A dryer fault diagnosis method based on deep reinforcement learning

Abstract

The application relates to the field of fault diagnosis, and discloses a drying machine fault diagnosis method based on deep reinforcement learning, which is used for improving the accuracy and control efficiency of drying machine fault diagnosis. By constructing a physical dynamic model combining thermodynamics and mechanical dynamics, the system stability is quantitatively evaluated in real time by combining Lyapunov functions, and the fault propagation path is intuitively displayed by using a directed topological network. First, the dynamic model is established by acquiring the multi-source parameters of the drying machine; the fault source component is located and its network influence degree is calculated; adaptive control instructions and topological inhibition instructions are generated, the control parameters are dynamically adjusted, the fault propagation is inhibited, the control instructions are converted into actuator driving signals, and the precise control of the drying machine is completed. Through the deep fusion of the physical model and the data-driven model, the fault diagnosis accuracy and the control efficiency are significantly improved, the fault troubleshooting time is effectively shortened, the production safety is ensured, and the application is suitable for intelligent operation of the drying machine under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and in particular to a method for diagnosing dryer faults based on deep reinforcement learning. Background Technology

[0002] In modern industrial production, dryers are key equipment widely used in food processing, chemical industry, pharmaceutical industry, and many other fields. Their operating status directly affects production efficiency and product quality. With the continuous improvement of industrial automation, the complexity and integration of dryers are increasing, leading to a significant increase in the difficulty of fault diagnosis and control. Traditional dryer fault diagnosis methods mainly rely on human experience, regular maintenance, or monitoring systems based on simple rules. These methods often exhibit limitations such as low diagnostic efficiency, poor accuracy, and slow response speed when faced with complex and changing operating conditions and potential fault modes.

[0003] The existing technology has the following shortcomings:

[0004] Traditional methods are often based on simplified physical models or empirical formulas, which are difficult to accurately reflect the complex thermodynamic and mechanical dynamic processes inside the dryer, thus limiting the accuracy of fault diagnosis.

[0005] The lack of an effective stability assessment mechanism makes it difficult to judge the system's operating status in real time and accurately, and makes it impossible to provide timely warnings or take control measures before a failure occurs.

[0006] Traditional methods are not good at visually demonstrating the propagation path of faults among the various components inside the dryer, which is not conducive to quickly locating the source of the fault and taking targeted measures.

[0007] Control commands are often generated based on fixed rules or preset parameters, lacking adaptability and flexibility, and are difficult to cope with complex and ever-changing working conditions and fault modes.

[0008] When processing multi-source sensor data, traditional methods often lack effective signal filtering, fusion, and analysis capabilities, resulting in low information utilization and low fault diagnosis efficiency.

[0009] Therefore, we propose a dryer fault diagnosis method based on deep reinforcement learning to solve the above problems. Summary of the Invention

[0010] This invention provides a dryer fault diagnosis method based on deep reinforcement learning, which can improve the accuracy of dryer fault diagnosis and control efficiency.

[0011] The first aspect of this invention provides a dryer fault diagnosis method based on deep reinforcement learning. The method includes: acquiring the thermodynamic and mechanical parameters of the dryer; establishing a system of differential equations for state variables to obtain a physical dynamic model; constructing a scalar function based on the sum of squared deviations between the state variables and the target value, according to the physical dynamic model and a preset target value, to obtain a Lyapunov function; acquiring the mechanical structural connection relationships of the dryer, mapping bearings, rotors, fans, and heaters as nodes, and mapping fault propagation paths as directed edges, to obtain a directed topology network; acquiring real-time data from temperature sensors, humidity sensors, and vibration sensors; calculating the time rate of change of the Lyapunov function; outputting the original DRL instruction when the time rate of change is less than a preset threshold, otherwise triggering a stability compensator to generate a correction instruction, to obtain an adaptive control instruction; locating the fault source component in the directed topology network, calculating the network influence of the fault source component, generating suppression actions for high-influence components, to obtain a topology suppression instruction; and fusing the adaptive control instruction and the topology suppression instruction, converting them into actuator drive signals to obtain a physical control signal.

[0012] Optionally, in a first implementation of the first aspect of the present invention, the method includes: establishing a temperature change rate equation and a humidity change rate equation based on the thermodynamic parameters of the dryer to obtain a heat conduction sub-model and a proton transfer model; establishing a rotational speed change rate equation and a wind speed correlation equation based on the mechanical parameters of the dryer to obtain a mechanical dynamics sub-model; and correlating the thermo-mechanical equations and coupling the humidity-wind speed equations based on the heat conduction sub-model, the proton transfer model, and the mechanical dynamics sub-model to obtain a complete set of dynamic equations containing four-variable coupling relationships.

[0013] Optionally, in the second implementation of the first aspect of the present invention, the method includes: calculating temperature deviation, humidity deviation, and rotational speed deviation based on the state variables and preset target values ​​in the physical dynamic model; squaring each deviation value based on the temperature deviation, humidity deviation, and rotational speed deviation, and then summing them by weight to obtain an energy-type scalar function V; associating the physical dynamic model with the energy-type scalar function V: using V as a measure of the system's energy state; defining a stability condition: the system is stable when V decreases over time; and obtaining a Lyapunov function for stability determination.

[0014] Optionally, in the third implementation of the first aspect of the present invention, the method includes: based on the mechanical structure connection diagram of the dryer, extracting bearing components as first-type nodes, rotor components as second-type nodes, fan components as third-type nodes, and heater components as fourth-type nodes to obtain a set of component nodes with physical identifiers; based on the dryer's fault history records, determining the transmission direction of rotor vibration enhancement caused by bearing wear, the transmission direction of fan noise caused by rotor imbalance, and the transmission direction of bearing temperature rise caused by heater overheating to obtain directed edge definition rules describing the physical path of fault propagation; based on the set of component nodes and the directed edge definition rules, sorting the nodes according to their physical positions, marking the connection relationships in the matrix according to the transmission direction to obtain a binary adjacency matrix representing the fault propagation path; based on the binary adjacency matrix, converting the adjacency matrix into a graph data structure, labeling each directed edge with a fault type identifier to obtain a directed topology network for executable path analysis.

[0015] Optionally, in the fourth implementation of the first aspect of the present invention, the method includes: synchronously reading data from three sensors to obtain real-time measurement values ​​from a temperature sensor, a humidity sensor, and a vibration sensor; performing signal filtering to obtain a pre-processed sensor data packet; substituting the sensor data into the Lyapunov function based on the pre-processed sensor data packet and the Lyapunov function to calculate the system energy value at the current moment, obtaining the real-time energy state value V(t); calculating the energy change per unit time based on the real-time energy state value V(t) at the current moment and the energy state value V(t-1) stored at the previous moment, obtaining the energy change rate; comparing the energy change rate with a preset stability coefficient λ to determine the system stability state, obtaining a stability determination flag; and based on the stability determination flag and the original instructions output by the DRL agent, directly outputting the original DRL instructions when the flag indicates a stable state, and triggering the stability compensator to generate correction instructions when the flag indicates an unstable state, obtaining an adaptive control instruction set.

