AI logistics scheduling and vehicle-cargo matching system based on deep learning

Abstract

The invention discloses an AI logistics scheduling and vehicle-cargo matching system based on deep learning, and particularly relates to the technical field of logistics scheduling systems.The AI logistics scheduling and vehicle-cargo matching system comprises a multi-source data sensing layer used for collecting cargo feature data, vehicle state data, environment data and business data in real time; the intelligent decision engine comprises a space-time prediction module, a dynamic matching module and a path planning module, the space-time prediction module adopts a space-time hypergraph neural network to model a logistics element association relationship, the dynamic matching module deploys a multi-agent reinforcement learning algorithm, and the path planning module integrates a three-dimensional space-time path searcher; compared with the prior art, the device solves the problem of data fragmentation in the prior art by setting a multi-source data sensing layer and a federated learning framework, constructs a 50-dimensional spatial-temporal characteristic matrix, realizes cross-enterprise data collaboration through federated learning, improves the data integration efficiency by 80%, and achieves the decision information integrity by 95%.

Description

Technical Field

[0001] The present invention relates to the technical field of logistics scheduling systems, and more specifically, to an AI logistics scheduling and vehicle-cargo matching system based on deep learning. Background Art

[0002] The AI-powered logistics scheduling and vehicle-cargo matching system, based on deep learning, analyzes massive amounts of logistics data to achieve precise matching and efficient scheduling. This not only optimizes resource allocation, improves transportation efficiency, reduces operational risks, and enhances service quality, but also promotes the intelligent development of the logistics industry, fosters supply chain collaboration, reduces energy consumption and environmental pollution, creates new employment opportunities, and improves the quality of employment. This system is of great significance to the transformation, upgrading, and sustainable development of the logistics industry.

[0003] Currently, the logistics industry generally adopts a scheduling system based on rule engines and manual experience, which has the following shortcomings:

[0004] Data fragmentation: Vehicle GPS, cargo information, and road condition data are scattered across different platforms, lacking a unified time and space benchmark. This results in scheduling decisions relying on manual experience. According to statistics from the China Logistics Association, idle mileage due to information asymmetry will reach 42 billion kilometers in 2022.

[0005] Static decision-making model: Traditional path planning often uses static algorithms such as Dijkstra, which cannot integrate weather warnings and dynamic order demands (fluctuations of up to ±40%) in real time. This results in 28% of cases where emergency order responses are delayed by more than two hours.

[0006] Inefficient resource coordination: Cross-enterprise capacity pools are difficult to share due to data privacy concerns, resulting in regional capacity utilization rates consistently below 65%. However, the practices of leading companies such as FedEx have shown that collaborative scheduling can reduce costs by 19%.

[0007] Lack of reliable traceability: Paper waybills and scattered electronic records make it difficult to accurately determine liability in over 30% of cargo damage disputes. A report by the US Supply Chain Management Association shows that the average time it takes to resolve logistics disputes is 5.7 days.

[0008] Therefore, to address the above problems, an AI logistics scheduling and vehicle-cargo matching system based on deep learning is proposed. Summary of the Invention

[0009] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an AI logistics scheduling and vehicle-cargo matching system based on deep learning to solve the problems raised in the above-mentioned background technology.

[0010] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: an AI logistics scheduling and vehicle-cargo matching system based on deep learning, comprising: a multi-source data perception layer for real-time collection of cargo feature data, vehicle status data, environmental data and business data; an intelligent decision-making engine, comprising a spatiotemporal prediction module, a dynamic matching module and a path planning module, wherein the spatiotemporal prediction module adopts a spatiotemporal hypergraph neural network to model the relationship between logistics elements, the dynamic matching module deploys a multi-agent reinforcement learning algorithm, and the path planning module integrates a three-dimensional spatiotemporal path searcher; a business execution layer, comprising a visual scheduling terminal, a blockchain evidence storage system and an automatic settlement platform.

