate diagnostic model is selected from multiple candidate diagnostic models. After calculating the total dynamic processing delay corresponding to the candidate diagnostic model and multiple associated monitoring nodes, a target diagnostic model is selected from at least one candidate diagnostic model based on at least the total dynamic processing delay, and a fault diagnosis configuration result is generated. Determining the maximum feature parsing capability of each candidate diagnostic model under the corresponding device compatibility parameters includes: Obtain the resolution capability parameter table, wherein the table includes equipment compatibility parameters, hidden layer activation threshold, reference distance to the core components of the target machine tool, and the relationship between the three; Based on the deployment coordinates of the candidate diagnostic model, calculate its deviation value from the baseline distance, and query the hidden layer activation threshold corresponding to the deviation value and device compatibility parameters in the table; The maximum feature parsing capability is generated based on the activation threshold.

2. The CNC tool fault diagnosis method based on deep learning according to claim 1, characterized in that, The fault diagnosis configuration results include all selected target diagnostic models, the maximum feature resolution capability of each target diagnostic model under the target device compatibility parameters, and deployment coordinates.

3. The CNC tool fault diagnosis method based on deep learning according to claim 2, characterized in that, After generating the fault diagnosis configuration results, the method further includes: Obtain second feature parameters of multiple candidate auxiliary monitoring nodes, wherein the second feature parameters include monitoring channel identifier and auxiliary resolution range, and the auxiliary resolution range is used to characterize the redundant feature coverage area of ​​the candidate auxiliary monitoring nodes; Based on the maximum feature parsing capability of each target diagnostic model under the target device compatibility parameters, the redundant feature compensation amount information corresponding to each target diagnostic model is determined, wherein the redundant feature compensation amount information is mapped to the corresponding auxiliary parsing coverage information; Based on the redundant feature compensation information and the auxiliary parsing range corresponding to each candidate auxiliary monitoring node, at least one alternative auxiliary monitoring node that satisfies the matching of the auxiliary parsing coverage information is selected from multiple candidate auxiliary monitoring nodes. Based on the monitoring channel identifier corresponding to each of the candidate auxiliary monitoring nodes and the deployment coordinates of the corresponding target diagnostic model, the total single-transmission delay of the feature compensation between the target diagnostic model and at least one of the candidate auxiliary monitoring nodes is calculated. The comprehensive parsing cost is determined according to the deployment delay corresponding to each of the candidate auxiliary monitoring nodes and the total single-transmission delay. The target auxiliary monitoring node is selected from at least one of the candidate auxiliary monitoring nodes, and the auxiliary parsing network configuration result is generated.

4. The CNC tool fault diagnosis method based on deep learning according to claim 3, characterized in that, The comprehensive analysis cost is determined based on the deployment delay corresponding to each of the candidate auxiliary monitoring nodes and the total delay for a single instance, including: Obtain a preset comprehensive parsing cost threshold, wherein the threshold is used to characterize the maximum allowable latency of the redundant parsing network covering all the monitoring nodes; Based on the comprehensive analysis cost, the total network latency corresponding to the configured candidate auxiliary monitoring node combinations is calculated, and the combination with the smallest difference between the total network latency and the comprehensive analysis cost threshold is selected to generate the target auxiliary monitoring node combination.

5. The CNC tool fault diagnosis method based on deep learning according to claim 1, characterized in that, Calculate the total dynamic processing latency of the candidate diagnostic model and the associated multiple monitoring nodes, including: Based on the data transmission path parameters of each monitoring node and the corresponding candidate diagnostic model, the single-point parsing delay of the candidate diagnostic model in processing the feature data of the monitoring node is determined, and the total parsing delay corresponding to the candidate diagnostic model is obtained by summarizing. Determine the activation response parameter associated with the maximum feature resolution capability corresponding to each of the candidate diagnostic models, and determine the periodic computation cost of the hidden layer activation based on the response parameter; The total dynamic processing latency is generated based on the cycle calculation cost, model update cost, and total parsing latency.

6. The CNC tool fault diagnosis method based on deep learning according to claim 1, characterized in that, Selecting a target diagnostic model from at least one of the candidate diagnostic models, based at least on the total dynamic processing latency, includes: Based on the benchmark analysis model corresponding to each candidate diagnostic model, the required amount of training data and storage adaptation amount are determined, and the node with the shortest transmission path and meeting the required amount of training data is selected from multiple candidate data source nodes to generate the target data source node. The scheme with the shortest adaptation path and meeting the required storage adaptation amount is selected from multiple candidate storage configuration schemes to generate the target storage configuration. Based on the transmission path, adaptation path, and corresponding unit processing coefficient, the data base cost of the candidate diagnostic model is calculated, and the data base cost and the total dynamic processing delay are summed to generate the comprehensive analysis cost of the candidate diagnostic model. The target diagnostic model is selected from at least one of the candidate diagnostic models in order of increasing comprehensive analysis cost.

7. The CNC tool fault diagnosis method based on deep learning according to claim 1, characterized in that, Based on the maximum feature parsing capability and the benchmark analysis model, at least one candidate diagnostic model is selected from multiple candidate diagnostic models, including: Determine the feature coverage of the candidate diagnostic model under the maximum feature parsing capability corresponding to each of the device compatibility parameters; Based on the matching degree between the coverage and the benchmark analysis model, a set of configuration parameters containing at least one candidate diagnostic model is generated. Based on the adaptability score corresponding to the device compatibility parameters of each model in the set, the set with the highest total score is selected as the candidate diagnostic model combination.

8. The CNC tool fault diagnosis method based on deep learning according to claim 3, characterized in that, The acquisition of the monitoring channel identifier includes: The operating parameters of the monitoring node are collected in real time through a preset sensor network. The operating parameters include vibration amplitude, acoustic emission frequency and cutting force.

9. The CNC tool fault diagnosis method based on deep learning according to claim 8, characterized in that, The acquisition frequency of the operating parameters is adaptively adjusted according to the fluctuation range of the real-time feature offset.

Images