[0016] Optionally, in the fifth implementation of the first aspect of the present invention, the method includes: based on the fault type signal and the directed topology network, matching the associated nodes in the topology network according to the fault type to determine the starting node of the fault propagation path and obtaining the fault source node identifier; based on the fault source node identifier and the directed topology network, performing graph traversal with the fault source node as the starting point, calculating the node betweenness centrality, and obtaining the node betweenness centrality value; based on the node betweenness centrality value and a preset influence threshold, selecting key nodes with betweenness centrality ≥ the preset influence threshold, generating suppression action parameters for each key node, and obtaining a topology suppression instruction set.

[0017] Optionally, in the sixth implementation of the first aspect of the present invention, the method includes: converting power commands into voltage signals, speed commands into pulse frequency signals, and suppression parameters into switching signals based on an adaptive control command set and a topology suppression command set, thereby obtaining analog drive signals and digital switching signals; based on the analog drive signals and digital switching signals, when a suppression command exists, overwriting the original control command of the corresponding node, and maintaining the original control signal when no suppression command exists, thereby obtaining an arbitration-adjusted execution signal set; based on the arbitration-adjusted execution signal set, outputting analog signals to the heater power regulator through a DAC module, outputting pulse signals to the fan speed controller through a PWM module, and outputting switching signals to the emergency braking unit through a DO module, thereby obtaining physical control signals for driving actuator operation.

[0018] The mechanism of this invention is as follows:

[0019] By modeling with physical laws (replacing data-driven approaches) and proving mathematical stability (replacing empirical parameter tuning), a dryer fault diagnosis and control system with strict stability guarantees is constructed, breaking through the data dependence and security bottlenecks of traditional DRL in industrial scenarios.

[0020] Beneficial effects:

[0021] Traditional methods often rely on a single physical model or a data-driven model, while this invention deeply integrates the two. The physical dynamic model provides the underlying constraints on system operation, while DRL optimizes control strategies through data learning, significantly improving the accuracy of fault diagnosis and the adaptability of control strategies.

[0022] By constructing a Lyapunov function and calculating its rate of change over time, a real-time quantitative assessment of system stability was achieved. This breaks through the traditional stability judgment methods based on thresholds or rules, and provides a more accurate basis for stability determination.

[0023] The dryer's mechanical structure is mapped as a directed topology network, visually displaying the fault propagation path and locating the fault source component based on network analysis. This overcomes the limitations of traditional fault diagnosis that relies on human experience, enabling rapid fault location and propagation path analysis.

[0024] By combining Lyapunov function stability assessment and directed topology network analysis, adaptive control commands and topology suppression commands are generated. The adaptive control commands dynamically adjust control parameters based on the real-time system state, while the topology suppression commands generate suppression actions for high-impact fault source components.

[0025] By simultaneously reading data from multiple sensors, including temperature, humidity, and vibration sensors, and performing signal filtering and fusion processing, the information utilization rate and fault diagnosis efficiency are improved. This overcomes the limitations of traditional methods that rely on a single sensor or simple data processing.

[0026] The system converts control commands such as power and speed into analog drive signals and digital switching signals, and outputs them to the actuator through modules such as DAC, PWM, and DO, thus realizing the intelligent generation and execution of control commands. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of an embodiment of the dryer fault diagnosis method based on deep reinforcement learning in this invention.

[0028] Figure 2 This is a schematic diagram of another embodiment of the dryer fault diagnosis method based on deep reinforcement learning in this invention. Detailed Implementation

[0029] This invention provides a dryer fault diagnosis method based on deep reinforcement learning, which improves the accuracy of dryer fault diagnosis and control efficiency. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the dryer fault diagnosis method based on deep reinforcement learning in this invention includes:

[0031] 101. Construct a physical dynamic model of the dryer. Input sources: thermodynamic parameters (heat capacity coefficient, mass transfer coefficient) and mechanical parameters (moment of inertia) of the dryer. Specific operations: establish a set of differential equations for state variables based on the laws of conservation of energy and mass. Define the product: a physical dynamic model describing the dynamic changes of temperature, humidity, wind speed, and rotational speed.

[0032] It is understood that the executing entity of this invention can be a dryer fault diagnosis device based on deep reinforcement learning, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0033] It should be noted that, taking the fluidized bed dryer as an example, a physical dynamic model is constructed based on the principles of thermodynamics and mechanodynamics, and the specific implementation is as follows:

[0034] Model input parameters, thermodynamic parameters: specific heat capacity of the material, c p = 2.1 kJ / kg·K; heat transfer coefficient (gas-solid) h p =85W / m 2 ·K; Mass transfer coefficient (water evaporation) k m =0.025m / s;

[0035] Mechanical parameters: Fan impeller moment of inertia J = 0.65 kg·m 2 The duct resistance coefficient ζ = 0.15 Pa·s 2 / m 6 ;

[0036] State variables are defined as follows: T(t): drying chamber temperature (°C); H(t): average moisture content of the material (kg water / kg dry material); ω(t): fan speed (rad / s);

[0037] The dynamic equation system is constructed by establishing a system of differential equations based on the laws of conservation of energy and mass.

[0038] 102. Generate the system stability criterion function. Input source: physical dynamic model and preset target values ​​(target temperature value, target humidity value, rated speed value) from step 101. Specific operation: construct a scalar function based on the sum of squared deviations between the state variables and the target values. Defining the product: Lyapunov function that quantifies the energy state of the system.

[0039] It should be noted that the system stability criterion function is generated from the following inputs: physical dynamic model (from step 101): state variables: drying chamber temperature T(t), material humidity H(t), and fan speed ω(t).

[0040] System of differential equations:

[0041]

[0042] Where k1 = 0.02, k2 = 0.5, k3 = 0.03, and k4 = 0.8 represent the coefficients of heat transfer, evaporation, and mechanical inertia, respectively, and Q... in Input power to the heater.

[0043] Preset target value: Target temperature T ref =75℃, target humidity H ref =0.18kg water / kg dry material, rated speed ω ref =3000RPM.

[0044] Lyapunov function construction process, defining state deviation: temperature deviation: e T =TT ref Humidity deviation: e H =HH ref Speed ​​deviation: e ω =ω-ω ref ;

[0045] Construct a scalar function (energy function):

[0046] Weighting design basis: Humidity deviation e H The weight (100) is relatively large, and because the rate of humidity change is slow (k3 = 0.03), its impact on stability needs to be amplified; rotational speed deviation e ω The weight (0.01) is relatively small, and small fluctuations are acceptable because the mechanical response is fast (k4 = 0.8).

[0047] Verify stability condition: Calculate the time derivative of V:

[0048]

[0049] Substitute into the physical model equations:

[0050]

[0051] Stability criterion: If If energy decays, the system becomes stable.

[0052] Practical application scenario, disturbance test: when the heating power Q in A sudden 10% drop (simulated fault) caused temperature T to fall from 75°C to 70°C:

[0053] Initial deviation: e T =-5,e H =0,e ω =0→V=(-5) 2 =25;

[0054] Dynamic response: Temperature recovery period (Stablize);

[0055] like This triggers the stability compensator in step 104.

[0056] Function characteristics and output, product: Lyapunov function

[0057] Function: Quantify the "energy" of the system deviating from the target state (the system is optimal when V=0);

[0058] pass The stability of sign judgment provides mathematical constraints for the generation of subsequent control instructions.