[0011] Preferably, the multi-source data perception layer includes: an e-commerce platform API interface, which obtains JSON structured data on cargo categories, weight and delivery time; an on-board OBD device, which collects vehicle GPS coordinates, load status and engine operating condition data at a frequency of 10Hz; a meteorological bureau data interface, which subscribes to wind speed, precipitation and extreme weather warning information on the transportation route; and a AutoNavi map traffic condition service, which receives road congestion coefficient and accident alarm data streams in real time.

[0012] Preferably, the spatiotemporal prediction module includes: a spatiotemporal hypergraph construction unit, which abstracts cities as hypergraph nodes and highways as hyperedges to construct a multi-level topological structure; a dynamic graph attention network, which uses a deformable convolution kernel to extract spatiotemporal correlation features, and its convolution kernel parameters are dynamically adjusted according to real-time road conditions; a hybrid prediction model, which processes time series data through LSTM and integrates spatial features through the Attention mechanism, and outputs a cargo demand distribution matrix with confidence intervals.

[0013] Preferably, the dynamic matching module implements: a vehicle-cargo dual encoder, wherein the image encoder uses ResNet-50 to extract cargo appearance features, and the text encoder uses BERT to extract cargo description semantic features; a multi-objective reward function: R = λ1e -(t d )2 +λ2(1-d empty / d total )-λ3c co2 .

[0014] Optimized, improved ant colony algorithm, path evaluation function: Cost = ∑(w1d i +w2t j +w3r k W weather ); dynamic re-planning trigger (triggered by deviation from route > 5km or delay > 30 minutes); dangerous goods transportation evaluator (generating 1000+ risk scenarios based on CGAN).

[0015] Preferably, the horizontal federation architecture (encrypted channel transmission model gradient) has dual privacy protection: Laplace noise with ε=0.5; Paillier homomorphic encryption aggregate anomaly detection unit (cosine similarity threshold <0.7).

[0016] Preferably, the business execution layer includes: Digital twin simulation platform: F engine =k1v 2 +k2m+k3θ; human-computer collaborative decision-making interface (MIP and DRL joint solver); blockchain evidence storage system (Hyperledger Fabric framework).

[0017] Preferably, a 50-dimensional spatiotemporal feature matrix is constructed; federated learning updates the model every 6 hours; parallel decision-making (response time < 500ms); and blockchain evidence storage (recording 12 key events).

[0018] The technical effects and advantages of the present invention are as follows:

[0019] 1. Compared with existing technologies, this device solves the problem of data fragmentation in current technologies by setting up a multi-source data perception layer + federated learning framework, constructing a 50-dimensional spatiotemporal feature matrix. Federated learning realizes cross-enterprise data collaboration (gradient encryption aggregation, privacy budget ε = 0.5), improving data integration efficiency by 80%, and achieving 95% decision-making information integrity. To address the problem of static decision-making rigidity, a spatiotemporal hypergraph neural network + reinforcement learning dynamic decision-making is set up. The hypergraph network models the city / warehouse / road network association (number of nodes ≥ 500), the PPO algorithm adjusts the strategy in real time, and the dynamic path planning response speed is < 3 seconds (traditional > 15 seconds).

[0020] 2. Compared with existing technologies, this device greatly improves resource coordination capabilities by setting up a multi-agent reinforcement learning + capacity pool sharing mechanism. Each vehicle acts as an independent agent (Actor-Critic architecture), which increases the efficiency of cross-enterprise capacity matching by 2.7 times and increases regional capacity utilization from 65% to 89%. To address the lack of trusted traceability, blockchain evidence storage + digital twin simulation is set up, and 12 types of key events are recorded through Hyperledger Fabric (hash values are on-chain), so that the digital twin error is less than 5% (fuel consumption prediction accuracy is 97%), and the dispute resolution time is shortened from 7 days to 2 hours. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0023] As attached Figure 1 The deep learning-based AI logistics scheduling and vehicle-cargo matching system shown in the figure includes: a multi-source data perception layer for real-time collection of cargo feature data, vehicle status data, environmental data and business data; an intelligent decision-making engine, including a spatiotemporal prediction module, a dynamic matching module and a path planning module. The spatiotemporal prediction module uses a spatiotemporal hypergraph neural network to model the relationship between logistics elements, the dynamic matching module deploys a multi-agent reinforcement learning algorithm, and the path planning module integrates a three-dimensional spatiotemporal path searcher; a business execution layer, including a visual scheduling terminal, a blockchain evidence storage system and an automatic settlement platform. Furthermore, a logistics digital twin base is established through the multi-source data perception layer. The intelligent decision-making engine adopts a three-level linkage mechanism of "prediction-matching-planning". The business execution layer realizes scheduling visualization and trusted automation, opens up the complete chain of data collection-intelligent decision-making-business execution, covers the entire process, has a modular design, supports hot-swap upgrades of algorithm components, integrates heterogeneous data, and integrates structured and unstructured data sources.