[0059] 103. Establish a fault propagation topology graph. Input source: the mechanical structural connection relationship of the dryer. Specific operation: map the bearings, rotor, fan, and heater as nodes, and map the fault propagation path as directed edges. Define the product: the directed topology network representing the fault propagation path.

[0060] It should be noted that the input source for the fault propagation topology diagram construction is the mechanical structure connection relationship of the dryer (taking a certain type of fluidized bed dryer as an example): core components: main drive motor, transmission bearings (2 sets), fan impeller, heater array (4 sets), temperature sensor, humidity sensor.

[0061] Physical connections: Motor output shaft → Transmission bearing 1 → Transmission bearing 2 → Fan impeller; Fan impeller → Air duct → Heater array → Drying chamber; Heater power cord → Control cabinet → Temperature / humidity sensor feedback loop;

[0062] The topology graph construction process involves node mapping (component → node):

[0063] Component Name Node symbol Property Examples Main drive motor M Rated power 7.5kW, failure rate 0.03 / 1000 hours Transmission bearing 1 <![CDATA[B1]]> <![CDATA[Axial stiffness 1.2×10 8 N / m, wear threshold 0.1 mm]]> Transmission bearing 2 <![CDATA[B2]]> <![CDATA[Same as B1 (symmetrical design)]]> fan impeller F Number of blades: 12; dynamic balance tolerance: 5 g·mm Heater group 1 <![CDATA[H1]]> Resistance value 18Ω, maximum temperature rise 120℃ Temperature sensor <![CDATA[T s ]]> Accuracy ±0.5℃, response time 2s

[0064] Directed edge mapping (fault propagation path):

[0065] Fault source node Fault type Affecting the target node Propagation probability Propagation delay B1 bearing wear B2 0.85 3min B1 Excessive vibration F 0.90 1s F Impeller dynamic imbalance M 0.75 10s H1 heating element rupture Ts 0.95 5s M Overload coil burnout control cabinet 0.60 2min

[0066] Output of topology network (directed graph structure): Node relationships: M→B1→B2→F→airway; H1||H2||H3||H4→T s ;

[0067] Critical path: Mechanical transmission chain: M→B1→B2→F (carrying 85% of mechanical failure propagation);

[0068] Thermodynamic coupling chain: H1→T s →Control cabinet (core path of temperature control failure);

[0069] Actual fault propagation verification, scenario: Bearing B1 wear 0.15mm (exceeding the threshold):

[0070] A fault signal is triggered at node B1 and propagates to B2 via edge (B1, B2) within 3 minutes.

[0071] B2 vibration intensifies → within 1 second, the impeller F becomes unbalanced due to the edge (B2,F);

[0072] The loss of measurement at node F by 8g·mm → the current fluctuation of motor M caused by the edge (F,M) is 22%.

[0073] Topology location: Tracing along the path B1→B2→F→M, the fault source is identified as B1.

[0074] 104. Generate stability constraint control command. Input sources: real-time temperature sensor, humidity sensor, vibration sensor data, Lyapunov function from step 102. Specific operation: calculate the time change rate of the Lyapunov function. When the time change rate is less than the preset threshold, output the original DRL command. Otherwise, trigger the stability compensator to generate a correction command. Definite product: an adaptive control command with stability guarantee.

[0075] It should be noted that the stability constraint control command generation input source: real-time sensor data (sampling period 0.5s): temperature sensor: current drying chamber temperature T = 73℃ (target T) ref =75℃); Humidity sensor: Material humidity H = 0.20 kg water / kg dry material (target H) ref =0.18); Vibration sensor: Fan speed ω = 2950 RPM (target ω ref =3000);

[0076] Lyapunov function (from step 102):

[0077] Where e T =TT ref ,e H =HH ref ,e ω =ω-ω ref .

[0078] Control command generation logic, real-time deviation calculation:

[0079] e T = -2℃,e H = +0.02kg water / kg dry matter, e ω = -50 RPM;

[0080] Current energy state V = (-2) 2 +100×(0.02) 2 +0.01×(-50) 2 =4 + 0.04 + 25 = 29.04;

[0081] Rate of change of Lyapunov function Calculation: Based on the trend of sensor data change (past 1 second):

[0082] ΔT / Δt = -0.8℃ / s (temperature continues to decrease);

[0083] ΔH / Δt = +0.005 (kg water / kg dry material) / s (humidity rises slowly);

[0084] Δω / Δt = -10 RPM / s (speed drops rapidly);

[0085] Substitute into the derivative formula:

[0086]

[0087] in conclusion: This indicates that the system energy is continuously increasing (deviating from a steady state).

[0088] Stability compensator triggered (preset threshold) ):current → Triggers the compensator to generate correction instructions;

[0089] Compensation strategy (proportional compensation + rate suppression): Heating power compensation amount: ΔQ = K p ·|e T |=1.8×2=3.6kW(K p =1.8 is the proportionality coefficient;

[0090] Fan speed correction: Δω = 50 RPM (to offset the current downward trend); Humidity suppression: increase exhaust damper opening by 15% (to accelerate moisture removal);

[0091] Output commands and effect verification, product: Adaptive control commands: Heater target power: 8.0 + 3.6 = 11.6 kW; Fan target speed: 3000 + 50 = 3050 RPM; Exhaust damper opening: 60% → 75%;

[0092] Stability recovery verification (3 seconds after command execution): The energy level dropped to -1.2 (meeting the stability condition of <-0.5); the V value decreased from 29.04 to 18.6 (energy decay rate of 35%), and the system returned to a stable trajectory.

[0093] 105. Generate fault suppression instructions. Input sources: the directed topology network in step 103 and the fault type signal identified in step 104. Specific operations: locate the fault source component in the topology network, calculate the network influence of the fault source component, generate suppression actions for high-influence components, and define the product: a topology suppression instruction to block the spread of faults.

[0094] It should be noted that the fault suppression command is generated from the following input source: fault type signal (from step 104): abnormal vibration of the wind turbine bearing (fault code F103, characteristic frequency 4.5kHz, amplitude 7.3mm / s). 2 ).

[0095] Directed topology network (from step 103): Nodes: main motor (M), bearing 1 (B1), bearing 2 (B2), fan impeller (F), heater group (H1-H4).

[0096] Directed edges: M→B1→B2→F (mechanical transmission chain), H1→F (thermal coupling edge).

[0097] Edge attributes: B1→B2 propagation probability 0.85, B1→F propagation probability 0.90, delay ≤3s.

[0098] The fault suppression command generation process involves fault source localization: based on the vibration signal frequency (4.5kHz), the topology network is matched to identify the fault source as bearing node B1 (which accounts for 62% of historical fault frequencies).

[0099] Related paths: B1→B2→F (main propagation path), B1→F (direct vibration transmission path).

[0100] Network impact calculation: Scope of impact: B1 directly affects B2, F, and M (a total of 3 nodes).

[0101] Comprehensive impact formula: Impact = α·propagation probability + β·number of downstream nodes + γ·failure frequency; Substituting parameters (α=0.6, β=0.3, γ=0.1): Impact B1=0.6×0.85+0.3×3+0.1×0.62=1.97 (high impact level);

[0102] The order is: B1(1.97)>F(1.2)>H1(0.8), so B1 is the first to be suppressed.

[0103] Suppressing action generation: For fault source B1: reduce the fan speed to 2800 RPM (originally 3000 RPM) to reduce the mechanical load on the bearing; inject lubricating coolant (flow rate 0.2 L / min) to suppress wear propagation.

[0104] For high-impact node F: activate dynamic balance compensation (counterweight +8g) to counteract the vibration transmitted by B1; limit the wind speed limit to 2.0m / s (originally 2.5m / s) to reduce the aerodynamic load on the impeller.

[0105] For coupling node H1: Increase the heating power to 9.0kW (from 8.0kW) to compensate for the heat loss caused by the speed reduction.

[0106] Data after suppression (within 3 seconds): B1 vibration amplitude decreased from 7.3 mm / s 2 Reduced to 2.1 mm / s 2 (Reduction of 71%); Fan speed stabilized at 2800±10RPM, temperature fluctuation <±1.5℃; the activity of fault propagation path B1→B2 was reduced to zero (successfully blocked).

[0107] Product: Topology Suppression Instruction Set

[0108] {Command 1: Fan speed = 2800 RPM, for 5 minutes};

[0109] {Command 2: Impeller dynamic balance compensation +8g};

[0110] {Instruction 3: Heater H1 power = 9.0kW}.

[0111] 106. Synthesize the final execution signal. Input sources: the adaptive control command from step 104 and the topology suppression command from step 105. Specific operation: fuse the commands and convert them into actuator drive signals. Defining the product: the physical control signals output to the heater and fan.

[0112] It should be noted that the final signal synthesis input sources are: adaptive control commands (from step 104): heater target power: 11.6kW (originally 8.0kW, +3.6kW due to temperature deviation compensation); fan target speed: 3050RPM (originally 3000RPM, +50RPM to compensate for speed drop); exhaust damper opening: 75% (originally 60%, to accelerate moisture discharge); topology suppression commands (from step 105): fan speed limit: 2800RPM (to suppress bearing vibration); impeller dynamic balance compensation: +8g (to offset vibration); lubrication cooling flow rate: 0.2L / min (to reduce bearing wear); heater power increase: 9.0kW (to compensate for heat loss caused by speed reduction); command fusion and conversion process, conflict resolution and priority allocation:

[0113] Stability priority: When control commands and suppression commands conflict (fan speed: control command requires 3050 RPM vs. suppression command limit of 2800 RPM), the speed limit of the suppression command (2800 RPM) shall prevail to avoid the risk of mechanical overload.

[0114] Compensation and coordination: The heater power is set to 9.0kW of the suppression command (higher than the control command of 11.6kW). Additional thermal compensation is required due to the speed reduction, but it shall not exceed the equipment limit of 12.0kW.

[0115] Parameter fusion rules: Superimposed parameters: Lubrication and cooling flow rate (0.2L / min) is directly superimposed onto the execution signal. Overriding parameters: Impeller dynamic balance compensation (+8g) overrides the original counterweight setting. Range constraints: Exhaust damper opening of 75% is within the safe range (<80%) and requires no adjustment.

[0116] Physical signal conversion: Heater drive signal: Power command → Current signal (9.0kW → 37.5A, based on heater resistance 24Ω).

[0117] Fan drive signals: Speed ​​command → Inverter frequency (2800RPM → 46.67Hz, based on the number of motor pole pairs 4). Dynamic balancing command → Electromagnetic compensator current (+8g → 1.2A, based on the calibration coefficient 6.67g / A).

[0118] Auxiliary execution signals: Lubrication flow rate → proportional valve opening (0.2L / min → 15% opening, based on the flow curve). Exhaust damper → servo motor angle (75% → 67.5°).

[0119] Output instructions and execution results, final physical control signals:

[0120] Actuator signal type numerical values heater Current signal 37.5A Fan frequency converter frequency signal 46.67Hz Dynamic balancing compensator Current signal 1.2A Lubricated proportional valve Opening signal 15％ Exhaust servo motor angle signal 67.5°

[0121] Effect verification (5 seconds after command issuance): Bearing vibration amplitude increased from 7.3 mm / s. 2 Reduced to 1.8 mm / s 2 (75% reduction); the temperature of the drying chamber was stabilized at 74.2℃ (target 75±1℃), and the humidity was reduced to 0.19kg / kg (target 0.18kg / kg); the fan speed was stabilized at 2800±5RPM, with no abnormal fluctuations.

[0122] In this embodiment of the invention, key steps are covered, including physical dynamic model construction, generation of Lyapunov function stability criteria, establishment of fault propagation topology, generation of stability constraint control instructions, generation of fault suppression instructions, and synthesis of final execution signals. By integrating physical models and data-driven methods, precise diagnosis and efficient control of the dryer's operating status are achieved. Combining thermodynamics and mechanical dynamics principles, a precise physical dynamic model is constructed, significantly improving the accuracy of fault diagnosis. By constructing a Lyapunov function to quantify the system's energy state, real-time quantitative assessment of system stability is realized, providing mathematical constraints for control command generation and enhancing system stability. Mapping the dryer's mechanical structure to a directed topological network visually displays the fault propagation path, enabling rapid fault location and propagation path analysis, shortening fault troubleshooting time. Based on the rate of change of the Lyapunov function, adaptive control commands with stability guarantees are generated, ensuring stable system operation under complex conditions. By locating fault source components through the topological network and calculating their network influence, topology suppression commands to block fault propagation are generated, improving system reliability and maintenance efficiency. The integration of adaptive control commands and topology suppression commands, converted into actuator drive signals, achieves precise control of the dryer, improving production efficiency and product quality.

[0123] Please see Figure 2 Another embodiment of the dryer fault diagnosis method based on deep reinforcement learning in this invention includes:

[0124] 201. Construct a physical dynamic model of the dryer. Input sources: thermodynamic parameters (heat capacity coefficient, mass transfer coefficient) and mechanical parameters (moment of inertia) of the dryer. Specific operations: establish a set of differential equations for state variables based on the laws of conservation of energy and mass. Define the product: a physical dynamic model describing the dynamic changes of temperature, humidity, wind speed, and rotational speed.

[0125] Specifically: Sub-step S110: Define the thermodynamic state change equation. Input source: thermodynamic parameters of the dryer (heat capacity coefficient, mass transfer coefficient). Specific operation: Establish the temperature change rate equation, establish the humidity change rate equation, and define the products: heat conduction sub-model and proton transfer model.

[0126] Sub-step S120: Define the mechanical dynamics equations. Input sources: mechanical parameters of the dryer (moment of inertia, coefficient of friction); specific operations: establish the rotational speed change rate equation; establish the wind speed correlation equation; define the product: mechanical dynamics sub-model;

[0127] Sub-step S130: Construct a coupled dynamic equation set. Input sources: heat conduction sub-model and proton transfer sub-model of sub-step S110, and mechanical dynamics sub-model of sub-step S120. Specific operations: correlate the thermo-mechanical equations; couple the humidity-wind speed equations; define the product: a complete dynamic equation set containing four-variable coupling relationships.