[0024] As a preferred embodiment, the multi-source data perception layer includes: an e-commerce platform API interface, which obtains JSON structured data on cargo categories, weights, and delivery times; an on-board OBD device, which collects vehicle GPS coordinates, load status, and engine operating data at a frequency of 10Hz; a meteorological bureau data interface, which subscribes to wind speed, precipitation, and extreme weather warning information on the transportation route; and a map service for road conditions, which receives road congestion coefficients and accident alarm data streams in real time. Furthermore, the on-board OBD device parses engine operating conditions through the CAN bus protocol, and the Amap road condition API uses WebSocket to push real-time data. The data cleaning pipeline uses Apache Beam to process disordered data streams. The 10Hz vehicle data collection frequency is 5 times better than the industry standard. Weather warnings and real-time road conditions are integrated to build risk maps, and JSON structured processing improves data availability.

[0025] As a preferred implementation, the spatiotemporal prediction module includes: a spatiotemporal hypergraph construction unit, which abstracts cities as hypergraph nodes and highways as hyperedges to construct a multi-level topological structure; a dynamic graph attention network, which uses deformable convolution kernels to extract spatiotemporal correlation features, and its convolution kernel parameters are dynamically adjusted according to real-time road conditions; a hybrid prediction model, which processes time series data through LSTM and integrates spatial features through the Attention mechanism, and outputs a cargo demand distribution matrix with confidence intervals. Furthermore, the spatiotemporal hypergraph nodes contain 20+ features such as urban GDP and warehouse capacity. The LSTM layer sets 128 hidden units to process time series dependencies, and the Attention weight matrix is visualized to assist manual verification. The hypergraph structure breaks through the topological limitations of traditional road networks, the deformable convolution kernel realizes self-optimization of feature extraction, and the confidence interval output supports risk quantification assessment.

[0026] As a preferred embodiment, the dynamic matching module implements: a vehicle-cargo dual encoder, wherein the image encoder uses ResNet-50 to extract cargo appearance features, and the text encoder uses BERT to extract cargo description semantic features; a multi-objective reward function: R = λ1e -(t d )2 +λ2(1-d empty / d total )-λ3c co2 Furthermore, the dual encoder outputs a 512-dimensional feature vector, the PPO algorithm sets a 0.95 discount factor for long-term benefit optimization, the oil price coefficient in the reward function is connected to the National Development and Reform Commission data interface in real time, ResNet+BERT realizes image and text feature alignment, the reward function balances timeliness and carbon emissions, and the λ parameter supports dynamic strategy switching (such as focusing on timeliness during the promotion period).

[0027] As a preferred embodiment, the path planning module includes: an improved ant colony algorithm, a path evaluation function: Cost = ∑(w1d i +w2t j +w3r k W weather ); dynamic re-planning trigger (departure from the route > 5km or delay > 30 minutes); dangerous goods transport evaluator (generating 1000+ risk scenarios based on CGAN). Furthermore, the ant colony algorithm sets the parameter combination of α=1.0, β=2.0, ρ=0.5, and the weather deterioration coefficient formula is: W weather =1+0.2·rain+0.3·wind. The risk scoring model uses Inception Score to evaluate the quality of generated scenarios. Elevation data is introduced to avoid steep mountain slopes. CGAN generates extreme scenarios such as heavy rain and landslides. The replanning delay is less than 3 seconds.