[0128] It should be noted that, taking a certain model of dual-energy complementary grain dryer as an example, its core parameters are as follows:

[0129] Thermodynamic parameter: heat capacity coefficient C P = 8.5 kJ / kg\cdotpK, mass transfer coefficient k m =0.15min -1 ;

[0130] Mechanical parameters: Fan rotational inertia J = 12.5 kg / pm 2 The coefficient of friction f = 1.2 N·m·ps / rad, and the wind speed conversion factor k v =0.8m\cdotps -1 \cdotprad -1 .

[0131] Sub-step S110: Define the thermodynamic state change equation and the temperature change rate equation:

[0132] Input heat Q in =25kW (heater power);

[0133] Heat loss Q loss Includes convective heat dissipation and mechanical energy consumption:

[0134] Qloss =hA(TT) env )+ηω 3 = (35×15)×(60-25)+0.02×1200 3 =18.375 + 34.56 ≈ 52.935 kW;

[0135] (h=35W / m 2 \cdotpK, A = 15m 2 Ambient temperature T env =25℃, mechanical efficiency η=0.02, rotational speed ω=1200rpm).

[0136] Temperature dynamic equation:

[0137]

[0138] Humidity change rate equation:

[0139] Saturated humidity H sat =0.18 kg / kg (equilibrium moisture content of grain);

[0140] Current humidity H = 0.22 kg / kg:

[0141]

[0142] Products: Thermal conductor model (temperature change) and proton transfer model (humidity change).

[0143] Sub-step S120: Define the mechanical dynamics equations and the equation for the rate of change of rotational speed:

[0144] Motor torque τ motor =65N\cdotpm;

[0145] Dynamic equation of rotational speed:

[0146] (Current rotational speed ω=20rad / s≈191rpm).

[0147] Wind speed correlation equation: v = 0.8 × 20 = 16 m / s

[0148] Product: Mechanical dynamics sub-model (relationship between rotational speed and wind speed).

[0149] Sub-step S130: Construct the coupled dynamic equation system, thermo-mechanical coupling:

[0150] Mechanical energy consumption ηω 3 The heat loss item (see S110) is included to reflect the impact of motor power consumption on temperature rise.

[0151] Humidity-wind speed coupling:

[0152] The mass transfer equation is modified by introducing a wind speed of v = 16 m / s:

[0153]

[0154] (Wind speed accelerates moisture diffusion, increasing the drying rate by 16 times).

[0155] Complete dynamic equation system:

[0156]

[0157] v = k v ω

[0158] Parameter table:

[0159] variable value Physical meaning T 60℃ Drying chamber temperature H 22kg / kg Current moisture content of grain ω 20 rad / s Fan speed v 16m / s Drying wind speed

[0160] 202. Generate the system stability criterion function. Input source: physical dynamic model and preset target values ​​(target temperature value, target humidity value, rated speed value) from step 201. Specific operation: construct a scalar function based on the sum of squared deviations between the state variables and the target values. Defining the product: Lyapunov function that quantifies the energy state of the system.

[0161] Specifically: Sub-step S210: Calculate the deviation of state variables. Input source: State variables (temperature, humidity, speed) in the physical dynamic model of step 201, preset target values ​​(target temperature, target humidity, rated speed). Specific operations: Calculate temperature deviation; calculate humidity deviation; calculate speed deviation; define the products: temperature deviation value, humidity deviation value, speed deviation value.

[0162] Sub-step S220: Construct an energy scalar function. Input sources: temperature deviation, humidity deviation, and rotational speed deviation values ​​from sub-step S210. Specific operations: Square each deviation value; perform a weighted summation; where the weighting coefficients are determined by thermodynamic sensitivity; define the product: an energy-type scalar function V.

[0163] Sub-step S230: Generate the stability criterion function. Input source: energy scalar function V from sub-step S220. Specific operations: Associate with physical dynamic model: use V as a measure of the system's energy state; define the stability condition: the system is stable when V decreases with time; define the product: Lyapunov function used for stability determination.

[0164] It should be noted that, taking a certain model of grain dryer as an example, its preset target operating state is as follows:

[0165] Target value: Temperature T ref =60℃, humidity H ref =15%, fan rated speed ωnom = 1200 rpm (equivalent to 125.66 rad / s).

[0166] Real-time status (sensor data): Temperature T = 65℃, humidity H = 18%, rotational speed ω = 1300 rpm (136.14 rad / s).

[0167] Sub-step S210: Calculate the state variable deviation, temperature deviation: e T =TT ref =65-60=5℃;

[0168] Humidity deviation: e H =HH ref =18-15=3%;

[0169] Speed ​​deviation: e ω =ω-ω non =136.14-125.66=10.48rad / s;

[0170] Product: Deviation value e T =5,e H =3,e ω =10.48.

[0171] Sub-step S220: Construct the energy scalar function and assign weighting coefficients (based on thermodynamic sensitivity):

[0172] Temperature has the highest sensitivity (temperature rise can easily cause equipment deformation), with a weight of ω1 = 0.6; humidity is the second highest (affects drying efficiency), with a weight of ω2 = 0.3; and rotational speed has the lowest sensitivity (strong mechanical inertia buffering), with a weight of ω3 = 0.1.

[0173] Energy function calculation:

[0174]

[0175] Product: Scalar function V = 14.34 (dimensionless energy value).

[0176] Sub-step S230: Generate stability criterion function and associate it with the physical model:

[0177] V is considered as the system's "energy state": the larger the value, the further it deviates from the target;

[0178] Stability condition is defined as (Energy decreases over time.)

[0179] Stability determination logic:

[0180] like (λ = 0.5 is the preset threshold), the system is stable;

[0181] like Trigger the stability compensator.

[0182] Product: Lyapunov function V and its derivative criterion

[0183] 203. Establish a fault propagation topology graph. Input source: the mechanical structure connection relationship of the dryer. Specific operation: map the bearings, rotor, fan, and heater as nodes, and map the fault propagation path as directed edges. Define the product: the directed topology network representing the fault propagation path.

[0184] Specifically: Sub-step S310: Identify key physical component nodes. Input source: Mechanical structure connection diagram of the dryer. Specific operations: Extract bearing components as first-type nodes; extract rotor components as second-type nodes; extract fan components as third-type nodes; extract heater components as fourth-type nodes; Determine the product: A set of component nodes with physical identifiers.

[0185] Sub-step S320: Define the fault propagation direction. Input source: Dryer fault history record (non-training data); Specific operations: Determine the propagation direction of increased rotor vibration caused by bearing wear; Determine the propagation direction of abnormal fan noise caused by rotor imbalance; Determine the propagation direction of bearing temperature rise caused by heater overheating; Define the product: Describe the directed edge definition rules for the physical path of fault propagation.

[0186] Sub-step S330: Construct a topological adjacency matrix. Input source: the set of component nodes from sub-step S310; the directed edge definition rules from sub-step S320; specific operations: sort the nodes by physical location; mark the connection relationship in the matrix according to the propagation direction (connection = 1, no connection = 0); product: a binary adjacency matrix representing the fault propagation path.

[0187] Sub-step S340: Generate the fault propagation network. Input source: the binary adjacency matrix from sub-step S330. Specific operations: convert the adjacency matrix into a graph data structure; label each directed edge with a fault type identifier; output: a directed topology network for which path analysis can be performed.

[0188] It should be noted that, taking a certain model of grain dryer as an example, its mechanical structure includes the following core components:

[0189] Bearing assemblies (2 sets): supporting rotor system, models NJ214 (front axle) and NJ216 (rear axle);

[0190] Rotor assembly: mass balancing core, moment of inertia J = 12.5 kg 2 ;

[0191] Fan assembly: Centrifugal fan, rated speed 1450 rpm;

[0192] Heater assembly: Resistance heater, power 25kW;

[0193] Historical fault records show:

[0194] The bearing wear failure rate was 38% (front axle) and 27% (rear axle).

[0195] Rotor imbalance caused the fan vibration to exceed the standard by 12 times per year;

[0196] Nine cases were reported where heater overheating led to a bearing temperature rise of ≥70℃.

[0197] Sub-step S310: Identify key physical component nodes, node mapping rules:

[0198] First type of node: Bearing assembly → Nodes B1 (front axle), B2 (rear axle);

[0199] Second type of node: Rotor assembly → Node R;

[0200] Third type of node: Wind turbine component → Node F;

[0201] Fourth type of node: Heater assembly → Node H;

[0202] Product: Node set {B1,B2,R,F,H};

[0203] Sub-step S320: Define the fault propagation direction and establish directed edge rules based on historical fault records (non-training data):

[0204] Bearing → Rotor conduction:

[0205] B1 wear (gap ≥ 0.3 mm) → increased rotor vibration (amplitude ≥ 0.5 mm / s);

[0206] Transmission path: B1→R, fault type identifier F01;

[0207] Rotor → Fan conduction:

[0208] R imbalance (eccentricity ≥ 0.2 mm) → abnormal fan noise (sound pressure level ≥ 85 dB);

[0209] Transmission path: R→F, fault type identifier F02;

[0210] Heater → Bearing conduction:

[0211] H overheating (temperature ≥150℃) → bearing temperature rise (temperature rise rate ≥5℃ / min);

[0212] Transmission path: H→B1, H→B2, fault type identifier F03;

[0213] Product: Set of directed edge rules

[0214]

[0215] Sub-step S330: Construct the topological adjacency matrix, and sort the nodes by physical location from front to back: H→B1→R→F→B2;

[0216] Adjacency matrix (connected = 1, unconnected = 0):

[0217] node H B1 R F B2 H 0 1 0 0 1 B1 0 0 1 0 0 R 0 0 0 1 0 F 0 0 0 0 0 B2 0 0 0 0 0

[0218] Product: Binary adjacency matrix (reflecting the physical path of fault propagation).

[0219] Sub-step S340: Generate the fault propagation network, graph structure transformation:

[0220] The adjacency matrix is ​​converted into a directed graph, and the distance between nodes reflects the actual installation location (the distance between H and B1 is 0.5m).

[0221] Directed edge labeling:

[0222] Side H→B1: Marked F03 (heat conduction fault);

[0223] Side B1→R: Marked F01 (Mechanical Vibration Fault);

[0224] Side R→F: Marked F02 (Power transmission fault);

[0225] Product: A directed topology network for which path analysis can be performed.

[0226] 204. Generate stability constraint control command. Input sources: real-time temperature sensor, humidity sensor, vibration sensor data, Lyapunov function from step 202; specific operation: calculate the time change rate of the Lyapunov function. When the time change rate is less than the preset threshold, output the original DRL command. Otherwise, trigger the stability compensator to generate a correction command. Definite product: an adaptive control command with stability guarantee.

[0227] Specifically: Sub-step S410: Acquire real-time sensor data. Input sources: real-time measurement values ​​from temperature sensor, humidity sensor, and vibration sensor. Specific operations: Simultaneously read data from the three sensors; perform signal filtering processing; identify the product: pre-processed sensor data package.

[0228] Sub-step S420: Calculate the current energy state value. Input source: the preprocessed sensor data packet from sub-step S410; the Lyapunov function from step 202. Specific operation: substitute the sensor data into the Lyapunov function; calculate the system energy value at the current moment; output: real-time energy state value V(t).

[0229] Sub-step S430: Calculate the rate of energy change. Input sources: the real-time energy state value V(t) at the current moment; the energy state value V(t-1) stored at the previous moment; specific operations: calculate the energy change per unit time; obtain the rate of energy change; product: rate of energy change.

[0230] Sub-step S440: Perform stability determination. Input source: energy change rate from sub-step S430; preset stability coefficient λ. Specific operation: compare energy change rate with stability coefficient threshold; determine system stability state; identify product: stability determination flag.

[0231] Sub-step S450: Generate control instructions. Input sources: stability determination flag from sub-step S440; raw instructions output by the DRL agent. Specific operations: When the flag is in a stable state, directly output the raw DRL instructions; when the flag is in an unstable state, trigger the stability compensator to generate correction instructions. Defining output: adaptive control instruction set.

[0232] It should be noted that, taking a certain model of grain dryer as an example, the target operating state is: temperature T ref

[0233] =60℃, humidity H ref =15%, fan rated speed ω nom = 1200 rpm (equivalent to 125.66 rad / s).

[0234] Real-time sensor data (filtered): Temperature T = 65℃; Humidity H = 18%; Rotation speed ω = 1300 rpm (136.14 rad / s);

[0235] Lyapunov function weighting coefficients: ω1 = 0.6 (highest temperature sensitivity), ω2 = 0.3, ω3 = 0.1; stability threshold λ = -0.5.

[0236] Sub-step S410: Sensor data acquisition and filtering, synchronous data acquisition:

[0237] Temperature sensor real-time value: 65.3℃ → 65℃ after median filtering;

[0238] Humidity sensor real-time value: 17.9% → 18% after moving average filtering;

[0239] Vibration sensor converted speed: 1295rpm → 1300rpm after Kalman filtering;

[0240] Product: Data packet {T=65,H=18,ω=1300}.

[0241] Sub-step S420: Real-time energy state calculation, deviation calculation:

[0242] e T =65-60=5℃;

[0243] e H =18-15=3%;

[0244] e ω =136.14-125.66=10.48rad / s;

[0245] Lyapunov function value:

[0246]

[0247] Product: Current energy value V(t) = 14.34.

[0248] Sub-step S430: Calculation of energy change rate, comparison with historical data:

[0249] At the previous moment, V(t-1) = 12.80 (stored value);

[0250] Sampling interval Δt = 1 min;

[0251] Calculation of rate of change:

[0252] Product: Rate of change of energy

[0253] Sub-step S440: Stability determination, threshold comparison:

[0254] Judgment condition: If (i.e., ≤-0.5) indicates stability;

[0255] current → Unstable;

[0256] Product: Stability label "Unstable".

[0257] Sub-step S450: Control instruction generation, DRL raw instructions:

[0258] Heater power increase command: +10% (output by deep reinforcement learning agent);

[0259] Fan speed command: +5%;

[0260] Stability compensation trigger: The compensator is triggered by the instability flag, generating a correction command.