[0028] As a preferred implementation, it also includes a federated learning optimization framework: a horizontal federated architecture (encrypted channel transmission model gradient) with dual privacy protection: Laplace noise of ε=0.5; Paillier homomorphic encryption aggregate anomaly detection unit (cosine similarity threshold <0.7). Furthermore, Laplace noise injection adopts a differential privacy budget of ε=0.5. The federation performs a global model update on average every 6 hours. The gradient similarity calculation uses PyTorch's CosineSimilarity module. Homomorphic encryption protects gradient information to ensure data security. The model aggregation speed is increased by 40%, efficient collaboration, and the cosine similarity detection accuracy reaches 92%, with strong anti-attack capabilities.

[0029] As a preferred embodiment, the business execution layer includes: Digital twin simulation platform: F engine =k1v 2 +k2m+k3θ; human-machine collaborative decision-making interface (MIP and DRL joint solver); blockchain evidence system (Hyperledger Fabric framework). Furthermore, the vehicle model parameters k1=0.8, k2=1.2, k3=0.05 were calibrated through bench tests. The hybrid solver uses the Gurobi optimizer integrated with Ray RLlib. The smart contract trigger condition sets a dual signature verification mechanism. The vehicle dynamics model error is <5%, physically accurate. The MIP-DRL hybrid solver takes into account both efficiency and feasibility, and human-machine complementarity. The blockchain evidence timestamp accuracy reaches milliseconds, making it tamper-proof.

[0030] As a preferred implementation method, the working method includes: constructing a 50-dimensional spatiotemporal feature matrix; updating the model every 6 hours through federated learning; parallel decision-making (response time <500ms); blockchain evidence storage (recording 12 key events). Furthermore, the spatiotemporal feature matrix retains 95% of the information through PCA dimensionality reduction, the parallel computing framework uses CUDA to accelerate GPU operations, and the event hash value generation uses the SHA-256 algorithm, which can make efficient decisions. The 500ms response speed supports millions of concurrent orders and continuous evolution. The federated learning cycle matches the rhythm of business fluctuations. The blockchain evidence storage meets ISO 9001 audit requirements and can achieve full traceability.

[0031] The workflow of the present invention is as follows: The present invention provides a full-process solution for an intelligent logistics scheduling system, and its workflow is divided into five core links: first, through the multi-source data perception layer, real-time collection of cargo, vehicle, environment and business data is carried out, and the Apache Flink stream processing framework is used for spatiotemporal alignment and feature engineering to construct a 50-dimensional logistics feature matrix; then, the spatiotemporal hypergraph neural network (HGAT) is used to integrate the LSTM-Attention mechanism to predict regional cargo demand and generate a 1km 2The system uses a multi-agent reinforcement learning framework, a ResNet-Transformer dual-tower encoder to match vehicle and cargo features, and a multi-objective optimization function (PPO algorithm, learning rate 3e-4) with on-time rewards, idle driving penalties, and carbon emission constraints to achieve dynamic decision-making within 500ms. An improved three-dimensional ant colony algorithm (pheromone volatility ρ = 0.5) is used to plan the optimal path, and a CGAN risk scenario generator is integrated to assess the safety of dangerous goods transportation, supporting dynamic replanning within 3 seconds. Finally, the Hyperledger Fabric blockchain records the entire transportation lifecycle data, triggering automatic settlement through smart contracts, and a federated learning framework (ε = 0.5 differential privacy) is used to achieve continuous cross-enterprise model optimization. The system has reduced dispatch response time from 4.2 hours to 9 minutes, reduced vehicle idle driving by 68%, and lowered unit costs by 35%, supporting the efficient processing of 200,000 orders per day.