[0261] Heater power reduced to -8% (to suppress temperature rise);

[0262] The fan speed should be kept constant (to avoid mechanical overload);

[0263] Product: Adaptive control instruction set:

[0264] Actuator Original instructions Correction instructions Heater power +10％ -8％ Fan speed +5％ 0％

[0265] 205. Generate fault suppression instructions. Input sources: the directed topology network in step 203 and the fault type signal identified in step 204. Specific operations: locate the fault source component in the topology network, calculate the network influence of the fault source component, generate suppression actions for high-influence components, and define the product: a topology suppression instruction to block the spread of faults.

[0266] Specifically: Sub-step S510: Locate the fault source component, input source: fault type signal identified in step 204; directed topology network in step 203; specific operation: match associated nodes in the topology network according to the fault type; determine the starting node of the fault propagation path; define the output: fault source node identifier;

[0267] Sub-step S520: Calculate node influence degree. Input source: fault source node identifier from sub-step S510; directed topology network from step 203. Specific operation: perform graph traversal starting from the fault source node; calculate node betweenness centrality; product: node betweenness centrality value.

[0268] Sub-step S530: Generate suppression instructions. Input source: node betweenness centrality value from sub-step S520; preset influence threshold (0.8); specific operation: filter key nodes with betweenness centrality ≥ 0.8; generate suppression action parameters for each key node; product: topology suppression instruction set.

[0269] It should be noted that the fault propagation topology network of a certain type of grain dryer has been constructed (product of step 203), containing 5 key nodes:

[0270] Nodes: B1 (front bearing), B2 (rear bearing), R (rotor), F (fan), H (heater);

[0271] Directed edge relation:

[0272] B1→R (Fault type F01: Rotor vibration caused by bearing wear);

[0273] R→F (Fault type F02: Rotor imbalance causes abnormal fan noise);

[0274] H→B1, H→B2 (Fault type F03: Heater overheating causes bearing temperature rise);

[0275] The real-time fault diagnosis system (step 204) identified abnormal vibration at wind turbine node F (vibration value 4.2 mm / s, exceeding the threshold of 3.5 mm / s), and the fault type signal was F02.

[0276] Sub-step S510: Locate the fault source component and match the fault type:

[0277] The fault type signal F02 corresponds to the transmission path R→F in the topology network (rotor imbalance causes abnormal fan noise).

[0278] Path backtracking: Tracing back along the directed edge: F←R←B1; The fault source is determined to be the front bearing node B1 (bearing wear triggers a chain reaction).

[0279] Product: Fault source node identifier B1.

[0280] Sub-step S520: Calculate node influence degree, graph traversal and betweenness centrality calculation: Starting from B1, traverse all its propagation paths: B1→R→F (affects the fan); B1→R (indirectly affects the rotor);

[0281] Calculate node betweenness centrality (a measure of a node's pivotal role in a propagation path):

[0282]

[0283]

[0284] Product: Node betweenness centrality values ​​(B1 = 0.92, R = 0.85, F = 0.32).

[0285] Sub-step S530: Generate suppression instructions and filter key nodes:

[0286] Preset influence threshold 0.8 → Filter B1 and R (both ≥ 0.8).

[0287] Suppression action design:

[0288] node Inhibit movement physical mechanism B1 Reduce rotor speed by 20% Reduce bearing load and inhibit wear deterioration R Activate the vibration compensator (amplitude -30%). To counteract the resonance effect caused by rotor imbalance

[0289] Product: Topology Suppression Instruction Set

[0290] {Bearing node B1: Speed ​​reduction of 20%, Rotor node R: Vibration compensation of -30%}.

[0291] 206. Synthesize the final execution signal. Input sources: the adaptive control command in step 204 and the topology suppression command in step 205. Specific operation: fuse the commands and convert them into actuator drive signals. Defining the product: the physical control signals output to the heater and fan.

[0292] Specifically: Sub-step S610: Command signal conversion, input source: the adaptive control command set of step 204 (including power / speed values); the topology suppression command set of step 205 (including node action parameters); specific operation: convert the power command into a voltage signal (0-10V); convert the speed command into a pulse frequency signal (4-20kHz); convert the suppression parameters into a switching signal; output: analog drive signal and digital switching signal;

[0293] Sub-step S620: Instruction priority arbitration; input source: analog drive signal and digital switch signal from sub-step S610; specific operation: when a suppression instruction exists, the original control instruction of the corresponding node is overwritten; when there is no suppression instruction, the original control signal is maintained; output: the set of execution signals after arbitration.

[0294] Sub-step S630: Generate physical control signals. Input source: the arbitrated execution signal set of sub-step S620. Specific operation: Output analog signal to heater power regulator through DAC module; output pulse signal to fan speed controller through PWM module; output switch signal to emergency braking unit through DO module; Definite product: physical control signal to drive actuator operation.

[0295] It should be noted that a certain model of grain dryer triggered the following command during operation:

[0296] Adaptive control command (output at step 204):

[0297] Heater power adjustment: +15% (corresponding to target power 25kW→28.75kW);

[0298] Fan speed adjustment: -10% (target speed 1200rpm → 1080rpm);

[0299] Topology suppression instruction (output at step 205):

[0300] An overheating fault was detected at bearing node B1 (between centrality 0.92 ≥ 0.8), triggering an emergency braking command: cut off the heater power supply (switching input 1);

[0301] Sub-step S610: Command signal conversion, power command conversion: power range 0–30kW mapped to 0–10V voltage signal.

[0302] (Original instruction);

[0303] After the suppression command is overridden: power supply is cut off → voltage = 0V;

[0304] Speed ​​command conversion: Speed ​​range 0–2000 rpm mapped to 4–20 kHz pulse frequency:

[0305] (No suppression instruction overridden);

[0306] Suppression parameter conversion:

[0307] Emergency braking command → Switch quantity 1 (high level triggers braking);

[0308] Products: analog signal [0V, 12.64kHz], digital signal [1];

[0309] Sub-step S620: Instruction priority arbitration, arbitration logic:

[0310] Suppress command present (switching quantity 1) → Overwrite heater's original power command;

[0311] Fan no suppression command → Maintain original speed command;

[0312] Arbitration Result:

[0313]

[0314] Sub-step S630: Generate physical control signals, signal output:

[0315] DAC module: Outputs 0V to heater power regulator → heater power off;

[0316] PWM module: Outputs a 12.64kHz pulse to the fan speed controller → fan speed decreases to 1080rpm;

[0317] DO module: Outputs high level 1 to the emergency braking unit → forced bearing cooling start;

[0318] Real-time effect:

[0319] Five seconds after the heater was de-energized, the bearing temperature dropped from 72℃ to 58℃ (cooling rate 2.8℃ / s).

[0320] The fan speed stabilized at 1080 rpm (vibration value changed from 4.2 mm / s to 1.8 mm / s).