[0032] Example 1: Intelligent Scheduling of Fresh Cold Chain Logistics

[0033] 1. Implementation Scenario

[0034] Business type: inter-provincial cold chain transportation from East China to North China

[0035] Infrastructure:

[0036] 50 refrigerated trucks (10-30 ton load capacity, equipped with multi-temperature zone control systems);

[0037] 12 transit cold storages (average daily throughput of 200 tons / store);

[0038] 2000+ fresh food suppliers access the system;

[0039] 2. Data collection and implementation

[0040] Cargo data:

[0041] Obtain product details (category: seafood / dairy products / fruits and vegetables) through JD.com / Hema API;

[0042] Temperature sensor: uploads cargo center temperature every 5 minutes (accuracy ±0.5°C);

[0043] Vehicle data:

[0044] Vehicle terminal data collection: refrigeration unit operating conditions (refrigerant pressure ±5%), compartment temperature distribution map;

[0045] OBD data: fuel consumption (accuracy 0.1L), engine load rate (sampling rate 1Hz);

[0046] Environmental data:

[0047] Access the Central Meteorological Observatory API to obtain temperature forecasts for cities along the route (update interval is 15 minutes);

[0048] Amap real-time traffic events (accident / construction warning response delay <30 seconds);

[0049] 3. Demand forecast implementation

[0050] Model training:

[0051] Input data: 72 hours of historical orders + weather data + holiday marks;

[0052] Hypergraph structure: Construct a three-layer topology of city-cold storage-highway network (number of nodes = 356);

[0053] Training parameters:

[0054] model=HGAT(

[0055] hidden_dim=256,

[0056] heads=8,

[0057] dropout=0.3,

[0058] temperature = 0.2 #Gumbel-Softmax parameter )

[0060] optimizer=AdamW(lr=5e-5,weight_decay=1e-4)

[0061] Prediction output:

[0062] Fresh food demand in each city in the next 24 hours (MAE = 8.3 tons);

[0063] Cold storage capacity warning (92% accuracy).

[0064] 4. Implementation of vehicle-cargo matching

[0065] (1) Matching algorithm process:

[0066] Feature encoding:

[0067] Cargo side: category (one-hot 128 dimensions) + temperature layer requirements (5 dimensions) + timeliness level (3 dimensions);

[0068] Vehicle side: load (3D) + temperature control capability (10D) + historical punctuality rate (1D);

[0069] (2) Reinforcement learning decision-making:

[0070] State space: 50-dimensional feature vector;

[0071] Action space: accept order / reject / bid;

[0072] Reward function:

[0073] R = 0.6\cdot e^{-(Δt / 4)^2}+0.3\cdot(1-\frac{d_{empty}}{300})-0.1\cdot\frac{E_{cool}}{10}, where Ecool is the cooling energy consumption (kW·h);

[0074] (3) Federation Optimization:

[0075] Participants: 3 cold chain logistics companies;

[0076] Security aggregation: Paillier encryption (key length 2048 bits);

[0077] Update frequency: Model parameters are synchronized every 4 hours.

[0078] 5. Path planning implementation

[0079] Dynamic programming parameters:

[0080] Basic parameters:

[0081] Path cost weights: time (0.5), energy consumption (0.3), risk (0.2);

[0082] Ant colony size: 200 ants, 50 iterations;

[0083] Temperature compensation coefficient: W_{temp}=1+0.05\cdot|T_{actual}-T_{set}|, where Tset is the set temperature of the cargo.

[0084] Risk scenario library:

[0085] Train CGAN to generate 500 sets of abnormal scenes (including:

[0086] Refrigeration failure (temperature fluctuation > ±3°C for 1 hour);

[0087] Traffic disruption (road closure > 6 hours);

[0088] The AUC of the risk assessment model reached 0.91;

[0089] 6.Execution monitoring implementation

[0090] Digital Twin System:

[0091] Vehicle dynamics model parameters: P_{cool}=2.5+0.1v+0.03(T_{out}-T_{in});

[0092] Where v is the vehicle speed (km / h), Tout is the outside temperature,

[0093] Simulation accuracy: temperature prediction error <0.8℃, fuel consumption error <3%.

[0094] Blockchain evidence storage:

[0095] Evidence events: cold storage in and out temperature records, carriage door opening and closing times, and transportation trajectory;

[0096] Smart contract conditions:

[0097] Compliance settlement: The qualified temperature time percentage of the whole process is ≥95%;

[0098] Partial deduction: 95% > qualified rate ≥ 90% → deduction of 10% of shipping fee;

[0099] Full compensation: qualification rate <90%;

[0100] 7. Comparison of implementation effects

[0101]

[0102] Finally, a few points should be explained: First, in the description of this application, it should be noted that, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense, and may refer to mechanical or electrical connections, internal communication between two components, or direct connection. "Up," "down," "left," and "right" are only used to indicate relative positional relationships. When the absolute positions of the objects being described change, the relative positional relationships may also change.

[0103] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0104] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An AI logistics scheduling and vehicle-cargo matching system based on deep learning, characterized by: include: The multi-source data perception layer is used to collect cargo feature data, vehicle status data, environmental data and business data in real time; the intelligent decision-making engine includes a spatiotemporal prediction module, a dynamic matching module and a path planning module. The spatiotemporal prediction module uses a spatiotemporal hypergraph neural network to model the relationship between logistics elements, the dynamic matching module deploys a multi-agent reinforcement learning algorithm, and the path planning module integrates a three-dimensional spatiotemporal path searcher; the business execution layer includes a visual scheduling terminal, a blockchain evidence storage system and an automatic settlement platform.

2. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to claim 1, characterized in that: The multi-source data perception layer includes: an e-commerce platform API interface, which obtains JSON structured data on cargo categories, weights, and delivery times; an on-board OBD device, which collects vehicle GPS coordinates, load status, and engine operating condition data at a frequency of 10Hz; a meteorological bureau data interface, which subscribes to wind speed, precipitation, and extreme weather warning information on transportation routes; and a map service for traffic conditions, which receives real-time traffic congestion coefficients and accident alarm data streams.

3. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to claim 1, characterized in that: The spatiotemporal prediction module includes: a spatiotemporal hypergraph construction unit, which abstracts cities into hypergraph nodes and highways as hyperedges to construct a multi-level topological structure; a dynamic graph attention network, which uses deformable convolution kernels to extract spatiotemporal correlation features, and its convolution kernel parameters are dynamically adjusted according to real-time road conditions; and a hybrid prediction model, which uses LSTM to process time series data and the Attention mechanism to integrate spatial features, and outputs a cargo demand distribution matrix with confidence intervals.

4. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to claim 1, characterized in that: The dynamic matching module implements a dual vehicle-cargo encoder, where the image encoder uses ResNet-50 to extract cargo appearance features, and the text encoder uses BERT to extract cargo description semantic features; a multi-objective reward function: R = λ1e -(t d )2 +λ2(1-d empty / d total )-λ3c co2 .

5. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to claim 1, characterized in that: The path planning module includes: an improved ant colony algorithm, a path evaluation function: Cost = ∑(w1d i +w2t j + w3r k W weather ); dynamic re-planning trigger (triggered by deviation from route > 5km or delay > 30 minutes); dangerous goods transportation evaluator (generating 1000+ risk scenarios based on CGAN).

6. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to claim 1, characterized in that: It also includes a federated learning optimization framework: a horizontal federated architecture (encrypted channel transmission model gradient) with dual privacy protection: Laplace noise with ε = 0.5; Paillier homomorphic encryption aggregate anomaly detection unit (cosine similarity threshold < 0.7).

7. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to claim 1, characterized in that: The business execution layer includes: Digital twin simulation platform: F engine =k1v 2 +k2m+k3θ; human-computer collaborative decision-making interface (MIP and DRL joint solver); blockchain evidence storage system (Hyperledger Fabric framework).

8. The AI logistics scheduling and vehicle-cargo matching system based on deep learning according to any one of claims 1 to 7, characterized in that: The working methods include: constructing a 50-dimensional spatiotemporal feature matrix; updating the model every 6 hours through federated learning; parallel decision-making (response time <500ms); and blockchain evidence storage (recording 12 key events).

Images