[0321] This invention encompasses steps such as constructing a physical dynamic model, generating a system stability criterion function, establishing a fault propagation topology, generating stability constraint control commands, generating fault suppression commands, and synthesizing the final execution signal. By integrating thermodynamics, mechanical dynamics principles, and deep reinforcement learning, it achieves precise monitoring and efficient control of the dryer's operating status. Variables such as temperature, humidity, wind speed, and rotational speed are incorporated into a unified dynamic model. Through thermo-mechanical coupling and humidity-wind speed coupling equations, it more comprehensively reflects the complex internal processes of the dryer, significantly improving the accuracy of fault diagnosis. An energy scalar function is constructed based on the sum of squared deviations of state variables to quantify the system's energy state, providing a mathematical basis for stability determination. By calculating the time change rate of the Lyapunov function in real time, dynamic evaluation of system stability is achieved, ensuring the system operates within a safe range. The dryer's mechanical structure is mapped as a directed topological network, visually demonstrating the fault propagation path. By calculating the betweenness centrality of nodes, the network influence of fault source components is quantified, providing a scientific basis for generating fault suppression commands and effectively blocking fault propagation. Combining real-time sensor data and the Lyapunov function stability criterion, adaptive control commands with stability guarantees are generated. When the system becomes unstable, a stability compensator is triggered to generate correction commands, ensuring the system quickly returns to a stable state. Adaptive control commands and topology suppression commands are fused, and a priority arbitration mechanism ensures that suppression commands are executed first when available, preventing fault exacerbation. This fusion mechanism improves the flexibility and response speed of the control system. The fused commands are converted into actuator drive signals and output to actuators such as heaters and fans through modules such as DAC, PWM, and DO, achieving precise control of the dryer. This method of generating and executing physical control signals improves the reliability and stability of the control system.

[0322] The present invention also provides a dryer fault diagnosis device based on deep reinforcement learning. The dryer fault diagnosis device based on deep reinforcement learning includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor performs the steps of the dryer fault diagnosis method based on deep reinforcement learning in the above embodiments.

[0323] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the dryer fault diagnosis method based on deep reinforcement learning.

[0324] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0325] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0326] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dryer fault diagnosis method based on deep reinforcement learning, characterized in that, The dryer fault diagnosis method based on deep reinforcement learning includes: The thermodynamic and mechanical parameters of the dryer are obtained, and a set of differential equations for state variables is established to obtain a physical dynamic model. Based on the physical dynamic model and the preset target value, a scalar function is constructed based on the sum of squared deviations between the state variables and the target value, thus obtaining the Lyapunov function; Obtain the mechanical structural connection relationship of the dryer, map the bearings, rotor, fan and heater as nodes, and map the fault propagation path as directed edges to obtain a directed topology network; Acquire real-time data from temperature, humidity, and vibration sensors, calculate the time rate of change of the Lyapunov function, and output the original deep reinforcement learning instruction when the time rate of change is less than a preset threshold; otherwise, trigger the stability compensator to generate correction instructions to obtain an adaptive control instruction set. Based on the fault type signal and the directed topology network, the associated nodes in the directed topology network are matched according to the fault type to determine the starting node of the fault propagation path and obtain the fault source node identifier; based on the fault source node identifier and the directed topology network, a graph traversal is performed starting from the fault source node to calculate the node betweenness centrality and obtain the node betweenness centrality value; based on the node betweenness centrality value and a preset influence threshold, key nodes with betweenness centrality values ​​≥ the preset influence threshold are selected, and suppression action parameters are generated for each key node to obtain a topology suppression instruction set; Based on the adaptive control instruction set and the topology suppression instruction set, the power command is converted into a voltage signal, the speed command is converted into a pulse frequency signal, and the suppression parameter is converted into a switching signal, thus obtaining analog drive signals and digital switching signals. Based on the analog drive signals and digital switching signals, when a suppression command exists, the original control command of the corresponding node is overridden; when no suppression command exists, the original control signal is maintained, thus obtaining an arbitrated execution signal set. Based on the arbitrated execution signal set, an analog signal is output to the heater power regulator through the DAC module, a pulse signal is output to the fan speed controller through the PWM module, and a switching signal is output to the emergency braking unit through the DO module, thus obtaining the physical control signals for driving the actuator operation.

2. The dryer fault diagnosis method based on deep reinforcement learning according to claim 1, characterized in that, include: Based on the thermodynamic parameters of the dryer, temperature change rate equations and humidity change rate equations are established, resulting in heat conduction model and proton transfer model. Based on the mechanical parameters of the dryer, equations for the rate of change of rotational speed and wind speed correlation are established to obtain a mechanical dynamics sub-model; Based on the heat conduction sub-model, proton transfer model, and mechanical dynamics sub-model, the thermo-mechanical equations are correlated and the humidity-wind speed equations are coupled to obtain a complete set of dynamic equations containing four-variable coupling relationships.

3. The dryer fault diagnosis method based on deep reinforcement learning according to claim 2, characterized in that, include: Based on the state variables and preset target values ​​in the physical dynamic model, the temperature deviation, humidity deviation, and rotational speed deviation are calculated. Based on the temperature deviation, humidity deviation, and rotational speed deviation, the square of each deviation value is taken and the weighted sum is obtained to obtain the energy-type scalar function V; Based on the energy-type scalar function V, and associated with the physical dynamic model: V is used as a measure of the system's energy state; a stability condition is defined: the system is stable when V decreases with time; and the Lyapunov function used for stability determination is obtained.

4. The dryer fault diagnosis method based on deep reinforcement learning according to claim 3, characterized in that, include: Based on the mechanical structure connection diagram of the dryer, the bearing assembly is extracted as the first type of node, the rotor assembly as the second type of node, the fan assembly as the third type of node, and the heater assembly as the fourth type of node, thus obtaining a set of component nodes with physical identification. Based on the dryer's fault history, the transmission direction of increased rotor vibration caused by bearing wear, the transmission direction of abnormal fan noise caused by rotor imbalance, and the transmission direction of bearing temperature rise caused by heater overheating are determined, thus obtaining the directed edge definition rules describing the physical path of fault propagation. Based on the component node set and the rules for defining directed edges, the nodes are sorted by physical location, and the connection relationships are marked in the matrix according to the direction of propagation to obtain a binary adjacency matrix representing the fault propagation path. Based on the binary adjacency matrix, the adjacency matrix is ​​converted into a graph data structure, and each directed edge is labeled with a fault type identifier to obtain a directed topology network for executable path analysis.

5. The dryer fault diagnosis method based on deep reinforcement learning according to claim 4, characterized in that, include: Simultaneously read data from three sensors to obtain real-time measurement values ​​from the temperature sensor, humidity sensor, and vibration sensor. Perform signal filtering to obtain a pre-processed sensor data package. Based on the preprocessed sensor data packet and Lyapunov function, the sensor data is substituted into the Lyapunov function to calculate the system energy value at the current time, and the real-time energy state value V(t) is obtained. Based on the real-time energy state value V(t) at the current moment and the energy state value V(t-1) stored at the previous moment, calculate the energy change per unit time and obtain the energy change rate; Based on the energy change rate and the preset stability coefficient λ, the energy change rate is compared with the stability coefficient threshold to determine the system stability state and obtain the stability judgment flag. Based on the stability determination flag and the original instructions output by the deep reinforcement learning agent, when the flag indicates a stable state, the original instructions of the deep reinforcement learning agent are directly output; when the flag indicates an unstable state, the stability compensator is triggered to generate correction instructions, thus obtaining an adaptive control instruction set.

Images